Study of Hybrid Machine Learning Multiaxial Low-Cycle Fatigue Life Prediction Model of CP-Ti

Abstract

Symmetric and asymmetric multiaxial low-cycle fatigue tests were conducted on commercially pure titanium under different control modes and multiaxial strain/stress ratios to establish a reliable hybrid physics and data-driven method. Optimized analysis formula–based models are proposed to provide reliable physical information first. Based on the dataset enhanced by the nonlinear variational autoencoder method, a hybrid VAE-ANN model is established and trained, developed using the Pearson correlation coefficient analysis and Leaky ReLU activation function. Through a series of fatigue life prediction validations under both symmetric and asymmetric loading conditions, the VAE-ANN model demonstrates excellent prediction accuracy, broad generalization capability, and strong compatibility, achieving the lowest average absolute relative error of 6.76% under symmetric and 22.61% under asymmetric loading conditions.