一种探索高熵合金相形成的端到端机器学习框架  

An end-to-end machine learning framework exploring phase formation for high entropy alloys

作　　者：张惠然[1,2,3] 胡瑞 刘茜 李盛洲 张光捷 钱权 丁广太[1] 戴东波[1] Hui-ran ZHANG;Rui HU;Xi LIU;Sheng-zhou LI;Guang-jie ZHANG;Quan QIAN;Guang-tai DING;Dong-bo DAI(School of Computer Engineering and Science,Shanghai University,Shanghai 200444,China;Materials Genome Institute,Shanghai University,Shanghai 200444,China;Zhejiang Laboratory,Hangzhou 311100,China;Department of Computer Science,University of Tsukuba,Tsukuba,Ibaraki 305-8573,Japan)

机构地区：[1]上海大学计算机工程与科学学院,上海200444 [2]上海大学材料基因研究院,上海200444 [3]之江实验室,杭州311100 [4]Department of Computer Science,University of Tsukuba,Tsukuba,Ibaraki 305-8573,Japan

出　　处：《Transactions of Nonferrous Metals Society of China》2023年第7期2110-2120,共11页中国有色金属学报（英文版）

基　　金：sponsored by the National Key Research and Development Program of China(No.2018YFB0704400);Key Research Project of Zhejiang Laboratory,China(No.2021PE0AC02);Key Program of Science and Technology of Yunnan Province,China(Nos.202002AB080001-2,202102AB080019-3);Key Project of Shanghai Zhangjiang National Independent Innovation Demonstration Zone,China(No.ZJ2021-ZD-006)。

摘　　要：探索高熵合金(HEAs)的相形成规则对于新型合金的设计具有明确的指导意义。提出一种端到端的框架用来从特征池和模型池中分别选择特征子集和机器学习(ML)模型。在该框架中,模型池中的模型基于其获得的特征重要性来选择适合自身的特征子集;通过评估每个模型和其对应的特征子集的拟合结果,用于建立目标任务的预测模型;最终,获得影响HEAs相形成的重要因素。研究结果显示,建立的相预测模型可将430种HEAs分成5种相,测试准确度达到87.8%,并且通过分析模型发现,当原子尺寸差异大于8.295%时,HEAs的单相固溶体的形成受到抑制。Exploring the rules of high entropy alloys(HEAs)phase formation has clear guiding significance for the design of new alloys.An end-to-end framework was proposed to select the feature subset and machine learning(ML)model from the feature pool and model pool,respectively.In this framework,each model in the pool is to determine its materials feature subset based on the feature importance.The final model was confirmed through the evaluation of the fitting result of every model and its feature subset.This method extracts important factors affecting the phase formation of HEAs.The results show that the chosen model could classify 430 HEAs into five phases,with test accuracy of 87.8%.And the model analysis suggests that the formation of single-phase solid solution is often inhibited when the atomic size difference is greater than 8.295%.

关 键 词：特征选择 高熵合金 机器学习 相预测 Hume-Rothery规则 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TG139[自动化与计算机技术—控制科学与工程]