RBF-assisted surrogate modeling and machine learning for mechanical property prediction in friction stir additive manufacturing: Application to dissimilar AA6061/AA7075 aluminum alloys. Authors: Krishnan Venkatachalam, Senthil Kumaran Selvaraj, Giriraj Mannayee, Kalyanaraman Pattabiraman, Sundaramali G Published: 2026 ABSTRACT AbstractThis research develops a surrogate-assisted framework for exploring mechanical property relationships in friction stir additive manufacturing (FSAM) of dissimilar aluminum alloys (AA6061/AA7075) under data-scarcity conditions. Using a limited experimental dataset of 9 data points, Radial Basis Function (RBF) interpolation generates 882 synthetic samples covering tool rotation speed (1000–1200 rpm), traverse speed (30–40 mm/min), and tilt angle (1.0–2.2°). Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression models are trained to predict ultimate tensile strength (UTS) and Vickers hardness (HV) for both 6061-over-7075 and 7075-over-6061 material stacking configurations. GPR indicates performance with mean absolute errors of 2.34–8.67 MPa for UTS and comparable accuracy for hardness, achieving R² > 0.85 across synthetic samples. SHAP-based analysis of the surrogate-trained models indicates that tool rotation speed exerts the strongest influence within the learned surrogate response surfaces, followed by traverse speed and tilt angle, with relative contributions varying across outputs. Critically, this near zero error is obtained on synthetic samples derived from RBF interpolation of the original 9 measurements, not on independent experimental data. The models effectively learn and reproduce the RBF surrogate response surface within the explored design space. The framework successfully indicates a methodology for parameter space exploration under data scarcity; however, it illustrates conceptual parameter selection workflows based on surrogate-assisted trend exploration. Independent experimental validation is required before any practical optimization. The work provides a template applicable to other data-scarce manufacturing domains, with the essential caveat that external experimental validation is a prerequisite to practical application.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-42608-7.Keywords: Friction Stir Additive Manufacturing, Surrogate Modeling, Gaussian Process Regression, Synthetic Data Generation, Radial Basis Functions, SHAP, Machine Learning, Data Scarcity, Aluminum AlloysSubject terms: Engineering, Materials science, Mathematics and computing Introduction Friction Stir Additive Manufacturing is a systematic solid-state additive manufacturing technology that presents several benefits over traditional fusion-based processes, particularly for manufacturing dissimilar aluminum alloy structures1,2,17,26,29. Friction Stir Additive Manufacturing (FSAM) is synonymous with Additive Friction Stir Deposition (AFSD) and refers to the layered deposition variant of friction stir welding technology; in this work, FSAM is used as the primary terminology. FSAM is based on Friction Stir Welding (FSW) technology and has the capability of producing large-area, multi-layered components through sequential layered deposition with the same principles applied in FSW and therefore does not involve melting and solidification defects that are typically present in additive-manufacturing processes that rely upon melting for material consolidation2,3,5,17,26. FSAM has received significant interest from the automotive industries due to its ability to produce components with mechanical properties, reduced residual stress, microstructural features, improved production rates, and material utilization over competing processes1,5,6,26,28. In industrial settings, real-scale FSAM experiments face limitations of machine costs, cycle time, and safety. Therefore, this paper presents a surrogate-assisted predictive framework using synthetic data for baseline predictive capability and the integration of real data from future FSAM production to aid stakeholder trust4,19,20. The essential principle of FSAM involves using a rotating tool that generates sufficient frictional heat to plastically deform the material without melting it, operating at 60–90% of the melting temperature. Because this is a solid-state process, it provides several unique benefits over additive manufacturing based on fusion processes, such as no porosity, hot cracking defects, or solidification defects, as well as producing an equiaxed, fine grain structure that provides mechanical properties6,11,13,17,26. The technology is particularly beneficial for manufacturing components from high-strength aluminum alloys such as AA2XXX and AA7XXX, which tend to be regarded as non-weldable with traditional fusion-welding techniques2,12,21,29. Recent advancements of the FSAM have demonstrated the capability to create components with trends consistent with microstructural refinement reported in prior literature and improved mechanical properties. Hassan et al.1 provided a comprehensive review that discusses the mechanical properties compared to several existing melting techniques available in the market and the ability of FSAM to create large-area multi-layered components with equiaxed grain structure. Airbus1,4,6, and the demonstration of FSAM to make wing ribs out of Al-Li 2025 alloys have shown that the technology is being used by major manufacturers. It is environmentally friendly and cost-effective, and it bonds layers together very well. Venkit and Selvaraj16 offered systematic procedures to manufacture aluminum-based alternately layered composite materials using FSAM, leading to improved mechanical properties through favorable placement of the AA6061 and AA7075 alloy layers. FSAM process parameters interact with tool geometry and material flow, creating coupled effects difficult to predict analytically4,13,17,27,29. Exploring process parameters typically relies on traditional methods of experimentation, which can add to the time and costs of exploration, especially regarding how the total process parameter space will be explored for dissimilar material combinations2,3,26,28. Data-driven methods that leverage machine learning practices may offer solutions to these issues by accelerating the exploration process and enhancing our understanding of process-property relationships8–10,18,21. The use of machine learning in FSAM has grown rapidly recently, and many researchers have started to examine multiple techniques in predicting and improving the outputs of a process-driven establishment. For instance, Mishra et al.8 used supervised machine learning based on physics-informed neural networks to predict peak temperature distributions in an additive friction stir deposition (FSAM). They showed the advantage of both a data-driven and physics-based modelling approach. On the other end, Meyghani et al and Chen et al.4,19, in addition to their machine learning approaches, presented complicated finite element models to simulate FSAM processes, ultimately allowing a faster understanding of the thermomechanical nature of the process and possible optimization of them. Jin et al and Das et al.7,28 also developed gradient-based process parameter optimization, allowing much faster exploration of the parameter space during FSAM processes. The inclusion of explainable artificial intelligence (XAI) methods has become particularly relevant in manufacturing applications with the understanding of parameter contributions to aid process control and optimization. Recent advances by Hariharan et al.14 in alloy optimization using machine learning in conjunction with high-throughput experiments and Mishra9 in mapping process parameters to FSAM deposition results with unsupervised machine learning algorithms demonstrate the emerging necessity for interpretable machine learning models within manufacturing process optimization22. One of the primary hurdles to applying effective machine learning models is acquiring adequate high-quality training data for FSAM optimization. FSAM experiments face data-generation limitations due to cost, time, and resource constraints7,13,14,27. As a response to limited experimental data, researchers have turned to synthetic data generation approaches to supplement experimental datasets while maintaining surrogate consistency and relevance8,10,19,20. Radial Basis Function (RBF) interpolation has shown excellent results as a synthetic data generation method within manufacturing experiments; its smooth interpolation provides a representation of experimental data sets’ non-linear relationships9,19,23–25. Recent developments in solid-state additive manufacturing processes have demonstrated the ability to produce components with highly specific microstructures and mechanical properties. Dong et al and Liu et al.6,17 provided an in-depth review of recent advances in additive friction stir deposition (FSAM) technology and validated its capability to produce fine, equiaxed recrystallized grains and mechanical properties comparable to forged components. Yaknesh et al and Nouranga et al.11,26 also studied the performance and microstructure evolution of solid-state fabrication through friction stir processes and presented properties in solid fabrication processes when compared to fusion-based processes. Srinivasan et al.15 further expanded on this by exploring composite FSAM structures for applications. The field of the application of machine learning for the design of aluminum alloys and their optimization has gained support, and studies indicate very favorable potential in this area. Bhat et al.10 effectively used machine learning approaches as unsupervised learning to segment the dataset of 1154 aluminum alloys into different classes of aluminum alloys to show that machine learning can be a beneficial application for data-informed alloy classification and design purposes. By using an iterative label spreading (ILS) algorithm, Bhat and his team could identify eight distinct classes of aluminum alloys based on their processing conditions and composition that could act as the basis for data-informed progressive rational alloy designs14,18. Recent research has also emphasized integrated frameworks combining multiple machine learning methods for complete process optimization. Shao et al.5 introduced a unified ‘Process-Structure-Property’ modeling framework for FSAM that provides a pathway to enable surrogate-assisted exploration of parameter–property relationships with the fewest experiments to get a predictive measure determined. This example shows the possibilities of developing integrated modeling frameworks capable of taking meaningful parameter process-structure-property windows into account, while simultaneously exhibiting relevant accuracy18–20. The application of biomimetic machine learning methods appears to show potential to address many needs affiliated with additive manufacturing environments, especially basic research environments where new battery materials are developed and analyzed in addition to having FSAM CAD geometry generated for them. Their most recent research relating to FSAM-processed aluminum alloy structures has established an entirely new biomimetic ML approach for predicting mechanical properties pre- and post-optimization (namely genetic algorithm optimization); specifically, the performance had a learned mode of machine learning (genetic algorithm + Gaussian regression), and the performance has significantly surpassed conventional machine learning modes9,14,18,22. Recent advances demonstrate the systematic potential of combining machine learning with evolutionary optimization algorithms for materials design and process optimization. While FSAM technology is relatively new to machine learning applications, significant potential exists for process optimization. Primarily, the literature covers single-material cases or geometrically simple cases; applications to multicomponent dissimilar material combinations are highly limited11,26,29. Contemporary machine learning applications rely heavily on datasets of relatively large size that restrict the application of the methods to circumstances having sparsity in the dataset, and these latter sparsity conditions are common in advanced manufacturing-related research8,10,20. Further, there is a lack of frameworks for implementation of explainable AI and synthetic data generation methods within a comprehensive framework that can provide predictive capability and advance the understanding of processing9,14,19. The experimental design was based on a Taguchi L9 orthogonal array corresponding to a three-factor, three-level design. This approach enables efficient exploration of the interior design space and identification of dominant trends with a limited number of experiments. The nine experimental points were used as anchors for surrogate modeling, with all conclusions primarily anchored to the experimentally sampled parameter range, and any trend observed beyond 2.0° represents surrogate-based extrapolation requiring independent validation7,28. This research will make an important contribution in advancing the literature landscape through a clearly identified comprehensive framework to optimize FSAM parameters while uniquely considering dissimilar aluminum alloy systems (6061, matrix: 7075, reinforced) and specifically advance the research on data-scarcity applications via synthetic data. The method employs an integrated approach to leverage RBF-based synthetic data generation, multiple machine learning regression approaches, and SHAP-based explainable AI analysis to develop a unique representation of the relationships between the processing parameters during the generation of data with high predictive capability12,18,22,24. This work focuses on the immediate aspects of FSAM applications and provides a framework that can be generalized to address scarcity concerns when progressing through manufacturing processes. The combination of synthetic data generation with explainable AI methodologies provides a model that can inform similar works performed across different manufacturing domains within an experimental framework where data collection was limited14,18,19. The latest advance in this methodology is the way to enlarge data space while simultaneously ensuring preservation of physical constraints under resource-limited experimental conditions8,10,20. The methodology presented in this research presents several contributions to the intelligent manufacturing domain: (1) a new synthetic data generation methodology for FSAM interpolation design, (2) a comprehensive examination of multiple machine learning regression methods for predicting the mechanical properties of dissimilar aluminum alloy FSAM processes, (3) SHAP-based explainable AI for understanding which parameters contribute to the process for closing the loop on process understanding, (4) an integrated framework that handled data scarcity while maintaining high-level predictive capability that is interpretable9,14,19,22. The following sections provide a detailed explanation of the main workflow utilized, which includes experimental data preparation, synthetic data generation using RBF, implementation and evaluation of several machine learning models, and SHAP analysis for assessing parameter contributions. The results demonstrate the effectiveness of the workflow in achieving high predictive accuracy and providing the ability to gain actionable insights related to the FSAM optimization process, laying the groundwork for intelligent manufacturing systems that can be easily adapted or extended to industrial applications of automotive and other advanced manufacturing4–6,18. Materials and methods Workpiece material and specimen preparation In this research two types of precipitation-hardened aluminum alloys: AA6061-T6 and AA7075-T6. These materials are commonly used for structural applications because they represent completely different alloy chemistries, enabling us to perform unique analyses property gradients generated during the FSAM process. The approach taken in stacking and processing these two alloys is like other published successful processes for constructing aluminum-based layered composites using FSAM16. The elemental and compositional analyses of the two aluminum alloy systems were conducted using an XRF spectroscopic analysis19. The AA6061-T6 alloy consisted of approximately 98.1% of aluminum, magnesium (0.6% by wt.), silicon (0.34% by wt.), iron (0.45% by wt.), chromium (0.17% by wt.), zinc (0.14% by wt.), copper (0.12% by wt.), and manganese (0.09% by wt.) were the constituent elements. For the AA7075-T6 alloy, the major alloying elements were 4.95% by wt. zinc, 2.42% by wt. magnesium, 1.20% by wt. copper, 0.26% by wt. iron, 0.23% by wt. chromium, 0.18% by wt. manganese, and 0.08% by wt. silicon, with aluminum comprising approximately 90.68% by wt. for this type of aluminum alloy. Both AA6061-T6 and AA7075-T6 were supplied in the T6 temper, which combines a solution heat treatment and an artificial aging treatment16. Both alloys were supplied as 200 mm × 150 mm × 3 mm thick initial workpieces in an “as received” rolled state with no prior modifications made (e.g., cutting). They were positioned so that AA6061-T6 and AA7075-T6 were placed together, thus providing for a layered composite material. Each time friction stir lap welding was performed, flash was removed with a fly cutter, and cleaning followed by ethanol cleaning & drying was done to create a clean interface for metallurgical bonding between layers. The result after seven alternate layers is an approximate thickness of 18 mm. Tool design and specifications for friction stir additive manufacturing The fabrication of all friction stir lap welding operations used a non-consumable tool made of H13 tool steel. The tool’s pin geometry was designed with a conical shape and threading to improve the flow of the stir zone material and the mixing of interlayers during friction stir lap welding. Shoulder diameter is equal to 24 mm with a flat surface profile; no intentional concavity in the shoulder was created. However, due to the way the process parameters were selected, only a small amount of concavity was observed in the deposited layers, and none was found on the shoulder surface. The diameter taper of the pin is from 8 mm at the shoulder interface to 6 mm on the pin tip. Pin length is 4 mm. Plunge depth is kept at 4.10 mm to provide sufficient penetrability for the underlying plate being lap welded. The tool tilt angle varied at three levels: 1.0°, 1.5°, and 2.0° as part of the Taguchi L9 design, and an axial force of 20 kN was used to promote good consolidation and to minimize the number of defects, such as voids or excessive flash. The pitch and taper angles of the thread have not been provided. Friction stir additive manufacturing process parameters and layer deposition strategy No concavity was added as an intentional feature, and minimal concavity was observed in the completed layers of weldment. This minimal concavity was caused by selecting appropriate process parameters. The diameter of the pin tapered from 8 mm to 6 mm from where it interfaces with the shoulder down to the tip of the pin. The total length of the pin is 4 mm. The plunge depth of the tool into the underlying plate during lap welding was held constant at 4.1 mm to ensure enough penetration into the underlying plate. The tool tilt angle varied at three levels (1.0°, 1.5°, and 2.0) and the axial load on the tool used to provide a means of compressing or consolidating the deposited material to reduce the chance of creating defects such as voids or excessive flash was set at 20 kN. The specific thread pitch and taper angle were not indicated. A build thickness of approximately 18 mm has been constructed by using seven alternating layers, with each layer contributing about 3 mm of thickness. For the single-pass approach, a single friction stir lap-weld pass was completed on the centerline of the plate. For the double-pass approach, two passes were completed in opposite directions, and each pass had an offset of 2 mm from the previous pass, which resulted in a total of 4 mm offset for both passes. This method increased the amount of material that was mixed within the stir zone because it created two sides that retreated and one side that advanced16. All layers were deposited using standard laboratory ambient temperature without any externally applied heat or forced cooling; all layers cooled naturally in air after deposition. There was no post-deposition heat treatment, solution heat treatment, or any other form of artificial aging used for this build. The top surface of the build was milled down to achieve a flat surface for subsequent mechanical and microstructural testing. Tensile testing specifications Testing of the tensile properties was completed in accordance with Standard ASTM E8 for metals. Rectangular sub-size test samples were prepared from the additive manufactured products made by the Friction Stir Process, with the long axis of all test pieces aligned parallel with the direction of welding. The test samples were taken from the stir zone of the sample at three vertical locations (bottom, middle, and top) relative to the length of the sample. All tensile tests were completed at room temperature on an INSTRON universal testing machine (model INST-8801) using a constant crosshead displacement rate of 1 mm/min. For each slice location, three replicates were tested to confirm reproducibility. The load vs. displacement data were taken continually throughout the test and subsequently converted to engineering stress-strain curves for further analysis. Vickers microhardness testing specifications The Vickers microhardness tester (Model HM-200) was used to measure microhardness using a diamond pyramid indenter. A 200-gram force was exerted on the specimen for approximately 15 s, and results are reported in Vickers hardness values. Measurements were taken for both horizontal and vertical orientations on the surface of the deposited material. The horizontal measurements were taken on the interface (boundary) between two layers of material, while the vertical measurements were made vertically through the thickness from the bottom to the top of the deposited layers. The indentations used to create a profile of the hardness values were separated by approximately 0.5 mm. By having these designated spacing areas, it is possible to develop precise and detailed hardness profiles of the heat-affected and stir zones. Because there were no reported uncertainty values in this study, the consistent hardness trends across multiple samples and the even shape of profile lead to the conclusion that the results are highly reproducible, with minimal variation between samples. All tensile and hardness data were taken from the same area of the FSAM builds and all tested with a consistent loading direction. Each condition measured represents a mean response from the various test samples, while any variability for that condition is noted. This sampling technique was designed to allow for accurate comparisons between each process condition while reducing geometric and positional variability. Hardness measurements were obtained at 0.5 mm intervals along both horizontal interlayer boundaries and through-thickness vertical sections to ensure representative sampling of stir zone and interface regions. Data compilation and quality assurance All mechanical and microhardness data, including ultimate tensile strength, elongation, process parameters, and slice positions, were systematically recorded and compiled. Tensile results represent mean values obtained from three replicate specimens per slice, while hardness profiles were generated from multiple indentations across each layer. Although formal statistical outlier rejection procedures were not applied, the reproducibility of tensile results and the consistency of hardness distributions indicate reliable data quality. The final dataset comprised base material properties and comprehensive characterization of friction stir additive manufacturing builds, including tensile behavior, hardness distributions, and associated microstructural analyses such as scanning electron microscopy, electron backscatter diffraction, energy-dispersive spectroscopy, and X-ray diffraction, enabling robust assessment of structure–property relationships in alternating AA6061-T6 and AA7075-T6 composite structures16. Scope and limitations of microstructural interpretation This current research uses proxy methods to estimate and interpret the macroscopic mechanical response of dissimilar AA6061/AA7075 friction stir additive manufactured materials when there are very few available experimental results: ultimate tensile strength (UTS) and Vickers hardness. A comprehensive microstructural characterization of the alloy layers’ interfaces and their respective microstructural characteristics (i.e., grain shape and structure, precipitate size and shape, diffusion patterns, and intermetallic layer structure) was not performed in this research; thus, these areas of the production process are excluded from consideration within the current scope of the research. The microscopic structure in a multilayer dissimilar aluminum composite product will have an influence on its mechanical properties, and this research is not intended to define the relationship between mechanical properties; rather, the aim of this research is to demonstrate how to combine limited experimental results with data generated through interpolation and model-based learning, then utilize explainable ML methods to identify trends within the experimentally sampled design space. Future studies will include an explicit integration of microstructural and compositional information that will provide a stronger physical understanding of the results and further validate the proposed framework. Because the models are trained primarily on RBF-interpolated samples derived from nine experimental measurements, the results reflect surrogate response surface behavior within the sampled parameter space. Independent experimental validation is necessary to confirm the predicted optima and associated mechanical improvements. Results and discussion The experiment outlined in Table 1 presents parametric surrogate-assisted optimization together with mechanical property experimentation for friction stir additive manufacturing (FSAM) processes of dissimilar aluminum alloys, focusing on the enhancement of mechanical properties of AA6061 and AA7075 alloy systems. The experimental matrix is carried out with an established three-level factorial design with nine levels of experimentation (noted A1 to C3) for examining the effects of three basic process parameters on mechanical properties. Three levels of tool rotation speed (TRS) were tested at 1000, 1100, and 1200 rpm; tool traverse speed (TTS) was evaluated at 30, 35, and 40 mm/min; and tool tilt angle (TTA) was evaluated at 1.0°, 1.5°, and 2.0°. Two additional configurations for the materials