Data-driven design of hybrid graphene oxide-MXene polymeric membranes for water purification Authors: Rauf Khan, Nadeem Baig, Zaher Mundher Yaseen, Mohammad Al-Suwaiyan Published: 2026 ABSTRACT SummaryTwo-dimensional (2D) materials offer promising platforms for water purification, particularly for challenging oil-in-water emulsions. Here, a hybrid membrane combining graphene oxide (GO) and MXene nanosheets was fabricated on a nylon support to explore the relationship between membrane composition and separation performance. Four membranes with varying GO loadings were evaluated, and machine learning models were applied to predict rejection efficiency using key operational parameters. Among the tested models, support vector regression achieved the highest predictive accuracy (R2 = 0.952 for training, 0.936 for testing, and 0.922 for validation), effectively capturing nonlinear relationships between membrane structure and operating conditions. Feature-importance analyses identified membrane composition and pressure as dominant factors governing performance. This study demonstrates how integrating 2D materials with data-driven modeling can enable predictive design of membrane systems, supporting more efficient development of water purification technologies. Future work will expand datasets and hybrid learning approaches to further enhance predictive robustness and generalization.Subject areas: Applied sciences, Materials science, Polymers Introduction Population growth, urbanization, industrialization, agricultural intensification, and climate change contribute to the global freshwater crisis.1,2 Addressing the global water crisis has become an urgent priority, elevating freshwater to a strategic resource. Among the major contributors to water pollution, oily and organic contaminants pose significant challenges to water quality and treatment processes.3,4 Many cases, such as the 2018 Sanchi oil tanker spill, which discharged up to 1,900 tons of oil into Chinese waters5 highlight the catastrophic impacts of oil contamination. Transportation-connected oil spills alone account for around 13% of global marine oil pollution.6 Furthermore, the World Health Organization (WHO) (2018–2015) reports that 2.1 billion people lack access to clean water and 2.3 billion lack adequate sanitation.7,8 Also, more than 5 million people have died annually due to waterborne diseases.9 These catastrophic figures underscore the need to develop sustainable technologies to mitigate water-related challenges and ensure the availability of clean water resources for future generations. Additionally, oil-water emulsions are particularly hazardous due to their persistence and difficulty of separation, which affects the water quality and sustainability.10 Old treatment methods, such as gravity separation, flotation, and biological processes, often fall short in effectively addressing these complex mixtures, thereby underscoring the urgent need for more advanced and sustainable purification strategies.11 In separation science, membrane-based materials are continually evolving, and recently, there has been a surge in macrocycling12,13 and two-dimensional (2D) membranes. The 2D materials are attracting attention for water purification.14 Pristine polymeric membranes face several challenges and are prone to rapid fouling, limiting their effectiveness in separating complex oil-in-water emulsions.15,16 These limitations highlight the pressing need to design membranes that combine the durability of PM with the advanced separation capabilities offered by novel lamellar 2D materials.17 Graphene oxide (GO) and MXene-based membranes have been extensively investigated individually for oil-water separation. GO membranes typically form densely stacked laminar structures due to strong interlayer interactions, resulting in high selectivity but inherently limited permeability during oil-in-water emulsion separation.18 Various intercalation and structural modification strategies have been proposed to overcome this limitation14; however, these approaches often increase fabrication complexity and reduce structural stability. In contrast, MXene nanosheets (e.g., Ti3C2Tx) exhibit higher in-plane stiffness19 and bending rigidity,20 which helps maintain more open transport channels and enhances water permeability. However, MXene layers alone may suffer from structural instability, weak adhesion to supports,21 and reduced selectivity due to enlarged nanochannels. These contrasting characteristics suggest that a 2D-2D (MXene-GO) hybrid architecture could overcome the intrinsic trade-off between permeability and selectivity. Specifically, the relatively flexible GO nanosheets can form compact, selective nanochannels, while the stiffer MXene nanosheets resist structural collapse and help preserve interlayer spacing, thereby enhancing water transport. The combination is therefore expected to generate a synergistic laminar structure with interconnected nanochannels, improved wettability, and reduced propensity for fouling. Despite this potential, the interfacial interactions and mechanisms of nanochannel formation in GO-MXene (2D-2D) hybrid membranes remain poorly understood, particularly on polymeric supports. Moreover, predictive frameworks linking membrane composition, operating conditions, and separation performance are lacking. Although ML has been applied to various membrane materials, its use in emerging 2D-2D hybrid membranes remains largely unexplored. ML models offer powerful tools to unlock the full potential of 2D membranes by enabling data-driven material selection, performance prediction, and structural optimization. However, the integration of ML models with 2D materials and the identification of optimal 2D membrane architectures remain under investigation to advance the development of hybrid membrane systems. Addressing these gaps requires integrating experimental design with data-driven modeling to identify key governing factors and enable predictive optimization. This study addresses the pressing challenge of oil-in-water emulsions by developing a 2D-2D hybrid membrane formed by depositing a synergistic layer of GO and MXene nanosheets onto a nylon membrane (NM) support. GO contributes abundant oxygen-containing functional groups, high hydrophilicity, and molecular-level tunability, while MXene nanosheets enlarge nanochannel pathways, enhance water flux, and improve antifouling performance. Beyond experimental fabrication and characterization, this work integrates ML-powered predictive modeling to estimate membrane rejection efficiency, establishing a lab-to-ML framework that reduces experimental cost and accelerates membrane optimization. To the best of our knowledge, this is among the first studies to combine experimentally validated GO-MXene hybrid membranes with ML models for oil-water separation. Specifically, this study aims to fabricate and characterize a stable and selective GO-MX/NM membrane, predict its separation performance using advanced ML models, and compare multiple predictive algorithms to identify the most reliable approach for forecasting membrane rejection. Results and discussion Characterization of materials and membranes MXene was synthesized by etching the parent MAX phase with HF; the disappearance of the aluminum peak at ∼39°, corresponding to the (104) plane, confirmed the successful conversion of MAX to MXene.22 The characteristic MXene (002) diffraction peak appears at a lower angle, around 9°, as shown in Figure 1A. Compared to the parent MAX phase, this (002) peak shifts to a lower angle, indicating successful MXene formation.21 The shift to a lower diffraction angle indicates an increase in interlayer spacing, consistent with the formation of few-layer MXene