From data to alloys predicting and screening high entropy alloys for high hardness using machine learning

Abstract

The growing need for structural materials with strength, mechanical stability, and durability in extreme environments is driving the development of high entropy alloys. These are materials with near equiatomic mixing of five or more principal elements, and such compositional complexity often leads to improvements in mechanical properties and high thermal stability, etc. Thus, high-entropy alloys have found their applications in domains like aerospace, biomedical, energy storage, catalysis, electronics, etc. However, the vast compositional design and experimental exploration of high-entropy alloys are both time consuming and expensive and require a large number of resources. Machine learning techniques have thus become essential for accelerating high entropy alloys discovery using data driven predictions of promising alloy combinations and their properties. Hence, this work employs a machine learning framework that predicts high entropy alloy hardness from elemental descriptors such as atomic radius, valence electron count, bond strength, etc. Machine learning regression models, like LightGBM, Gradient Boosting Regressor, and Transformer encoder, were trained on experimental data. Additionally, a language model was also fine tuned to predict hardness from elemental descriptor strings. The results indicate that LightGBM has better accuracy in predicting the hardness of high entropy alloys compared to other models used in this study. Further, a combinatorial technique was used to generate over 9 million virtual high entropy alloy candidates, and the trained machine learning models were used to predict their hardness. This study shows how machine learning-driven high throughput screening and language modelling approaches can accelerate the development of next generation high entropy alloys.