Constituent-Material-Anchored Continual Learning for Full Stress–Strain Prediction of Multi-Material PETG/PC-ABS MEX Laminates Authors: Ramachandran Avala Subramanian, Mahalingam Nainaragaram Ramasamy, Michal Prauzek, Quoc-Phu Ma, Jaromir Konecny, Ales Sliva Published: 2026 ABSTRACT Predicting the tensile response of multi-material parts produced by material extrusion (MEX) remains difficult because the final behavior depends on both the constituent polymers and the quality and arrangement of dissimilar interfaces. This study introduces a constituent-material-anchored, phase-aware continual-learning framework for full stress–strain curve prediction of PETG/PC-ABS laminate coupons. Experimentally measured PETG and PC-ABS reference curves were combined through a rule-of-mixtures baseline; an XGBoost residual model then learned pointwise corrections using strain, baseline stress, mechanical phase label, and PETG thickness fraction as inputs. Validation used five PETG reference coupons, five PC-ABS reference coupons, five C1 laminate coupons, two C2 out-of-distribution coupons, and three coupons for each model-suggested Rank 1–3 architecture. UTS agreement alone was not sufficient: Rank 2 had a zero-shot UTS error of only −0.18% but a full-curve RMSE of 20.74%. After the first architecture-specific coupon was introduced, RMSE decreased from 12.34% to 2.72% for C1, from 18.60% to 6.38% for C2, from 21.04% to 6.93% for Rank 1, from 20.74% to 7.50% for Rank 2, and from 19.40% to 7.48% for Rank 3. The framework therefore provides a data-efficient, interpretable proof of concept for laminate screening and tensile-curve prediction, while its broader statistical robustness and extension to other loading modes require larger datasets. 1. Introduction Material extrusion (MEX), including fused filament fabrication (FFF) and fused deposition modeling (FDM), is widely used for polymer additive manufacturing because it enables low-cost fabrication of complex thermoplastic parts with relatively simple equipment. Its mechanical performance is nevertheless difficult to predict because printed coupons are affected by process-sensitive variables such as layer height, raster orientation, nozzle temperature, print speed, thermal history, porosity, inter-road fusion, and interlayer bonding. These factors can produce anisotropy, weak interlayer adhesion, void formation, and stress–strain responses that differ from conventionally manufactured materials [1,2,3,4]. The prediction problem becomes more demanding in multi-material MEX. In addition to the process effects present in single-material prints, a laminate made from dissimilar polymers is governed by thermal compatibility, wetting, interdiffusion, load transfer across material transitions, and the number and location of interfaces. Consequently, the tensile curve of a PETG/PC-ABS laminate cannot be assumed to be a simple average of the two constituent responses. This material pair was selected because PETG and PC-ABS combine different stiffness, ductility, thermal resistance, and impact-related characteristics, and because earlier characterization of PETG/PC-ABS MEX laminates showed measurable tensile penalties relative to ideal constituent-material load sharing [5,6,7]. Machine learning has increasingly been used to reduce experimental trial-and-error in MEX and related polymer additive-manufacturing problems. Review studies identify data-driven models as useful tools for nonlinear process–property relationships [1,2,3]. Earlier tensile-property studies used group-method-of-data-handling models, artificial neural networks, adaptive neuro-fuzzy inference systems, fuzzy logic, and hybrid optimization strategies to predict tensile strength or optimize printing parameters [8,9,10,11,12,13,14,15,16]. More recent work has adopted deep learning, Gaussian-process regression, random forests, boosting, support-vector methods, and ensemble learning for tensile or related mechanical-property prediction [17,18,19,20,21,22,23,24,25,26,27,28]. Recent studies have also expanded mechanical-capacity prediction beyond standard polymer-MEX tensile benchmarks, including PETG-specific models, multi-material ABS/PETG optimization, reinforced or fiber-filled polymers, and broader data-driven capacity prediction in lattice structures, castings, UHPC-strengthened columns, UHPC compressive strength, and additively manufactured Ti-6Al-4V fatigue strength [29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47]. These studies confirm the value of data-driven capacity prediction, but most remain focused on scalar outputs such as strength, modulus, fatigue strength, or load capacity rather than the full tensile trajectory. Three limitations