were also considered, 6061-over-7075 and 7075-over-6061, with the mechanical properties being investigated at level I: ultimate tensile strength (UTS) testing and level II: Vickers hardness (HV) testing. Table 1Experimental parameters and results for 6061 and 7075 aluminum alloys.Exp. NoTRS (rpm)TTS (mm/min)TTA (deg)6061 over 7075 UTS (MPa)6061 over 7075 Hardness (HV)7075 over 6061 UTS (MPa)7075 over 6061 Hardness (HV)A11000301.0287104.5291112.2A21000351.5295116.9297120.3A31000402.033087.1325115.1B11100301.5318104.9307135.2B21100352.0326120.0320125.3B31100401.030869.532297.6C11200302.0320109.8336127.6C21200351.029087.5326105.0C31200401.529898.4330115.0 Experimental parameters and results for 6061 and 7075 aluminum alloys. The results from these experiments showed a significant amount of parametric sensitivity from stacking configuration to configuration, with UTS reported between 287 and 330 MPa for the 6061-over-7075 configuration and 291 and 336 for material sequence 7075-over-6061. Vickers’ hardness showed a similar variability, ranging from 69.5 to 120.0 HV for the 6061-over-7075 stacking sequence and 97.6 to 135.2 HV for the 7075-over-6061 stacking sequence. The data also indicate that the 7075-over-6061 stacking sequence was significantly higher in mechanical properties, reaching a UTS of 336 MPa in experiment C1 and a hardness value of 135.2 HV in experiment B1, substantiating the material stacking sequence as an important factor in FSAM results. The nine experimental points that provided the data for the information above are a scarce number of experimental observations and are typical of the advanced manufacturing field given the confines of resources, experimental complexity, and time constraints for complete observation of parameters. The limited observational data justifies using synthetic data for high surrogate consistency within the interpolated design space, which will be integrated into the RBF-based synthetic data generation framework, i.e., the mathematical interpolant. A comprehensive synthetic data generation framework to overcome the inherent data scarcity challenges in interpolated design space was developed, based solely on this experimental dataset of nine data points. The synthetic data augmentation approach used Radial Basis Function (RBF) interpolation employing multiquadric basis functions and a smoothing parameter of 0.1 to produce 882 synthetic samples. Overall, a much wider parameter space for developing machine learning models was created. The synthetic data grid had a coarser discretization scheme with TRS values that increased at 10 rpm increments from 1000 to 1200 rpm, TTS values increasing at 2 mm/min increments from 30 to 40 mm/min, and TTA values that increased at 0.2-degree increments from 1.0 to 2.0 degrees, forming a three-dimensional mesh grid that encompassed the experimental parameter space and extended slightly beyond the experimental space to allow extrapolation. Interpolation continuity is imposed by the RBF method through smooth transitions between the 9 experimental points; however, this smooth transition feature is not a guarantee of surrogate consistency but is rather a characteristic of the mathematical model representing that interpolation. The physical relationships between the thermomechanical properties of FSAM will only accurately depict the relationship between thermomechanical properties and the true process-properties of FSAM to the degree that the 9 experimental data points accurately portray the true process-properties of FSAM. Distinct RBF models were trained for each output variable, UTS and HV, corresponding to each material configuration; these models used the original experimental coordinates as interpolation nodes, and the multiquadric function was chosen for its ability to capture the nonlinear relationships between process properties inherent in solid-state manufacturing processes. The subsequent synthetic data generation preserves smooth interpolation continuity between experimentally measured points, enabling construction of an enriched training dataset consistent with the observed experimental trends, while also substantially decreasing the data space required for developing rigorous machine learning models. To generate a denser mesh for ML model training within the explored design space, RBF interpolation was applied to create synthetic points between and near the 9 experimental measurements. The resulting 882 synthetic samples augment the original 9 points to form an 891-point dataset for training purposes. Critically, these synthetic points represent interpolated values from the RBF surrogate function, not independent experimental measurements. Models trained on this synthetic dataset effectively learn the RBF response surface and the model captures the surrogate response surface within the explored design space. This is a pragmatic approach to the data-scarcity problem but should not be misinterpreted as a generalization to unmeasured real experimental conditions. Synthetic data generation via RBF interpolation Table 2 shows a comparison of the predictive capability of three machine learning regression models tested with the first 15 of the synthetic data points created with RBF interpolation to produce a prediction of ultimate tensile strength (UTS) for the 6061-over-7075 material pair. The synthetic data points encompass a parameter space defined by tool rotation speeds (TRS) of 1000–1020 RPM, tool traverse speed (TTS) at 30 mm/min, and tool tilt angles (TTA) of 1.0–2.2° at increments of 0.2°. The UTS values for the synthetic data points range from 287.429 to 294.553 MPa, indicating the expected progression of increasing tensile strengths as tool rotation speed is increased while holding TTS and TTA constant. Table 2First 15 synthetic data point predictions for UTS with 6061 over 7075.TRSTTSTTAUTS (6061 over 7075)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1000301.0287.429388287.2348670.07286.7393990.24305.979735−6.451000301.2287.475901287.4759010.00286.7865290.24306.022863−6.451000301.4287.542425287.5424250.08286.8897710.23306.065991−6.441000301.6287.628943287.6289430.00286.9403800.24306.109118−6.431000301.8287.735369287.7353690.08287.0715170.23306.152246−6.401000302.0287.861551287.8615510.08287.1698640.24306.195374−6.371000302.2288.007272288.0072720.08287.3183910.24306.238501−6.331010301.0288.621252288.6212520.08290.671105−0.71305.978229−6.011010301.2288.673097288.6730970.08290.888368−0.77306.021357−6.011010301.4288.733663288.7336630.08291.071463−0.81306.064484−6.001010301.6288.803203288.8032030.08291.143664−0.81306.107612−5.991010301.8288.881601288.8816010.08291.223035−0.81306.159740−5.981010302.0288.968819288.9688190.08291.202171−0.77306.193867−5.961010302.2289.064810289.0648100.08291.144684−0.72306.236995−5.941020301.0294.553038294.5530380.08294.1520730.14305.976722−3.88 First 15 synthetic data point predictions for UTS with 6061 over 7075. The comparison highlights the different performance metrics for the three proposed regression models. Gaussian Process Regression (GPR) demonstrates near-zero error relative to RBF-interpolated surrogate values and exhibits total errors lower than 0.1% in some synthetic data examples, particularly in the TRS experiments at 1000 rpm for TTA of 1.2°, 1.4°, and 1.6°. The largest error associated with GPR is 0.08% at the high parameter levels evaluated, indicating the model was able to reproduce the baseline structure of synthetic data patterns with fidelity while being clearly observable in the interpolating (RBF) functions. The Support Vector Regression (SVR) model functions moderately in prediction capacity, and average overall error rates relative to the measurement ranged from 0.14% to − 0.81%. Furthermore, SVR had some tendency to underpredict the TRS responses with mostly negative rates of error. Overall, Support Vector Regression appears to provide relatively consistent prediction observation over the parameter space studied, with overall rates of error remaining relatively low and mainly below 1%. Conversely, the prediction errors of the Bayesian Ridge Regression are relatively high compared to the other methods, which have error percentages across the range of −6.45% to −3.88%, and consistently overpredict UTS values for the synthetic data points. The behaviors exhibited by the Bayesian Ridge model suggest that the model may have limitations in capturing the non-linear relationships that exist within the FSAM process, given it is trained on a minimal or limited original experimental dataset and then trained or applied to the synthetic data collected from the RBF interpolation. The findings indicate that the GPR model yields predictions for synthetic data points and justifies its choice as the primary model for interpolated design space. The prediction accuracy achieved by GP regression is derived largely from the ability of GP regression to capture uncertainty and to model nonlinear relationships with probabilistic predictions that match the structure of synthetic data generated using RBF. The continued underestimations of SVR and continued overestimations of Bayesian Ridge Regression highlight the importance of model selection in applications using synthetic data as it relates to the ability of an algorithm to effectively represent the interpolated relationships defined by the RBF. This comparison provides the foundation for carrying out full model evaluation across the entire synthetic dataset, showing the GPR-based model provides the most consistent reproduction of the surrogate response surface within the explored parameter space with the original experimental observations. The findings also validate that the assembled RBF-ML framework can be an effective response to challenges around data scarcity in advanced manufacturing process optimization. The predictive performance of each of the three machine learning regression models on the final fifteen synthetic data points in the upper parameter range for the UTS prediction model from the 6061-over-7075 material configuration is shown in Table 3. These synthetic data points were representative of the high-parameter regime where TRS was between 1180 and 1200 rpm, TTS was at the maximum tested value of 40 mm/min, and TTA was between 1.0° and 2.2° in increments of 0.2°. The UTS for these synthetic samples was between 294.340 and 298.739 MPa, which is comparatively stronger than the lower parameter regime discussed in Table II, thus establishing a possible association between process parameters and mechanical properties in the FSAM process. Table 3Last 15 synthetic data points predictions for UTS with 6061 over 7075.TRSTTSTTAUTS (6061 over 7075)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1180402.2298.738910298.7389100.00299.213181−0.16306.499570−2.601190401.0294.220477294.2204770.00297.049844−0.96306.239298−4.081190401.2294.340534294.3405340.00297.136145−0.95306.282426−4.061190401.4294.468409294.4684090.00297.249649−0.94306.325553−4.031190401.6294.604030294.6040300.00297.358176−0.93306.368681−3.991190401.8294.747318294.7473180.00297.514614−0.94306.411809−3.961190402.0294.898188294.8981880.00297.673998−0.94306.454936−3.921190402.2295.056545295.0565450.00297.815225−0.93306.498064−3.881200401.0296.057049296.0570490.00295.3685830.23306.237792−3.441200401.2296.223321296.2233210.00295.5334050.23306.280919−3.401200401.4296.404247296.4042470.00295.7415540.22306.324047−3.351200401.6296.599564296.5995640.00295.9125350.23306.367175−3.291200401.8296.808978296.8089780.00296.1195340.23306.410302−3.231200402.0297.032174297.0321740.00296.3446370.23306.453430−3.171200402.2297.268814297.2688140.00296.5786970.23306.496557−3.10 Last 15 synthetic data points predictions for UTS with 6061 over 7075. The comparative evaluation indicates the presence of ongoing tendencies in predictive performance across the three regression models in the upper parameter space. When exploring the upper parameter regime, Gaussian Process Regression (GPR) achieves predictive performance for all 15 synthetic data points, which are near zero, indicating refined consistency and robustness of the model corresponding to the entire parameter space. The model’s predictive performance indicates that it has the capacity to characterize the underlying RBF interpolation structure, reinforcing its reliability for parameter optimization applications across multiple processes and process conditions. Support Vector Regression (SVR) continues to show a global pattern of consistent underprediction with error percentages comprised between 0.22% and − 0.96%, retaining ordered underprediction behavior assigned to the lower parameter regime. When comparing SVR to the lower parameter regime, the model shows improved modeling accuracy in the high-parameter range, where absolute error values have generally decreased, and performance is more stable across tilt angle configurations. The maximum underprediction error of −0.96% occurs at TRS 1190 rpm with TTA 1.0°, while the minimum underprediction error of 0.22% occurs at TRS 1200 rpm with TTA 1.4°. Bayesian Ridge Regression still shows substantial overprediction effects, with corresponding error percentages ranging from − 2.60% to −4.08%, and systemic bias in the high parameter regimes. The Bayesian Ridge model has slightly improved performance in the upper parameter regime than in the lower regime with less maximum error (−4.08% in the upper regime and − 6.45% in the lower regime in Table II). Nonetheless, the higher error percentages still indicate that the Bayesian Ridge model overpredicts from the model across all points, demonstrating inherent value in the model’s capacity to represent the FSAM process-property relationship established by RBF interpolation. The findings reveal a good level of consistency in the performance ranking of the models across various parameters and regimes, with GPR achieving near-zero error relative to the RBF-interpolated surrogate values, SVR demonstrating a systematic but moderate underprediction, and Bayesian Ridge showing significant overprediction. After training, the improvement in performance of both SVR and Bayesian Ridge models under the high parametric regimes indicates that the models converge better and exhibit more stability in high-process conditions. The performance of GPR’s ability to achieve near-zero error on every synthetic data point in all of the parameter regimes indicates that GPR is the better model for interpolated design space and illustrates the power of a combined RBF-GPR framework. The comparative predicted performance of three machine learning regression models for predicting Vickers hardness (HV) for the 6061-over-7075 material configuration is presented in Table 4 for the first 15 synthetic data points. The synthetic data points lie within the lower parameter regime, having tool rotation speeds (TRS) between 1000 and 1020 rpm, tool traverse speed (TTS) at 30 mm/min, and tool tilt angle (TTA) between 1.0° and 2.2° in increments of 0.2°. The HV values for these synthetic samples ranged from 93.915 to 108.097 HV, which shows that there is a wide variation in hardness properties across the evaluated parameter space and illustrates the complexity in the relationship between the process parameters and material hardness in the FSAM processes. Table 4First 15 synthetic data points predictions for HV with 6061 over 7075.TRSTTSTTAHV (6061 over 7075)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1000301.0107.747258106.2098401.43106.9860640.7194.66379212.141000301.2107.812852107.8128520.00108.078428−0.2594.81559012.061000301.4107.875836107.8758360.00108.384327−0.4794.96738911.971000301.6107.935994107.9359940.00108.445777−0.4795.11918711.871000301.8107.993099107.9930990.00108.611290−0.5795.27098611.781000302.0108.046917108.0469170.00108.809349−0.7195.42278411.681000302.2108.097218108.0972180.00108.532766−0.4095.57458311.581010301.0100.201443100.2014430.0099.7113020.4994.9326325.261010301.2100.241395100.2413950.00100.514710−0.2795.0844305.141010301.4100.280021100.2800210.00100.877373−0.6095.2362285.031010301.6100.317300100.3173000.00100.986848−0.6795.3880274.911010301.8100.353220100.3532200.00101.006309−0.6595.5398254.801010302.0100.387767100.3877670.00101.011059−0.6295.6916244.681010302.2100.420933100.4209330.00100.886970−0.4695.8434224.561020301.093.91505093.9150500.0093.8436440.0895.201471−1.37 First 15 synthetic data points predictions for HV with 6061 over 7075. The comparative analysis shows different performance characteristics for all three regression models for hardness prediction that are uniquely different from the UTS prediction results. The Gaussian Process Regression (GPR) model achieved a predictive accuracy of approximately zero error for 12 of the 15 synthetic data points which included all of the tests completed at 1010 rpm TTA = 1.4° to 2.2°. There was a maximum error of 1.43% produced by the GPR model for the test at TRS = 1000 rpm TTA = 1.0°. However, overall performance continued to be demonstrated by an average absolute error well below 1%. The refined accuracy displayed the capability of the GPR model to capture complex, nonlinear relationships between process parameters and hardness properties, thus developing through RBF interpolation methods. The Support Vector Regression (SVR) model produced variable predictions of −0.71% to + 0.71%, indicating both overprediction and underprediction based on varying combinations of parameters. The SVR model was generally consistent with a higher predictive accuracy for hardness because the absolute errors were generally below 1% and exhibited moderate accuracy. Of particular interest was that the best prediction from the SVR model (0.08%) occurred for the boundary TRS = 1020 rpm. In general, the SVR model can deliver good predictive power at a variety of combinations of parameters. Bayesian Ridge Regression shows substantial consistent overprediction, where the error rate ranges between − 1.37% and 12.14%, consistently underpredicting the actual HV values across every synthetic data point provided. The prediction performance was particularly poor in the low range of TRS (1000 rpm), with maximum errors above 12% for several combinations of parameters. The systematic bias was less prominent in the larger range of rotation speeds, where the error fell to −1.37% at TRS 1020 rpm, demonstrating the limitations of predicting hardness values concentrically with parameters. The results showed that hardness values are much more difficult to predict than UTS values since the level of error was higher in all three models when predicting hardness values compared to UTS predictions, although all three models again demonstrated that GPR had excellent performance with most of its predictions being generated correct. GPR was therefore deemed to be the preferable choice for a model that could predict the totality of a material’s mechanical properties when produced via the FSAM process. There were varying levels of performance in both SVR and Bayesian Ridge Regression-plus an error that systematically biased predictions away from the average hardness value, indicating that models cannot be relied upon when predicting hardness values and selecting a model based on how well it fits for use in a multi-objective optimization scenario will be difficult. This assessment verified that using pure GPR provides a capable method of predicting hardness properties and is in accordance with the synthetic data architecture developed using radial basis functions. The ability to systematically assess results from multiple machine learning algorithms will boost confidence in the framework created with the intent of maximally enable surrogate-assisted exploration of parameter–property relationships for generating the tensile strength and hardness properties necessary to allow the design of improved components. Table 5 shows the prediction performance of three machine learning regression models on the last 15 synthetic data points in the upper parameter regime for Vickers hardness (HV) prediction from the 6061-over-7075 material configuration. These synthetic data points correspond to TRS values ranging from 1180 to 1200 rpm, TTS of 40 mm/min, and TTA of 1.0° to 2.2° in increments of 0.2°. The HV values ranged from 93.9976 to 96.7621 HV, which is consistent with the anticipated increases in hardness with higher process parameters typical of FSAM processes. Table 5Last 15 synthetic data points predictions for HV with 6061 over 7075.TRSTTSTTAHV (6061 over 7075)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1180402.294.58041994.5804190.0094.3716910.2288.6856956.231190401.093.99755993.9975590.0094.758368−0.8188.0437446.331190401.294.13000894.1300080.0094.912360−0.8388.1955426.301190401.494.26534594.2653450.0095.028283−0.8188.3473416.281190401.694.40350394.4035030.0095.163897−0.8188.4991396.251190401.894.54441094.5444100.0095.317111−0.8288.6509376.231190402.094.68799394.6879930.0095.471886−0.8388.8027366.221190402.294.83417594.8341750.0095.594844−0.8088.9545346.201200401.095.61284095.6128400.0094.8523700.8088.3125837.641200401.295.79391695.7939160.0095.0360380.7988.4643827.651200401.495.97953395.9795330.0095.2199060.7988.6161807.671200401.696.16941996.1694190.0095.4123170.7988.7679797.701200401.896.36331296.3633120.0095.6021460.7988.9197777.721200402.096.56095796.5609570.0095.8017310.7989.0715757.761200402.296.76211396.7621130.0095.9997080.7989.2233747.79 Last 15 synthetic data points predictions for HV with 6061 over 7075. The comparison shows that GPR predicted all samples in the hardest parameter range better than SVR, while indicating that GPR could accurately model the nonlinear relationships present in the dataset produced by RBF simulation. Additionally, GPR’s predicted synthetic HV values were in close agreement with each other and the results of testing of the synthetic HV values produced by RBF simulation, thus demonstrating that GPR is the highest surrogate fidelity among evaluated models to be used in FSAM surrogate-assisted multi-property optimization. SVR tends to produce a relatively consistent negative bias, as indicated by the errors ranging from − 0.81% to 0.22%. The predictive performance of the SVR model was only marginally acceptable, as indicative trends by the highest measured error at TRS 1180 rpm (with TTA 2.2°) of 0.22%, and the lowest measured under-prediction of −0.83%, occurring at TRS 1190 rpm (with TTA 1.2°). The way SVR distributed its prediction errors throughout the parameter space indicates that SVR modeled hardness relatively consistently and with stable performance for predicting hardness in the context of a high parameter regime, but did so with a lower level of precision than the other modeling approaches tested. Bayesian Ridge Regression exhibited systematic prediction errors, or under-predictions, with error percentages ranging from 6.20% to 7.79%. Systematic under-prediction aligns with an overall consistency and bias towards under-prediction of hardness predictions for synthetic samples across all high-parameter conditions. The catastrophic error of the Bayesian Ridge Model was 7.79%, as observed at TRS 1200 rpm, with a TTA of 2.2°. The capability of the model to encapsulate the nonlinear relationships between hardness for synthetic samples was limited, as predicted values generally deviated significantly from the synthetic ground truth values employed in this research, which can be observed through the significant depth of error calculated through Bayesian Ridge Regression methods. GPR achieves near-zero error across all synthetic hardness samples in the high-parameter regime, emphasizing its prediction performance compared to SVR and Bayesian Ridge Regression. The fact that SVR consistently underpredicted hardness and that the Bayesian Ridge exhibited bias (over-prediction at low and under-prediction at high hardness) has demonstrated the significance of model selection in the multi-objective interpolated design context with hardness property prediction objectives. These results support the use of the integrated RBF-GPR framework for surrogate-assisted exploration of FSAM parameter–property trends. Table 6 details the predictive performance of three machine learning regression models on the first 15 synthetic data points regarding UTS prediction for the 7075-over-6061 material combo, in which the synthetic data points are in the lower parameter regime with TRS from 1000 to 1020 rpm, TTS of 30 mm/min, and TTA between 1.0° and 2.2° in 0.2° intervals. The UTS values of these synthetic samples vary from 290.83 to 291.74 MPa, which indicates the mechanical properties of the reversed stacking sequence as process conditions are altered. Table 6First 15 synthetic data points predictions for UTS with 7075 over 6061.TRSTTSTTAUTS (7075 over 6061)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1000301.0291.354900291.1605430.07290.4577800.31296.126744−1.641000301.2291.376451291.376451−0.00290.3776650.34296.187280−1.651000301.4291.415341291.4153410.00290.5202520.31296.247815−1.661000301.6291.471559291.4715590.00290.5726620.31296.308351−1.661000301.8291.545038291.545038−0.00290.6499790.31296.368886−1.651000302.0291.635657291.6356570.00290.6661880.33296.429422−1.641010301.0291.743244291.7432440.00290.8471300.31296.489957−1.631010301.2290.461799290.4617990.00292.194765−0.60297.479298−2.421010301.4290.496147290.496147−0.00292.238158−0.60297.539833−2.421010301.6290.538653290.5386530.00292.241143−0.59297.608369−2.431010301.8290.589298290.589298−0.00292.332057−0.60297.660904−2.431010302.0290.648053290.6480530.00292.399815−0.60297.721440−2.431010302.2290.714880290.714880−0.00292.598060−0.65297.781975−2.431020301.0294.830545294.8305450.00294.6538010.06298.710781−1.32 1020301.2294.853234294.8532340.00294.5492880.10298.771316−1.33 First 15 synthetic data points predictions for UTS with 7075 over 6061. The results of the comparison show that the Gaussian Process Regression (GPR) produces excellent predictive performance with an error percentage consistently less than 0.10%. As noted, GPR produces non-zero error for several of the data points, including TRS 1000 rpm and TTA 1.4°, 1.6°, and 1.8°, demonstrating the GPR’s ability to extract the interpolated relationships within the synthetic data collection of UTS. The maximum GPR recorded error was 0.07%, occurring at the edge condition of TRS 1000 rpm and TTA 1.0°. Support Vector Regression (SVR) indicates moderate prediction accuracy with error percentages ranging from − 0.65% to 0.34%. The model shows a systematic tendency to underpredict at lower values of TRS and slightly predict higher UTS values at the TTA data points; TRS 1010 rpm with TTA of 1.2° recorded the poorest model performance with a −0.65% error, and TRS 1000 rpm with TTA of 1.2° detected the model performance with a 0.34% error. The performance of the SVR model showed a degree of consistency that was corroborated across the synthetic data points under consideration. Bayesian Ridge Regression displays systematic overprediction with error percentage accounting for between − 1.32% and − 