Mechanics-informed risk-aware learning for multiaxial structural reliability with Bayesian calibration

Reliable prediction of structural vulnerability under complex multiaxial loading remains challenging due to nonlinear load couplings, imbalanced failure-critical samples, and limited model interpretability. Here, we present a mechanics-informed, risk-aware learning framework integrating polynomial-harmonic feature augmentation, Weibull-based risk reweighting, and a unified Degradation Risk Score that combines stress margins and bolt pretension loss. Demonstrated on a bolted steering-knuckle assembly with finite-element-derived multiaxial load-response data, the framework improves multiple linear regression from R² = 0.37 to 0.96 with a 67% reduction in RMSE, while enhancing robustness and sensitivity in high-risk regimes across ensemble and neural network models. SHAP analysis confirms that physically meaningful multiaxial interaction features dominate predictions, revealing critical load paths associated with structural vulnerability. Finally, Bayesian logistic and Weibull calibration provide a route to link Degradation Risk Score to component-level failure probabilities, enabling probabilistic risk assessment and reliability-centred decision-making in practical engineering systems.