Physics-informed Gaussian process regression model for predicting the fatigue life of welded joints

Abstract

Fatigue failure in welded joints substantially threatens the reliability of engineering structures. To address this issue, this study proposes a novel hybrid physics-informed Gaussian process regression (Pi-GPR) model to predict the fatigue life of welded joints. The Pi-GPR model is advantageous in reducing the model's dependency on extensive experimental datasets by integrating physical features from fatigue fracture mechanics. Unlike previously developed fatigue life prediction models, the Pi-GPR model uniquely addresses nonlinear characteristics of welding and fatigue testing while simultaneously quantifying the prediction uncertainty stemming from the variability of testing parameters. Spearman's rank correlation analysis method identified cross-sectional geometry features highly correlated with fatigue life, incorporating these physical features into the Pi-GPR model. Notably, the Pi-GPR model used easily measurable length-related physical features to provide comprehensive geometrical information, demonstrating exceptional prediction performance and offering confidence intervals for each result. Furthermore, the Pi-GPR model maintained superior prediction accuracy even with minimal training data, thus confirming its low data dependency.