A physics‐informed generative adversarial network framework for multiaxial fatigue life prediction

Abstract

Deep learning has achieved great success in multiaxial fatigue life prediction. However, when data‐driven models are used to describe data from physical processes, the relationship between inputs and outputs is agnostic. This paper proposes a deep learning framework combining generative adversarial networks and physical models to predict multiaxial fatigue life. This framework incorporates three life prediction equations in the loss function of generator, respectively. The results show that models with suitable physical constraints outperform neural networks in predicting results. Introducing the Smith–Watson–Topper model as a physical loss degrades the predictive performance of the physics‐informed network. On the contrary, introducing Fatemi–Socie and Shang–Wang model as physical loss improves the predictive performance of physics‐informed network. Learning using physics knowledge can lead to the ability of model to generate data that satisfy the governing equations of physics.

A physics‐informed neural network framework based on fatigue indicator parameters for very high cycle fatigue life prediction of an additively manufactured titanium alloy

“…He et al propose a deep learning framework combining generative adversarial networks and PINN to predict multiaxial fatigue life. 49 The prediction error is the smallest over several machine learning methods. Alessio Centola et al use the process parameters, the thermal treatments, the surface treatments, and the stress amplitude as the input parameters of the PINN model to predict the fatigue life of AM Ti6Al4V components.…”