Machine Learning-Based Prediction of the Martensite Start Temperature

Abstract

The prediction of the martensite start temperature ($M_s$) for steels based on their chemical compositions is a complex problem. Previous work has developed empirical, thermodynamic, and machine learning models to estimate $M_s$. However, the empirical models are limited to specific steel grades, the thermodynamic models rely on different model assumptions, and the machine learning models are based on a small number of data, are limited to specific steel grades, as well or are not available for easy use to the public. Herein, a new machine learning model for the prediction of $M_s$ is developed on the basis of two publicly available datasets consisting of 1800 steels from different steel grades. Extensive hyperparameter tuning is performed to find the best artificial neural network for the dataset. The best model improves prediction accuracy compared to previous state of the art. Despite a very good prediction accuracy of the model, unexpected behavior is observed in specific unseen data. This opens up the discussion for the requirements of new metrics. The dataset and the model are freely available at https://github.com/EAH-Materials. An easy-to-use web tool to estimate $M_s$ without the need of programming based on the chemical composition can be found at https://eah-jena-ms-predictor.streamlit.app/.