structures that enhance ion and molecular transport. Furthermore, it is believed that no oxidation occurs during MXene synthesis, as evidenced by the absence of additional reflections attributable to the TiO2 phase. Additionally, the interpeak spacing increases, further confirming MXene’s sheet-like morphology. All these indications confirm the synthesis of MXene from the MAX phase. The FTIR spectra of GO, MXene, and the GO-MX hybrid membrane are shown in Figure 1B. Furthermore, a strong absorption band at 3,400 cm−1 was assigned to -OH stretching vibrations, along with peaks near 1,720 and 1,620 cm−1, assigned to C=O stretching and C-OH bending, indicating the presence of various oxygenated functional groups.23,24 On the other hand, the Ti-O vibration was observed at 800 cm−1, and signals at 1,100 cm−1 indicated the formation of a C-F bond during etching.25 On the other hand, in the GO-MX hybrid membrane, the Spectra combined features of both materials with intensified -OH and C=O signals, suggesting strong interfacial interactions and improved surface functionalization. Therefore, the coexistence of Ti-O, and C=O bands confirms the successful integration of GO and MXene, providing abundant light-harvesting and hydrophilic functionalities that are favorable for water purification applications. XRD and FTIR characterization of MAX phase, MXene, and GO-MXene membranes (A) XRD of (a) Ti3AlC2 MAX phase and (b) Ti3C2Tx MXene nanosheet. (B) FTIR of (a) PM, (b) MXene, (c) 0.1 mg GO mix with 2 mg MXene, and (d) 0.2 mg GO mix with 2 mg MXene. The surface morphology of the synthesized GO was investigated using a scanning electron microscope (SEM), as shown in Figure 2A. The micrograph reveals a characteristic wrinkled, layered texture, a typical feature of GO nanosheets.26 Therefore, the sheet-like structures appear thin and crumpled, indicating successful exfoliation of graphite oxide into nanoscale lamellae. Energy dispersive X-ray spectroscopy (EDX) also provides useful information, as shown in Figure 2B. Moreover, the carbon peak indicates a graphite backbone, and the oxygen peak confirms the presence of functional groups, such as hydroxyl, epoxy, and carboxyl. Such functional groups increase the likelihood of GO-MXene compatibility during membrane fabrication. The surface is clearly smooth, with large, compact grains and distinct lamellar features.27 The MAX phase exhibits a crystalline layered structure, in which aluminum layers are interspersed between sheets of titanium carbide (Figure 2C). Furthermore, the EDX peaks corresponding to Ti, Al, and C were detected in agreement with the expected stoichiometry of Ti3AlC2 (Figure 2D). In accordance with the experiment, SEM images of the etched sample revealed the usual accordion-like structure of MXene, in contrast to the compact structure of the MAX phase.28 The layered shape, single-layer sheets, and expanded interlayer spacing further confirmed the effective removal of Al layers, producing nanosheets with large surface area and high accessibility, which is ideal for membrane integration (Figure 2E). The EDX analysis further supports this transformation, showing dominant Ti and C peaks with a greatly diminished Al signal, verifying successful etching.29 However, the XRD has not shown any Al peaks in the MXene spectra.30 The appearance of Al in the EDX might be due to AlF residue left behind after washing the MXene sheets. On the other hand, EDX mapping indicates that the O and F peaks correspond to surface termination during HF treatment31 (Figure 2F). Morphological and elemental characterization of GO, MAX phase, and MXene SEM images of (A) GO, (C) MAX phase, (E) MXene, and EDX of (B) GO, (D) MAX phase, and (F) MXene. Scale bars: 10 μm. The morphological analysis indicates successful lamination of GO-MX onto the pristine membrane. For this purpose, characterization is performed on all synthesized membranes. Thus, SEM of the NM reveals a porous surface with irregular, larger voids, confirming that the surface is untreated. On the other hand, EDX confirmed the organic composition dominated by carbon, oxygen, and nitrogen. On this basic morphology of the NM, subsequent lamination of the 2D material sheets has developed the separating layer (Figures 3A and 3B). The membrane laminated with 2 mg of MXene on the NM support exhibits a well-defined layered morphology, confirming the sheet-like structure of MXene nanosheets. The presence of Ti, C, and O, characteristic elements of MXene, was further verified by EDX elemental mapping (Figures 3C and 3D). SEM images, EDX, elemental mapping, and cross-section of the membranes SEM images of membranes include (A) PM, (C) MXene, (E) 0.1 mg GO with MXene, and (G) 0.2 mg GO with MXene. EDX images show (B) PM, (D) MXene, (F) 0.1 mg GO with MXene, and (H) 0.2 mg GO with MXene, along with (I) cross-sectional SEM image of the hybrid GO-MXene. Additionally, (J and K) EDX spectrum and elemental mapping illustrates the spatial distribution of different elements, and Elemental mapping of (L) Titanium, (M) Carbon, (N) Oxygen, (O) Aluminium, (P) Flourine on the surface of the hybrid GO-MXene membrane. Scale bars: SEM:10 μm, and cross-section SEM: 50 μm. Upon incorporation of 0.1 mg of GO into 2 mg of MXene, the hybrid membrane surface exhibits layered features and strong interfacial interactions between GO and MXene nanosheets. Correspondingly, EDX analysis reveals an increased oxygen content, attributable to oxygen-containing functional groups introduced by GO (Figures 3E and 3F). Increasing the GO loading to 0.2 mg yields a densely stacked lamellar structure with uniform surface coverage and well-aligned, sheet-like flakes, indicating effective lamination and improved structural integrity. EDX mapping further confirms enriched oxygen distribution, reinforcing the synergistic role of GO within the hybrid architecture. Notably, among all fabricated membranes, this GO-MXene hybrid exhibited the superior separation performance, identifying it as the optimal membrane configuration (Figures 3G and 3H). The combination of material validation methods, such as SEM and EDX, with quantitative performance data makes this dataset distinctive by directly linking structural features to functional outcomes. SEM and EDX not only confirm the dataset’s quality but also strengthen its suitability for developing robust ML models to predict membrane rejection performance. The spatial distribution and integration of GO-MXene within the nylon substrate were investigated using cross-sectional SEM analysis. As shown in Figures 3I and 3J, a well-defined layered structure is observed, in which the porous morphology of the underlying nylon substrate is preserved, while a continuous, dense nanosheet layer forms on the membrane surface. This layer indicates the formation of a compact and integrated 2D-2D laminated structure, arising from strong interfacial interactions between GO and MXene nanosheets. The corresponding EDX elemental mapping (Figures 3K–3P) further confirms the compositional uniformity of the hybrid layer. In particular, the homogeneous distribution of Ti across the top surface demonstrates that MXene nanosheets are uniformly embedded within the GO matrix rather than forming isolated aggregates. Moreover, the Ti signal is predominantly present at the membrane surface with minimal penetration into the substrate pores, indicating that MXene, together with GO, forms a clear surface barrier layer. In contrast, carbon is detected throughout the membrane due to the nylon support, with additional contribution from GO at the top layer. Importantly, the clear compositional distinction between the surface layer and the porous substrate confirms that the hybrid nanosheets remain primarily confined to