Physics-guided machine learning for the failure prediction of a novel Z-beam composite wing structural concept

Abstract

To enhance the structural and aerodynamic performance of next-generation civil aircraft, novel composite wing concepts are being explored for aeroelastic tailoring and weight reduction. However, conventional analytical and empirical tools struggle to evaluate these complex configurations at early design stages, limiting the exploration of their full potential. This study develops physics-guided machine learning models for the prediction of mechanical response and failure behaviour of innovative Z-beam-based composite wing structures. The Z-beam concept provides increased geometric freedom, enabling tailored bending–torsion coupling. A dataset of representative Z-beam elements is generated using finite element simulations with periodic boundary conditions. Local stress concentrations are analysed using refined sub-models with LaRC05 failure criteria to compute detailed failure indices. A physics-guided learning framework is proposed to predict failure indices, dominant failure modes, and critical locations within the wing box. The framework embeds the linearity of elastic response and the monotonic relation between strain and failure indices to enhance data efficiency and prediction robustness. A second model is trained to estimate bending, torsional, and coupling stiffness properties. Results show that the proposed models deliver high predictive accuracy, providing an effective tool for early-stage design exploration and optimisation of composite wing structures.