Predicting the Storage Modulus of Chromium‐Crosslinked Polymer Gels Through Advanced Machine Learning Techniques

ABSTRACT The accurate prediction of the storage modulus of chromium‐crosslinked polymer gels presents a complex but critical challenge due to the interplay of structural, chemical, and mechanical factors. This study implements a robust ensemble learning framework, combining Decision Trees (DT), Random Forests (RF), AdaBoost (AB), and Ensemble Learning (EL) to model and predict the storage modulus with high accuracy. A dataset incorporating key features, such as polymer and chromium concentrations, strain, temperature, and pH, was used to train these models. Among them, RF presented superior accuracy with R 2 of 0.7995 and mean squared error (MSE) of approximately 170.7 million in the test phase, outperforming its counterparts in generalizability. However, Decision Trees and AdaBoost exhibited severe overfitting, with AdaBoost reaching a training R 2 of 0.9989 but failing in testing ( R 2 = 0.8989, MSE = ~86 million). To address overfitting, hyperparameter optimization, and regularization strategies were implemented, including limiting tree depth and controlling estimator numbers. Feature importance analysis using SHAP (SHapley Additive exPlanations) identified strain (%) as the most significant negative factor, while chromium concentration positively influenced modulus. These findings align with physical and chemical principles, as strain disrupts gel networks while increased crosslink density enhances storage modulus. The novelty of this methodology lies in its departure from conventional reinforcement learning‐based optimization approaches. Unlike reward‐driven route optimization frameworks, this ensemble learning model integrates experimental physicochemical parameters with explainable inference. This enables direct mechanistic insight into the structural behaviors governing gel elasticity, establishing a distinctive and interpretable computational pathway for polymer‐gel modulus prediction.