Abstract
Fiber-reinforced polymer (FRP) composites offer high specific strength, high stiffness, and good corrosion resistance, which suits them to aerospace, automotive, marine, and wind energy structures. Their interlaminar fracture behaviour is harder to pin down, because delamination testing is slow and costly and the response shifts with loading mode. This study presents an explainable hybrid machine learning framework for interlaminar fracture toughness under Mode-I, Mode-II, and Mixed-Mode loading. A weighted ensemble pairs Categorical Boosting (CatBoost), a gradient boosting method, with the Tabular Prior-Data Fitted Network (TabPFN), a pretrained transformer built for tabular data. A Tree-structured Parzen Estimator (TPE) tunes the CatBoost hyperparameters and the ensemble weight against specimen-grouped cross-validation. The work draws on the public D-STANDART experimental dataset and treats DCB, ELS, and MMB as three separate prediction problems for GIC, GIIC, and GC rather than pooling them into one model. Partitioning runs at specimen level, so no coupon supplies records to more than one subset, and feature selection uses training data alone. SHAP then measures how each input shifts predicted toughness in each mode. The framework reached R² of 0.9280 for DCB, 0.9870 for ELS, and 0.9627 for MMB, with RMSE of 5.1759, 1.5041, and 2.8809 and MAE of 1.6502, 0.5372, and 0.7000. Bringing a tabular foundation model into delamination analysis, and keeping the three loading modes apart, lets the framework report mode-specific feature behaviour instead of one averaged picture.