motivate the present work. First, scalar endpoints such as UTS or elongation do not fully describe stiffness evolution, nonlinear transition, peak-stress development, or pre-failure curve shape. Second, many studies validate by internal train–test splitting within a known design space; this is not equivalent to testing genuinely unseen laminate architectures. Third, most models do not explicitly represent different mechanical regimes along a tensile curve, even though elastic, hardening, UTS-neighborhood, and post-UTS regions can have different error characteristics. These limitations are especially important for interface-sensitive multi-material laminates, where two stacks with similar strength may still have different stress–strain trajectories and different design implications. The principal methodological novelty of this study is the combination of four elements in one experimentally validated workflow: (i) constituent-material reference curves are used as physically interpretable anchors; (ii) a rule-of-mixtures (RoM) baseline is corrected by a learned pointwise residual rather than replaced by a purely black-box model; (iii) the residual model is phase-aware and is trained on an adaptive strain grid refined near yield, high-curvature, and UTS regions; and (iv) the model is updated sequentially as laminate coupons become available and is evaluated on unseen architectures. In the present implementation, the residual regressor uses strain, RoM stress, phase label, and PETG thickness fraction as inputs. The method is therefore designed for data-efficient tensile-curve learning, not for claiming universal generalization across all polymer systems. Table 1 summarizes the position of the present work relative to representative machine-learning-based mechanical prediction studies. Accordingly, this study addresses the following research questions:1.To what extent can constituent-material tensile anchors predict the full tensile stress–strain curve of PETG/PC-ABS laminates fabricated by MEX?2.How much does phase-aware residual correction improve full-curve prediction relative to the initial RoM baseline, particularly near mechanically important regions such as yield, high-curvature transitions, and UTS?3.Can the workflow support out-of-distribution laminate screening by predicting previously unseen PETG/PC-ABS architectures that are subsequently fabricated and tested experimentally? 1.To what extent can constituent-material tensile anchors predict the full tensile stress–strain curve of PETG/PC-ABS laminates fabricated by MEX? To what extent can constituent-material tensile anchors predict the full tensile stress–strain curve of PETG/PC-ABS laminates fabricated by MEX? 2.How much does phase-aware residual correction improve full-curve prediction relative to the initial RoM baseline, particularly near mechanically important regions such as yield, high-curvature transitions, and UTS? How much does phase-aware residual correction improve full-curve prediction relative to the initial RoM baseline, particularly near mechanically important regions such as yield, high-curvature transitions, and UTS? 3.Can the workflow support out-of-distribution laminate screening by predicting previously unseen PETG/PC-ABS architectures that are subsequently fabricated and tested experimentally? Can the workflow support out-of-distribution laminate screening by predicting previously unseen PETG/PC-ABS architectures that are subsequently fabricated and tested experimentally? 2. Materials and Methods This study followed a constituent-material-anchored, physics-informed, phase-aware residual-learning workflow for predicting the full tensile stress–strain response of PETG/PC-ABS laminates fabricated by material extrusion. The methodology was organized in sequential stages: experimental acquisition of pure PETG and pure PC-ABS reference curves, construction of constituent-material anchor curves, generation of an initial RoM baseline, preprocessing and quality checking of laminate coupon data, mechanical phase identification, adaptive strain-grid refinement, residual correction using a trained regression model, continual model updating, and out-of-distribution (OOD) laminate screening. Pure PETG and pure PC-ABS coupons were used only as single-material reference inputs for constructing the baseline response, while laminate coupons were used to learn the deviation between the RoM prediction and the measured full tensile curve. The workflow was designed as a data-efficient experimental strategy for the present PETG/PC-ABS system, not as a claim of universal robustness across all polymer pairs or loading modes. Figure 1 summarizes the complete methodological workflow used