the membrane surface, preserving the structural integrity of the support. The slight roughness observed in the cross-sectional images is consistent with the atomic force microscopy (AFM) surface morphology, further supporting the presence of stacked and interlocked nanosheets. Additionally, the EDX spectra confirm the presence of characteristic elements (Ti, C, O, Al, and F), indicating the successful incorporation of MXene into the GO framework. Collectively, these results provide strong evidence for the formation of a well-integrated GO–MXene laminated architecture with interfacial contact and dominant surface coverage, rather than loosely aggregated structures. The morphology and structure of the GO-MXene nanosheets were further examined using TEM and AFM to validate their 2D characteristics. As shown in Figures 4A and 4B, the TEM images reveal thin, sheet-like structures with lateral dimensions in the sub-micron range, confirming successful exfoliation of both GO and MXene nanosheets. The GO sheets appear larger and more continuous, whereas the MXene nanosheets are smaller and more uniformly distributed across the GO surface, indicating the formation of a hybrid 2D-2D architecture. AFM analysis (Figures 4C and 4D) further confirms the nanosheet morphology and surface topology of the hybrid membrane. The surface exhibits peaks and valleys, indicating heterogeneous stacking of GO and MXene nanosheets. Quantitative roughness analysis shows an average roughness (Sa) of approximately 202 nm and root-mean-square roughness (Sq) of ∼245 nm over a 10 × 10 μm scan area. This increased roughness is attributed to differences in the lateral size and stacking behavior of GO and MXene sheets, with smaller MXene nanosheets intercalated within larger GO layers. The combined TEM and AFM results confirm the formation of exfoliated 2D nanosheets and their heterogeneous assembly, which is expected to influence interlayer spacing and transport pathways within the hybrid membrane. Morphological characterization of GO-MXene hybrids and membranes TEM images of (A) 0.1 GO with 2 mg MXene, (B) 0.2 GO with 2 mg MXene, and AFM images of (C) 2D topography and (D) 3D surface profile of the GO-MXene hybrid membrane. Scale bars: TEM: 500 nm and AFM: area 10 × 10 μm. The GO-MXene hybrid represents an effective combination for constructing a selective barrier layer that enables excellent oil-in-water emulsion separation. GO contains abundant oxygenated functional groups (-OH, -COOH, and epoxy), while MXene possesses surface terminations such as -OH, -O, and -F. These functionalities promote strong hydrogen bonding and electrostatic interactions with water molecules, leading to the formation of a stable hydration layer at the membrane surface. This hydrated interface plays a critical role in demulsification by facilitating the rejection of emulsified oil droplets while allowing efficient water transport. The hydration layer also plays a critical role in reducing oil adhesion and enhancing antifouling properties. Thus, various characterization techniques, including SEM, EDX with elemental mapping, AFM, and FTIR, were used to confirm the successful formation and structural features of the GO-MXene selective layer. In addition to verifying material synthesis, these analyses offer important insights into membrane performance. Specifically, SEM and cross-sectional imaging reveal the development of a continuous, dense, laminated layer responsible for size-selective transport and rejection. EDX mapping confirms the uniform distribution of MXene within the GO matrix, indicating a well-integrated hybrid structure that contributes to consistent separation behavior. AFM analysis provides information on surface roughness and nanoscale morphology, which influence wettability and fouling resistance. Furthermore, FTIR confirms the presence of functional groups closely associated with hydrophilicity and transport pathways. These structural and chemical characteristics collectively govern water permeability and oil-rejection efficiency. Therefore, although the characterization data were not directly used as input features in the ML model, they establish the fundamental relationships needed to determine the membrane structure that supports the experimentally measured separation performance. Predictive ML model performance The model performs well in predicting the rejection efficiency of the multiple membranes. The model achieved an R2 of 0.918, with error metrics (MSE = 5.351, RMSE = 2.313, and MAE = 1.731), indicating that the ANN effectively captures the relationships between all operational parameters and membrane performance. During the testing phase, the model maintained good performance, achieving an R2 of 0.871, with moderate increases in MSE (9.029), RMSE (3.005), and MAE (2.176). This model shows good predictive performance on unseen data; during the validation phase, it achieved an R2 of 0.861 and the lowest MAE of 1.435 across all three phases (Figure 5A). The strong predictive power across all synthesized membranes indicates that the ANN model is a reliable approach for predicting rejection efficiency under diverse operating conditions. On the other hand, SVR performs exceptionally, as shown in Figure 5B. The model achieved an excellent R2 of 0.952 during training, with error metrics (MSE = 3.004, RMSE = 1.733, and MAE = 0.937), indicating that it predicts observed performance from process variables by capturing their nonlinear relationship. Also, it achieved a good result of R2 of 0.936 during testing, with the error metrics being (RMSE = 1.984, MAE = 1.117). The validation also confirms the model’s soundness, with an excellent R2 = 0.922. The SVR model’s strong predictive performance is attributed to its kernel-based regression architecture, which can capture nonlinear interactions between the experimental and predicted variables. Additionally, the model’s performance on the new dataset highlights its flexibility. The boxplot of MAE across cross-validation folds shows that the SVR model exhibits the lowest error and minimal variance, indicating superior stability and generalization (Figure S1). On the other hand, the GBoost model exhibits strong predictive performance and achieves a calculated rejection on multiple membrane systems, with an R2 of 0.889 and error metrics (MSE = 6.925, RMSE = 2.632, and MAE = 1.975), as shown in Figure 5C. In testing (R2 = 0.855) and validation (R2 = 0.849), the model demonstrates good predictive consistency but is somewhat below that of the SVR. Additionally, this decrease in results is due to the fact that GBoost models are learned sequentially, which makes them effective at reducing bias but susceptible to noise and small changes in experimental data. In contrast, kernel optimization was further refined, yielding more fluid generalization and enabling handling of nonlinear and heterogeneous relations. Moreover, because the error-minimization process is repeated, it is an effective predictive method, particularly for nonlinear data. This model was highly predictive during training, with an R2 of 0.916 and low error metrics (MSE = 5.241, RMSE = 2.289, and MAE = 1.708), as shown in Figure 5D. These findings indicate that decision trees were an appropriate method for describing the correlation between operational parameters and the rejection performance of multiple membrane systems. However, the weakness of the Random Forest model could be explained by the limited number of estimators and the small dataset size, which may prevent the model from fully capturing complex nonlinear relationships. Some hyperparameters may be better at increasing the number of trees, and turning on more hyperparameters could help, although our current research did not