摘要
硬度是工模具钢性能的一个重要指标.本文通过机器学习中的层次聚类和LASSO回归方法,给出了工模具钢硬度-成分关系的解析表达式.层次聚类将79种牌号工模具钢的合金元素成分-硬度数据集按欧式距离分为了4簇,其中高铬钢和低铬钢两簇包含了几乎所有的数据.然后我们对高铬钢和低铬钢两簇数据分别进行LASSO回归,画出LASSO路径,并利用留一法交叉验证得到具有最佳泛化预测能力的成分-硬度的解析公式.之后引入电负性、原子半径变化率、价电子数、电子亲合能和第一电离能等原子尺度的特征,重新进行LASSO回归,得到了新的硬度公式.结果显示机器学习方法可以成功预报工模具钢的硬度.同时,使用原子尺度的特征为钢铁材料的性能研究提供了新的思路.
Hardness is a major indicator of the quality of tool and mold steels(TMS). Analytic formulas of hardness versus compositions were proposed here by using hierarchical clustering(HC) and LASSO regressions, based on data of 79 brands of TMS. HC presents two large groups of TMS, which contain high and low concentrations of chromium. Then, LASSO regressions were applied on each group and the regression result with the lowest root mean square error in leave-one-out cross-validation was taken out as the theoretical prediction formula of hardness versus chemical composition feature. Furthermore, atomic features were selected from electronegativity, atomic radius, valence electron number, electron affinity, first ionization energy, etc. LASSO regressions of the data with the atomic features give another prediction formula. These results demonstrate the powerful ability of machine learning in the design of TMS.
作者
王家豪
孙升
何燕霖
张统一
WANG JiaHao;SUN Sheng;HE YanLin;ZHANG TongYi(Materials Genome Institute,Shanghai University,Shanghai 200444,China;School of Materials Science&Engineering,Shanghai University,Shanghai 200444,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2019年第10期1148-1158,共11页
Scientia Sinica(Technologica)
基金
国家重点研发计划(编号:2017YFB0701604,2018YFB0704400)资助项目
关键词
机器学习
工模具钢
硬度预测
machine learning
tool and mold steels
hardness prediction
