EMPIRICAL BENCHMARKING OF GRADIENT BOOSTING FOR THERMAL DEGRADATION PREDICTION IN MINERAL–POLYMER COMPOSITES
TL;DR
The proposed preprocessing pipeline including phr to wt% conversion, label encoding, and median imputation provides a reproducible strategy for heterogeneous TGA literature datasets.
Abstract
This study presents a mathematical and empirical analysis of gradient boosting algorithms applied to thermogravimetric analysis (TGA) data for composites based on vinyl ether matrices and epoxides with mineral fillers. The gradient boosting algorithm is output step by step with a numerical example of TGA. XGBoost and CatBoost are then analyzed formally. 12 algorithms were compared on 478 data from the peer-reviewed literature using 5-repeated 5-fold cross-validation. Extra Trees and CatBoost achieved statistically equivalent top performance (mean R² = 0.7430 ± 0.0879 and 0.7313 ± 0.0793, respectively; Nadeau & Bengio corrected t = 0.483, p = 0.634), outperforming all linear baselines, SVR, and MLP. CatBoost is recommended for practical deployment: its ordered boosting produces unbiased gradient estimates on small datasets (n < 1000), and SHAP analysis of its gradient tree structure reveals physically interpretable feature attributions aligned with known TGA mechanisms. SHAP values identified TGA peak temperature (T_max) and matrix type as the dominant predictors. The proposed preprocessing pipeline including phr to wt% conversion, label encoding, and median imputation provides a reproducible strategy for heterogeneous TGA literature datasets.