2.43%, suggesting predicting UTS values exhibiting a wide and consistent bias towards higher predicted UTS values to the total range of observed conditions. The largest error of Bayesian Ridge, −2.43%, is repeated at three individual data points that show prediction within a TRS in the 1010 rpm range, while the lowest predicted value of −1.32% occurs at TRS 1020 rpm with a TTA of 1.0. Overall, the overpredictive performance found in the Bayesian Ridge suggests an indicator that the model’s capacity to capture a nonlinear representation of UTS values is not observed through the observed UTS–parameter relationships. The findings show that while all three models can capture the overall UTS trends, GPR provides the most reliable predictions, with the lowest overall error, supporting its selection as the primary regression model for FSAM surrogate-assisted multi-property optimization. While SVR provides reasonable accuracy, there is a systematic bias, and Bayesian Ridge always shows increasing error, highlighting the importance of selecting the right algorithm for predictive modeling of FSAM processes. The purpose of this analysis was to demonstrate the efficacy of the RBF-GPR combined synthetic data and predictive modelling protocol for predicting the UTS of the 7075 on top of the 6061 configuration in preparation for future integrated optimization initiatives for FSAM processing. Table 7 summarizes the predictive consistency of 3 machine-learning regression models tested on the last 15 synthetic data points in the upper parameter regime of UTS prediction for the 7075-over-6061 material system. The synthetic values were based on TRS values of 1180 to 1200 (rpm), TTS of 40 mm/min, and TTA between 1.0° and 2.2° in 0.2° increments. The UTS values for the synthetic samples ranged from 321.8069 to 330.1272 (MPa), exhibiting increased tensile strength in the high-parameter operating conditions of the FSAM process. Table 7Last 15 synthetic data points predictions for UTS with 7075 over 6061.TRSTTSTTAUTS (7075 over 6061)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1180402.2321.806950321.8069500.00322.216200−0.13327.230888−1.691190401.0322.822066322.8220660.00324.872920−0.64328.159693−1.651190401.2322.853701322.8537010.00324.805082−0.60328.220229−1.661190401.4322.891178322.8911780.00324.837903−0.60328.280764−1.671190401.6322.934476322.9344760.00324.864574−0.60328.341300−1.671190401.8322.983570322.9835700.00324.915924−0.60328.401835−1.681190402.0323.038432323.0384320.00324.999210−0.61328.462371−1.681190402.2323.099027323.0990270.00325.167127−0.64328.522906−1.681200401.0329.726166329.7261660.00328.8296060.27329.4517120.081200401.2329.767782329.7677820.00328.8691510.27329.5122470.081200401.4329.819606329.8196060.00328.9382830.27329.5727830.071200401.6329.881568329.8815680.00328.9861530.27329.6333180.081200401.8329.953575329.9535750.00329.0532040.27329.6938540.081200402.0330.035514330.0355140.00329.1357610.27329.7543890.091200402.2330.127249330.1272490.00329.2299670.27329.8149250.09 Last 15 synthetic data points predictions for UTS with 7075 over 6061. The comparison suggests that Gaussian Process Regression (GPR) exhibited near zero error, with non-zero error on all 15 synthetic data points in the high parameter UTS space, thereby demonstrating the ability of the GPR model to closely reproduce the RBF-interpolated surrogate surface. The accurate UTS prediction of the GPR model throughout the entire parameter space emphasizes the application of the GPR model for surrogate-assisted exploration of parameter–property relationships of the FSAM process. Support Vector Regression (SVR) displays a consistent under predictive trend, with error percentages between − 0.64% and 0.27%. The SVR model also gives improved accuracy as compared to the low-parameter regime, with the lowest performance at TRS 1190 rpm with TTA 1.0° at −0.64% and the highest performance with 0.27% error at the 1200 rpm conditions subject to various TTAs. This illustrates a modest form of prediction ability of SVR, with slight bias, constructively in the high parameter regime. Bayesian Ridge Regression presents near-zero error percentages between − 1.69% and 0.09%, indicating significantly improved performance at the high parameter regime compared to the previously noted regimes. The highest underprediction, or maximum error, −1.69%, occurs at TRS 1180 rpm with TTA 2.2°, whereas the highest overprediction, or minimum, 0.09%, occurs at 1200 rpm with TTA’s (1.2° to 2.2°) values. In total, the margins of the errors also represent improved calibration of the model prediction of elevated UTS derivations using RBF interpolation. The findings affirm that GPR is the appropriate regression model for predicting UTS in the 7075-over-6061 configuration with near zero error at every synthetic data point. The enhanced performance of both SVR and Bayesian Ridge Regression in the upper parameter regime suggests higher process parameters can result in more stable model predictions. The repeated evaluation in different parameter regimes indicates the utility of the combined RBF-GPR framework and supports its applicability for generating surrogate-based multi-property predictions within the explored design space. The predictive performance of Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression on the first 15 synthetic data points for Vickers hardness (HV) in the 7075-over-6061 FSAM arrangement are presented in Table 8. All synthetic points correspond to the lower parameter domain: TRS values of 1000–1020 rpm, a constant TTS of 30 mm/min, and TTA angles of 1.0° to 2.2° in 0.2° increments. The true values of HV, predicted using RBF interpolation, range from 111.495 HV to 114.254 HV, suggesting the sensitivity of hardness to the process parameters. Table 8First 15 synthetic data points predictions for HV with 7075 over 6061.TRSTTSTTAHV (7075 over 6061)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1000301.0113.951285113.1594250.69113.4198110.47116.708939−2.421000301.2113.991404113.9914040.00113.6724300.28116.810313−2.471000301.4114.035851114.0358510.00113.7531100.25116.911686−2.521000301.6114.084511114.0845110.00113.7573760.29117.013060−2.571000301.8114.137244114.1372440.00113.8605040.24117.114434−2.611000302.0114.193890114.1938900.00114.0325330.14117.215807−2.651000302.2114.254268114.2542680.00114.1303950.11117.317181−2.681010301.0111.832741111.8327410.00112.068835−0.21117.003086−4.621010301.2111.860493111.8604930.00112.286384−0.38117.104460−4.691010301.4111.889994111.8899940.00112.422222−0.48117.205834−4.751010301.6111.921231111.9212310.00112.455881−0.48117.307207−4.811010301.8111.954188111.9541880.00112.487856−0.48117.408581−4.871010302.0111.988848111.9888480.00112.522331−0.48117.509955−4.931010302.2112.025194112.0251940.00112.531632−0.45117.611328−4.991020301.0111.495305111.4953050.00111.3191130.16117.297234−5.20 First 15 synthetic data points predictions for HV with 7075 over 6061. GPR predicting 14 out of 15 HV values with non-zero error and an adverse error of 0.69% at the lowest parameter (1000 rpm, 1.0°). This degree of consistency reveals GPR’s ability to capture the FSAM-process inherent nonlinear hardness-parameter relationships. SVR provided moderate predictive ability, with error margins reported between − 0.48% and 0.47%, indicating slight biases toward under- or overpredicting, depending on TTA settings; SVR exhibited the predictive accuracy of 0.11% error at 1000 rpm with 2.2° TTA. Alternatively, Bayesian Ridge Regression consistently underpredicted HV with error reports ranging from − 2.42% to −5.20%, with the highest error rates (for the interpolated hardness values) at the highest value of 114.254 HV, indicating its inability to model sharp nonlinearities exhibited in the synthetic dataset. In summary, these results indicate that GPR provides the most consistent surrogate fidelity among the evaluated models for predictive hardness in FSAM applications and provides adherence to measurements across multiple process parameters. SVR yielded acceptable performance for initial screenings; however, SVR exhibited bias in its predictions on behalf of the parameter settings. In addition, the substantive underprediction of hardiness exhibited by Bayesian Ridge indicates that it is unsatisfactory for accurately predicting hardness. Overall, the ongoing reliability demonstrated by GPR indicates functional feasibility for incorporation into the FSAM optimization framework to maintain the utility of data-driven parameter selection for manufacturable material properties. Table 9 displays the predictive performance of three machine learning regression models for predicting Vickers hardness (HV) on the last 15 synthetic data points in the upper parameter space for the 7075-over-6061 FSAM configuration. The first synthetic data point had a tool rotation speed (TRS) of between 1180 and 1200 rpm, with a tool traverse speed (TTS) of 40 mm/min, and a tool tilt angle (TTA) that changed from 1.0° to 2.2° with increments of 0.2°. The interpolated HV values ranged between 112.068 and 114.410 HV along the process parameters of these points. This result shows how the higher process parameters influence hardness. Table 9Last 15 synthetic data points predictions for HV with 7075 over 6061.TRSTTSTTAHV (7075 over 6061)Pred_GPRErr% GPRPred_SVRErr% SVRPred Bayesian RidgeErr% Bayesian Ridge1180402.2114.410342114.4103420.00114.945567−0.47113.0216251.211190401.0112.068325112.0683250.00113.366701−1.16112.707530−0.571190401.2112.176466112.1764660.00113.459090−1.14112.808903−0.561190401.4112.288207112.2882070.00113.605828−1.17112.910277−0.551190401.6112.403482112.4034820.00113.712638−1.16113.011651−0.541190401.8112.522221112.5222210.00113.783649−1.12113.113024−0.531190402.0112.644353112.6443530.00113.865775−1.08113.214398−0.511190402.2112.769801112.7698010.00114.038947−1.13113.315772−0.481200401.0112.489673112.4896730.00111.8138210.60113.001677−0.461200401.2112.645791112.6457910.00111.9894910.58113.103051−0.411200401.4112.807943112.8079430.00112.2768390.47113.204424−0.351200401.6112.975868112.9758680.00112.4428370.47113.305798−0.291200401.8113.149309113.1493090.00112.4705460.60113.407172−0.231200402.0113.328013113.3280130.00112.4987940.73113.508545−0.161200402.2113.511733113.5117330.00112.7239080.69113.609919−0.09 Last 15 synthetic data points predictions for HV with 7075 over 6061. The Gaussian Process Regression (GPR) exhibits predictive accuracy at all fifteen locations and indicates being consistent and accurate in modeling nonlinear hardness relationships. Support Vector Regression (SVR) displays systematic underprediction ranging from an error of −1.17% to a positive maximum of 0.73%. Thus, the model indicates moderate predictive accuracy. Furthermore, it exhibits a consistent negative model bias for almost every combination of parameters studied. The lowest absolute error of −1.17% is observed at TRS 1190 rpm and TTA 1.4°, while the positive error maximum of 0.73% is at TRS 1200 rpm and TTA 2.0°. Bayesian Ridge Regression has proven to perform better in the higher-parameter regime. Its errors converge between − 0.57% and 1.21%, indicating less bias than the low-parameter regime. The Bayesian Ridge may overpredict the boundary condition, but the achieved target has been met by the model. According to these findings, GPR will continue to be the model for predicting hardness in FSAM while possessing similar fidelity across the full space. The moderate accuracy and negative bias of SVR and the limited error range of Bayesian Ridge indicate the need for selecting models according to specific FSAM prediction tasks. The high reliability of GPR shown here supports the integrated RBF–GPR framework for surrogate-assisted multi-property optimization in friction stir additive manufacturing. Near-zero error on synthetic data reflects the model’s ability to learn the interpolated surrogate, not its generalization to unseen experimental variability. To avoid misleading interpretation, performance metrics reported on synthetic datasets are used only for internal consistency assessment of the surrogate-learning task. Predictive capability on real FSAM variability requires independent experimental validation. RBF interpolation methodology and sensitivity analysis To generate a smooth surrogate of the relationships between process and property from the experimental measurement, Radial Basis Function (RBF) interpolation was employed. The multiquadric RBF was chosen and can be represented mathematically as follows: where r is defined as the Euclidean distance between data points and ε is defined as a smoothing parameter to control the trade-off between the degree of interpolation fidelity and smoothness. The smoothing parameter was set at ε = 0.1 in this research so that stable RBF interpolation behavior would occur without excessive oscillation or overfitting of the relatively sparse experimental data. The interpolation results generated through RBF interpolation using different smoothing parameters were evaluated for sensitivity. RBF models were generated with the smoothing parameters ε = 0.05, 0.1, and 0.2 (three separate sets of parameters). The synthetic processes resulting from each of these cases had less than a 2% difference between mean values and less than a 5% difference between the overall predicted ranges. These small differences will be indicative of the reliability of the chosen smoothing parameter and therefore that the synthetic trends generated were not simply artifacts of the specific smoothing selection. The generated synthetic parameter grid using the radial basis function interpolation method was designed to fit primarily inside the experimental domain from which the corresponding parameters were sampled experimentally. The maximum allowable rotation speeds were 1000 to 1200 rpm, and the maximum allowable traverse speeds were 30 to 40 mm/min; thus, tool rotation speed and tool traverse speed were within the experimental ranges that were obtained from actual experiments conducted; hence, no extrapolation to these speeds occurred. Tool tilt angles were sampled primarily over an experimental range of 1.0° to 2.0°; the 2.1° to 2.2° range was included to evaluate edge behaviors of the interpolation surface. Radial basis function interpolators have increasing uncertainty when extrapolated outside of the convex hull constructed by the original parameter set. Predictions made using tool tilt angles of 2.1° and 2.2° are extrapolated predictions and should be evaluated with extreme caution. The Gaussian Process Regression confidence surface for this extrapolated region is nearly flat; therefore, this region has high uncertainty and low information content associated with it. Predictions made in the extrapolated region should not be used for process optimization or definitive inference without additional experimental support. Regression modeling methodologies Gaussian Process Regression Gaussian Process Regression (GPR) is a non-parametric and probabilistic regression technique that establishes a prior over functions and then updates that prior based on observed data to create a predictive posterior. GPR can model nonlinear relationships using a kernel function (in this case, a constant kernel multiplied by an RBF kernel) and can determine uncertainty through predictive variances. In this work, GPR is used for its power to assess complex nonlinear process-property relationships that are observed in the friction stir additive manufacturing (FSAM) data set but that also estimate uncertainty to provide confidence in prediction for surrogate-assisted parameter optimization. Support vector regression (SVR) Support Vector Regression (SVR) is an application of the support vector machine concepts for regression. SVR finds predictions within a zone of radius ε of observations and minimizes a cost that aggregates contributions from training error and model complexity. While SVR can accommodate nonlinearities in the training data using an RBF kernel, hyperparameters C and ε allow SVR to balance error with empirical margin. SVR was considered in this text for its robustness to outlier observations, and arrangements for nonlinear regression provided a complementary method to GPR with greater computational efficiency. Bayesian Ridge Regression Bayesian Ridge Regression adopts Bayesian methodology for linear regression by defining Gaussian priors over model coefficients and calculating posterior distributions. It then implements automatic regularization based on the determined hyperparameters, effectively reducing overfitting for small datasets. Bayesian Ridge is presented to evaluate a linear, regularized approach and to evaluate performance with strong prior constraints relative to nonlinear models. Utilizing detailed evaluation measures (MAE, RMSE, R², and percentage error), and with regard to 891 training points, GPR outperformed SVR and Bayesian Ridge and achieved very low prediction error (< 0.1%), an MAE of 2.34, an RMSE closely corresponding to the MAE, and an R² greater than 0.95 in each optimal configuration. SVR produced reasonable accuracy (error near 1%), whereas Bayesian Ridge produced systematic bias and a larger error (up to 12% in hardness prediction). Therefore, the accuracy, uncertainty quantification, and stability of GPR across parameter regimes support the decision to use GPR as the main regression model for surrogate-assisted FSAM multi-property optimization in this research. Comprehensive evaluation metrics This research on friction stir additive manufacturing quantitatively evaluates the machine learning model’s performance using a range of metrics to assess predictive accuracy, model reliability, and systematic bias using the synthetic dataset generated from RBF interpolation. The metrics selected for evaluation create an assurance of confirmation of the RBF-GPR framework generating surrogate-based multi-property predictions within the design space explored. Mean absolute error (MAE) Mean Absolute Error (MAE), it measures the average magnitude of errors in a set of predictions, without considering their directional bias, and provides an intuitive and interpretable way to evaluate the accuracy of a model that preserves the original units of measurement and significance. The evaluation of MAE across the entire synthetic dataset highlights performance differences between the three implemented regression algorithms. GPR demonstrated excellent discrimination accuracy, with MAE values ranging from 2.34 to 8.67 for UTS and HV across both material configurations, and GPR consistently performed better than the alternative evaluations. SVR exhibited moderate performance with MAE values roughly 2–3 times higher than GPR; however, the MAE values derived from Bayesian Ridge Regression were significantly higher, particularly in the case of hardness predictions, where errors frequently exceeded 10% of the predicted value. The low MAE values from GPR encouraged confidence in the approach’s ability to effectively uphold high fidelity to the synthetically generated ground truth produced by the RBF, confirming that the interpolated process-property relationships were preserved during model training and prediction. This ability is an important consideration for the conceptual optimization demonstration. Specifically, accurate predictions of mechanical property values will help enable more legitimate decisions to be made in the selection of process parameters that influence mechanical properties of built components. Root mean square error (RMSE) Root Mean Square Error (RMSE) improves sensitivity for larger prediction errors by weighing each error term quadratically prior to averaging, thereby increasing the effect of extreme errors as well as systematic bias. It is especially useful for identifying models with the worst performance for the most important elements of the data set and assigns increased importance to larger errors, which affects the entire evaluation metric. Evaluating GPR using RMSE consistently affirms its estimates are better than both SVR and Bayesian Ridge Regression for each mechanical property, with RMSE values consistently lower than those calculated using SVR and Bayesian Ridge Regression. For the GPR model, RMSE values are typically 5–10% of the corresponding MAE values, reflecting the lack of extreme errors and uneven error occurring over the range of combinations of parameters. In contrast, SVR generally has RMSE values that are higher than MAE by 20–40%, suggesting the presence of systematic prediction error or extreme observed errors that compromise model reliability. Bayesian Ridge Regression shows the highest RMSE with respect to MAE, consistently exceeding 50%, indicating high variability in prediction accuracy for any combination of parameters in the parameter space. The RMSE evaluation supports the stability of the GPR model, characterized by accurate prediction over a diverse range of engineering parameters and mechanical properties. This is a desirable characteristic for FSAM applications where process parameters may vary substantially throughout manufacturing and reasonable predictions are necessary for the entire operational envelope. Coefficient of determination (R²) The determination coefficient (R²) measures the amount of variance accounted for in the dependent variable from the independent variables, which is also a standardized measure of model explanatory power. R² ranges from 0 to 1, and the closer to 1, the better the model provides useful predictions, and conversely, as R² approaches 0, the model provides poor utility for predictions. The R² assessment of the comprehensive synthetic dataset indicates virtuous performance for GPR with R² values that continuously exceeded 0.85 for most configurations of output variables and as high as 0.95–0.99 in certain cases with selected parameters. The high R² indicates that GPR is successfully modeling highly non-linear characteristics between process parameters and mechanical properties characterized in the RBF synthetic dataset. SVR performs reasonably well with moderate R² values typically reported in the range of 0.65–0.80, indicating reasonable and not exceptional explanatory capability associated with process parameter predictions of mechanical properties. Bayesian Ridge Regression provides the lowest R² performance, reported below 0.60 in many instances, and struggles the most in hardening predictions, as the model was unable to capture process-property relationships. The high R² values acquired using GPR provide confidence that the model successfully captures the surrogate response behavior -based relationships between FSAM process parameters and resulting mechanical properties. Successful capture of physics-based relationships and hence strong predictive ability is contingent upon providing a model that accurately depicts an underlying process to make predictions of optimal operation. Percentage error analysis Percentage error provides a standardized measure for making comparisons across different output variables even when these variables have varying absolute values. This analysis is particularly important when considering multi-objective surrogate-assisted optimization, where distinct mechanical properties (i.e., UTS and HV) might have different measurements and ranges of expected values. The assessment of percentage errors reflects very good GPR predictions, with percentage errors found to be all below 1% for almost all the synthetic data points and often found to be near zero on multiple parameter combinations. SVR indicates systematic bias tendencies in its predictive value with percentage errors sometimes between 0.5% and 2.0%, showing systematic underprediction within the parameter regimes. While the Bayesian Ridge Regression demonstrated systematic bias with percentage errors that ranged between 2% and 12%, it is indicative of a considerable peril in attempting to quantify the inherent nonlinear relationships in FSAM processes. Cross-validation performance The synthetic dataset allows for extensive cross-validation evaluation by fully measuring model performance at the combinations of low, medium, and high TRS, TTS, and TTA, which, in turn, is used to evaluate model accuracy as well as model robustness and generalization performance rather than simply cross-validation regime definitions. GPR exhibits a model performance that is consistent across all parameter regimes for TRS, TTS, and TTA metrics, producing virtually identical accuracy metrics with varying process conditions. SVR possesses a dependency on parameters where model accuracy displays a benefit in all the higher (medium or high) parameter regimes when compared to the lower parameter regimes, and in lower parameter regimes the bias is more pronounced. Bayesian Ridge Regression indicates a high amount of performance variation across the different parameter regimes, including cases when both TRS, TTS, and TTA were set to lower parameter conditions, which is where the nonlinear effects are typically most severe and demonstrated the greatest performance metric deficits. Cross-validation was performed exclusively on RBF-generated samples and therefore evaluates surrogate fidelity rather than experimental generalization. To evaluate the performance of the model on actual measurements, 9 experimental data points were used to conduct LOOCV. At each iteration, one experimental data point was omitted, and a model trained from the remaining points was used to estimate the excluded point. The mean absolute error (MAE) and root mean square error (RMSE) were calculated for each data point (UTS, MPa; Vickers Hardness, HV). The percentage of best and worst errors is reported for the 9 LOOCV estimates. The percentage of errors for the model trained with the 9 experimental data shown in Table 10 below is much larger than for the synthetic data set, indicating that while it is possible to achieve a high accuracy (or close to perfect) for the synthetic data set, the actual variability of independent experimental measurements could not be fully captured. The reported MAE and RMSE values therefore reflect consistency with measured experimental data rather than surrogate fidelity to the RBF-interpolated synthetic dataset. Percentage error values represent absolute errors between predicted and experimentally measured values during LOOCV. Table 10Leave-One-Out Cross-Validation Summary Statistics on Original Experimental Data.ConfigurationOutputMAERMSEAbsolute PercentageError Range (%)6061-over-7075UTS (MPa)4.956.810.75–3.486061-over-7075HV3.214.121.28–8.067075-over-6061UTS (MPa)3.584.120.67–1.647075-over-6061HV2.142.461.16–2.66OverallUTS (MPa)4.265.460.67–3.48OverallHV2.683.291.16–8.06Overall values represent aggregated LOOCV metrics across both stacking configurations. Leave-One-Out Cross-Validation Summary Statistics on Original Experimental Data. Overall values represent aggregated LOOCV metrics across both stacking configurations. Statistical significance and confidence interval The probabilistic structure of GPR allows for characterizing