Physics-Embedded Machine Learning for Fatigue Cumulative Damage Prediction

Abstract

Fatigue damage accumulation is critical to the safety and reliability of mechanical structures, yet accurate prediction remains challenging, especially under small-sample conditions. This study proposes an innovative physics-embedded machine learning (ML) framework to enhance residual fatigue damage prediction by integrating the Manson–Halford (MH) physical model with data-driven algorithms. The framework employs a dual-regressor approach: One regressor embeds the MH model to predict the interaction coefficient, while the other is purely data driven to directly predict residual fatigue damage, with a customized loss function enforcing physical consistency between the two outputs. A compiled dataset of 14 materials demonstrates the framework's superiority over six baseline ML models. Notably, the model retains high accuracy even with 30% fewer training data, showcasing its robustness in data-scarce scenarios. By harmonizing physical mechanisms with ML, this work provides a generalizable and efficient strategy for fatigue damage prediction.