Machine-Learning-Guided Design of Soy Protein Isolate Film Mechanics

Soy protein isolate (SPI) films are promising renewable alternatives to petroleum-derived plastics, yet their mechanical properties arise from complex interactions between formulation and processing variables that are difficult to resolve through isolated experiments or empirical models. Here, we establish a machine-learning framework integrating literature-derived data with prospective experiments for the rational design of SPI film mechanics. A curated database was constructed from published studies, and five regression algorithms were evaluated. Extremely randomized trees, random forest, and gradient boosting achieved the best performance for tensile strength, Young's modulus, and elongation at break, with random five-fold cross-validation R² values of 0.895, 0.835, and 0.673, respectively. Shapely additive explanations (SHAP) analysis identified SPI and plasticizer concentrations as dominant compositional factors, while processing effects were property-dependent. Model-guided experiments then investigated the underexplored coupling between pH and heating temperature at fixed composition. Experiments reproduced broad processing-property trends but revealed reduced predictive accuracy under acidic and high-temperature conditions. Multiscale characterization linked these deviations to pH-dependent protein association, secondary-structure evolution, and network morphology. This literature-to-experiment workflow transforms heterogeneous published data into experimentally testable, mechanistically interpretable design principles for data-guided development of protein-based materials.