Machine learning-driven design of multiscale reinforced horse chestnut starch-based composite films for biodegradable...

To address the poor mechanical strength, limited thermal stability, and inadequate UV-shielding capability of starch films, as well as to overcome the inefficiency of conventional trial-and-error optimization, this study fabricated multiscale hybrid composite films via solution casting and established a data-driven framework for mechanical-property prediction and formulation optimization. Horse chestnut starch (HS) was used as the biopolymer matrix, into which graphitic carbon nitride (g-C₃N₄), microcrystalline cellulose (MCC), and zinc oxide (ZnO) were incorporated as multiscale reinforcing agents. Experimental results revealed a pronounced synergistic effect among the three fillers; the composite containing g-C₃N₄: MCC: ZnO = 2:1:1 wt% achieved the highest tensile strength (37.73 MPa), attributed to the hierarchical 2D-1D-0D hybrid network formed by layered g-C₃N₄, fibrous MCC, and nanosized ZnO. To predict mechanical behavior, three machine-learning algorithms were evaluated, namely recurrent neural network (RNN), random forest (RF), and extreme gradient boosting (XGBoost). XGBoost exhibited the highest accuracy (R² = 0.998; MSE = 0.233) and was further interpreted using SHapley Additive exPlanations (SHAP), which clarified the contributions of each filler to tensile strength and elongation at break. Using the optimal ML model, response-surface functions were built to quantify the joint influence of MCC and ZnO, and were further applied to multi-objective optimization. The optimized formulation region revealed that moderate MCC levels combined with low ZnO contents yield the best balance between tensile strength and elongation at break. This study develops a data-driven framework that integrates experiments, explainable ML, and optimization to enable the rational design of high-performance starch-based films.