A Physics‐Informed Neural Network Approach to Fatigue Life Prediction Under Multi‐Level Loads

Abstract

A fatigue life prediction model under multi‐level loading was developed to address reduced prediction accuracy associated with load sequence effects and load interactions. First, an improved nonlinear damage accumulation model was proposed based on the Pavlou damage accumulation framework. It was then embedded in the loss function of an artificial neural network to construct a physics‐informed neural network (PINN) for fatigue test data. A data structure termed “Former‐Preceding Inheritance/Load‐Zero Padding” (F‐PI/L‐ZP) was designed for the input layer to accommodate variable numbers of load levels. It was found that purely data‐driven models without physical information exhibited poor accuracy, and linear damage models provided only marginal improvements. Trained using fatigue test data under loading sequences with different numbers of levels, the proposed model demonstrated higher prediction accuracy, broader applicability, and a stronger ability to quantitatively describe fatigue‐damage evolution than the PINNs based on the M‐H, Pavlou, and improved M‐H models.