in this study. The process begins with tensile testing of pure PETG and pure PC-ABS constituent materials, which provide the experimental reference curves for the constituent-anchor construction. These constituent anchors are then combined using a material-fraction-based rule-of-mixtures formulation to generate an initial physics-informed OOD prediction for PETG/PC-ABS laminate architectures. In parallel, experimentally tested laminate coupons are acquired and passed through quality assurance and preprocessing, including strain standardization, interpolation, curve cleaning, and consistency checks. The processed curves are then segmented into mechanical phases, including elastic, hardening, and post-UTS/necking regions, and the strain grid is adaptively refined near mechanically important regions. The residual error between the measured laminate response and the rule-of-mixtures baseline is calculated pointwise and used to train the phase-aware residual-learning model. The trained model corrects the initial baseline prediction and produces full stress–strain curve predictions for unseen laminate architectures. These predictions are then evaluated, used for design screening, and ranked according to the implemented selection score. The top-ranked laminate designs are finally selected for experimental validation, and the newly obtained coupon data can be added back into the workflow for sequential model updating. 2.1. Definition of OOD Laminate Validation In this study, an OOD laminate is formally defined at the architecture level. A prediction is considered OOD when the ordered PETG/PC-ABS layer stack being predicted has not contributed any coupon to the residual-learning memory before that prediction is made. This definition differs from a random train–test split of points from the same curve or coupons from the same architecture. Laminate C2 and the Rank 1–3 architectures therefore represent OOD cases at the time of their first prediction because their stack sequences were unseen by the correction model. After a coupon from a given architecture is accepted into memory, later predictions for that same architecture are described as architecture-specific updates rather than purely blind OOD predictions. 2.2. Materials, Coupon Fabrication, and Tensile Testing This section defines the experimental material system and coupon sets used to support the constituent-anchor and incremental-learning workflow. Pure PETG and pure PC-ABS coupons were used as single-material reference coupons for constructing the reference tensile curves, whereas PETG/PC-ABS laminate coupons were treated as laminate architectures for zero-shot comparison, out-of-distribution validation, model-guided design assessment, and subsequent incremental updating. The material grades, coupon geometry, fabrication settings, and tensile testing protocol were based on the previously reported PETG/PC-ABS material extrusion study, while the present work reorganizes these experimental inputs for full-curve prediction and residual-learning methodology [5,6,7]. 2.2.1. Materials Two commercial thermoplastic filaments supplied by Fiberlogy, Brzezie, Poland, were used as the constituent materials: Fiberlogy EASY PET-G and Fiberlogy PC/ABS. Both materials were used with a nominal filament diameter of 1.75 mm. The Fiberlogy EASY PET-G product line is described by the supplier as a glycol-modified PET material intended for mechanically and chemically resistant printed parts, whereas the Fiberlogy PC/ABS product line combines polycarbonate with ABS for improved impact resistance, chemical resistance, and thermal stability [5,6]. Pure PETG and pure PC-ABS coupons were fabricated and tested separately before the laminate-learning workflow. Five repeated tensile coupons were used for each constituent material,(1)nPETG=5,nPC-ABS=5, where nPETG and nPC-ABS denote the number of valid single-material reference coupons used for PETG and PC-ABS, respectively. These single-material reference coupons were not treated as laminate prediction targets. Instead, their full stress–strain curves were used later to construct the pointwise constituent-anchor mean curves and standard-deviation envelopes described in Section 2.4. The complete coupon dataset was divided according to its role in the workflow, as summarized in Table 2. Monolithic PETG and PC-ABS coupons were used to define the constituent-material anchors. Laminate C1 and Laminate C2 were used to evaluate zero-shot prediction, laminate learning, and out-of-distribution (OOD) validation. The Rank 1–3 coupons were used in the final model-guided architecture validation stage. The first PETG/PC-ABS laminate introduced into the learning workflow, denoted Laminate