explore whether other models are comparable. This kind of behavior is typically observed in tree-based ensemble models when handling heterogeneous experimental datasets with relatively small sample sizes. Future work will investigate optimized hyperparameter tuning and expanded datasets to further enhance the performance of ensemble-based models. Overall, R2 alone was not used to select the model; RMSE and MAE were also evaluated on training, testing, and validation data, and a balanced evaluation was conducted between predictive accuracy and error minimization. The SVR model had lower error values (RMSE and MAE) and higher R2 values, proving that it performs better than all other models. Statistical comparison of model errors across datasets demonstrated similar model performance patterns, leading to the choice of the SVR model. RF delivered stable results but performed slightly below SVR, which achieved a better balance between accuracy and robustness through its kernel-based optimization. In addition to predictive accuracy, the computational cost and scalability of the models were also considered. Although the SVR model achieves the highest accuracy, it is computationally expensive due to kernel-based optimization, especially for large datasets. Compared to other models, ensemble models like Random Forest and Gradient Boosting are more scalable and can handle larger amounts of data. The ANN model is also quite scalable, though it requires further tuning and more training time. Consequently, although SVR offered better predictive accuracy in this research, it may require further optimization for large-scale systems. Additionally, RF remains a valuable comparative approach offering meaningful insights into how process parameters influence rejection performance in multi-membrane configurations. Finally, the similarity of the results obtained in the training, testing, and validation periods, as well as the slightly lower R2 values in testing compared to training, indicates that the models do not overfit and retain generalization capability. This observation can also be justified by analysis of the Taylor diagram itself, which shows that correlation and variance alignment are stable across all data subsets. Also, to assess the practical relevance of the ML predictions, prediction errors were compared with the ±2%–3% range of experimental unpredictability for turbidity-based measurements and operating conditions. The MAE and RMSE values fall within this range, indicating that the models reflect meaningful trends rather than being driven by experimental noise. Also, experimental observations and SHAP analysis presented the predicted relationships, which prove the physical validity of the model. Scatter plots for ML models evaluation Rejection prediction ML models: (A) ANN, (B) SVR, (C) GBoost, (D) RF model. Taylor diagram for model performance assessment The Taylor diagram is another model assessment method used to assess a model’s robustness by showing the correlation coefficient, variability, and centered RMSE simultaneously. Therefore, all models show that the ANN achieved high accuracy: R2 = 0.98 in training, 0.871 in testing, and 0.861 in validation. Using a Taylor diagram, we obtained a good correlation coefficient of 0.94–0.97, with standard deviation values (σ = 0.9–1.1) nearly aligned with the reference normalized line (Figure 6A). On the other hand, a robust visual illustration of the SVR model achieved superior predictive accuracy compared with all evaluated models. Additionally, it had a substantial correlation coefficient value near 0.99 across training, testing, and validation. Furthermore, the standard deviation (0.8–1.0) supports the model’s validity, consistent with the observed dataset. The compacted data points near the normalized reference line show high model robustness and capture nonlinear relationships while minimizing error and variance. The Taylor diagram thus reinforces the robustness and generalization capability of the SVR model, establishing it as the most reliable and accurate predictive framework for novel multi-membrane projection performance (Figure 6B). After that, the GBoost model also yields strong results, with a correlation coefficient of 0.96–0.97, indicating a strong relationship between the predicted and experimental data. Its standard deviation is near 0.7–0.8, slightly lower than the reference, suggesting that the model effectively captured the trend while slightly underestimating data variability. Furthermore, the GBoost model shows a low RMSE and compact clustering near the reference but still lower precision than the SVR model (Figure 6C). Taylor diagram of the RF model reveals high correlation coefficients of 0.93–0.97, also indicating a strong relationship with the experimental data. Unfortunately, their standard deviation is 0.6–0.9, showing that the model caught the majority of the experimental variability while retaining prediction stability (Figure 6D). According to the Taylor diagram, the SVR model demonstrated the highest predictive accuracy and stability, with the closest alignment with the reference in both correlation and variability. ML models performance comparison using Taylor diagrams Taylor diagram for comparison of ML models: (A) ANN, (B) SVR, (C) GBoost, and (D) RF. Assessment of feature significance Permutation variable importance The feature-importance assessment of the ANN model indicated that membrane type is the most critical parameter for predicting rejection efficiency. In the present study, membrane type is treated as a categorical variable representing different GO-MXene compositions. However, it indirectly reflects a combination of underlying physicochemical properties, including hydrophilicity, interlayer spacing, surface functional groups, and charge characteristics, which collectively influence separation performance. Pressure is the second most crucial parameter, given its direct role in driving permeate flux and enhancing separation efficiency. Subsequently, filtration time had a moderate influence, likely attributable to transient fouling and stabilization phenomena during operation. At the same time, feed concentration contributed minimally, indicating that structural and operational parameters exerted greater influence on performance (Figure 7A). On the other hand, SVR also finds that membrane type is the most dominant predictor of rejection performance, consistent with the ANN results, but with a clear distinction in feature weighting. This strong dominance demonstrates that membrane surface morphology is essential for any design experiment. Additionally, SVR has a strong ability to learn nonlinear features accurately, which helps make this combination an exceptional performer in terms of rejection. Surprisingly, feed concentration has a higher value than time and pressure, which may be because the SVR model relies more heavily on nonlinear features, whereas the remaining models primarily use structure averaging or thresholds. Filtration time and pressure had relatively lower importance values than membrane type. The SVR model’s kernel-based optimization effectively captures structural dependence, reinforcing its superior predictive accuracy (Figure 7B). As observed by ANN and SVR, the GBoost model also identified membrane type as the most dominant factor influencing rejection performance. The trend outcome reinforces the pivotal role of membrane materials and structural design in governing selectivity and interfacial transport. Furthermore, pressure ranked second, reflecting its contribution to transmembrane flux and solid transport dynamics. The GBoost model placed slightly greater emphasis on operational pressure, indicating that the boosting framework captured the