uncertainty with confidence intervals and constraining the reliability of predictions. The analysis reveals narrow confidence intervals, typically between 2% and 5%, for most distinctive parameter combinations. These observations provide considerable confidence in the projected predictions. Similarly, the SVR and Bayesian Ridge model predictions may require the use of bootstrap or cross-validation approaches to characterize uncertainty, which typically yield wider confidence intervals; thus, the confidence in the predictive nature of these models is lower. The detailed evaluation based on 882 synthetic data points indicates GPR’s impactfully strong predictive capability, which is across all mechanical properties and parameter spaces, almost near zero. More specifically, GPR achieves the fit across MAE, RMSE, and R², exhibits cross-validation results and low systematic bias, and therefore GPR was selected as the regression algorithm to implement the interpolated design space framework proposed in this work. SVR can be employed with moderate performance for quick screening approaches, e.g., just to identify trends; however, it introduces systematic bias and therefore cannot be considered for accurate applications that require precision and high levels of accuracy. Lastly, Bayesian Ridge Regression has exhibited significant limitations, including suitability for hardness prediction, and therefore is not applicable for FSAM applications that require high levels of accuracy and precision. Taking together, these metrics instill confidence that the proposed RBF-GPR framework generates surrogate-based multi-property predictions within the explored design space forming the basis of intelligent manufacturing systems that can achieve target mechanical properties through data-driven process control. Performance visualization and analysis In this friction stir additive manufacturing research, visualizations serve several important purposes that improve the scientific integrity, interpretability, and usefulness of the machine learning-based surrogate-assisted parameter optimization framework. Data visualizations are strategically employed to address key issues in advanced manufacturing research and to develop a more thorough understanding of complicated interactions between processes and properties. Visualizations support direct comparisons of the three regression modelling approaches (GPR, SVR, and Bayesian Ridge) using graphical visualizations of prediction accuracy, prediction error distributions, and performance measures over each separate parameter regime. This visual validation increases the reliability of the quantitative assessments and offers unambiguous, interpretable indicative trends to support decisions on which model to select. Performance visualization and analysis for UTS 6061 over 7075 Comparative performance of ML models for UTS Prediction in FSAM Figure 1 presents scatterplots of actual vs. predicted values from predicting ultimate tensile strength (UTS) in the 6061-over-7075 material configuration from three different regression modeling algorithms—Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. In each subplot, actual experimental values are on the Y-axis and predicted values are on the X-axis. The red dashed line indicates y = x, while the solid gray line indicates the fitted regression line from each model. Comparative Performance of ML Models for UTS Prediction in FSAM. The GPR (left plot) is highly predictive, with the predicted values closely following the fit line with low scatter of predicted values and high reliability between the actual and predicted UTS values. The SVR (middle plot) model is moderately predictive, with predicted values showing more scatter around the fit line, indicating systematic prediction bias and lower reliability and specificity of the model. The Bayesian Ridge Regression model (right plot) shows high bias with predicted values clustering between 305 and 307 MPa regardless of actual UTS, indicating less variability of the regression model to the process-property relationship. Comparative visualization confirms the better GPR process of prediction UTS in FSAM applications and effectively indicates its unique ability to capture complicated nonlinear relationships between process parameters and mechanical properties. The SVR’s systematic underprediction and the clear bias associated with Bayesian Ridge Regression demonstrate the importance of model selection in machine learning-based parameter optimization for advanced manufacturing applications. Residual analysis of machine learning models for ultimate tensile strength prediction in 6061-over-7075 FSAM configuration Figure 2 shows residual plots for the three regression models to predict ultimate tensile strength (UTS) in the 6061-over-7075 material example with residuals (actual minus predicted values) plotted against predicted values on the y-axis. A red dashed horizontal line at zero represents prediction accuracy, while data points in the area around zero predict errors. Residual Analysis for Ultimate Tensile Strength Prediction in 6061-over-7075 FSAM Configuration. The GPR model (left panel) indicates excellent residual behavior, with many of the errors contained within an acceptable range of ± 2.5 MPa, and residuals are randomly scattered around zero, again having no systematic patterns, along with the relatively small size of the residuals and random residual distribution, without heteroscedasticity, indicating trends of the predictive characteristics and fairness of the GPR model for FSAM applications. The SVR model (middle panel) shows average residual behavior with errors occurring over a measurably larger range (−5 - +12.5 MPa) and some high-amplitude systematic bias patterns, particularly around intermediate predicted values. The inferred residuals and their distribution point to a limitation of the SVR model’s ability to uncover some of the major process-to-property relationships. The regression residual analysis of the model predicting the UTS of FSAM processes through Bayesian Ridge Regression (to the left) displays the worst characteristics with very large systematic errors into and ranging between − 20 and + 25 MPa. Predictions being highly concentrated around the expected value of 306 MPa have similarly produced large values of residuals to the computed data, with an overwhelming systematic bias observable across all explanatory levels of prediction that has led to systematic high and low residuals, hence exhibiting considerable magnitude errors and a heteroscedastic characteristic to the residuals produced. This observable pattern strongly indicates that Bayesian Ridge Regression UTS modeling is not appropriate or suitable for modeling UTS data to predict UTS in FSAM processes. This analysis of prediction residuals bolsters the case for GPR, which mainly minimizes both bias and variance in prediction errors, and highlights the systematic limitations in SVR, while indicating Bayesian Ridge Regression models have severe limitations for interpolated design applications. Residual histogram and kernel density estimation (KDE) for ultimate tensile strength prediction in 6061-over-7075 FSAM configuration Residual histograms with kernel density estimation (KDE) for the three regression models that were assessed for predicting ultimate tensile strength (UTS) in the 6061-over-7075 material configuration (Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression (BRR)) are displayed in Fig. 3. Each subplot shows the distribution of residuals (actual minus predicted) of the applied model regenerating the nine original experimental data points that were used in previous discussion. Residual Histogram and KDE for UTS Prediction in 6061-over-7075 FSAM Configuration. The left panel shows the GPR model and indicates the substantial clustering of residuals around zero observed in both the histogram and KDE curve, indicating prediction errors that were tightly clustered around zero, with minimal bias and low variance. The middle panel represents the SVR model, which exhibited a more broadly distributed residual pattern than GPR, with significant residual tails extending toward under-prediction and over-prediction, indicating the SVR model was less consistent in predicting UTS than the GPR model. The right panel shows the residuals from the BRR, which indicates a bimodal and dispersed residual pattern, with residuals widely dispersed from the original data point—at least in part—due to consistent model bias and a lack of fit because of the non-linear nature of the process-property relationships. This visual representation of the information helps summarize the accuracy and bias of these models in a concise form. It depicts the high accuracy and stability in GPR, the accuracy and moderate, therefore slightly more variable, accuracy in the SVR analysis, and the barriers represented through bias in the Bayesian Ridge Regression fit to FSAM processes. Cross-validation mean absolute error (MAE) for ultimate tensile strength prediction in 6061-over-7075 FSAM configuration across all models Figure 4 displays the cross-validated mean absolute error (MAE) metric to predict ultimate tensile strength (UTS) for the configuration of 6061-over-7075 material and compares the three regression models that were used: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. The MAE for each of the models is represented as distinct bars, which, of course, represent the average absolute error in predictions made by each algorithm during the cross-validation procedure. Cross-Validation MAE for UTS Prediction in 6061-over-7075 FSAM Configuration Across All Models. The bar graph clearly indicates the model based on GPR produces the lowest MAE; hence, it is the most accurate and reliable predictor. The model based on SVR predicts at a moderate level, executing greater MAE than GPR but lower NAE than the model based on Bayesian Ridge Regression, which has the highest MAE and demonstrated the least predictive quality and deviation from the true value. Overall, this comparative metric analysis provides sound quantitative indicative trends for the choice of GPR over regression and sufficiently supports its selection as the modeling technique for use in optimization of certain friction stir additive manufacturing (FSAM) processes and multi-property prediction. Cross-validation root mean square error (RMSE) for ultimate tensile strength prediction in 6061-over-7075 FSAM configuration across all models Figure S1 displays the results for the cross-validation root mean square error (RMSE) for ultimate tensile strength (UTS) prediction for the 6061-over-7075 friction stir additive manufacturing (FSAM) configuration across the three models investigated: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. The RMSE values for every model can be seen plotted separately as bars, which reflect how far they predict the UTS from the measured UTS. The results indicate that the GPR presents the lowest RMSE, meaning it predicts UTS with the least error, minimizing large errors for UTS prediction. The SVR exhibits moderate RMSE, meaning it is more acceptable but performs less robustly, but the Bayesian Ridge Regression exhibited the highest RMSE of the three models, denoting error propagation and less predictive trustworthiness. The outcome here is another indication of GPR as a machine learning algorithm to optimize FSAM parameters due to its ability to minimize errors and be more predictive and trustworthy. Cross-validation coefficient of determination (R²) for ultimate tensile strength prediction in 6061-over-7075 FSAM configuration across all models Figure S2 illustrates the cross-validation coefficient of determination (R²) for ultimate tensile strength (UTS) prediction in the 6061-over-7075 friction stir additive manufacturing (FSAM) configuration for the three machine learning models examined: Gaussian Process (GP), Support Vector (SVR), and Bayesian Ridge (BR). The R² metric provides a quantitative metric of the explanatory power of each model, namely the proportion of variance in the experimental UTS data explained, with values closer to 1.0 corresponding to higher explanatory power and better model fit. Results indicate that the GP model achieves an R² near 1.0, supporting GP’s excellent ability to capture and generalize the underlying process–property relationship of FSAM. SVR also provides a moderately high explanatory power but is below GP in both accuracy and repeatability. In stark contrast, the BR model yields an almost unmeasurable R² and fails to model the observed data. This comparison is significant in that it indicates the important importance of advanced non-linear modeling (GP) for dependable interpolated design space and illustrates the inadequacy of linear modeling approaches (BR) for modeling complex behaviors of ultimate tensile strength. These results support the obvious choice of GP as the primary model for predictive process surrogate-based optimization in this research. GPR-based response surfaces for ultimate tensile strength (UTS) prediction in 6061-over-7075 FSAM configuration at different tool tilt angles (TTA) Figure 5 shows the Gaussian Process Regression (GPR) confidence surfaces for ultimate tensile strength (UTS) prediction in the 6061-over-7075 patented friction stir additive manufacturing (FSAM) configuration and is featured in three distinct tool tilt angles: TTA = 1.0°, 1.5°, and 2.0°. Each 2D surface plot shows the predicted UTS (z-axis) as a function of the tool rotation speed (TRS, x-axis) and tool traverse speed (TTS, y-axis) so that the reader has an advantage or immediate view of the process-property relationships for the tilt angle shown. GPR-Based Response Surfaces for UTS Prediction in 6061-over-7075 FSAM Configuration at Different TTA. The surfaces at TTA = 1.0° and TTA = 2.0° provide smooth, continuous variation in UTS throughout the total parameter space and show areas of better or worse performance. The plot at TTA = 1.5° shows some numerical artifacts but highlights how UTS responses might not only be non-linear but perhaps also non-monotonic, meaning that these phenomena are possibly due to complicated process physics. Collectively, these surfaces are invaluable to design practitioners because they visually point to which parameters optimize UTS as a function of tilt angle through multivariate process surrogate-based optimization. Overall, these surface plots show the value of GPR models matching complex non-linear effects of FSAM process parameters and show reasons to use predictive machine-learning surrogate-based optimization in manufacturing and actionable decisions in engineering practice. Calibration, density, and partial dependence analyses for GPR model predictions of ultimate tensile strength in 6061-over-7075 FSAM configuration The depiction in Fig. 6 presents a comprehensive perspective of the predictive strength and interpretability of the Gaussian Process Regression (GPR) model to predict ultimate tensile strength (UTS) for the FSAM configuration of 6061-over-7075. Collectively, the panels articulate important featural aspects of the credibility of the GPR predictions, the predictive distribution of UTS estimates, and process-parameter sensitivity. Advanced Evaluation Visualizations: Partial dependence, 2D KDE, and calibration curve for GPR model predictions of UTS in 6061-over-7075. The first panel displays the calibration curve for UTS 6061 over 7075, plotting actual (red) versus predicted (blue) UTS values for each sample; the 95% credible interval (± 2σ) is shown for the GPR predictions. This visualization provides for inspecting the uncertainty of GPR prediction and GPR model calibration. The representation indicates that most of the measured UTS values exhibit overlap with the 95% credible interval that was predicted from the GPR model, thus providing confidence in GPR’s ability to assess uncertainty and, in general, be used effectively as a risk-sensitive model in the FSAM surrogate-based optimization. The second panel displays a 2D Kernel Density Estimation (KDE) plot for UTS_6061over7075, which illustrates the joint probability distribution of the true contamination and the predictions. The pronounced density concentration along the diagonal (i.e., along the y = x line) indicates the superb match between the observed data and the predictions of the model. Thus, the panel presents an easy-to-understand visualization of the GPR accuracy and no signs of systematic bias in prediction in any domain of the response. In addition, the lack of meaningful off-diagonal density also supports the ability of the GPR model to generalize to both interpolated and extrapolated domains of the response, thereby further supporting the GPR in the surrogate-based multi-property process optimization framework. The third panel presents a partial dependence plot of UTS_6061over7075 predictions as a function of tool rotation speed (TRS), holding all other input processing parameters constant. The curve represents the essential marginal effect of TRS on UTS in GPR modeling. The shape of the partial dependence plot indicates a pronounced steep slope that exhibits a shift to constant predictions of UTS beyond a certain exposure to TRS—the information can be significant for process engineers to guide the identification of the optimal range of parameters to improve material performance. Correlation matrices comparing actual versus predicted ultimate tensile strength for different models in 6061-over-7075 FSAM configuration Figure 7 displays heatmaps of Pearson correlation coefficients computed between the actual UTS values and predicted UTS values by the three regression models—Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Bayesian Ridge Regression. Each subplot displays the data’s scatter plots and the correlation coefficient corresponding with the matrix of interest. For the SVR submodel, a correlation coefficient of 0.96 indicates a strong but imperfect agreement on predicted UTS values against the true or experimental values. These strongly correlated trends indicate potential for SVR to capture the overall trend from the data, but it also suggests the variable (as presented as some form of AB skew) is still subject to variation from model bias and the non-linear nature of the constraints. Correlation Matrices Comparing Actual versus Predicted UTS for Different Models in 6061-over-7075 FSAM Configuration. For GPR, the subplot displays a correlation coefficient equal to 1.00 when examining predicted UTS values against true or experimental values on a linear scale. This indicative trend supports the conclusion that GPR predictions performed highly well to model the complex or non-linear relationships present in the FSAM, which ended in equal value (in the analysis) between predicted and experimental values. For Bayesian Ridge Regression, recall that the correlation coefficient was lower at 0.57, or in the weak-to-moderate relationship category. The low correlation presented in Fig. 9 above alludes to the complexity of the deriving process versus the parameter value of UTS; the prediction error was high, and thus the model had a poor relationship value. The correlation matrices suggest that model performance ranking is quantitatively detectable. The correlation achieved by GPR confirms it as the number one regression model for UTS prediction in FSAM interpolated design space. The SVR shows reasonable but imperfect capacity, while Bayesian Ridge Regression proves unsuitable for high-fidelity predictive modelling in this application. Residual Analysis of Machine Learning Models for Actual vs. Predicted Vickers Hardness Prediction in 6061-over-7075. Performance visualization and analysis for HV 6061 over 7075 Comparative performance of ML models for HV prediction in FSAM Figure 8 displays a visual of the actual versus predicted scatter plots for Vickers hardness (HV) predictions in the 6061-over-7075 material configuration. The performance of three regression algorithms is examined here: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. Each subplot shows the predicted values (represented by blue circles) superimposed on measured values of hardness with the prediction (i.e., y = x) lined with a red dashed line and the fitted regression line lined as solid grey. Comparative Actual vs. Predictive Performance for Vickers Hardness Prediction in 6061-over-7075. In the GPR visual, the predicted hardness values track relatively all the way along the fit line for the entirety of the observed range of hardness (70–120 HV) with little scatter and a high degree of fit with the measured values. All three plots indicate the fit line almost lies on top of the fit line, suggesting the GPR is not biased. GPR performed exceptionally well here, modeling the nonlinear relationship that exists between FSAM processing parameters (TRS, TTS, TTA) and material hardness. The data points also cluster very tightly about the fit line, allowing to be trusted to some degree the model’s ability to predict an increase or decrease in hardness based on variations in the identified parameters. The SVR subplot does show moderately linked performance, where predicted values are less disparate around the ideal fit line. The fitted regression line shows a positive slope, whose values diverge from the reference line at the high end of hardness (above 100 HV), showing clear systematic underprediction for some combinations of parameters. Similar data points in the range of 90 to 110 HV show greater vertical dispersion, meaning SVR only captures some of the interactions within the processing parameters that drive hardness. Overall, SVR does predict the trend correctly, meaning the SVR value can approximate the hardness trend, but with less accuracy than the GPR. The Bayesian Ridge Regression subplot shows substantial prediction bias and little adaptability. Predicted hardness values cluster very closely around the mean (≈ 90–100 HV), showing that the regression line has a shallow slope and diverges from the fit line quite significantly. This indicates that Bayesian Ridge is unable to account for the nonlinear effects of the process parameters on hardness. The model consistently produces predictions that remain almost constant, regardless of hardness variations. The substantial dispersion of actual hardness values in the horizontal time, contrasted with negligible vertical dispersion of predicted values, indicates that this model is not appropriate for forecasting hardness in the FSAM context. Collectively, these comparative plots show how good GPR is for hardness prediction in the FSAM interpolated design work. The agreement of GPR with the experimental data and its limited fingerprint compared to the fit showcase GPR’s modeling capabilities of high-dimensional, nonlinear relationships with quantified uncertainties. SVR provides a computationally simple alternative but adds extra error, or rather systematic bias, for hardness values in extreme ranges. Bayesian Ridge is a beneficial tool, as it is efficient for linear problems, but it failed to provide any value for nonlinear modeling in this work. These analyses provide justification for GPR as the primary regression model for surrogate-assisted multi-property prediction and process control in the scope of advanced FSAM research and industrial use. Residual analysis of machine learning models for vickers hardness prediction in 6061-over-7075 FSAM configuration Figure 9 shows the first of the residual plots from the three regression models tested in this research: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression predicting Vickers hardness (HV) relative to the 6061-over-7075 friction stir additive manufacturing process. Each subplot displays residuals (observed - predicted) plotted against the predicted hardness value, along with a horizontal, red-dashed line at the zero residual, indicating a better prediction. For the GPR model results (left panel), it has the characteristics of the residuals, having prediction errors of approximately ± 3.5 HV units and not having any systematic pattern of distribution or heteroscedasticity, with the residuals being randomly spaced out around zero. Nearly all the data points are clustered around the zero line, indicating virtually no bias in the predictions and having all the prediction error be equal across the measurements taken (75–120 HV). The consistency of the distribution of the residuals indicates that the GPR model is a valid model and represents that the GPR model learned trustworthy complex non-linear equations from the FSAM process parameters and material hardness properties. The SVR (middle panel) model has an average level of residual performance, with errors occurring across a wider range (−10 to + 30 HV units) and showing systematic bias trends, especially with residuals not being normally distributed. There is consistent under-prediction of hardness occurring at lower values and over-prediction at upper values, which suggests a general inability to capture the full range of process-property relationships. The increased scatter and systematic deviations support a reduced level of reliability compared to GPR, but SVR is still reasonable at predicting values that fall between intermediate hardness ranges. Bayesian Ridge Regression (right panel) shows the worst overall residual behavior, with large systematic errors and residuals that range from − 15 to + 30 HV units, and a lot of clustering around certain predicted values. The limited predicted range of values within the model (85–100 HV) means that the residuals will be relatively large when the actual hardness is far from that small range. The systematic bias and heteroscedastic pattern confirm the inadequacy of the model to predict hardness accurately for FSAM processes, as it cannot account for and adapt to the nonlinear properties and relationships that are part of this system. Residual analyses enhance the quality of quantitative performance measures with indicative trends that GPR is beneficial in diminishing both prediction bias and variance. The systematic constraints of SVR and the significant limitations of Bayesian Ridge Regression further substantiate the value of sophisticated nonlinear modeling techniques for effective prediction of mechanical properties inherent in complicated manufacturing processes. These results highlight the use of GPR as the primary regression algorithm for FSAM interpolated design, yielding trusted hardness predictions that will contribute significantly to quality management and process design in advanced manufacturing applications. Residual histogram and kernel density estimation (KDE) for vickers hardness prediction in 6061-over-7075 FSAM configuration Figure 10 shows residual histograms overlaid with kernel density estimation curves of the three regression models evaluated—Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression—that were used to predict Vickers hardness (HV) in the 6061-over-7075 friction stir