Predicting strain-induced martensite in austenitic steels by combining physical modelling and machine learning

Abstract

Computational materials design has made significant progress lately. However, one underexploited opportunity lies in the combination of physically based modelling and machine learning (ML). In the present work we exploit this combination for modelling of strain-induced martensitic phase transformation (SIMT) in austenitic steels. A fully predictive model for SIMT, responsible for the TRIP effect in many steels, is devised. An experimental dataset correlating SIMT with composition, temperature and strain is collected from the open literature firstly. Secondly, the Olson-Cohen model is applied to make physically based predictions on temperature and strain dependence of SIMT in order to expand the database to the final size of 16,500 entries relating the features and the target. Thirdly, ensemble ML methods are applied to model the data and the final model is validated on a holdout dataset, including also dual-phase alloys. The final model provides accurate predictions of SIMT in a temperature range from −196 to 100 °C and from 0 to 1 in strain. The model can readily be extended to consider further factors such as strain rate and stress state. Moreover, it can be used together with thermodynamic and kinetic calculations, or thermomechanical simulations, for the design of steels and components, respectively.