机器学习辅助的高通量实验加速硬质高熵合金CoxCryTizMouWv成分设计

Machine Learning Assisted High-Throughput Experiments Accelerates the Composition Design of Hard High-Entropy Alloy CoxCryTizMouWv

针对目标性能的多元合金成分设计因具有巨大的成分参数空间而极具挑战,而且传统的试错实验由于效率低能探索的合金成分有限。提出利用高通量实验结合机器学习方法加速非等摩尔比的硬质高熵合金Co x Cr y Ti z Mo u W v的成分设计。首先通过自主研发的全流程高通量合金制备系统制备了138个不同成分的高熵合金铸态样品。然后根据测量的维氏硬度(HV)数据,使用随机森林法和支持向量机法进行机器学习建模,并预测了五元合金体系内潜在的3876个不同成分合金的硬度。随机森林机器学习模型的预测结果在高(HV>800 MPa)、中(600<HV<800 MPa)、低(HV<600 MPa)硬度区域的平均误差分别为2.87%,3.30%和6.70%,实验硬度值在对应区域的测量误差分别为1.69%,1.88%和1.87%。根据机器学习模型预测结果建立的“成分-硬度”与“描述因子-硬度”关系图谱展示了全成分空间内高熵合金的硬度变化规律及影响硬度的重要描述因子——原子半径差。研究结果表明,高通量实验与机器学习相结合可使多元合金成分优化效率提高百倍以上。此外,建议未来研究应在“机器学习”基础上加强“向机器学习”,在更高层次上获得新的专业知识认知。

The composition design of multi-component alloy for the target performance is extremely challenging due to the enormous potential composition.The traditional trial-and-error experiments can only explore limited alloy compositions because of its low efficiency.In this work,the composition design of non-equimolar hard high-entropy alloy Co x Cr y Ti z Mo u W v was accelerated via combining the high-throughput experiment with machine learning.Firstly,138 as-cast high-entropy alloys were prepared by a home-developed all-process high-throughput alloy synthesis system.Then,the machine learning models were built based on the measured Vickers hardness(HV)by using random forest(RF)and supporting vector machine methods.And,they made the prediction of HV values for 3876 potential alloys in the five-component alloy system.The HV values predicted by RF machine learning models have the averaged errors of 2.87%,3.30%and 6.70%,respectively in high(HV>800 MPa),medium(600<HV<800 MPa),and low(HV<600 MPa)hardness regions.The measurement errors of experimental data at the corresponding hardness regions were 1.69%,1.88%and 1.87%,respectively.The“composition-hardness”and“descriptor-hardness”relationship maps,constructed by using the prediction of machine learning models,show the variation tendency of alloy hardness at the full composition space and the important descriptors to alloy hardness,namely atom size difference.This work demonstrates that the high-throughput experiments combined with machine learning can accelerate the composition design of multi-component alloy more than hundred times.Additionally,future research should emphasize on“learning from machine”based on“machine learning”to acquire the new domain knowledge at a higher level.