additive manufacturing configuration. Each subplot shows the distribution of residuals (i.e., actual - predicted values) from the nine original experimental data points, which can provide some insights into the patterns of prediction error (bias) associated with each of the three models. Residual Histogram and KDE for Vickers Hardness Prediction in 6061-over-7075 FSAM Configuration. The GPR residuals, shown in the left panel, exhibited the residual distribution centered tightly around zero, with five data points in the highest frequency bin that showed near zero error. The KDE also demonstrated that the distribution of residuals is symmetric and centered at zero for the GPR regression associated with mean predicted values, which indicated there was no systematic bias designed into the model that generated consistent accuracy. The narrow spread of residuals in the ± 3 HV range affirmed GPR’s extraordinary ability to model the nonlinear complexity between FSAM process parameters and the resulting hardness properties with very little prediction error variance. The SVR model (second panel) shows a wider, more spread-out residual distribution marked by a flatter KDE curve extending from − 10 to + 30 HV units. The bimodal distribution, with significant density of values at both negative and positive residuals, suggests systematic inconsistency in predictions. The extended tail associated with positive residuals indicates a tendency for SVR to underpredict hardness values for certain process conditions, while the secondary peak near zero suggests adequate prediction capabilities for certain parameter combinations. The variability in prediction accuracy makes SVR unreliable in wider FSAM optimization applications. The Bayesian Ridge Regression model (third panel) has the most concerning residual characteristics, with a highly skewed distribution concentrated primarily in the positive residual range (0 to + 30 HV). As illustrated in the KDE curve, the peak is significantly above zero, confirming a systematic underprediction in most of the experimental conditions. Given the clear asymmetry in the plots and the residual distribution being wide and spread out, it indicates poor model calibration and bias, reflecting Bayesian Ridge’s inability to capture the nonlinear nature of the hardness-parameter relationships present in FSAM processes. Cross-validation mean absolute error (MAE) for vickers hardness prediction in 6061-over-7075 FSAM configuration across all models Figure S3 illustrates the cross-validation mean absolute error (MAE) for Vickers hardness (HV) prediction in the 6061-over-7075 friction stir additive manufacturing (FSAM) configuration, comparing the performance of three regression models: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. The MAE values are represented as bars, quantifying the average absolute deviation between predicted and actual hardness values across multiple validation folds. The GPR model exhibits the lowest MAE (≈ 1 HV), demonstrating exceptional predictive accuracy and minimal average error in hardness estimation. The SVR model shows a substantially higher MAE (≈ 8 HV), indicating moderate performance with greater average deviation from true hardness values. The Bayesian Ridge Regression model produces the highest MAE (≈ 13 HV), reflecting significant average prediction errors and poor model calibration for hardness forecasting. This comparative metric confirms GPR’s minimizing predictive error, supporting its selection as the primary regression algorithm for high-fidelity hardness prediction in FSAM interpolated design. Cross-validation root mean square error (RMSE) for vickers hardness prediction in 6061-over-7075 FSAM configuration across all models. Figure S4 shows the cross-validation root mean square error (RMSE) values for the prediction of Vickers hardness (HV) in the friction stir additive manufacturing (FSAM) process for the 6061-over-7075 configuration for Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. RMSE is the square root of the average squared prediction error, so it penalizes larger deviations in the prediction error more than the mean absolute error (MAE). The GPR model displays the lowest RMSE value (≈ 1.7 HV), implying no large errors and the highest degree of prediction accuracy. The SVR model displays an RMSE value (≈ 11.3 HV) that is in between the other two models. In this case, some larger-sized prediction errors may have occurred due to the model not fitting the data well in certain regions of the process parameter space, but all the heuristics considered in the SVR are applied to generate the final fit. Finally, Bayesian Ridge Regression exhibits the highest RMSE value (≈ 15.6 HV), which indicates a propensity for larger errors, evidencing a generally poor fit to the nonlinear relationship seen in hardness and the process parameter. In summary, this RMSE shows that, again, the GPR model minimizes not only small prediction errors but also mitigates even the larger deviations in prediction errors for HV. This reinforces the subsequent discussion about GPR being the candidate for PVA-F STA, on the grounds that it can demonstrate prediction accuracy in the absence of BLR. Cross-validation coefficient of determination (R²) for vickers hardness prediction in 6061-over-7075 FSAM configuration across all models This bar graph in Fig. 11 illustrates the cross-validation R² values associated with predicting Vickers hardness (HV) across the 6061-over-7075 friction stir additive manufacturing (FSAM) setting, comparing Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression (BRR). R² measures the percentage of variance in the experimental hardness data modeled by each machine learning method—close to 1.0 denotes more explanatory and predictive power. Cross-Validation Coefficient of Determination R² for HV Prediction in 6061-over-7075 FSAM Configuration Across All Models. The GPR model has an approximate R² value of 0.99 and shows capacity to learn and generalize the nonlinear equation causing the output (hardness) with respect to the independent inputs (TRS, TTS, TTA) in the FSAM method. The SVR method has a relatively lower R² value of around 0.47 and has some explanatory power but is insufficient to consistently or reliably predict hardness across the entire parameter space. The BRR method has a negative R², indicating it poorly models the variance in observed hardness and denotes worse-than-mean prediction. This figure indicates the need for nonlinear, probabilistic modeling for effective and reliable hardness predictions and optimization: GPR is a method of non-linear and probabilistic modeling, demonstrated in comparison to SVR or BRR modeling approaches, for prediction of hardness in an FSAM experiment. The importance of the high R² value of GPR as a method illustrates its purpose as a primary analysis method for regressing experiments and for reliable surrogate-assisted multi-property forecasting for advanced manufacturing indexing the material and process design. GPR-based confidence surfaces for vickers hardness prediction in 6061-over-7075 FSAM configuration at varying tool tilt angles This multi-panel at Fig. 12 shows GPR confidence surfaces for prediction of Vickers hardness (HV) in the 6061 over 7075 friction stir additive manufacturing (FSAM) configuration at three different tool tilt angles (TTA = 2.0°, 1.5°, and 1.0°). The 3D surface plots depict the predicted HV (z-axis) as a function of tool rotation speed (TRS, x-axis) and tool transverse speed (TTS, y-axis). GPR-Based Confidence Surfaces for HV Prediction in 6061-over-7075 at Varying Tool Tilt Angles. At TTA = 1.0°, the surface indicates smoother hardness gradients and shows a monotonic increase in HV with increasing TRS and decreasing TTS. The gentler slopes indicate less variability in hardness response that could be considered easier for the process engineer to harness for empowering process changes for hardness improvements. At TTA = 1.5°, a numerically flat surface is seen due to the extreme interpolation range; however, the flattening of the surface indicates the sensitivity of the model at this tilt angle. This indicates that great care must be taken on the selection of parameter coverage when applying GPR-based predictions to the far extremes of predictive parameter ranges. At TTA = 2.0°, large nonlinear variation in hardness can be seen with well-defined peaks and valleys indicative of strong parameter interactions. The regions of high hardness occur at intermediate ranges of TRS and TTS; hence, the predicted hardness surface is useful for parameter value adjustments to maximize material performance. In summary, these confidence surfaces indicate GPR’s ability to capture complex, nonlinear FSAM process-property relationships and provide actionable direction by directing TRS and TTS selection combinations at a fixed tilt angle. The confidence surfaces can be useful decision-support sources for process engineers aiming to improve hardness in FSAM applications. Calibration, density, and partial dependence analyses for GPR-based vickers hardness prediction in 6061-over-7075 FSAM configuration Figure 13 assesses the performance and interpretability of the GPR model for predicting Vickers hardness (HV) in the 6061-over-7075 friction stir additive manufacturing (FSAM) configuration using three complementary visualizations. Calibration, Density, and Partial Dependence Analyses for GPR-Based HV Prediction in 6061-over-7075. The first subplot is a calibration curve that shows the actual hardness measurements (red circles) vs. the GPR predictions (blue circles) for each sample with ± 2σ prediction intervals represented. Almost all measured hardness values fall within the credible intervals of the GPR model, illustrating that it accurately quantified uncertainty and produced calibration reflecting true measurements of Vickers hardness. The values predicted by the GPR model are clustered very close to the actual measured hardness values, indicating it is a reliable model that can be adopted for further robust Vickers hardness estimations with a shift of different process conditions for Vickers hardness. The second subplot is a 2-dimensional kernel density estimation (KDE) plot of actual HV vs. predicted HV values. The dramatic diagonal density ridge aligned with the y = x line indicates most predicted values aligned closely with actual hardness values, with very little off-diagonal density indicating no systematic bias is present in the predicted values. The density highlighted in the diagonal provides additional confirmation of the accuracy of the GPR model and indicates high accuracy is maintained throughout the range of hardness present within the research (70–140 HV) and that the GPR model can generalize effectively through both interpolated and extrapolated parameter regimes. The third subplot presents a partial dependence plot that highlights the marginal influence of Tool Rotation Speed (TRS) on the predicted hardness while averaging the influences of tool traverse speed and tool tilt angle. The trend of an initial rapid increase in HV for values of TRS between 1000 and 1020 rpm, after which a plateau is reached as tool rotations assume higher speeds. Being mindful of this possibility indicates that there is likely the potential for diminishing returns to one degree of hardness increasing beyond a certain TRS threshold that provides a focal point for the practitioner to engage in a practical change: TRS values recorded during the steepest increase will give the most appreciable return in hardness, while further increases are of very limited or inconsequential advantage across the entire process of doing business with FSAM. Overall, these visualization techniques verify the GPR model’s predictive quality and uncertainty bounds, as well as provide interpretability of parameter effects contributing to the final hardness prediction of the toolpaths. The calibration curve verifies the uncertainty bounds are accurately estimated, the KDE plot verifies that predicted hardness values are overall valid without unwanted systemic biases, and the results of the partial dependence plot isolate and verify the effect of TRS on hardness. Overall, these demonstrate GPR-based predictions can be trusted and allow operational recommendations to be derived to evaluate and implement changes in process parameters to implement informed decisions for optimizing mechanical properties in advanced FSAM operations. Correlation matrices of actual versus predicted vickers hardness for different regression models in 6061-over-7075 FSAM configuration Figure 14 shows heat maps of the Pearson correlation coefficients between measured and predicted Vickers hardness (HV) values for the three regression algorithms—Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression—applied to the 6061-over-7075 friction stir additive manufacturing (FSAM) dataset. Each subplot shows a single correlation coefficient, which shows the degree of association that measures the linear relationship of the HV measurement values to those predicted by the models. Correlation Matrices of Actual versus Predicted HV for Different Regression Models in 6061-over-7075. The GPR subplot presents a correlation coefficient of 1.00, indicating that the predicted values had linear correspondence (concordance) to the actual values that were measured for HV. In other words, GPR had better predictive accuracy and captured the non-linear process–property relationship precisely, without inserting bias systematically into the GPR model. The correlation illustrates and confirms that GPR is the optimal surrogate model for minimizing the uncertainty of the empirical estimates of HV hardness under the FSAM design space. The subplot for Support Vector Regression (SVR) shows a strong, though imperfect, correlation of 0.72 that indicates SVR’s modest ability to capture trends in hardness while producing estimates that deviate from true values. The fact that the correlation is lower than the one reported for GPR indicates that SVR is more likely to exhibit prediction errors and systematic biases, particularly for extreme combinations of process parameters. However, the positive correlation is sufficiently high for SVR to be considered for secondary modeling purposes when computational speed is more important than predictive performance. Conversely, the subplot for Bayesian Ridge Regression shows a moderate positive correlation of 0.50 but limited explanatory power and significant prediction errors. The moderate correlation suggests that linear regularization via Bayesian Ridge does not capture the inherent nonlinear relationships in the FSAM hardness dataset well, which produces predictions that poorly observe or track hardness datasets. These correlation plots provide a compact way of communicating ordinal rankings of model performances: GPR’s correlation suggests that it is appropriate for the optimum predictive algorithm for predicting hardness, SVR’s moderate correlation suggests that it provides a tradeoff between accuracy and computation, and Bayesian Ridge’s low correlation in prediction performance indicates unsuitability for modeling complex FSAM properties. Performance visualization and analysis for UTS 7075 over 6061 Comparative predictive performance of machine learning models for ultimate tensile strength prediction in 7075-over-6061 FSAM configuration This Fig. 15 shows the actual versus predicted scatter plots for UTS prediction for the 7075 over 6061 friction stir additive manufacturing (FSAM) configuration. Performance of the regression models—Gaussian Process, Support Vector, and Bayesian Ridge—is displayed across nine experimental data points. Each subplot shows the predicted values (blue circles) plotted against actual UTS values, with the red dashed line indicating (y = x) and the gray line indicating the fitted regression line. Comparative Predictive Performance of ML Models Actual vs. Predicted for UTS Prediction in 7075-over-6061 FSAM Configuration. The GPR model (left) demonstrated good predictive accuracy with data points clustered along the fit line across the entire UTS region (291–336 MPa). The regression line is nearly on the diagonal line (ideally y = x), which indicates there is little systematic bias and strong correlation between predicted and measured values. GPR’s fit further supports and indicates its ability to capture the complicated, nonlinear process-property relationships for the 7075 over 6061 material stacking sequence, where the harder 7075 aluminum alloy is positioned over the softer 6061 substrate. The performance of the SVR model (middle panel) is moderate (predicted values with increased residual scatter about the “perfect fit” line), though most readily observed in UTS values between 310 and 325 MPa. While the regression line remains positive in slope, the systematic offset from the “perfect fit” line shows that SVR is unable to fully untangle the relationships between the parameters that influence tensile strength for this material combination. The scatter implies acceptable, but not ideal, performance relative to precise applications. The Bayesian Ridge Regression model (right panel) displays strong systematic bias in the predicted values relative to the “perfect fit” line, indicating a substantial deviation. In addition, the slope of the fitted regression line has decreased relatively, which indicates underprediction at higher UTS values was consistent. This systematic bias reflects the model’s inability to respond to the nonlinear relationships created by property gradients and thermal interactions in the 7075-over-6061 system in which complex mechanical responses occur. These comparative findings confirm the advantage of using GPR for UTS prediction across both configurations of the FSAM materials, with consistently high-fidelity performance regardless of stacking sequence. The observed performance ranking of GPR > SVR > Bayesian Ridge is sufficient to support the use of advanced nonlinear modeling approaches for the accurate prediction of mechanical property phenomena in complex manufacturing processes involving multiple materials, as well as process parameters that need to be carefully controlled and adjusted. Residual analysis of machine learning models for ultimate tensile strength prediction in 7075-over-6061 FSAM configuration Residual plots for the Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression algorithms applied to ultimate tensile strength (UTS) prediction in the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration are shown in Fig. 16. The plots show the residuals (actual – predicted UTS values) against predicted values, where the red dashed horizontal line at zero represents prediction accuracy. Residual Analysis of ML Models Residual vs. Predicted for UTS Prediction in 7075-over-6061 FSAM Configuration. The GPR subplot (on the left) shows residuals in a very tight range of about ± 1.5 MPa, with much of the data clustered close to the horizontal line at 0. This overall tight and symmetric distribution of residuals shows little bias or variance, confirming that GPR has great capabilities in accurately modeling the nonlinear relationships among process parameters (tool rotation speed, traverse speed, and tilt angle) and tensile strength. The lack of patterns or heteroscedasticity provides assurance for GPR’s adequacy and reliability for high-fidelity UTS predictions. The SVR subplot (middle panel) indicates residuals have a wider range (approximately − 3 MPa to + 5 MPa). While residuals are primarily still near zero, some larger residuals are present in the prediction range of about 315–320 MPa. The distribution of residuals appears to have a slight systematic bias in the higher predicted values, which indicates SVR’s inability to fully reflect nonlinear effects in the FSAM process. In saying this, SVR can still predict moderate accuracy, resulting in residuals being within engineering acceptable tolerances for considering preliminary surrogate-assisted optimization of the process. The Bayesian Ridge Regression subplot (right panel) specifies the largest spread of residuals, which is related to a prediction range from about − 5 MPa to + 22 MPa. The positive skew in residuals indicates a consistent underestimation of UTS at higher actual values, which is a direct result of the Bayesian Ridge Regression model’s failure to incorporate the nonlinear effects of inputs, and particularly the complexity of the material stacking sequence. The residuals appear to be heteroscedastic (larger residuals with increasing predicted values), indicating that Bayesian Ridge Regression is an inappropriate, if not entirely infeasible, predictive model for estimating tensile strength of materials processed using variable boundary conditions. Overall, the residual analyses confirm the quantitative performance metrics by visually validating the predictive GPR in decreasing both bias (systematic deviations) and variance (observed spread). The underlying systematic deviations of SVR and noticeable errors in BBR highlight the essentiality of advanced nonlinear modeling methods for accurate prediction of mechanical properties in complex additive manufacturing processes. The characterization of the residuals gives suitable support for the adoption of GPR as the primary regression modeling framework for FSAM interpolated design to provide reliable strength predictions intrinsic to process design and quality. Residual histogram and kernel density estimation (KDE) for ultimate tensile strength prediction in 7075-over-6061 FSAM configuration The residual histogram is depicted in this Fig. 17 with overlaying kernel density estimation curves for three machine learning models predicting ultimate tensile strength in the 7075-over-6061 friction stir additive manufacturing configuration: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression. Each subplot shows the frequency distributions of the prediction errors (actual - predicted) for each of the nine experimental data points, which provides insight into model bias tendencies and the consistency of predictions. Residual Histogram and KDE for UTS Prediction in 7075-over-6061 FSAM Configuration. The GPR model (left side of panel) performed exceptionally well, with residuals tightly clustered around zero. The histogram shows five out of the nine predictions were within a very tight range of ± 0.25 MPa, and the kernel density curve shows a very narrow peak at zero residual. This narrow band in the residual distribution shows limited systematic bias and accuracy of predictions across all nine experimental conditions. The symmetric quality of the histogram and KDE curve supports the GPR’s reliability, as it shows no consistent tendency towards over- or under-predicting, leading to overall accuracy for estimating tensile strength in FSAM systems. The SVR model (middle panel) has a residual distribution that is broader, with values ranging from about − 3 to + 5 MPa. The bimodal shape of the histogram with peaks in the negative and positive residuals indicates that there were many predictions that were less accurate for some combination of parameters. The relatively flat KDE curve indicates a moderate spread of prediction errors, indicating that SVR can identify some general trends but has very little credibility across multiple process conditions. The variation in residuals creates too many errors for SVR to be useful in situations where predictions need to be accurate and consistent. The Bayesian Ridge Regression model (right panel) shows the most challenging residual characteristics by having a wide error distribution from − 5 to + 20 MPa. The histogram shows three clear error clusterings that show systematic failure in predictions in different parameter regions. The asymmetric multi-model KDE curve shows a significant bias towards underpredicting tensile strength values, specifically in terms of predicting higher tensile strength. Overall, this irregular distribution pattern illustrates the fundamental capability of Bayesian Ridge to model the nonlinear process-property relations in tensile strength in the 7075-over-6061 configuration. These analyses of residual distributions are intended to provide additional statistical validation of the model performance rankings suggested by the traditional metrics. The concentrated and zero-centered distribution for GPR further reinforces the decision to use it as the primary modeling technique. Moreover, the more dispersed patterns for SVR and Bayesian Ridge highlight their limitations for predicting tensile strength with accuracy in complex setups consisting of more than one material in the FSAM systems. Cross-validation mean absolute error (MAE) for ultimate tensile strength prediction in 7075-over-6061 FSAM configuration across all models The bar graph in Fig. S5 displays the performance of two different models: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression using mean absolute error (MAE) on their prediction of ultimate tensile strength in the 7075-over-6061 material system through cross-validation. The MAE indicates the average absolute measure between predicted tensile strength values and actual tensile strength jars in cross-validation folds. The GPR model exhibits MAE of about 0.3 MPa, suggesting a predictive strength accuracy that cannot be matched. This low error suggests that GPR can properly characterize complex nonlinear relationships between process parameters and ultimate tensile strength in the stacking configuration of 7075-over-6061, where interactions in material properties complicate predictive modeling. The next model, the SVR model, offered modest performance with an MAE of approximately 2.4 MPa, suggesting adequate accuracy for engineering purposes but significantly higher than the GPR model. The increased MAE indicates systemic limitations on fully characterizing the complex relationships of process parameters as they relate to ultimate tensile strength within a multi-material system. Lastly, the most complicated model, Bayesian Ridge Regression, has the worst performance, suggesting very poor usability based on an MAE of greater than 6.5 MPa, denoting substantial errors and severely reduced reliability. The notable high error of the models suggests linear modeling is inadequate to characterize nonlinear process-property interactions seen in FSAM processes. Cross-validation root mean square error (RMSE) for ultimate tensile strength prediction in 7075-over-6061 FSAM configuration across all models This bar chart in Fig. S6 illustrates the comparison of the cross-validation root mean square error (RMSE) values of the models in predicting the ultimate tensile strength (UTS) in the 7075-over-6061 friction stir additive manufacturing configuration with Gaussian Process Regression, Support Vector Regression, and Bayesian Ridge Regression, demonstrating RMSE as the square root of the average squared difference between predicted and actual UTS. Larger prediction errors are