Element-level finite-element-informed machine-learning surrogates for failure-index prediction in a composite wing box with intralaminar damage

This study develops and numerically assesses finite-element-informed machine-learning surrogates for predicting five classical failure indices in a composite wing-box substructure containing prescribed intralaminar stiffness degradation. A damage-oriented database is generated by varying damage position, extent, and intensity over rib, skin, and spar regions of an equivalent-orthotropic finite-element model. Random Forest, XGBoost, LightGBM, and a multilayer perceptron are trained and evaluated using a common data-partitioning and cross-validation framework, with separate attention to the upper end of the failure-index range. The tree-based ensembles consistently outperform the neural-network benchmark; the Max Stress index is the most accurately reproduced, whereas the Christensen index and spar-dominated cases remain more difficult, particularly for failure-index values above 0.7. The results establish the feasibility of rapid numerical screening within the adopted model assumptions.