Deep Learning of Temperature – Dependent Stress – Strain Hardening Curves

Abstract

In this study, structure – property relationships (SPR) have been investigated using machine learning methods (ML). The research objective was to create a ML model that can predict the stress – strain response of materials at different temperatures from a given microstructure with industrially acceptable accuracy and high computational efficiency. Automated microstructure generation techniques and numerical simulations were developed to create a dataset for the ML model. Two – phase 3D representative volume elements (RVEs) were analyzed using finite element analysis (FEA) to obtain the stress – strain responses of the RVEs. The phase arrangement of the RVEs, the temperature, and the stress – strain responses were used to train the ML model. The microstructure arrangement and the temperature – dependent mechanical properties of each phase were known parameters, while the output parameter was the stress – strain response of the two – phase RVE. The ML model has shown excellent prediction accuracy in the temperature range from 20 °C to 250 °C. In addition, the model showed very high computational efficiency compared to FEA, allowing much faster prediction of the stress – strain curves at specific temperatures.