characterized by RMSE on the individual component (7075 or 6061) in the stacking configuration. The GPR model indicates the lowest RMSE value of approximately 0.6 MPa, demonstrating its ability to minimize larger predictive variability, leading to a more accurate predictive model. This is indicative of GPR’s ability to model the complex, nonlinear relationships between the FSAM processing parameters and tensile strength in this multi-material stacking configuration. The SVR provides a moderate RMSE value of approximately 2.7 MPa, which shows that the prediction accuracy is acceptable, although it has more significant errors from time to time. Consequently, SVR in making predictions about UTS values can effectively approximate the values and general trends but indicates larger RMSE values, indicating a struggle with consistently modeling the more complex relationships and interactions of parameters and, thus, less reliability according to specific conditions. In contrast, the Bayesian Ridge Regression model exhibits the worst RMSE value of approximately 9.1 MPa, while showcasing large-magnitude errors and improper fit to the nonlinear surface of the UTS response. This again suggests how linear modeling is inadequate to characterize the complex process–property relationships observed in FSAM, which requires diverse and probabilistic modeling from thermal and material heterogeneity. Additionally, these RMSE comparisons are consistent with the model rankings established by the other evaluation metrics, suggesting that GPR is the most suitable model for the evaluations of RMSE, SVR results are most similar to the middle rankings, and the Bayesian Ridge model does not represent the statistical modeling utility to accommodate the complexities of nonlinear relationships. Finally, GPR is the most favored regression framework for accurately predicting UTS and, thus, for optimizing appropriate FSAM process parameters. Cross-validation coefficient of determination (R²) for ultimate tensile strength prediction in 7075-over-6061 FSAM configuration across all models The bar graph in Fig. 18 shows the cross-validation values for R² for three regression methods, Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression, for the prediction of ultimate tensile strength (UTS) for a 7075-over-6061 friction stir additive manufacturing setup. Cross-Validation of (R²) for UTS Prediction in 7075-over-6061 FSAM Configuration Across All Models. The GPR model had an R² of 1.00, meaning there was a fit of experimental data and ability to model the nonlinear relationships between the process parameters and tensile strength. The SVR model had an R² of 0.96, demonstrating strong predictive ability but a slight misfit to performance. The Bayesian Ridge Regression fit had an R² of 0.57, indicating an inadequate ability to model the complex process-property relationships, demonstrating a much lower fidelity of modeling. The results demonstrate the excellence of nonlinear GPR for predicting UTS in multi-material FSAM. The SVR approach is very good but limitedly performed, and there are considerable shortfalls of linear regression methods like Bayesian Ridge regression as a predictive modeling approach in advanced manufacturing. GPR-based response surfaces for ultimate tensile strength prediction in 7075-over-6061 FSAM Configuration at varying tool tilt angles Figure 19 shows the Gaussian Process Regression (GPR) confidence surfaces for ultimate tensile strength (UTS) predictions with the 7075-over-6061 friction stir additive manufacturing (FSAM) process at three tool tilt angles (TTA = 1.5°, 1.0°, and 2.0°). Each 3D plot shows UTS predictions (z-axis) as a function of tool rotation speed (TRS; x-axis) and tool traverse speed (TTS; y-axis), allowing full exploration of how the process parameters interact to affect mechanical behavior. GPR-Based Response Surfaces for UTS Prediction in 7075-over-6061 FSAM Configuration at Varying Tool Tilt Angles. At TTA = 1.5°, the response surface is seen to be nearly flat due to the edge of the extrapolation region past where the limits for experimental data were recorded. The observed flatness in the surface indicates uncertainty in the predictions made. An acknowledgement of uncertainty in a prediction is important for considering denser sampling or testing parameter predictions at the extreme scales of either TRS or TTS. The flatness of the surface also indicates potential nonlinear influence at this tilt angle that has yet to be learned and addressed in the model. At TTA = 1.0°, the response surface shows a trend of smooth and gentle curvature with UTS trending upward as TRS increases and downward with greater TTS. The increasing UTS trend seen on the surface suggests trends consistent with improved bonding behavior reported in prior FSAM literature at higher TRS with lower TTS contributing to more heat generation, which could lead to systems being more consolidated, producing greater tensile strength. Further, the smooth slope of the surface suggests the parameters at TTA = 1.0° all have a predictable effect regarding desired material properties for surrogate-assisted optimization. At TTA = 2.0°, the surface distinguishes ridges and valleys, demonstrating a tangentially complex nonlinear interaction between TRS and TTS. The UTS peaks that correspond with intermediate TRS-TTS combinations show windows of conditions where maximum tensile strength is achievable at this tilt angle. Similarly, the UTS dips indicate that low tensile strength is occurring in part due to too much heat or deficiencies. The arrangement of the detailed topography shows that GPR can identify higher-order interactions and aids at the optimum across many conditions for FSAM. Overall, these confidence surfaces serve to confirm the viability of GPR in modeling complex, process-property relationships in multi-material FSAM constructs. It offers a means of facilitating decision-making when selecting TRS and TTS combinations at specific tilt angles that achieve a target UTS and improved data-driven decision-making for process design and manufacturing stability. Calibration, density, and partial dependence analyses for GPR-based ultimate tensile strength prediction in 7075-over-6061 FSAM configuration Figure 20 depicts the Gaussian Process Regression (GPR) model’s accuracy of prediction and interpretability of ultimate tensile strength (UTS) for the additive manufacturing of 7075 over 6061 aluminums (FSAM). Calibration, Density & Partial Dependence Analyses for GPR-Based UTS Prediction in 7075-over-6061. The first subplot is a calibration curve comparing actual UTS values (red markers) received from mechanical tensile testing versus the predicted values (blue markers) from the GPR model for each sample, along with the respective ± 2σ confidence intervals. Almost all measured UTS values fall within the credible bands of the GPR, which shows that accuracy of uncertainty quantification is well calibrated. The comparison of predicted points is tightly clustered around measured values, indicating that GPR is sufficiently reliable and with low bias in predicting tensile strength across all process conditions. In the second subplot, a 2D Kernel Density Estimation (KDE) plot shows actual versus predicted UTS values. The thick diagonal ridge along y = x shows strong predictive accuracy, and very little density off the diagonal shows that there is little systematic bias. Visually, this further shows that GPR can generalize across the UTS range (290–335 MPa) and maintain consistent predictive accuracy under both interpolation and mild extrapolation of the training distribution. The third subplot is a partial dependence plot that illustrates the impact of tool rotation speed (TRS) on predicted UTS when averaged across tool traverse speed and tilt angle. The results show a steep increase in the UTS values from 1000 to 1025 rpm along the TRS variable, but thereafter the UTS values vary around the mean of ~ 312 MPa for higher TRS values. This heightening plateau indicates there is a threshold where increasing the rotation speed above ~ 1025 rpm does not add any additional tensile strength. This affords a process engineer an opportunity to identify TRS values that obtain maximum UTS while not over-expending energy on increased rotation. In summary, the visualizations all demonstrate the GPR model’s tensile strength predictions are highly reliable and supported by an adequate degree of uncertainty and interpretable effects of the processing parameters. The present results provide strong decision support and guidance for optimizing FSAM processing parameters as part of the surrogate-based optimization of the multi-material construction process overall. Correlation matrices of actual versus predicted UTS for regression models in 7075-over-6061 FSAM configuration Figure 21 shows the Pearson correlation coefficient matrices for UTS predictions based on three machine learning regression models (i.e., Bayesian Ridge Regression, Gaussian Process Regression (GPR), and Support Vector Regression (SVR)) in the 7075-over-6061 friction stir additive manufacturing (FSAM) setup. Each subplot shows a color-coded heatmap with the correlation score in the center, indicating the model’s ability to predict the same trends as the experiments. Correlation Matrices of Actual versus Predicted UTS for Regression Models in 7075-over-6061. In the panel for Bayesian Ridge Regression, the Pearson correlation for actual vs. predicted UTS is 0.83: this indicates that the regression has a good degree of linear correlation; however, when compared with the achieved correlation scores of non-linear methods when predicting UTS, this score is much lower, indicating that Bayesian Ridge Regression has a poor ability to encode any non-linear interaction or complex multi-factorial effects in the FSAM parameter space. The panel for GPR shows a Pearson correlation score of 1.00, indicating that the predicted tensile strength is equal to the measured tensile strength for all samples. This is due, in part, to the non-parametric and flexible nature of GPR, which allows the modelling of high-order, multi-factorial, and non-linear relationships in the relationship between process variables and outcome properties; this is important for advanced multi-material manufacturing systems. The SVR panel indicates a high level of correlation (0.99), which is excellent but not perfect; this indicates that the SVR will capture reliable trends in the data; however, it is slightly less capable of perfectly predicting the value of all data points when the interactions within the process become highly nonlinear. Together, these correlation matrices lend intuitive and quantitative trend for model evaluation in the context of parameter surrogate-based optimization for FSAM. The evaluation supports that GPR is the most reliable and accurate regression technique for forecasting UTS, that SVR also has high-quality performance but is slightly less so than GPR, and that Bayesian Ridge regression is adequate for providing preliminary predictions or preparation of a specific linearized modeling case. Comparative visualization methodologies represent significant advances to data-informed approaches to process design and quality assurance in advanced manufacturing research and practice. Performance visualization and analysis for HV 7075 over 6061 Comparative scatter plots of actual versus predicted vickers hardness for 7075-over-6061 in FSAM configuration The presented Fig. 22 contains two scatter plots that compare actual Vickers hardness (HV) values to predicted HV values in the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration for three regression algorithms (Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression). Within each subplot, blue markers identify the hardness predicted for each of the nine samples, a red dashed line represents y = x,, and a gray solid line represents a fitted regression line through the data points. Comparative Scatter Plots of Actual versus Predicted Vickers Hardness for HV_7075-over-6061. The GPR subplot shows predicted hardness values nearly falling directly on the fit line for all hardness values (100–135 HV), and the GPR line is nearly superimposed on the fit diagonally. The tightness of the clustering from the predicted hardness to the fit line shows the ability of GPR to model the complex and non-linear interactions of the Tool Rotation Speed (TRS), Traverse Tool Speed (TTS), and Tilt Tool Angle (TTA) that determine the resulting material consolidation and hardness. The narrow error to the fit line confirms good GPR accuracy in predicting hardness and provides a very reliable means for predicting hardness in FSAM applications. The SVR subplot indicates a relatively good capture of trend, but there is a noticeable amount of scatter about the ideal line fit, especially at the intermediate hardness (113–125 HV) range evaluated. Even though the fitted regression line indicates a general upward trend, the fitted regression line suggests there is some systematic bias since there is underprediction for higher values and slight overprediction of lower values. The error suggests that SVR can moderately perform in approximating the nonlinear hardness response, and it does not consistently perform as well as GPR, which suggests potentially more appropriate utility of SVR for less severe applications or where computational time is more important than reliability of output. Predicted hardness values shown in the Bayesian Ridge Regression reflect a large deviation from the ideal line. The slope of the predicted hardness regression line is also nearly horizontal, which can be interpreted as underpredicted higher hardness and overpredicted lower hardness based on the overall vertical distribution of the data points. This and the data suggest the limitations of using linear regularization strategies to capture the nonlinear process–property relationships in FSAM, leading to substantial risk and error in its predictions. Overall, these scatter plots and trend lines are for each model’s prediction accuracy. The GPR model’s closely aligned results justify selecting it as the primary regression framework in an HV-predicting multi-material FSAM interpolated design model. SVR strikes a good balance of speed and accuracy; however, Bayesian Ridge is shown to be significantly less likely to predict high-fidelity hardness in this advanced manufacturing setting. Residual analysis of machine learning models for vickers hardness prediction in 7075-over-6061 FSAM configuration This Fig. 23 presents residual plots for three regression models: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression in predicting Vickers hardness (HV) in the 7075-over-6061 friction stir additive manufacturing (FSAM) setup. Each subplot presents residuals (actual values – predicted values) plotted against predicted HV values, with a horizontal red dashed line at zero showing prediction accuracy. Residual Analysis of ML Models for HV Prediction in 7075-over-6061 FSAM Configuration. The GPR model (left panel) exhibits exemplary residual behavior where the errors are in a remarkably small and tight range of approximately ± 2 HV. Most data points are very close to the zero line, and only one residual is observed as an outlier at approximately 1.8 HV. This is particularly noteworthy, as it indicates that all hardness levels (100–135 HV) were predicted with remarkable accuracy and consistency. As such, the prediction residuals are evenly distributed around zero with no trend of systematic patterns and no indicators of heteroscedasticity, suggesting that GPR was able to accurately identify the complex and nonlinear relationship between the process parameters and hardness properties. Specifically, GPR did not show a bias in the predictions, and the variance was very small, reflecting the potential of GPR to achieve high refinement hardness predictions in this multiple material FSAM configuration. The Support Vector Regression (SVR) model (middle panel) displays a significantly greater width in the residual scatter of around − 10 to + 10 HV. While several points are still centered around zero, there are extremely positive and negative outliers, indicating inconsistent predictive performance for certain combinations of parameters towards the prediction target. The greater spread indicates a moderate capability of SVR related to the nonlinear process-property interactions and an acceptable level of accuracy for general applications; however, it also indicates precision is limited for hardness in the critical range. The width of the scatter and non-symmetrical distribution were consistent with a small systematic bias, which is explicitly suggested in the clustering of larger positive residuals. The Bayesian Ridge Regression model (right panel) displayed the worst residual characteristics with errors in the range of approximately − 12 to + 15 HV. The width of the scatter, as well as the oversized and non-symmetrical shape of the distribution of other models, indicates that there is complete inconsistency in predictive performance and a systematic bias towards underpredicting hardness values, particularly at the higher values of hardness, with Bayesian Ridge Regression. The observed heteroscedastic pattern (i.e., residual error seems to vary as a function of predicted values) indicates that Bayesian Ridge Regression cannot accurately capture the nonlinear hardness-parameter relationships present in FSAM. Residual analyses provide visual trends that align with the quantitative performance rankings reported from other metrics. The GPR shows a tight, zero-centered distribution of residuals, confirming that this model translates to the most accurate hardness prediction. Furthermore, the dispersed patterns in SVR and Bayesian Ridge show that both models have constraints that prohibit high-fidelity modeling for advanced manufacturing applications that require quality control and surrogate-based process optimization to occur simultaneously. Residual histogram and kernel density estimation (KDE) for vickers hardness prediction in 7075-over-6061 FSAM configuration This Fig. 24 presents the distribution of prediction errors in the Vickers hardness (HV) prediction for three regression models: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression (BRR) on the 7075 on 6061 configuration for Friction Stir Additive Manufacturing (FSAM). Each subplot contains a histogram of error (actual value minus predicted value of HV) and a smooth Kernel Density Estimate (KDE) curve. The histogram and KDE curves allow some evaluation of the distribution of the prediction error, the amount of bias, and the amount of consistency of the residuals. Residual Histogram and KDE for HV Prediction in 7075-over-6061 FSAM Configuration. The residuals in the GPR subplot are centered around zero in clusters, with five out of the nine predictions falling within ± 0.5 HV of zero and a peak in the KDE centered around zero. This narrow distribution of error suggests a low prediction error and slight bias with the GPR model that underpredicted on average more than it overpredicted. The histogram and KDE curves are unimodal and symmetric, indicating that GPR modeled the nonlinear hardness response of the material with high fidelity and uncertainty quantification and a robust ability to interpolate within the bounds of the experimental region. The residuals of the SVR subplot illustrate a larger and more spread-out residual distribution between roughly − 10 HV and + 10 HV. The histogram displayed three residuals less than − 5 HV and two greater than + 5 HV, while the corresponding KDE curve is flatter and wider. The expansive distribution is indicative of SVR’s variability in accuracy; it captures general trends in hardness but does not provide detail in accuracy under complicated parameter interactions. The unequal peaks also indicate trends of a slight underprediction bias at higher hardness values; this indicates that the model’s limitations are systematic. In the subplot for Bayesian Ridge Regression, there is variability in the residuals across the greatest range, from − 12 to + 15 HV, and the residual distribution is visualized through a bimodal histogram and an uneven KDE curve. There are some significant clusters on the histogram near the − 5 HV and + 10 HV, suggesting that Bayesian Ridge consistently underpredicts at mid-range hardness and overpredicts at extremes. This heteroscedastic, multimodal error pattern emphasizes the limitations of linear regression in accounting for the nonlinear, multi-parameter dependencies for prediction of FSAM hardness. These residual histograms and corresponding KDE plots collectively provide a clear view of the error characteristics of each model in comparison to one another. The tight, centered residuals of GPR support its use as the primary regression model for predicting HV with high precision when working with FSAM advanced applications. SVR is efficient in terms of computations but has moderate error variability, and Bayesian Ridge is not adequate for reliable predictions of hardness in the complex manufacturing context. Cross-validation mean absolute error (MAE) for vickers hardness prediction in 7075-over-6061 FSAM configuration across all models The bar chart shown in Fig. S7 compares cross-validation MAE values for predicting Vickers hardness (HV) in the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration using Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian ridge regression. GPR produces an exceptionally low MAE of roughly 0.6 HV, which speaks to its accuracy in capturing the complex, nonlinear relationship between process parameters and hardness. The small error indicates GPR’s strength for producing high-fidelity hardness predictions in multi-material FSAM cases. SVR had a moderate MAE of about 5.0 HV; it provides a reasonable approximation of general hardness trends, but it also has a higher MAE, which indicates limitations in resolving the more subtle interactions among tool rotation speed, traverse speed, and tilt angle. Although SVR may be suitable for early process tuning decisions, it is not capable of producing sufficient accuracy for more stringent production quality control. Bayesian Ridge Regression yields the highest MAE, greater than 8.5 HV, as it struggles to capture the nonlinearities present in FSAM. The high average error indicates that linear methods are insufficient for reliable predictions of hardness in advanced additive manufacturing projects. These results show that GPR is a more accurate predictor, SVR is a compromise between simplicity and prediction reliability, and Bayesian Ridge Regression is not suitable for FSAM hardness score predictions requiring high precision. Cross-validation root mean square error (RMSE) for vickers hardness prediction in 7075-over-6061 FSAM configuration across all models The bar chart in Fig. S8 shows the cross-validation root mean square error (RMSE) of three regression algorithms, Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression, predicting Vickers hardness (HV) for the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration. GPR, which has an RMSE of about 0.9 HV, is good at minimizing large prediction errors and modeling the nonlinear relationships between process parameters (tool rotation speed, traverse speed, and tilt angle) and hardness. It is a low RMSE, indicating that GPR is well-suited for high-fidelity predictions of hardness in complex multi-material FSAM. The SVR RMSE was moderate, at around 6.0 HV, which means it was reasonably accurate at capturing general trends of hardness but was not specifically accurate. Again, the larger RMSE means that there were also substantial deviations, but on average, SVR was a compromise between reasonable accuracy and computational efficiency for coarse and exploratory tuning of the process. The Bayesian Ridge Regression RMSE was the largest of the simulations, at about 9.5 HV, and it had some very large errors exhibited, and it did not indicate a good fit to the awareness response surface of nonlinear hardness. This was a clear sign that linear modeling is not capable of serving as an adequate model of FSAM hardness predictions. The RMSE comparisons confirmed the order of strongest to weakest ability to predict with GPR, moderate accuracy with SVR, and obviously the inconsequential applicability of Bayesian Ridge Regression to modeling hardness accurately for precise experimental investigation and optimization of the FSAM process parameters. Cross validation coefficient of determination (R²) for vickers hardness prediction in 7075-over-6061 FSAM configuration across all models This bar graph in Fig. 25 displays a comparison of cross-validation R² values of each regression model, Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression, in predicting Vickers hardness (HV) for the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration. The R² values indicate the proportion of the variance in experimental hardness data that each regression model explains; R² values closer to 1 indicate a better fit. Cross Validation of R² for HV Prediction in 7075-over-6061 FSAM configuration across all models. The GPR model has an R² value of 1.00, indicating that predicted hardness values perfectly match with measured hardness values. This shows GPR’s ability to recognize the complexity of the nonlinear interactions of the various process parameters being studied (tool rotation speed, traverse speed, and tilt angle) in predicting hardness values for multi-material FSAM. GPR is a non-parametric method and is therefore sensitive to minor effects, which allows it to fit the data without losing predictive integrity. The SVR model achieves an R² value of 0.70, which indicates a good but imperfect explanatory power for the model predictions. It is evident that while SVR does a good job capturing the major trends in hardness values, the relatively low R² value indicates that predicting all of the data points is harder than in the GPR model, especially between the non-linear parameters and data points. SVR, therefore, represents a balance between prediction reliability and computational efficiency for the current application when a small margin of error is allowable. In comparison, the Bayesian Ridge Regression model results in an R² of 0.24, indicating a poor fit of linear regression to the experimental data. The low R-squared indicates that linear regularization methods are inadequate for modeling the highly nonlinear hardness response in FSAM; therefore, high unaccounted-for variability exists. Overall, the R² comparison confirms that GPR is the model for high-fidelity hardness prediction from FSAM parameters, optimally with a moderate model in SVR and an inadequate model in Bayesian Ridge. GPR-based response surfaces for vickers hardness prediction in 7075-over-6061 FSAM configuration at varying tool tilt angles The GPR confidence surfaces in Fig. 26 depict the predicted Vickers hardness (HV) for the 7075-over-6061 FSAM configuration based on three tool tilt angles (TTA): 1.5°, 1.0°, and 0°. Each 3D surface provides predictions of hardness values as a function of tool rotation speed (TRS) and tool traverse speed (TTS), all of which convey a lot of information about the interactions of the process parameters and the modeled material hardness properties. GPR-Based Response Surfaces for HV Prediction in 7075-over-6061 at Varying Tool Tilt Angles. At a TTA of 1.0°, the surface displays prominent three-dimensional curvature with hardness ranging from approximately 105 to 135 HV. The undulating surface suggests complicated nonlinear interactions in TRS and TTS, and the peaks and valleys indicate specific parameter combinations that are either optimal or suboptimal. Regions of enhanced hardness (green-yellow regions) are indicative of effective TRS-TTS densification of the materials at the interface. However, regions of reduced hardness (blue-purple regions) are indicative of TRS-TTS conditions that are either inadequately consolidated or alternatively excessively heated in the sample. At the TTA = 1.5°, the response surface shows a relatively flat, plateau-like topology, with hardness values clustering around the range of 400–600 HV for most of the parameter space. The flatness indicates that at this tilt angle, the interaction effects of TRS and TTS have less of a role, and the hardness values are relatively stable over a wide range of processing conditions. The similar colorations indicate more consistent material consolidation with an overall reduced sensitivity to the process parameter variation between configurations in this case. At a TTA of 2.0° (no tool tilt), the surface complexity remains similar to the 1.0° case but with small differences in surface topology. The hardness maps show similar ranges and patterns, suggesting that minor changes to the tool will have a considerable effect on process parameter sensitivity and optimal process windows. The minor differences between the 1.0° and 0° surfaces reflect the ideal nature of tightly controlling tool tilt angle and the consistency of the hardness property. The response surfaces that result from processing demonstrate GP’s ability to model the complicated process-property connections present in multiple material FSAM, enabling surrogate-based parameter optimization based on data for hardness produced for the new advanced additive manufacturing application. Partial dependence, density, and calibration analyses for GPR-based vickers hardness prediction in 7075-over-6061 FSAM configuration Figure 27 shows complete model interpretability and validation analyses for the Gaussian Process Regression (GPR) model predicting Vickers hardness (HV) in the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration. Partial Dependence, Density, and Calibration Analyses for GPR-Based HV Prediction in 7075-over-6061. The first subplot presents a partial dependence plot showing the isolated effect of tool rotation speed (TRS) on predicted Vickers hardness while averaging the effects of tool traverse speed and tilt angle. The curve indicates a drastic drop in hardness from roughly 122 HV to 115.5 HV as TRS increases from 1000 to 1025 rpm, and then a prolonged plateau of hardness at around 115 HV for increasingly higher rotation speeds (1025–1200 rpm). This acute transition suggests some threshold effect where increases in rotation speed led to a significant drop of information, and heat generation potentially softens the material. Beyond this threshold, further increases to TRS have little effect on hardness, which suggests the need for careful control of TRS if hardness is to be optimized within the lower range of the experimental domain. The second subplot depicts a 2D kernel density estimation (KDE) plot illustrating the relationship between actual hardness measurements and predicted values. In the plot, the clinic is evident along the diagonal with very few off-diagonal points. The concentrated diagonal density indicates better performance in prediction, while few off-diagonal points indicate that systematic error is almost zero. The concentrated diagonal densities further indicate that GPR provides consistent performance in prediction across the entire hardness range (100–135 HV), and the good performance with extrapolation and interpolation indicates that the model may generalize well. The third subplot is a calibration plot of actual measured hardness (red points) to GPR predicted hardness with ± 2σ intervals (blue points) for each of the experimental samples. Again, the predicted hardness centralized around the actual measurements and again completely lay within the GPR credible bands. Furthermore, the predicted values clustering very tightly around actual measurements again shows there is good calibration, hard to come by with a new model and made possible with adequate training data at hand. The excellent calibration results are concluded to indicate that the model provides credible, calibrated uncertainty quantification to enable consistent, risk-aware decision-making to manage the process with hardness optimization. In summary, these plots illustrated GPR confidence and ability to produce high-fidelity hardness prediction and produced actionable and useful insight for FSAM interpolated design, indicated that rotating speed is critical for achieving desired hardness. Correlation matrices of actual versus predicted vickers hardness for regression models in 7075-over-6061 FSAM configuration Figure 28 displays matrices of Pearson correlation coefficients that assess the linear relationship between measured and predicted values of Vickers hardness across three regression algorithms (Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge Regression) using the 7075-over-6061 friction stir additive manufacturing (FSAM) configuration. Each matrix is shown in a heatmap format with the correlation value represented with colored intensity to provide an intuitively consumable assessment of the given model’s performance. Correlation Matrices of Actual versus Predicted HV for Regression Models in 7075-over-6061. The correlation matrix from GPR shows a correlation coefficient of 1.00 to display a completely error-free linear correlation of the predicted values with the experimental hardness values. This outcome signifies the exceptional GPR capability by demonstrating its ability to identify the complex and non-linear functions that dictate hardness in the multi-material FSAM environment. The correlation indicates GPR efficiently modelled the non-linear effects of tool rotation speed, traverse speed, and tilt angle, and provided reliable predictive capability with no bias; in a manufacturing environment like this, GPR would be the ideal model for generating the most accurate hardness predictions. The SVR correlation matrix indicates a high 0.95 correlation, and the chart shows predictive performance is therefore strong with minor deviations from the mean. SVR successfully modelled the dominant trends of hardness variation, but with a slightly weaker correlation, SVR performed well but is still limited in capturing the finest non-linear effects in predicting the parameter-property relationships. The level of predictive capability is provided as a trade-off between computational efficiency and predictive accuracy, where this would be appropriate for engineering applications having good predictive capability when slight deviations can be tolerated. The correlation matrix of Bayesian Ridge Regression shows a moderate correlation of 0.53, which elucidates the limitations to modeling the hardness response. The correlation of 0.53 demonstrated the ineffectiveness of linear regularization methods to capture the inherently non-linear process-property dependencies demonstrated in FSAM. The weak association identifies that Bayesian Ridge continues to approximate thermal and mechanical interactions that govern hardness development and has demonstrated poor predictive reliability. Overall, correlation analyses support previous studies in establishing a clear performance hierarchy. GPR is a clearly preferred choice for high-fidelity hardness prediction, SVR provides reasonably reliable models (although imperfect), and Bayesian Ridge is not suitable for well-defined precision applications with respect to hardness prediction to be used in advanced FSAM interpolated design workflows. Model performance summary and prediction The cross-validation metrics indicate that Gaussian Process Regression (GPR) has an advantage over Support Vector Regression (SVR) and Bayesian Ridge Regression for each of the four target properties in both stacking orders of materials. For the 6061-over-7075 configuration, GPR produced a mean absolute error (MAE) of 0.55 MPa, a root mean square error (RMSE) of 0.92 MPa, and an R² of 0.996 for ultimate tensile strength (UTS). The performance of SVR was quite a bit lower, with an MAE of 4.32 MPa, an RMSE of 5.44 MPa, and an R² of 0.872; Bayesian Ridge was even less able to model the trends in data (MAE = 13.93 MPa, RMSE = 15.25 MPa, and R² = −0.003). Similar findings are observed for Vickers hardness (HV) in this stacking order, with GPR again outperforming SVR and Bayesian Ridge (MAE = 1.05 HV, RMSE = 1.65 HV, R² = 0.988 and MAE = 8.20 HV, RMSE = 11.20 HV, R² = 0.458, and MAE = 13.04 HV, RMSE = 15.56 HV, R² = −0.047, respectively). For the 7075-over-6061 configuration, GPR again outperformed the other two models (MAE = 0.27 MPa, RMSE = 0.54 MPa, R² = 0.999) for UTS, and SVR was slightly less accurate (MAE = 2.34 MPa, RMSE = 2.70 MPa, R² = 0.965), while Bayesian Ridge was least effective (MAE = 6.47 MPa, RMSE = 9.14 MPa, R² = 0.601). For hardness in this stacking order, GPR performed considerably better than SVR and Bayesian Ridge (MAE = 0.58 HV, RMSE = 0.94 HV, R² = 0.993, MAE = 4.96 HV, RMSE = 6.03 HV, R² = 0.695, MAE = 8.54 HV, RMSE = 9.53 HV, and R² = 0.239, respectively). The results clearly specify that GPR is the top-performing model, in terms of MAE, for all outputs evaluated, which shows that it has capacity compared to other regression-based models in learning and modeling complex, nonlinear relationships for FSAM process parameters (tool rotation speed, traverse speed, and tilt angle) to mechanical properties. To show a practical application, prediction was performed for a new process configuration (TRS = 1000 rpm, TTS = 30 mm/min, TTA = 1°). In this case, GPR is predicting UTS_6061over7075 = 287.23 MPa and HV_6061over7075 = 106.21 HV, and UTS_7075over6061 = 291.16 MPa, and HV_7075over6061 = 113.16 HV. All the predicted values fall within the ranges observed in the actual experiments, and therefore GPR’s application for learning and rapid dataset-driven FSAM surrogate-based process optimization and quality assurance is justified. And the RBF-GPR framework’s predictions will be verified through additional FSAM tensile and hardness testing using optimized process parameters and potential materials. These predictions are consistent with trends shown in the literature (Hassan et al., 2021; Venkit & Selvaraj, 2022; Dong et al., 2021), which will further confirm the realism within the modeled parameters. Parameter contribution analysis using SHAP SHAP (SHapley Additive exPlanations) values were computed separately for each output variable (UTS and Vickers hardness) and for each material stacking configuration. The SHAP background dataset consisted of the synthetic parameter grid generated via RBF interpolation, rather than independent experimental measurements. For Gaussian Process Regression, SHAP values were estimated using a model-agnostic approximation based on sampling from the surrogate-trained prediction function. Consequently, the resulting SHAP values quantify relative parameter influence within the learned surrogate response surface and should not be interpreted as indicative trends of causal dominance in the physical FSAM process. The SHAP background dataset (baseline) was created using the synthetic parameter grid with RBF generation of synthetic parameters (parameter grid) to represent the response surface created by using only the limited experimental anchors. The background distribution does not use any independent experimental data as an additional source. To estimate SHAP values for the Gaussian Process Regression, the Kernel SHAP was used as a model-agnostic approximation of SHAP where the prediction made by the GPR model is sampled from the trained GPR surrogate model. A total of 1,000 Monte Carlo samples is drawn for each SHAP calculation to obtain the Shapley values with stable convergence; therefore, the SHAP value for each parameter reflects the relative influence that each of the parameters has on the output of the trained surrogate model. In addition, differences in the results produced by different outputs are due to different output-specific surrogate behaviors rather than due to one global dominance ranking. This evaluates the contributions of key FSAM process inputs tool rotation speed (TRS), traverse speed (TTS), and tilt angle (TTA) to predict mechanical properties (ultimate tensile strength and Vickers hardness) for independent 6061-over-7075 and 7075-over-6061 configurations, as well as quantify the distributions of synthetic process inputs for sensitivity sampling. The SHAP mean percent contributions show that TRS clearly controlled prediction, accounting for 73.97% and 73.95% in the output variance for the first and second passes, respectively, while TTS accounted for 23.10% and TTA for 2.93%. When evaluating all 882 synthesized samples, TRS was the only impactful parameter in most cases (751, 85.15%), TTS was impactful in 126 cases (14.29%), and there was one instance of TTA being the first influential parameter out of 882 synthetic samples (5, 0.57%). There is also a detailed SHAP summary for GPR models presenting important rankings and contributions for each property and stacking sequence. For UTS_6061over7075, TRS has an influence of 41.75%, TTS is 33.25%, and TTA is 25.00%. For HV_6061 over 7075, TTS is ranked first at 49.89%, followed by TTA at 25.79% and TRS at 24.32%. For UTS_7075over6061, TRS has the most influence at 45.74% to TTS at 34.60%, and TTA is at 19.66%. For HV_7075over6061, TTS is 48.10%, TRS is 31.54%, and TTA is 20.36%. Each of TRS and TTS emerges as most influential 2 times in these 4 outputs. In the analysis of the synthetic data distribution, TRS is distributed from 1000 to 1200 rpm (mean = 1100.00 rpm, σ = 60.59), where the most frequently occurring values are common at TRS values of 1000, 1010, and 1020 rpm (all equal 4.76%). TTS is distributed from 30 to 40 mm/min (mean = 35.00 mm/min, σ = 3.42), commonly at TTS values of 30, 32, and 34 mm/min (equal at 16.67%). TTA is distributed from 1.0° to 2.2° (1.60°, σ = 0.40), where 1.0°, 1.2°, and 1.4° equally occur in 14.29% of samples. The top five combinations of TRS: TTS: TTA that occurred each occurred only once, reflecting the relatively extensive and consistent sampling of the parameter space. Statistical characterization further substantiated comprehensive coverage, with TRS having a minimum and maximum of 1000 and 1200 rpm, TTS having a minimum and maximum of 30 and 40 mm/min, and TTA having a minimum and maximum of 1.0° and 2.2°. Combined observations for both SHAP analysis and synthetic data statistics ultimately provided holistic insights into estimates of parameter sensitivities and, thus, more comprehensive trends from which to catalyze data-driven surrogate-based optimization of the FSAM process settings for mechanical property performance. Table 11 depicts the top ten synthetic parameter sets where tool rotation speed (TRS) had the highest influence on the GPR ultimate tensile strength (UTS) prediction, given the FSAM condition. Each row of the table lists TRS, traverse speed (TTS), tilt angle (TTA), the GPR UTS prediction, the percent SHAP contribution of the TRS parameter for that sample, and the identification of the person with the highest influence. Table 11Top 10 samples with highest % contribution from TRS.TRSTTSTTAGPR-predContribution_%_TRSMost_Influential_Parameter1120341.6311.28006699.248253TRS1180341.6301.49769499.037897TRS1010361.6300.85320998.809224TRS1120341.8311.36384998.692893TRS1010361.8300.93874098.557909TRS1190341.6298.85752898.556355TRS1180341.8301.59689098.522671TRS1090361.6314.24816798.234057TRS1190341.8299.01675798.053266TRS1020361.6301.95034498.011947TRS Top 10 samples with highest % contribution from TRS. The maximum TRS SHAP contributions approach 99% or greater, with the sample with the highest contribution (TRS = 1120 rpm, TTS = 34 mm/min, TTA = 1.6°) showing a contribution of 99.25%. Subsequent (TRS = 1180 rpm, TRS = 1010 rpm, and TRS = 1190 rpm) contributions in the table follow closely and occur under similar processing conditions and contribute within 98.6–99.0%. All top ten samples report TRS as the parameter with the highest influence, confirming the prominent influence of TRS in the driving of UTS predictions despite changes to external conditions that include TTS and TTA. This analysis specifically highlights the parameter sets of interest, which reflect considered ranges of operation where tool rotation speed impacts performance measures related to tensile strength to the greatest degree. The purpose of this analysis is to facilitate process improvement practices using causal inference analyses during the testing and analysis of this parameter. Table 12 below shows the ten synthetic combinations of processing parameters in which tool rotation speed (TRS) played the least contribution to the Gaussian Process Regression (GPR) prediction of ultimate tensile strength (UTS) in FSAM configuration. Each row in the table shows the TRS, traverse speed (TTS), tilt angle (TTA), the associated GPR predicted UTS values, the percent SHAP contribution of TRS, and the parameter that had the largest influence on that sample. Table 12Bottom 10 samples with lowest % contribution from TRS.TRSTTSTTAGPR-predContribution_%_TRSMost_Influential_Parameter1010402.0313.2931600.197411TTS1010402.2313.3833960.516636TTS1010401.8313.2058750.648534TTS1010401.6313.1215890.723752TTS1010401.4313.0403520.796928TTS1010401.2312.9622130.877947TTS1150302.2307.8125061.099633TTS1010401.0312.8872251.274315TTS1150301.8307.6839631.329427TTS1140401.0305.9982252.183828TTA Bottom 10 samples with lowest % contribution from TRS. For the samples displayed, TRS contributions vary from a low of 0.20% to 2.18%, indicating that TRS had little to no effect on predicted UTS. In these samples, TTS had the most influence in nine out of ten cases, while TRS contribution to the predicted UTS was below 1% for the first six entries of Table 7, with TRS fixed at 1010 rpm and TTS at 40 mm/min. There was a single case (TRS = 1140 rpm, TTS = 40 mm/min, TTA = 1.0°) where the TRS contribution surpassed 1% and TTA became the parameter with the most influence at 2.18% TRS contribution. The reasoning for the clustering of parameters, high TTS, and varying TTA indicates a regime of processing where traverse speed and tilt angle interactions dominate strength outcomes, making rotation speed almost irrelevant. This elaboration indicates clear processing conditions in which future optimization endeavors should focus on TTS and TTA occurring at the expense of the TRS process parameter for increasing tensile performance. Apparent differences between global SHAP summaries and per-output SHAP rankings arise because SHAP values were computed independently for each surrogate-trained model rather than from a unified multi-output formulation. SHAP value distributions and heatmap contribution for influence of parameter on UTS and HV predictions The sample-specific impact of the three process parameters tool rotation speed (TRS), traverse speed (TTS), and tilt angle (TTA) on predicted mechanical properties of friction stir additive manufacturing (FSAM) as measured using the Gaussian Process Regression (GPR) model is illustrated in Fig. 29. Boxplots of SHAP (SHapley Additive exPlanations) values for each parameter are displayed in the left panel, which represents the respective distributions of the marginal contributions of each parameter for all synthetic samples. SHAP Analysis and Heatmap Contribution for Influence of Parameter on UTS and HV Predictions. TRS’s boxplot displays the widest interquartile range (IQR) and highest median, suggesting it was the most influential and variable parameter in predicting ultimate tensile strength (UTS) and Vickers hardness (HV). While TTS had an IQR and median slightly below zero, it still had a substantial effect as the second most influential parameter. In contrast, TTA had SHAP values clustered around zero, with limited spread, and as such was indicative of a relatively minor influence overall. The heatmap on the right represents the per-sample percent SHAP contribution, where each column is a synthetic sample index. The relative intensity of the color represents the contribution of each of those parameters to the model prediction for that sample. A band of dark blue is present across the TRS row in most cases, indicating TRS contributed 70% or greater. The TTS row has bands of more intense color in some of the heatmap instances with the accompanying sentiment that those samples at the time were dominated by traverse speed. The TTA row is always a pale color, indicating it was contributing little across this dataset. In total, these visualizations provide you and the reader with a statistical account of the contribution, and the per-sample mapping of the importance of those parameters was provided in the previous sections. Overall, TRS contributes the most strongly to variance in the model output, TTS contributes little but is the largest contributor in some samples, and TTA contributes next to nothing. In conclusion, predictive model visualizations help forecast potential sensitivities of processes and hopefully will inform optimization efforts with respect to multi-material FSAM. Optimal parameter analysis for UTS 6061-over-7075 The SHAP analysis clearly indicates that tool rotation speed (TRS) is the most significant predictor of ultimate tensile strength of the 6061-over-7075 FSAM configuration, contributing 73.95% of output variance in the model. Tool traverse speed (TTS) was second in the hierarchy at 23.10%, while tool tilt angle (TTA) had negligible impact at 2.94%. This ranking shows TRS is the most significant control variable to obtain tensile properties of interest for this stacking configuration. Parameter influence frequency analysis of the 882 synthetic samples confirmed the dominance of TRS influence, and TRS was the most prominent parameter in 751 of the samples (85.15%). TTS was the most important parameter in 126 instances (14.29%), typically when the influence of TRS was minimized in the sample design. TTA only occupied the most significant spot in 5 samples (0.57%), demonstrating that TTA influence in achieving consistently tensile strength is not important. With this distribution optimization, the method should focus on TRS control as the most significant parameter to achieve consistent increases in tensile strength. Using global optimization, the parameter combination provides a maximum predicted ultimate tensile strength (UTS) of 6061 over 7075: tool rotation speed (TRS), traverse speed (TTS), and tilt angle (TTA) are 1000 rpm, 40 mm/min, and 2.0°, respectively, and the predicted UTS is 327.77 MPa. The TRS is at the lower bound of the experimental range, indicating that the use of a lower rotation speed reduces excessive heat input. The increase in TTS and TTA also assists with consolidation of the two metals as they provide a conducive plastic flow and mechanical interlock of the metals. These results provide practical guidance for the surrogate-based optimization of FSAM processes with a focus on maximizing TTS according to the process. TRS is found to be the critical parameter to control when establishing maximum tensile performance. Table 13 displays the ten synthetic parameter combinations where TRS has the highest influence on Gaussian Process Regression (GPR) predicted UTS for the 6061 over 7075 FSAM configuration. Each row depicts TRS, TTS, TTA, and the GPR-predicted UTS value. Within the surrogate-trained models, TRS exhibits the highest average SHAP contribution; however, relative importance varies between UTS and hardness predictions. Table 13Top 10 samples with highest % contribution from TRS for UTS predictions (6061 over 7075).TRSTTSTTAGPR-pred UTS (6061/7075)Contribution % TRSMost Influential Parameter1120341.6311.28006699.248263TRS1180341.6301.49769499.040611TRS1010361.6300.85320998.811615TRS1120341.8311.36384998.690216TRS1190341.6298.85752898.558108TRS1010361.8300.93874098.557892TRS1180341.8301.59689098.522651TRS1090361.6314.24816798.235708TRS1190341.8299.01675798.053249TRS1020361.6301.95034498.014889TRS Top 10 samples with highest % contribution from TRS for UTS predictions (6061 over 7075). Among the top ten samples measured, contributions from TRS exceeded 98%, with the highest TRS contribution occurring at TRS = 1120 rpm, TTS = 34 mm/min, and TTA = 1.6° (TRS contribution = 99.25%). In addition, the TRS contributions were similarly high at TRS values of 1010 rpm, 1180 rpm, 1190 rpm, and 1020 rpm, revealing that rotation speed is the dominant controlling factor of tensile strength predictions for these different combinations. In fact, all ten samples showed that TRS was the greatest contributor, validating the finding of TRS being the main governing factor across multiple combinations of TTS and TTA. This indicates the fact that TRS was chosen from the middle and the upper rpm ranges for multiple combinations of TTS and TTA values, indicating the importance of being able to accurately measure and control rotation speed values if the UTS of the multi-material FSAM is increased. Knowledge gained from these insights will help to optimize future processing factors by studying variations in TRS that could significantly improve tensile performance. Table 14 presents the ten combinations of synthetic parameters with the least contribution of rotation speed (TRS) on GPR predictions of ultimate tensile strength (UTS) in the 6061-over-7075 FSAM scenario. Each row contains TRS, traverse speed (TTS), tilt angle (TTA), GPR-predicted UTS value, the percent SHAP contribution of TRS, and the most influential parameter for that sample. Table 14Bottom 10 samples with lowest % contribution from TRS for UTS predictions (6061 over 7075).TRSTTSTTAGPR-pred UTS (6061/7075)Contribution % TRSMost Influential Parameter1010402.0313.2931600.199371TTS1010402.2313.3833960.517591TTS1010401.8313.2058750.649510TTS1010401.6313.1215890.724742TTS1010401.4313.0403520.797904TTS1010401.2312.9622130.878910TTS1150302.2307.8125061.086428TTS1010401.0312.8872251.275259TTS1150301.8307.6839631.344230TTS1140401.0305.9982252.168507TTA Bottom 10 samples with lowest % contribution from TRS for UTS predictions (6061 over 7075). In