Data-driven High-throughput Experimental Alloy Design

Yi Liu*, Jiong Wang, Wanchen Zhao, Chen Zheng, and Bin Xiao

Materials Genome Institute, Shanghai University, Shanghai, China 200444

ABSTRACT: Materials design normally requires the optimization within the huge parameter spaces of composition and processing for given targeting properties. Conventional experiment approaches rely heavily on trial-and-error experiences causing the design process slow and costly, hard to explore efficiently the whole parameter spaces completely. This work combined high-throughput experiments (HTE) and machine learning (ML) methods to accelerate the process of materials design. HTE with the features of multi-station, parallelization, automation, and miniature, performances more efficiently than conventional experiments in terms of preparation time per sample. The data-driven ML methods can describe complex structure-property relationships and reduce the number of necessary experiments by guiding the multi-step experiments iteratively. HTE can provide more systematic high-quality consistent data than the conventional methods. On the other hand, ML can expand further the range and capability of predicting structure-property relationship of materials. This report demonstrated the whole R&D efficiency can be increased hundreds times by combing HTE and ML to accelerate composition design of multiple component alloys, e.g high-entropy alloy CoCrTiMoW and 6061 Al alloys. First we developed all-process high-throughput alloy synthesis systems that prepare samples automatically in multi-station batch modes, at least 10 times faster than the conventional method. Then we constructed several ML models by combining multiple features and algorithms. We designed the compositions of the alloys at next steps, guiding the iterative experiments until the reasonable match between ML prediction and experiments. Adopting active ML methods can reduce the number of experiments effectively, tens of times faster than the exhausted exploration of whole composition range. At last we constructed “composition-property” and “feature-property” relations or performed the importance analyses of features, transiting from “machine learning” to “learning from machine” that enables the data-driven design followed by the knowledge-driven design at the higher level. Based on the deep understanding of structure-property relation of materials, we can guide the property optimization and predictive design of new materials. This work demonstrated that the high-throughput experiments guided by machine learning (HTE-ML) may become a general acceleration strategy of composition optimization and design of multi-component materials.

Keywords: High-throughput experiments；Machine learning；Data-driven; Alloy design

* Corresponding author: yiliu@shu.edu.cn