Multi-fidelity physics-informed machine learning framework for fatigue life prediction of additive manufactured materials

Abstract

The development direction of high reliability and longer serviceable life for major equipment requires accurate fatigue life predictions of additively manufactured (AM) components. However, small samples and high scatter of fatigue performance have become significant challenges in accurately modeling the fatigue failure behavior of AM components. To overcome the limitation of traditional fatigue life prediction models, a multi-fidelity physics-informed machine learning (PIML) framework is proposed. In this framework, the uncertainty quantification of fatigue performance and the fitting low-fidelity fatigue data with physical consistency are achieved through a physics-guided Wasserstein generative adversarial network with gradient penalty (WGAN-GP). The introduced concept of transfer learning allows training a physics-informed neural network (PiNN) using multi-fidelity fatigue data during the training process. Embedding the effect of manufacturing defects on fatigue performance as physical constraints can ensure the physical consistency of the overall multi-fidelity framework. Compared with traditional neural network (NN) and PiNN, the multi-fidelity framework has significant advantages in strong prediction performance, generalization ability and effectiveness. Moreover, the results of deep feature transfer demonstrate that the proposed multi-fidelity framework is expected to be a unified fatigue life prediction framework for AM materials.