C1, was the 0.2 mm alternating-layer architecture starting with PETG on the build plate. This configuration was selected because 0.2 mm was the layer height used for the laminate slicing condition, and alternating the two materials at this layer scale produced the most interface-rich configuration within the 4 mm coupon thickness. As a result, the architecture contained the maximum number of PETG/PC-ABS material transitions available for the selected laminate build strategy. Introducing this case first allowed the algorithm to encounter an intentionally demanding interface-sensitive laminate response before being applied to lower-interface or model-suggested architectures. Therefore, Laminate C1 was used as the initial zero-shot comparison case and then retained in the sequential learning memory to provide the first laminate-specific residual correction beyond the constituent-material rule-of-mixtures baseline [7]. The experimental tensile records used for the reference and initial laminate data acquisition are illustrated in Figure 2. 2.2.2. Coupon Geometry All coupons used in this research were prepared according to ISO 527-2, the dog-bone-shaped Type 1B tensile coupons [48]. The nominal coupon thickness was maintained at 4 mm for all coupon groups. The principal coupon dimensions, including the 150 mm overall length, 50 mm gauge length, 10 mm reduced-section width, 20 mm end width, R60 transition radius, and 4 mm thickness, are shown in Figure 3. The coupon geometry shown in Figure 3 was adopted from the previously reported PETG/PC-ABS MEX laminate study and was retained in the present work as the common geometry for both single-material reference testing and subsequent laminate validation [7]. The use of a consistent coupon geometry is important here because the constituent-material anchor curves constructed later in Section 2.4 and the laminate curves used for residual learning are compared on the same tensile-coupon basis. 2.2.3. MEX Fabrication Conditions The material properties and fabrication settings used for the pure-material and PETG/PC-ABS laminate coupon groups are summarized in Table 3. The table is included here to define the material inputs and fabrication conditions used to generate the experimental coupon data; mechanical performance interpretation is not discussed in this section. 2.2.4. Tensile Testing Protocol The tensile-test data used in this study were taken from the repository associated with the previously published PETG/PC-ABS MEX laminate characterization study [49]. In that study, tensile testing was conducted according to ISO 527-2 using Type 1B dog-bone coupons printed from PETG, PC-ABS, and PETG/PC-ABS laminate configurations. The tests were performed on a Shimadzu EZ-LX universal testing machine equipped with a 5 kN load cell and operated under displacement control. All tensile tests used in the present work were conducted at a crosshead speed of 50 mm/min. The dataset used for model construction and validation included five monolithic PETG coupons and five monolithic PC-ABS coupons for anchor-curve construction. The 0.2 mm alternating PETG/PC-ABS laminate contained five coupons and was used for the initial zero-shot comparison and interface-rich learning case. The 1.33 mm PC-ABS/PETG/PC-ABS laminate contained two coupons and was used as the primary out-of-distribution validation case. The three model-suggested ranked laminate architectures were subsequently printed and tested using three coupons per architecture. The raw tensile records consisted of stress–strain data exported from the mechanical testing workflow. These full stress–strain curves, rather than only scalar mechanical properties, were used for anchor-curve construction, residual learning, and prediction validation. Strain was inferred from crosshead displacement in the original testing workflow, with no extensometer reported. Therefore, the present study treats the exported stress–strain records as machine-derived tensile response curves and applies the quality assurance and preprocessing steps described in Section 2.3 prior to model training or validation. 2.3. Data Acquisition, Quality Assurance, and Preprocessing The raw tensile records were not used directly for model training or validation. Instead, each coupon file was passed through a standardized preprocessing workflow to obtain clean and comparable stress–strain curves across monolithic PETG, monolithic PC-ABS, and PETG/PC-ABS laminate coupons. The workflow identified the available force, stress, and strain channels, removed non-numeric metadata rows, corrected strain units where required, removed the preload region, and organized each curve into a consistent strain-increasing format. The resulting cleaned tensile curves were then used for anchor-curve construction, residual learning, phase detection, and full-curve prediction assessment. 