Physics-Constrained and Explainable Machine Learning for Thermogravimetric Prediction of Wheat Straw-Epoxy Biocomposites: A Formulation-Level Validation Study

Machine-learning performance estimates for densely sampled thermogravimetric analysis (TGA) data depend on whether individual observations or complete thermograms are withheld. Six regression models were evaluated on seven thermograms from untreated, alkali-treated, and alkali-silane-treated wheat-straw/epoxy formulations, comparing a random 80:20 point-level split with Leave-One-Curve-Out (LOCO) validation. Under random splitting, 91.4% of test points lay within 0.25 °C of a same-curve training neighbour, and R2 reached 0.999999. Under LOCO validation, the composite-averaged R2 ranged from 0.975 to 0.986. A monotonicity-constrained histogram gradient-boosting model (PC-HGB) matched the best LOCO accuracy (R2 = 0.986; MAE = 2.36%) and produced no mass-increasing predictions, whereas Random Forest produced 119 such steps. However, a composition-blind monotone baseline reached R2 = 0.982 (MAE = 2.43%), showing that LOCO accuracy mainly reflects the shared degradation curve and that formulation-specific effects are only weakly resolved by seven curves. When neat epoxy was withheld, DNN and SVR gave R2 = −14.15 and −0.145, showing sensitivity to extrapolation. SHAP analysis showed that fibre loading and treatment became more influential in the high-temperature char-formation region; these attributions describe the fitted model, not demonstrated out-of-sample effects. Formulation-level validation, simple baselines, and physical constraints should accompany point-level accuracy in TGA machine-learning studies.