Machine learning for hierarchical prediction of elastic properties in Fe-Cr-Al system

Abstract

We apply extremely randomized trees and deep neural network to cluster expansion generated data to construct the hierarchical model that predicts the ternary properties using machine learning model trained by binary data. We focus on the elastic properties bulk modulus and shear modulus. By feeding composition and temperature as features and elastic property as target property into extremely randomized trees, the predictions of ternary alloys achieve the mean absolute errors of 0.56 GPa and 1.49 GPa in bulk modulus and shear modulus, respectively. The performance in shear modulus predictions can be improved by adding point probability that characters the ordering effect into feeding features. We find that the compositions and temperature are key features in bulk modulus, while compositions, temperature, and the ordering effect are important in shear modulus.