Stress–strain curve predictions by crystal plasticity simulations and machine learning

Abstract

The stress–strain curve (SSC) prediction for additively manufactured as-build metal materials with laser powder bed fusion (LPBF) is a lengthy and tedious process. It involves the sophisticated representative volume element (RVE) reconstruction of complex experimental microstructures for subsequent state-of-the-art crystal plasticity simulations with hyperparameter tunings in the appropriate physical model. However, even with a well-fitted model, simulations with different RVEs or temperatures, for example, are too time-consuming and computationally intensive. In recent years, several attempts were directed towards the SSC predictions with machine learning (ML) tools to speed up this process. Mainly, the artificial neural networks (ANN) were reported so far for this purpose. Here, we present our version to predict the temperature dependence of SSCs for LPBF fabricated industrially important Hastelloy X with various ML methods. Compared to previously reported studies on this matter with direct link between the microstructures and SSCs, we directly link only experimental conditions and predicted SSCs, which could be more preferable for some application scenarios discussed below. It was found that due to the structure and “small” size of our training dataset, the decision tree-based ML regressors worked better than other popular ML methods.