incremental effect of process-driven variations (Figure 7C). Lastly, the RF model also follows the trend, and the membrane type is the most effective feature in rejection efficiency. Additionally, the strong dominance shows the decisive influence of membrane surface chemistry and structural configuration on solute selectivity. Pressure was a secondary factor, indicating its role in modulating transmembrane transport, whereas time and feed concentration contributed marginally (Figure 7D). The overall comparison of the models GBoost and RF showed a similar ranking pattern, but with reduced sensitivity to process fluctuations, reflecting their averaging across multiple decision trees. Feature importance analysis. PVI-based feature significance of predictive models (A) ANN, (B) SVR, (C) GBoost, and (D) RF model. The error bar represents the standard deviation. SHAP-based feature contribution analysis SHAP analysis reveals that high SHAP values with strong positive signs indicate a dominant influence on separation performance. The SHAP analysis indicates that higher magnitudes and strong positive values confirm its dominant role in governing separation performance. Additionally, this assessment aligns with the PVI results and emphasizes that intrinsic membrane characteristics, such as surface energy, pore structure, and hydrophilicity, strongly dictate solute rejection efficiency. On the other hand, pressure, feed concentration, and filtration time contribute less than membrane type, suggesting secondary, context-dependent effects. Furthermore, moderate positive contributions from higher pressure suggest improved driving force and flux. At the same time, feed concentration and time exert mixed effects on rejection efficiency, reflecting the influence of concentration polarization and fouling during extended operation (Figure S2A). Furthermore, the SHAP analysis of the SVR model indicates that membrane type remains the most important positive factor for prediction performance. This highlights how improved membrane architecture enhances selective transport and minimizes fouling, thereby elevating rejection efficiency. Surprisingly, feed concentration in the SVR model shows better values than pressure and time, suggesting that controlled concentration gradients and optimized pressure enhance the driving force without causing flux instability. Subsequently, filtration time showed mixed effects: extended operation occasionally led to negative SHAP values, indicating fouling-induced declines. Summarizing the scenario, the SVR model effectively captured these nonlinear interactions, reaffirming its superior interpretability and predictive precision (Figure S2B). The SHAP analysis of the GBoost model indicates that membrane type is the most influential factor in predicting rejection efficiency. Compared with the SVR, the GBoost model captured similar features but with slightly broader SHAP variability. Additionally, it shows greater sensitivity to local data fluctuations. On the other hand, pressure exerted a moderate positive influence, whereas feed concentration and time exerted minimal effects, indicating both positive and negative effects linked to fouling and concentration polarization. In conclusion, GBoost also performs well, similar to SVR, but with less uniform feature impact, indicating greater sensitivity to experimental variation (Figure S2C). In the last SHAP value analysis for the RF, it follows the same trend as the other models: the RF showed a wider SHAP spread, suggesting greater sensitivity to local data variations and reduced generalization. The remaining features, such as pressure, contributed positively but less consistently, while feed concentration and time exerted limited effects due to their secondary influence on fouling and concentration gradients. The conclusion is that RF captured the same directional trends as SVR, but with lower precision, highlighting SVR’s superior ability to model nonlinear membrane-process interactions (Figure S2D). Practical implementations This research provides computational design support, enabling scientists to design and refine the membrane prior to laboratory testing to achieve targeted rejection efficiency. Although those experiments were used to train and validate the ML model, practical deployment reverses the sequence. The researcher will first design the membranes computationally and then select the optimal membrane from the set. This innovative approach will reduce the number of laboratory trials, lower chemical consumption, accelerate iteration from concept to pilot, and provide a more explicit rationale for each recommendation. Additionally, this approach turns experimental insight into a human-readable, decision-ready tool for membrane development. This technique applies to municipal water and desalination, food and beverage processing, pharmaceutical and bioprocess operations, semiconductor ultrapure-water systems, textiles, pulp and paper, and general chemical manufacturing. This study aligns with digital twins or supervisory control systems as plants become more ML-enabled. MXene-based hybrid membranes exhibit versatility across fields such as photocatalysis, sensing, and energy storage. Among all, most studies have not yet used ML models, although several other models have been explored; their use is limited. We specifically explored models for a novel MX-GO/NM hybrid membrane, advancing data-driven insights into membrane performance and design, and compared these models with other applied models, making this research unique (Table 1). Table 1Comparison of the current study with the literatureMembraneSynthesisApplicationML models usedReferenceMXene-based electrodes/composites (e.g., NiFe-LDH/MXene, MoO3/MXene)data-driven synthesis using literature data (∼7,500 points) + feature engineering + ML-guided optimization + experimental validation (fabrication of electrodes)supercapacitors (energy storage) – prediction of specific capacitance, cyclic stability, and capacitive retentionRandom Forest (best), CART, ANN, gradient boosting (GB), linear model (LM)32CO-MX/TiO2In situ cobalt modification; MXene/TiO₂ heterojunction formationMXene/TiO₂ heterojunction formationPhotocatalytic nitrogen fixation / ammonia productionN/A33Ti3C2O2, Ti3C2F2, Ti3C2data-driven design using MD simulations + literature data + ML + optimization algorithmsseawater desalination (water flux & salt rejection)XGBoost (main), Random Forest (RF), SVR + optimization algorithms (GA, PSO, DE, NSGA-II, NS-PSO)34Ti3C2Tz MXeneintensive layer delamination and acid etchinginexpensive precursor materialN/A35Nanocellulose-enabled membranesinterfacial polymerizationnanocellulose membranes provide an affordable platform for a range of pressure-driven filtration methodsANN, Random Forest, gradient boosting36Cr2CTx MXeneetchinghydrogen evolution reactionN/A37MXene and graphene nanoplatesHF etchingprediction of electrochemical properties in a supercapacitorANN, MLR, SVR38MXene-derived nanoflower-shapedin situ transformationphotocatalytic activityN/A39MXene/PVA nanofiber membraneelectrospinning + MXene synthesis + RSM optimizationbioethanol dehydration (pervaporation)response surface methodology (RSM)40Ti3C2Tx MXene/graphene compositehydrothermal methodelectrochemical energy storageN/A41CuO/MXene heterostructureMOF-derived + solvothermal + annealingCO gas sensorLSTM (deep learning)42MXene hybridsheterojunctionphotocatalyticN/A43MXene-GO/PM hybridHF etchingprediction of rejection efficiency of novel multiple hybrid membranesANN, SVR, GBoost, RFthis work Comparison of the current study with the literature Overall, this study introduces a novel GO-MX-NM hybrid membrane system developed and validated for efficient water purification. Four