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

Area of Science:

• Mechanical Engineering
• Machine Learning
• Reliability Engineering

Background:

• Predicting structural vulnerability under multiaxial loading is difficult due to complex load interactions and limited data.
• Existing models often lack interpretability and struggle with imbalanced failure data.

Purpose of the Study:

• To develop a mechanics-informed, risk-aware learning framework for enhanced structural vulnerability prediction.
• To improve model interpretability and enable probabilistic risk assessment for engineering components.

Main Methods:

• Integrated polynomial-harmonic feature augmentation and Weibull-based risk reweighting.
• Developed a unified Degradation Risk Score combining stress margins and bolt pretension loss.
• Applied the framework to finite-element-derived data from a bolted steering-knuckle assembly.

Main Results:

• Significantly improved multiple linear regression performance (R² from 0.37 to 0.96) with reduced RMSE.
• Enhanced robustness and sensitivity of ensemble and neural network models in high-risk scenarios.
• SHAP analysis revealed physically meaningful features driving predictions and identified critical load paths.

Conclusions:

• The framework offers a robust approach to predicting structural vulnerability and assessing component reliability.
• Physically informed features enhance model interpretability and identify key failure mechanisms.
• Enables probabilistic risk assessment for informed, reliability-centered engineering decisions.