Data-driven body-centered cubic phase prediction in cobalt free high-entropy alloys

TL;DR

This study develops a machine learning-based method, augmented with data generation, to predict the body-centered cubic phase stability in cobalt-free high-entropy alloys, aiding their design for advanced applications.

Contribution

It introduces a combined approach of semiempirical parameters, data augmentation via GANs, and Gaussian process classification to accurately predict phase formation in complex HEAs.

Findings

Achieved 84% accuracy in phase prediction.

Identified mixing enthalpy and atomic size difference as key descriptors.

Demonstrated the effectiveness of data augmentation in improving ML model performance.

Abstract

High-entropy alloys (HEAs) are known for superb combination of performance attributes, making them ideal for advanced applications, e.g., nuclear engineering. The concept of cobalt-free HEAs aims to mitigate concerns about cobalt's radioactivity, however, predicting their phase formation remains challenging due to their complex compositions. In this work, we integrate six semiempirical parameters, i.e., mixing entropy ({\Delta}Smix), mixing enthalpy ({\Delta}Hmix), atomic size difference ({\delta}), valence electron concentration (VEC), d-orbital energy level (Md), and the {\Omega} parameter, along with machine learning (ML) to predict the body-centered cubic phase stability in Co free HEAs. To address the limitations of experimental data, generative adversarial networks were used to augment the dataset, thus improving the accuracy of the Gaussian process classification model used for phase prediction. After dimensionality reduction to five principal components, the model achieved an accuracy of 84%, with {\Delta}Hmix and {\delta} identified as the key descriptors influencing phase formation. This approach highlights the synergy of ML and data augmentation in accelerating the design of HEAs for advanced applications.