membranes were fabricated; the hybrid comprising 0.2 mg GO and 2 mg MXene achieved the highest rejection efficiency and structural integrity when the input features were fed into the ML models. To complement experimental results, advanced ML models were applied to an innovative approach for multi-membrane prediction using a reduced set of key parameters. Despite restricted input parameters, the SVR model achieved outstanding accuracy (R2 = 0.952 training, R2 = 0.936 testing, and R2 = 0.922 validation), among all models, indicating robustness and the dataset’s quality. This integration of experimental validation of the GO-MX hybrid membrane with ML-driven prediction marks one of the first data-efficient ML applications on multiple membrane systems. The results demonstrate how combining materials designed with intelligent modeling can accelerate membrane optimization. The current ML framework is developed using a limited dataset comprising four membrane configurations within a single GO-MXene material system. Therefore, the model should be interpreted as a proof-of-concept demonstration of small-sample nonlinear regression, rather than a universally generalizable tool for membrane design. However, the absence of diverse membrane chemistries and physics-informed descriptors limits the model’s broader applicability. This study focused on establishing a preliminary correlation between material composition and separation performance within a controlled system, thereby demonstrating the feasibility of integrating ML with membrane development. Future work will focus on expanding the dataset to include a wider range of membrane materials, operating conditions, and physics-based descriptors (e.g., structural, transport, and surface properties) to improve model robustness and generalizability. These efforts will enable the development of a more comprehensive intelligent membrane design. Furthermore, future studies should expand the dataset and explore additional hybrid ML-driven techniques to further improve predictive accuracy and accelerate the design of next-generation membranes for sustainable water purification. Limitations of the study There are some limitations of this study. First, the machine learning models were trained on a limited number of experimental data using four membrane models, which may reduce model generalizability. Second, the predictive model is based on a minimal set of operating conditions and does not explicitly include physicochemical descriptors such as zeta potential and interlayer distance, which could aid model interpretation. Third, experiments were performed under laboratory conditions with a model oil-water emulsion and may not represent real-world industrial wastewater. Furthermore, the stability, anti-fouling properties over multiple cycles, and scalability of the GO-MXene membranes were not comprehensively investigated. Further work, including larger datasets, the development of models for a wider range of membrane materials, and testing under more realistic operating conditions, will be required to improve model reliability and universality. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Nadeem Baig (nadeembaig@kfupm.edu.sa). Materials availability This study did not generate new unique reagents. Data and code availability •The experimental data supporting the findings of this study are available from the lead contact upon reasonable request. The experimental data supporting the findings of this study are available from the lead contact upon reasonable request. •The machine learning code used in this study is available from the lead contact upon reasonable request. The machine learning code used in this study is available from the lead contact upon reasonable request. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request. Acknowledgments The authors acknowledge financial support from the Interdisciplinary Research Center for Membranes and Water Security, King Fahd University of Petroleum and Minerals (KFUPM), under grant number INMW2604. Also, the authors acknowledge support from the Civil and Environmental Engineering Department at KFUPM, Saudi Arabia. Author contributions R.K., conceptualization, investigation, visualization, data curation, modeling and software, and writing – original draft. N.B., conceptualization, investigation, visualization, validation, supervision, and writing – review & editing. Z.M.Y., conceptualization, validation, project leadership, review and edits, supervision, and writing – review & editing. M.A.-S., conceptualization, investigation, visualization, validation, supervision, and writing – review & editing. Declaration of interests The authors declare no conflict of interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERTi3AlC2 MAX phaseSigma-AldrichCas No: 196506-01-1Hydrofluoric acid (HF, 48 wt%)Sigma-AldrichCas No: 7664-39-3Sodium dodecyl sulfate (SDS)Sigma-AldrichCas No: 151-21-3Graphene oxide (GO, 4 mg/mL)Sigma-AldrichCas No: 777676Diesel oilIndustrial & Mechanical Services Co. Ltd.–Sodium hydroxide (NaOH)Fluka AGCAS No: 1310-73-2Nylon membranesLVDALAB–PythonPython Software FoundationPython 3.12 Experimental model and study participant details This study did not involve human participants, animal subjects, cell lines, or clinical samples. Method details Experimental work Materials Sigma-Aldrich supplied MAX phase 312, Ti3AlC2, with a minimum purity of 99.99%, HF Hydrofluoric acid (HF, 48% wt.), and Sodium dodecyl sulfate (SDS) (ACS reagent, ≥99.0%). Graphene Oxide (GO), 4 mg/mL, was also obtained from Sigma-Aldrich. Diesel oil (100% pure) was obtained from Industrial & Mechanical Services Co. Ltd. NaOH (99.9% pure) was obtained from Fluka AG. Nylon membranes were purchased from the LVDALAB. All experiments were conducted using distilled water. Synthesis of MXene Ti3C2Tx was synthesized from the MAX phase precursor Ti3AlC2 through the selective etching of the aluminum layers in a typical procedure. Ti3AlC2 powder was gradually added to a 48 wt% HF solution under continuous stirring at 35°C. The etching process was allowed to proceed for several hours to ensure the complete removal of the aluminum layers.44 Following etching, the resulting suspension was repeatedly washed with deionized water through centrifugation until the pH of the supernatant reached a neutral range (pH 6), confirming the removal of residual acid.45,46 The sediment obtained after the final washing step was then redispersed in deionized water to form a stable MXene colloidal suspension. Therefore, exfoliated MXene nanosheets were stored under inert conditions at 4°C until further use (see figure below). Shows the synthesis of MXene through HF etching and its exfoliation using DMSO Fabrication of GO-MX/NM The GO-MX hybrid membrane was fabricated by depositing the nanosheets onto a polymeric membrane in a dead-end cell. GO was combined with MXene nanosheets at specific MXene-to-GO ratios to fabricate hybrid membranes with varying GO loadings. A fixed amount of MXene (2 mg) was used, while GO was added at 0, 0.1, and 0.2 mg to obtain membranes with different compositions. The suspensions were ultrasonicated to ensure uniform dispersion and subsequently deposited onto nylon membranes using a dead-end filtration setup at a controlled pressure of 0.5 bar until complete surface coverage was achieved. The resulting GO-MXene laminated membranes were dried at room temperature and stored in deionized water prior to characterization and performance evaluation. The selected GO loadings were designed to systematically investigate the effect of compositional variation on membrane performance and to enable the ML models to capture the relationship between composition and rejection efficiency. Characterization of