Advancing Structural Failure Analysis with Physics-Informed Machine Learning in Engineering Applications

Abstract

While machine learning (ML) shows significant potential for structural-failure analysis, purely data-driven approaches face critical limitations, including data scarcity, lack of physical consistency, and poor interpretability in safety–critical applications. Physics-informed ML (PIML) addresses these challenges by integrating physical principles with data-driven methods, thereby enabling accurate and interpretable predictions, while maintaining physical consistency. This study presents a systematic categorization of PIML implementation strategies in structural-failure analysis, classifying the approaches into four distinct categories: physics-guided data manipulation, physics-inspired architectural design, physics-constrained loss functions, and hybrid physics–ML models. We examined the applications across the complete failure lifecycle, from mechanism analysis and fatigue-life prediction to structural-health monitoring and post-failure analysis, to demonstrate how different PIML strategies address specific engineering challenges. Through a critical evaluation of representative studies, we identified the current limitations, including data-integration complexities, physics-formalization difficulties, and computational trade-offs between accuracy and efficiency. Future research directions emphasize multisource knowledge fusion, transferable PIML frameworks, and enhanced post-failure analysis capabilities. This systematic framework provides clear guidance for selecting appropriate PIML strategies based on application requirements and available resources, thereby advancing the reliability and safety of engineering structures.