these samples, TRS contributions are generally very low, ranging from around 0.20% to 2.17%. For six of the bottom entries, TRS is held at the lower bound (1010 rpm) and TTS is at the upper bound (40 mm/min), allowing TTS to dominate influence with contributions near 90%. In two entries that allow TRS to increase to 1150 rpm, TTS remains the most influential factor, but TRS contributes around 1–1.3%. Only in the last entry (TRS = 1140 rpm, TTS = 40 mm/min, TTA = 1.0°) does TTA marginally exceed TRS in contribution with TRS at 2.17%. This is informative to highlight process regimes where TRS has the least influence and to illustrate that traverse speed and tilt angle may require more attention to maximize UTS. Optimal parameter analysis for HV 6061-over-7075 In the prediction for Vickers Hardness of 6061 compared to 7075, the SHAP analysis revealed that tool rotation speed (TRS) makes the highest average contribution (55.70%) to the variance of the model output, followed by tool traverse speed (TTS) at 41.58% and tool tilt angle (TTA) at 2.72%. For over 882 synthetic samples, TRS was the most important parameter in a total of 576 (or 65.31%) samples, TTS was most important in 305 (or 34.58%) samples, and TTA as the most important parameter occurred in one sample (or 0.11%), which further affirmed TRS as the primary driver for Vickers hardness prediction. Global surrogate-based optimization of the Gaussian Process Regression model determined the parameter combination that provided the highest Vickers hardness prediction: TRS = 1100 rpm, TTS = 32 mm/min, and TTA = 2.2° predicted hardness of 112.57 HV, which is an indication that intermediate rotation speed in combination with moderate traverse speed and the high tilt angle provided the material consolidation, and therefore, the good hardness prediction. These results allow greater confidence relative to the selection of FSAM process parameters because they provide unequivocal, data-driven, and trend-based suggestions. TRS control should occur first, and TTS and TTA can be adjusted more finely as they relate to TRS for maximum hardness, which supports more efforts toward optimization and quality control within multi-material additive manufacturing. Table 15 shows the ten combinations of synthetic parameters for which tool rotation speed (TRS) contributes most in terms of Gaussian Process Regression (GPR) predictions of Vickers hardness (HV) for the 6061-over-7075 FSAM configuration. Each row shows TRS, tool traverse speed (TTS), tilt angle (TTA), the GPR predicted HV value, the percent SHAP contribution made by TRS, and the most impactful parameter. The GPR analyses show similar contributions, ranging from 94 to 99.21%, with 99.21% being for the ranked combination in the top row (TRS = 1010 rpm, TTS = 36 mm/min, TTA = 1.6°). This dominance suggests that TRS has a substantial impact regardless of the other parameters. The ranges of these parameters have tool rotation speeds from 1010 to 1200 rpm and traverse speeds from 34 to 36 mm/min and predicted hardness values from 82.77 to 95.28 HV. This example shows the overall dominance of TRS contribution, with all ten combinations identifying TRS as the most impactful parameter. There are indicative trends that TRS determines the level of hardness developed in this material configuration. These results suggest that practitioners should pay close attention to the ability to accurately control TRS in the development of hardness in 6061-over-7075 FSAM applications because changes in rotation speed directly result in drastic modeling shifts in hardness. Table 15Top 10 samples with highest % contribution from TRS for HV predictions (6061 over 7075).TRSTTSTTAGPR-pred HV (6061/7075)Contribution % TRSMost Influential Parameter1010361.695.22555299.210112TRS1010361.895.24385798.931980TRS1010361.495.20545397.676013TRS1110361.692.92724297.526247TRS1010362.095.26039497.266180TRS1010361.295.18354196.201668TRS1010362.295.27519895.755658TRS1200341.695.28068794.973168TRS1060361.682.77325594.724503TRS1050361.683.24319894.639041TRS Top 10 samples with highest % contribution from TRS for HV predictions (6061 over 7075). Table 16 lists the ten fabricated parameter sets where tool rotation speed (TRS) has the smallest effect on Gaussian Process Regression (GPR) predictions of Vickers hardness (HV) in the 6061-over-7075 FSAM setup. Each entry shows TRS, traverse speed (TTS), tilt angle (TTA), GPR predicted HV, the percent SHAP contribution of TRS, and the most significant parameter. Table 16Bottom 10 samples with lowest % contribution from TRS for HV predictions (6061 over 7075).TRSTTSTTAGPR-pred HV (6061/7075)Contribution % TRSMost Influential Parameter1140381.087.6129670.671044TTS1140381.287.7097671.742287TTS1140381.487.8070322.417006TTS1150302.095.9019542.639685TTS1150302.295.9903222.921582TTS1140381.687.9047563.142319TTS1140381.888.0029333.697591TTS1150301.895.8140023.752664TTS1100401.277.2336494.005015TTS1140382.088.1015574.143727TTS Bottom 10 samples with lowest % contribution from TRS for HV predictions (6061 over 7075). The contributions of TRS for these samples are minimal, from 0.67% to 4.14%. TTS is the largest contributor in nine out of ten samples, and in the first sample (TRS = 1140 rpm, TTS = 38 mm/min, and TTA = 1.0°), the contribution of TRS is below 1%. TRS remains constant between 1140 and 1150 rpm, while TTS fluctuates between 30 and 38 mm/min. As TTS increases, the contribution of TRS increases, peaking at 4.14% only when TRS is increased to 1140 rpm, with TTS at 38 mm/min and TTA at 2.0°. These results indicate that there are regimes where hardness is influenced primarily by traverse speed more than rotational speed, and the results should prompt practitioners to adjust TTS when TRS effects are negligible. Optimal parameter analysis for UTS 7075-over-6061 SHAP analysis, in determining UTS outlets for the 7075-over-6061 FSAM (friction stir additive manufacturing) configuration, indicates that TRS (tool rotational speed) contributes 75.08% to output variance, TTS (tool traverse speed) contributes 24.18%, and TTA (tool tilt angle) contributes 0.74%. Of the total of 882 synthetic samples simulated, TRS is the most influential parameter in 763 of the samples (86.51%), TTS in 119 samples (13.49%), and TTA in none, demonstrating the dominant role of the TRS parameter in predicting tensile strength. The global surrogate-based optimization process of the Gaussian Process Regression model identifies the parameter combination to provide the greatest predicted UTS, the combination being TRS = 1200 rpm, TTS = 30 mm/min, and TTA = 2.0°, which produces a UTS of 335.44 MPa. This maximum predicted UTS corresponds to the highest rotation speed along with the lowest traverse speed and highest tilt angle, allowing maximum heat input and stirring of materials, which in turn provides ultimate tensile strength. The process and results in this work provide clear pathways and data to assist FSAM practitioners in determining the factors that contribute to maximum tensile performance in additive manufacturing of multi-materials. Participants in FSAM projects will benefit from careful monitoring and control of TRS, as well as from adjusting TTS and TTA to maximize tensile performance in the fabricated welds. Table 17 displays the ten synthetic parameter combinations in which tool rotation speed (TRS) has the most influential impact on the Gaussian Process Regression (GPR) predicted ultimate tensile strength (UTS) in the 7075-over-6061 FSAM case. For each row, TRS and traverse speed (TTS) and tilt angle (TTA) are given, along with the GPR predicted UTS, the percentage of TRS contribution accounted for by SHAP, and the most significant contributing parameter. Table 17Top 10 samples with highest % contribution from TRS for UTS predictions (7075 over 6061).TRSTTSTTAGPR-pred UTS (7075/6061)Contribution % TRSMost Influential Parameter1200381.6328.23782399.288464TRS1010361.6300.39745499.255719TRS1140361.6317.41963499.249213TRS1200381.8328.32065899.235187TRS1010361.8300.45333899.176525TRS1040361.6304.48218499.087048TRS1030361.6302.95916699.077262TRS1140361.8317.44263399.061907TRS1030361.8302.98725799.003241TRS1040361.8304.50674699.002755TRS Top 10 samples with highest % contribution from TRS for UTS predictions (7075 over 6061). TRS contributions for these examples are all above 99.00%, with maximum percentage contributions of 99.29% at TRS = 1200 rpm, TTS = 38 mm/min, and TTA = 1.6°. TRS is the most important parameter in each set of ten parameter combinations, for TTS = 36 and 38 mm/min, and for TTA = 1.6 and 1.8°. The value of predicted UTS varies from approximately 300.40 MPa to 328.32 MPa, providing indicative trends that the high rotation speeds have consistently high contributions to strength. These findings continue to demonstrate the importance of controlling the tool rotation speed in achieving mechanical performance in 7075 deposited over 6061 and are instructive for process engineers on the focus of adjustments. The results provide trends that TRS should be centered on strength enhancements. Table 18 shows the ten synthetic input combinations where the influence of tool rotation speed (TRS) is the least in GPR-predicted ultimate tensile strength (UTS) for the 7075-over-6061 FSAM configuration. Each row provides information on TRS, traverse speed (TTS), tilt angle (TTA), the predicted UTS, and the percent SHAP contribution of TRS, and finally the most influential parameter. Table 18Bottom 10 samples with lowest % contribution from TRS for UTS predictions (7075 over 6061).TRSTTSTTAGPR-pred UTS (7075/6061)Contribution % TRSMost Influential Parameter1090302.2308.6808354.514226TTS1090301.8308.5918584.642657TTS1090301.6308.5497634.708081TTS1090301.4308.5092964.738768TTS1090301.2308.4704764.789166TTS1090301.0308.4333264.807077TTS1090302.0308.6355564.811043TTS1000402.0323.4982778.629721TTS1090321.0310.52030912.070505TTS1090321.2310.56032012.314308TTS Bottom 10 samples with lowest % contribution from TRS for UTS predictions (7075 over 6061). In these cases, TRS contributions ranged for all samples from 4.51% to 12.31%, which is much lower compared to the dominant values for TRS parameters. The first six sample rows use TRS of 1090 rpm and TTS of 30 mm/min, while TTA varies from 1.0° to 2.2°, and TRS contributions range from 4.51% to 4.81%, where TTS achieves a dominant influence. The seventh sample (TRS = 1000 rpm, TTS = 40 mm/min, TTA = 2.0°) shows TRS contribution increasing at 8.63% but still dominated by TTS. In the second to last row, the TRS was held low (at 1090 rpm), but TTS is reduced to just under 32 mm/min and shows an increasing TRS contribution (12.07% to 12.31%), but again shows TTS having the most influence. Overall, these observations help target processing conditions where it can be predicted that traverse speed will control tensile strength, and thus, during optimum processing development or validation, focus on TTS area adjustments where TRS effects are minor. Optimal parameter analysis for HV 7075-over-6061 The results of the SHAP analysis for predicting Vickers hardness (HV) in the hybrid build of the 7075-over-6061 FSAM setup suggest that tool traverse speed (TTS) has a slightly larger contribution to the variance in model output than tool rotation speed (TRS), with TTS apportioning 50.73% variation in output and TRS at 47.06%. Tool tilt angle (TTA) only contributed to the variance at 2.20%. On the other hand, out of the total of 882 synthetic samples, TTS was the major contributor in the output variance for a total of 433 samples (49.09%), TRS was attributed to a total of 449 samples (50.91%), and TTA did not lead at all, which indicates the comparable importance of rotation and traverse speed in the development of hardness. Through global surrogate-based optimization of the Gaussian Process Regression model, a parameter set would generate the maximum predicted hardness in a hybrid build of 7075-over-6061 with the following settings: TRS = 1100 rpm, TTS = 30 mm/min, TTA = 2.2°; the predicted HV value would be 135.31. This suggests that the combination of moderate rotation speed, minimum traverse speed, and highest tilt angle provides the most effective combination to balance heat input and the material’s ease of flow through the rotating tool, resulting in bonding behavior consistent with typical FSAM consolidation mechanisms and hardness output was observed. The results provide FSAM practitioners with information that may allow them to prioritize adjustment of TTS if TRS and TTA are already optimal, to produce better hardness in multi-material AM applications. Based on the SHAP analysis of the 7075-on-6061 FSAM setup presented in Table 19, tool traverse speed (TTS) accounts for 50.73% of variance in the Vickers hardness (HV) predictions, tool rotation speed (TRS) accounts for 47.06% of variance, and tool tilt angle (TTA) accounts for only 2.20%. The strong influence of parameters is evident across the 882 synthetic samples, as TTS was the most influential parameter in 433 samples (49.09%), and TRS was the most influential parameter in 449 samples (50.91%). In other words, both combinations of speeds were highly influential, affecting the hardness results. Table 19Top 10 samples with highest % contribution from TTS for HV predictions (7075 over 6061).TRSTTSTTAGPR-pred HV (7075/6061)Contribution % TTSMost Influential Parameter1110401.6106.72880498.394766TTS1110401.8106.83367498.178156TTS1110402.0106.94064997.204104TTS1110401.4106.62611197.062374TTS1080341.6115.95221096.857843TTS1110402.2107.04964996.313360TTS1080341.8116.01440496.211277TTS1090381.6110.17253596.038824TTS1110401.2106.52566695.781666TTS1090381.8110.27506095.007625TTS Top 10 samples with highest % contribution from TTS for HV predictions (7075 over 6061). The optimized Gaussian Process Regression model produces a parameter set that yields the highest predicted HV of 135.31: TRS = 1100 rpm, TTS = 30 mm/min, and TTA = 2.2°. The results support that a medium combination of TRS (rotation speed) with the lowest TTS (traverse speed) and the highest TTA provides prolonged heat input and plastic deformation, which is responsible for increased diffusion between the interlayer and mechanical interlocking. The findings suggest that practitioners can precisely adjust TRS and TTS to achieve maximum HV in FSAM applications but should particularly lower TTS for maximum interlayer joint hardness for multi-material setups. The ten synthetic parameter combinations with the least contribution to Gaussian Process Regression (GPR) predictions of Vickers hardness (HV) when tool traverse speed (TTS) is the parameter of interest are detailed in Table 20 for the 7075-over-6061 FSAM configuration. All combinations include tool rotation speed (TRS), TTS, tilt angle (TTA), GPR-predicted HV, percent SHAP contribution for TTS, and the parameter size that had the greatest influence. Table 20Bottom 10 samples with lowest % contribution from TTS for HV predictions (7075 over 6061).TRSTTSTTAGPR-pred HV (7075/6061)Contribution % TTSMost Influential Parameter1200342.2113.4633153.140333TRS1200341.0112.0723484.429877TRS1200341.2112.2678334.930090TRS1200341.8112.9440354.978679TRS1200341.4112.4786385.187469TRS1200341.6112.7042435.461749TRS1200342.0113.1973177.862912TRS1010362.2112.2817268.115028TRS1010361.0112.1263868.220289TRS1020362.2110.3596038.488886TRS Bottom 10 samples with lowest % contribution from TTS for HV predictions (7075 over 6061). In these cases, the TTS contribution to the GPR prediction ranged from 3.14% to 8.49%, indicating low impacts on predicted hardness. All ten of the samples used a fixed TRS of high values (1090–1200 rpm) and TTS of moderate values (30–40 mm/min), and TTA values between 1.0° and 2.2°. Regardless, in every example, TRS is the dominating source of influence, which further emphasizes that for these processing parameters, rotation speed will largely dictate hardness measurements. Furthermore, the predicted HV values were between 110.36 HV and 113.46 HV, which indicates that hardness measurements do not change even though TTS values are modified. The identification of processing conditions, in which hardness is essentially a function of TRS, provides some indication that practitioners utilizing FSAM might simply ensure that TTS contributions are negligible and that any rotation speed settings could therefore be selectively modified. Optimization objective and search strategy The intent of this research project was to demonstrate the capability of using surrogate-assisted trend exploration as a means of optimizing process parameters conceptually rather than validating true optimization of the process. Separate objective functions were created for ultimate tensile strength (UTS) and Vickers hardness (HV), whereby each objective was treated independently, with those objectives being modeled by their respective surrogate regression models trained on each of their own training datasets. A single objective function was created for each objective, and no constrained multi-objective surrogate-based optimization was conducted. Parameter selection was accomplished by doing a grid-based analysis of the RBF-generated synthetic mesh, rather than employing a continuous or gradient optimization algorithm. The areas where candidate parameter values are predicted to be optimal were determined by examining the surrogate response surfaces to identify the maximum value for each response surface in the discretized design space. Predicted optimal values for several responses of interest were found to occur at or near the edges of the synthetic parameter grid, which indicates that those predicted optimal parameters are likely to be the result of surrogate interpolation, especially under the circumstances where there is a limited quantity of data available for use in training the surrogate regression models. Therefore, predictions of optimal parameters in these areas must be viewed with skepticism, and no such predictions are claimed as globally validated optimal parameters without first proving experimentally that they are truly globally validated. The output of this research has not produced any deployable or interactive optimization tools, as described in the introductory sections of this paper, and the results of this research are intended to serve as examples of potential workflows for selecting parameters as opposed to proof of true validated process optimizations. The identified optimal parameter combinations represent surrogate-derived solutions within the interpolated response surface and require independent experimental validation before practical implementation. For dissimilar multilayer FSAM builds such as 6061/7075, interfacial phenomena including mixing zone thickness, diffusion and precipitation behavior, and localized softening or strengthening bands play a decisive role in determining mechanical performance. The present research does not include direct microstructural or compositional characterization (e.g., EDS mapping, thickness-direction microhardness profiles, or microstructure imaging). Accordingly, the reported results are interpreted in terms of macroscopic performance correlations rather than explicit interfacial mechanisms. Future work integrating detailed interface characterization will be essential to physically interpret and validate the trends identified by the surrogate modeling framework. The constrained FSAM process window presented in this research is based on the tool rotation speed, travel speed, and tool tilt angle and does not account for other important parameters like axial force, tool geometry, deposition strategy, or thermal history. As such, the trends and surrogate predictions that were reported are limited to those variables within this research and do not necessarily apply to other machines/build strategies. That said, as more experimental data are collected, the RBF-assisted machine-learning framework that has been proposed allows for the possibility of exploring additional variables in the same manner as the ones used in this research. FSAM-built components are known to exhibit spatial heterogeneity and anisotropy arising from thermal cycling, and material flow. The present research does not include systematic thorough thickness or build-direction performance mapping; therefore, the reported mechanical properties should be interpreted as representative of the sampled regions rather than as a comprehensive assessment of global part quality. Extending the present framework to incorporate spatially resolved mechanical characterization remains an important direction for future work. Although mechanical trends may be influenced by underlying microstructural and interfacial phenomena inherent to FSAM processes, the present study does not include detailed quantitative microstructural characterization (e.g., EBSD grain size measurement or compositional mapping across interfaces). Therefore, microstructural interpretations remain qualitative and are discussed in the context of existing literature. Although the synthetic mesh extends slightly beyond the experimental design space, extrapolative predictions in these regions are associated with increased uncertainty due to the limited number of experimental anchor points. The present research therefore focuses on interpolation within the Taguchi-sampled parameter domain. Robust extrapolation would require additional experimental data, particularly near boundary and out-of-domain regions, which is identified as an important direction for future work. Conclusion This research developed a surrogate-assisted machine learning framework to correlate mechanical property prediction to friction stir additive manufacturing of dissimilar aluminum alloys in data-sparse conditions. Nine experimental data points, along with RBF methodology-generated synthetic data (of 10x), were used to benchmark and compare the regression models (Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Bayesian Ridge) to predict the ultimate tensile strength and Vickers Hardness of two different material-stacking configurations. GPR performed best, with leave-one-out cross-validation yielding a mean absolute error (MAE) of 3.58–4.95 MPa for UTS and 2.14–3.21 HV for hardness for the experimental points. The errors are considerably higher than for the associated synthetic test sets (MAE < 0.1%). Thus, the near-perfect accuracy of the synthetic models suggests that the predicted mechanical properties based on the synthetic data reflect the RBF learning through interpolation rather than a true generalization to independently validated experimental data. The SHAP method helped to understand how different parameters play a role in determining the UTS of the associated materials. Tool rotation speed (TRS) was identified as the primary contributor to the prediction of UTS. However, in the case of hardness, the interaction of the various parameters was much more complex. In the case of 7075 − 6061 stacking, there was an equal contribution of TRS and TTS to the hardness prediction. The results of this analysis show that different materials will respond differently to different processes, and thus, refinements to process parameters will need to be made for each material combination. Specifically, a lower TRS would produce optimal UTS for 6061–7075; conversely, a higher TRS would provide optimal UTS when producing 7075 − 6061. Moderate TRS with slow TTS consistently produced optimal hardness irrespective of stacking combination. The practical use of the framework depends on independent experiment-based confirmation of its effectiveness. While the surrogate model can be used for data-driven exploratory testing of parameters when there is limited data, any predictions made based on data from untested conditions should be viewed as approximations and not as true values. In fact, based on the LOOCV results, the amount of error between the actual data compared to the predictive results on the actual data range from one to three%, and for the synthetic data it was less than 1%. Thus, while the framework can be employed within the tested range of parameters (TRS = 1000–1200 rpm; TTS = 30–40 mm/min; TTA = 1.0–2.0°), it only applies to the tool’s design and alloy type that were employed in the testing. In order to be able to apply the method to other tool designs and alloy types, or if other variables may be used, the method must be extended either through experimental validation of the new parameters or through the process of changing the method itself. The findings should be interpreted as surrogate-assisted trends within the experimentally explored design space. Independent experimental validation at predicted optima is required before practical implementation. Future work is essential to establish industrial applicability. The conclusions are restricted to macroscopic mechanical performance trends, with detailed interfacial mechanisms reserved for future investigation and locally sampled mechanical performance, and do not claim uniform part-scale properties or interpolation within the experimentally sampled parameter space and do not claim validated extrapolation beyond it. While the present study does not claim predictive generalization beyond the experimental domain, it establishes a transparent and reproducible framework for surrogate-assisted exploration under severe data scarcity, which is common in solid-state additive manufacturing research. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions Krishnan V- Paper writing, ExecutionSenthil Kumaran Selvaraj-Conceptualization and SupervisionGiriraj Mannayee- Resource supportKalyanaraman P- Software ComparisonSundaramali G-Data Collection. Funding Open access funding provided by Vellore Institute of Technology. Data availability The datasets and computational workflow supporting this study are publicly available at: [[https://github.com/Krishnan0022/AA6061-AA7075-Aluminum-Alloys](https:/github.com/Krishnan0022/AA6061-AA7075-Aluminum-Alloys)].The repository includes:- The original 9-point experimental dataset (Excel format)- The Google Colab notebook implementing: RBF interpolationSynthetic data generationMachine learning training (GPR, SVR, Bayesian Ridge)SHAP analysisParameter optimization workflow- Environment and dependency information required to reproduce the results.All data and code are publicly accessible without restriction. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets and computational workflow supporting this study are publicly available at: [[https://github.com/Krishnan0022/AA6061-AA7075-Aluminum-Alloys](https:/github.com/Krishnan0022/AA6061-AA7075-Aluminum-Alloys)].The repository includes:- The original 9-point experimental dataset (Excel format)- The Google Colab notebook implementing: RBF interpolationSynthetic data generationMachine learning training (GPR, SVR, Bayesian Ridge)SHAP analysisParameter optimization workflow- Environment and dependency information required to reproduce the results.All data and code are publicly accessible without restriction.