membranes The surface morphology and structural features of the GO-MX-NM hybrid membranes were investigated using an advanced scanning electron microscope (SEM, EM-30AX, COXEM Co., Ltd., Korea). Additionally, it reveals a uniform dispersion of MXene nanosheets within the polymer matrix, and EDX confirms the elemental composition, illustrating the successful synthesis of the hybrid membrane. The X-ray diffraction (XRD; Miniflex 600, Rigaku, Japan) patterns exhibit distinct shifts in characteristic peaks. It shows an expansion in the interlayer spacing between the hybrid 2D materials. Additionally, atomic force microscopy (AFM) confirms interfacial interactions between membranes, and Fourier transform infrared (FTIR) spectroscopy confirms the presence of characteristic functional groups and chemical interactions between GO-MX and the nylon membrane. Data collection Data analysis is a crucial component of this research, in which 4 membranes were fabricated in the laboratory. A dead-end filtration setup (HP4750) with an active membrane area of 14.6 cm2 was used to assess the membrane’s performance. A 100 mL oil-water emulsion (diesel oil with surfactant) was added as the feed solution after the membrane was installed in the cell. Experiments on filtration were carried out at progressively higher pressures ranging from 0.2 to 3 bar. A fixed permeate volume of 15 mL was collected at each pressure, and the associated filtration time was noted. A turbidity meter was used to measure permeate turbidity to calculate the rejection efficiency. The process was repeated at various feed concentrations and operating pressures, yielding 180 data points used to develop ML models. The input parameters are pressure, time, feed concentration, and membrane type; rejection efficiency is the output parameter. The input features included operating pressure, filtration time, feed oil-water emulsion concentration, and membrane type, whereas the output variable was rejection efficiency (%). The rejection efficiency was measured using a turbidity meter (T100WL, Oakton Turbidity Meter kit). Figure S3 shows that turbidity and oil concentration have a linear relationship. Therefore, calculations based on turbidity were used for training ML models. The flux of multiple membranes was calculated using Equation 1. Where J indicates the permeation flux, A represents the surface area of the membranes (m2), Δt symbolizes the permeation time (h), and ΔV is the volume (L) of the permeate. On the other hand, the separation efficiency was measured using Equation 2. Where Cp indicates the oil concentration in the permeate, and Cf represents the oil concentration in the feed, whereas η denotes the separation efficiency. Additionally, the flux recovery ratio was calculated using Equation 3. Where FRR is flux recovery ratio, “Jf” is the flux of the membrane after cleaning the fouled membrane, and “Ji” is the initial flux of the cleaned membrane. Methodological overview Applied machine learning models Artificial neural network (ANN): Bio-inspired ML models attract attention by solving optimization problems with precision and accuracy. Additionally, a range of models with varying algorithms, hidden layers, and training data ratios was investigated to identify the optimal model. Furthermore, four membranes were fabricated to obtain good data for predicting the relationship between input features (pressure, time, feed concentration, and membrane type) and rejection efficiency. Hyperparameters such as the number of hidden layers, the number of neurons per layer, activation functions, and the learning rate were optimized to improve predictive accuracy while minimizing overfitting.47,48 The general formula used in the ANN background is shown in Equation 4. Where (yj) is the neuron output, f represents the activation energy, Bj is the bias, and wji are the weights at the last, xi are the inputs. Support vector regression (SVR): This model was first introduced by Vapnik in 1995.49 Furthermore, SVR based on the structural risk minimization principle generalizes better than other approximation approaches for nonlinear functions. Furthermore, by introducing a loss function that is insensitive to noise, SVR maintains robustness to noise while ensuring precise predictions. Kernel selection (radial basis function or polynomial), regularization parameters, and Yvalues were systematically tuned to optimize model performance.50 The general formula used in the SVR model background is shown in Equation 5. Where fx is the model output, and xi is the training data points, and αi∗ is the Lagrange multiplier for optimization, and the last one, b, is the bias term in the equation. Gradient Boosting regression (GBoost): This model produces robust predictions by iteratively combining weak learners, most commonly decision trees. Each new tree corrects the errors of the previous ones, progressively improving accuracy. In this study, the number of estimators, learning rate, and maximum tree depth were tuned to balance bias and variance. Thus, the iterative correction of residual errors made GBoost highly effective in modeling the rejection trends across different membranes and operational conditions.51 The goal of GBR is to construct a predictive model F(x) as a sum of weak learners as shown in Equation 6, where hm(x) are the weak learners (decision trees), γm are the coefficients, and M is the number of boosting stages. At each stage m, a new tree ℎm(x) is fit to the negative gradient of the loss function evaluated at the current model Fm−1(x), effectively reducing the model’s residual errors. The update rule for adding new learners is described by Equation 7. Additionally, Equation 8, where γm is chosen to minimize the overall loss L. Random Forest (RF): This deep learning model was used to improve prediction reliability by combining the outputs of multiple decision trees. Additionally, all trees were trained using a bootstrap sample, and overall predictions were computed as the average of all trees, minimizing sensitivity to noise and overfitting.52 Also, the key hyperparameters, such as the number of trees, maximum depth, and the number of features considered at each split, were optimized using grid search and cross-validation.53 The predictive model in RFR is defined as the average of the outputs of all the trees, as shown in Equation 9. Where, fm(x) shows the prediction of the mth tree and M is the total number of trees in the forest. Feature selection method Permutation variable importance (PVI) PVI was used to provide a deeper understanding of the training models. Furthermore, this technique quantifies the influence of each input parameter by assessing the decline in model performance when the feature values are randomly shuffled.54 A greater decline in efficiency indicates that the variable is more important in driving the prediction. By applying PVI, the relative influence of pressure, time, foot concentration, and membrane type was quantified, thereby identifying the most critical factors governing rejection performance. Shapley additive explanations (SHAP) SHAP values are commonly used to assess how each feature affects the overall ML model. Furthermore, this analysis will provide the specific value to every parameter used in the dataset.55 Additionally, this technique not only ranks features by overall importance but also shows how each variable influences predictions in positive or negative directions. Thus, analysis provided a transparent, interpretable view of how operational conditions and membrane characteristics jointly affected rejection efficiency, thereby complementing the results from PVI. Modeling development The research combined 2D material design with ML mdoels. Additionally, a novel hybrid membrane (GO-MC/NM) was synthesized, and new data were collected. Therefore, these syntheses were confirmed by characterizations such as XRD, FTIR, and SEM/EDX, which further validate the data. Subsequently, manual data are entered during data preprocessing and serve as the base field for the various model comparisons. After that, statistical models were employed to train the model for future prediction. At the same time, these models are further evaluated using feature-importance analysis. To quantify feature effects using permutation variable importance and interpret the directions of influence using SHAP values. Additionally, it not only provides the essential feature but also explains how each parameter affects it, such as whether it increases or decreases it. Another assessment, the Taylor diagram, was used as a statistical tool to evaluate model performance. Furthermore, shows the combined effects, such as the correlation coefficient and the data’s variability, relative to the normalized reference line. The combination of experimental work and predictive modeling makes this research unique (see figure below). A complete flow chart was adopted for the proposed method Data preprocessing Data preprocessing is an important step prior to applying ML models. Additionally, this step prepares the data for predictive analysis by eliminating redundant points in the raw data. This is achieved primarily through Python libraries that make the data sound and well-structured. The raw data were initially filtered to remove outliers, thereby strengthening the dataset. Outliers identified using statistical techniques were either fixed or removed to preserve data integrity. Subsequently, a Pearson correlation analysis was conducted to address multicollinearity and data leakage. Moreover, categorical variables like the type of membrane were put on the numbers, and all numeric variables were put on a similar scale so that the contribution of each variable during training was equal.56 Machine learning construction After data preprocessing, the cleaned data were uploaded to Python and split into three categories: training, testing, and validation. After that, give weights: training = 70%, testing = 20%, and validation = 10%. Additionally, 5-fold Cross-validation (Figure S1) was applied exclusively to the training dataset to optimize model performance and avoid overfitting. The testing and validation datasets were kept separate and used only for performance evaluation, ensuring an unbiased assessment of model generalization. All models were trained and evaluated using a 5-fold cross-validation framework. The ML models developed are trained on a small experimental dataset and thus can be used under the conditions investigated in this study. In particular, the model has been shown to be valid across the tested pressure range (0.2–3 bar), feed concentration changes, and the four modes of the GO-MXene membranes used in this study. Prediction reliability may be lower when results are extrapolated beyond these conditions. Furthermore, setting up the hyperparameter is also a crucial part in modeling. ANN model uses reLU activation with 2 hidden layers (50.49) and uses a low learning rate = 1 × 10−3 to enhance the efficiency. On the other hand, SVR utilized an RBF kernel-based predictive modeling with the parameters (C = 3.0, ε = 0.01, γ = “scale”). SVR performs exceptionally among all the ML models. After that, the GBoost model uses a slightly higher learning rate of 0.05 as compared to the ANN. Also, it tuned with the 50 estimators with a depth of 2. On the other hand, the RF model uses three estimators with the minimum sample leaf = 1, and sqrt feature selection, making the model robust and efficient. After tuning the ML models, cross-validation was employed to assess the models using techniques such as the Taylor diagram, PVI, and SHAP values. These steps all make the model more robust and improve prediction efficiency. The detailed hyperparameter settings used for all ML models are provided in the Supplementary Information (Table S1). Performance metrics To check the reliability of the model is a very important and difficult step. So, we used some statistical indicators that assess the model performance, such as mean absolute error (MAE),57 mean square error (MSE),58 root mean squared error (RMSE),59 and the coefficient of determination (R2).60 Additionally, all calculations were performed using the formulas shown in Equations 10, 11, 12, 13, and 14. Here, yi is the observed value, ŷi is the predicted value, ȳ is the mean of the observed values, and n is the number of observations. MAE represents the average magnitude of prediction errors. At the same time, RMSE provides a more stringent measure by assigning greater weight to larger deviations. The R2 value, on the other hand, reflects the proportion of experimental variants in rejection efficiency that the model successfully captures, with values closer to 1 indicating stronger predictive performance. Thus, these supportive metrics help assess the performance and predictive ability of the ML models across diverse experimental conditions. Additionally, a Taylor diagram was used as a statistical tool to evaluate model performance by combining correlation, standard deviation, and RMSE. In brief, the various ML paradigms were chosen to capture the nonlinear dependence between the membrane structure and operational parameters. In particular, ANN was used for its ability to characterize highly nonlinear trends; SVR was used because it is effective in small-sample regression with kernel-based optimization; and tree-based ensemble methods such as GBoost and RF were adopted for their adaptability in characterizing nonlinear interactions and minimizing variance. This choice allows for a comparative analysis of neural, kernel-based, and ensemble learning methods and provides a holistic assessment of model performance on small-sample nonlinear regression problems. Quantification and statistical analysis The permeate flux and rejection efficiency were calculated according to the formulas in the oil-in-water emulsion filtration experiments section. The machine learning model performance was assessed using R2, MSE, RMSE, and MAE. The experimental data were split into train, test, and validation sets in a 70%, 20%, and 10%, respectively. The training dataset was used for 5-fold cross-validated training to assess model stability and avoid overfitting. We compared model performance and accuracy using Taylor diagrams for correlation, standard deviation (SD), and centered RMSE. PVI and SHAP analyses were used to assess feature importance. For boxplots of cross-validation, the median is marked as a central line, the interquartile range as a box, the whiskers represent non-outlier data points, the mean value as a square, and any data point outside the whiskers is considered an outlier. Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.116271. Contributor Information Nadeem Baig, Email: nadeembaig@kfupm.edu.sa. Zaher Mundher Yaseen, Email: z.yaseen@kfupm.edu.sa. Supplemental information Document S1. Figures S1–S3 and Table S1 References Associated Data Supplementary Materials Document S1. Figures S1–S3 and Table S1 Data Availability Statement •The experimental data supporting the findings of this study are available from the lead contact upon reasonable request. The experimental data supporting the findings of this study are available from the lead contact upon reasonable request. •The machine learning code used in this study is available from the lead contact upon reasonable request. The machine learning code used in this study is available from the lead contact upon reasonable request. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.