Professor Xu Wei's team at the State Key Laboratory of Rolling Technology and Continuous Rolling Automation at Northeastern University recently published a research result entitled Physical metallurgy-guided machine learning and artificial intelligent design of ultrahigh-strength stainless steel in ActaMater.79 (2019) 201-214. The first author of the article is Shen Chunguang, a PhD student in the laboratory, and the co-corresponding authors are Wang Chenchong and Xu Wei.
Paper link: https://doi.org/10.1016/j.actamat.2019.08.033
Material genetic engineering is a frontier direction in the field of materials that has emerged in recent years. It aims to reduce the cost and cycle of material R & D through the comprehensive use of integrated computing and database technologies. With the development of integrated computing ideas and data mining technology, material genetic engineering has achieved remarkable results in the research and development of many new high-performance metal materials. However, because the composition process design of steel structural materials involves a large amount of analysis of complex coupling relationships and is constrained by many controversial mechanisms, the design of steel materials based on material genetic ideas has always been an international hot and difficult issue in the field.
With the advent of the era of big data, machine learning algorithms provide a feasible way to solve process design and optimization under complex systems. Although there are several successful cases of material design methods based on machine learning today, the design methods used in most studies are only using machine learning algorithms to directly build a strong correlation between "component / process-target performance" at the level of statistical mathematics. , While ignoring the physical metallurgical information such as microstructure characteristics that are most concerned in traditional material design ideas, this design method not only greatly hinders the understanding of the inherent physical mechanism of the design results, but also leads to the extreme existence of data in the design method. Large reliance has caused great difficulties in improving the design efficiency and universality of the model.
Aiming at the above bottlenecks in the design of steel materials based on traditional machine learning algorithms, Professor Xu Wei's team proposed a machine learning method guided by physical metallurgy from the perspective of the fusion of physical metallurgy and machine learning algorithms. In this design method, by using physical and metallurgical parameters that are highly related to strength to perform dimensionality enhancement processing on the original data, not only can the physical metallurgical information be integrated into the machine learning process, but also the underlying information of the original data can be fully tapped to improve data quality A prediction model with excellent generalization ability was obtained. The prediction model was then combined with high-throughput genetic algorithm optimization to form an efficient alloy computing design framework.
Based on this design framework, a new type of ultra-high-strength stainless steel was successfully designed under a small sample data set of 10 ^ 2. Compared with the original data set, the design of the alloy system not only achieved an increase in strength but also a significantly lower alloy content. In addition, by comparing the design process without physical metallurgy participation, it clearly reveals the ability of physical metallurgy information to improve model performance and design efficiency when participating in machine learning performance prediction. The research results provide feasible ideas for the design of steel materials based on small sample machine learning algorithms and the improvement of physical interpretability in machine learning algorithms.