Deep Learning-Based Framework for Efficient Design of Multicomponent High Hardness High Entropy Alloys

Machine learning (ML) has introduced a paradigm for researching high-entropy alloys (HEAs). However, a significant challenge in alloy design lies in finding the optimal balance between exploring HEAs systems and ensuring the reliability of their performance. Traditional experience-based methods for alloy system development limit the discovery of alloys, while purely data-driven approaches often struggle to guarantee the practical performance of the designs. We proposed a deep learning-based framework that integrates materials domain knowledge with data-driven techniques to optimize the design process for multicomponent, high-hardness HEAs to address this issue. A material concatenation embedding (MCE) module was first developed and integrated with a BiLSTM-CRF network to automate the analysis of 2698 papers published over the past 5 years, extracting 8067 data points. By incorporating materials domain knowledge into the data analysis, we identified high-potential elements and key processing conditions to guide the design and construction of the machine learning data set. After manually summarizing and organizing the target literature, we constructed a hardness data set of 13 elements. A two-stage design strategy for multicomponent HEAs was developed using a combination of genetic algorithm (GA) and particle swarm optimization (PSO). The first stage explores alloy systems, while the second refines composition proportions, facilitating both innovation and performance enhancement. Our analysis incorporated SHAP feature importance and Pearson correlation coefficients (PCC), complemented by materials domain knowledge, to validate the findings and guide alloy system selection. We successfully designed three HEAs that differ from those in existing data sets: Cr 20.6 Fe 22.5 Mo 20.6 Ti 18.3 V 18, Al 9.32 Cr 20.62 Fe 21.71 Mo 27.09 Ti 21.26, and Al 6 Cr 20.3 Fe 19.5 Mo 20.1 Nb 18.8 Ti 15.3 . The predicted average relative error in hardness is under 5%, and the hardness of the optimal alloy is only 38 HV lower than the historical record.