Alloys innovation through machine learning: a statistical literature review

Abstract

This review systematically analyzes over 200 publications to explore the growing role of data-driven methods and their potential benefits in accelerating alloy development. The review presents a comprehensive overview of different aspects of alloy innovation by machine learning and other computational approaches used in recent years. These methods harness the power of advanced simulation techniques and data analytics to expedite materials’ discovery, predict properties, and optimize performance. Through analysis, significant trends and disparities within the data discerned, while highlighting previously overlooked research gaps, thus underscoring areas that require further exploration. Machine Learning techniques are widely applied across various alloys, with a pronounced emphasis on steel and High Entropy Alloys. Notably, researchers primarily investigate the physical, mechanical, and catalytic properties of materials. In terms of methodology, while 68% of the examined papers rely on a single machine learning model, the remainder employ a range of 2 to 12 models, with Neural Network being the most prevalent choice. However, a notable concern arises as 53% of these papers do not share their dataset, and a staggering 81% do not provide access to their code. Paramount importance of adopting a systematic approach when scrutinizing machine learning methodologies is underscored. Analysis shows lack of consistency and diversity in the methods employed by researchers in the field of alloy development, highlighting the potential for improvement through standardization. The critical analysis of the literature not only reveals prevailing trends and patterns but also shines a light on the inherent limitations within the traditional trial-and-error paradigm.