2.3.1. Raw Coupon Data Acquisition For each tensile coupon, the exported mechanical test record was treated as the raw input to the computational workflow. The records contained tabulated tensile-response data, including stress and strain, with force also available in the relevant files. Only individual coupon-level records were used for curve-level analysis; summary files, statistical reports, and previously processed output files were excluded to avoid mixing raw experimental measurements with aggregated results. The full stress–strain curves were retained throughout preprocessing because the proposed framework learns and validates the complete tensile response rather than relying only on scalar outputs such as ultimate tensile strength. 2.3.2. Column Detection and File Loading Each coupon file was loaded using an automated column-identification step to account for possible differences in exported column names or ordering. Stress and strain channels were identified from the column headings using keyword matching, while derived quantities such as modulus were excluded from stress selection. When a force channel was present, it was also retained because it provided additional information for preload removal and quality checking. After the relevant channels were identified, the selected data were converted to numeric form, rows with missing or invalid values were removed, and files without usable stress–strain information were excluded from further analysis. 2.3.3. Strain-Unit Conversion and Interpolation After the stress and strain channels were extracted, the strain values were checked for unit consistency. Records reported in percentage strain were converted to decimal strain so that all coupons used the same strain representation. The stress–strain data were then sorted in increasing strain order, and duplicate strain entries were removed to produce a monotonic curve suitable for interpolation and comparison. The cleaned monolithic PETG and PC-ABS curves were interpolated onto a common strain grid for pointwise calculation of mean anchor curves and variability bands, while laminate coupon curves were later projected onto the model grid used for residual computation and validation. 2.3.4. Coupon Quality Checks and Rejection Criteria Quality assurance was applied before any coupon curve was used for anchor construction, learning, or validation. Files were rejected if they did not contain valid stress–strain data or if too few usable points remained after numerical cleaning. When force data were available, the initial preload region was removed using the early low-force portion of the curve, and the force–stress relationship was checked for consistency as an indicator of possible measurement instability. Coupons showing unrecoverable data-format problems were excluded, while coupons with usable stress–strain curves but minor warnings were retained after preprocessing. This ensured that the learning framework was trained and evaluated using physically meaningful tensile curves rather than corrupt or incomplete records. 2.4. Constituent-Material Anchors and Rule-of-Mixtures Baseline The cleaned PETG and PC-ABS single-material reference curves were interpolated to a common strain grid. Pointwise mean curves were used as the constituent-material anchor curves, and pointwise standard-deviation curves were retained to describe reference-coupon variability. For a laminate stack with layer thicknesses tl, the PETG fraction was calculated from the total PETG thickness divided by the total stack thickness:(2)r=∑lIlPETGtl∑ltl,IlPETG=1forPETG,0forPC-ABS. When mass ratio is required, it can be calculated from the same stack description by density weighting:(3)rm=ρPETG∑lIlPETGtlρPETG∑lIlPETGtl+ρPCABS∑l(1−IlPETG)tl, where ρPETG=1.27 g/cm3 and ρPCABS=1.07 g/cm3 are the density values used in Table 3. The implemented residual-learning model used the thickness fraction r as its composition input because the candidate stacks were generated by layer thickness. The density-weighted mass ratio rm was used only for mass-related derived quantities and is a direct extension for future versions of the model. The initial laminate prediction was the material-fraction-weighted RoM curve:(4)σRoM(ε;r)=rσpPETG(ε)+(1−r)σpPCABS(ε). The reference variability curves were propagated using the same fraction weighting to provide a RoM uncertainty band for visualization. Before any laminate data had been learned, mixed PETG/PC-ABS stacks were additionally multiplied by a fixed adhesion-efficiency fallback of 0.83 for 0.1