Using hybrid physics-informed neural networks to predict lifetime under multiaxial fatigue loading

Abstract

In this article, a machine learning approach is utilized to predict lifetime under multiaxial fatigue loading. A novel hybrid physics-informed neural network is proposed, where a combination of a LSTM/GRU cell and a fully connected layer is used to extract the damage parameter of a loading cycle. A newly proposed logarithmic activation function is then used to introduce the power law relationship between the damage parameter and the predicted fatigue life. In addition, the selected parameters of the suggested network are physically guided. Two data pre-processing methods are used to ascertain the rotational invariance of the axial–torsional loading conditions. The prediction capability of the suggested approach is demonstrated by the experimental datasets that consist of axial–torsional test results obtained for 42CrMo4 steel and for 2024-T3 aluminium alloy. A good correlation between the predicted and experimental data was achieved. Finally, the extrapolation capability of the proposed approach is demonstrated through modelling the stress-life curves for the data-points outside the experimental data range.