摘要
预报钢材的力学性能对其研发和应用都具有重要作用.本文对360条源自日本国立材料研究所(NIMS)数据库的钢材数据,使用正则化线性回归、随机森林和人工神经网络三种机器学习方法对钢材的疲劳强度、拉伸强度、断裂强度和硬度四种力学性能进行多目标机器学习,即使用一组特征变量来描述这四种目标函数,并使用MultiTaskLasso、特征重要性和复相关系数三种方法对特征变量的重要程度进行了排序,也用"面积法"来平衡模型复杂度(特征个数)与预测精度,且通过线性回归给出了基于化学成分、制备工艺参数、夹杂物参数的解析表达式,对四种力学性能具有较好的预测能力.作为对比,本文还使用原子半径、价电子数和鲍林电负性等9种原子尺度特征替代9种元素特征后进行多目标机器学习建模和预测,计算结果显示原子尺度特征同样具备较好的预测能力.
The prediction of mechanical properties of steels is highly significant for the development and applications of steels. In the present work, we used three multi-objective machine learning algorithms, Regularized linear regression, Random forest and Artificial neural network, to predict four mechanical properties of steels, including fatigue strength, tensile strength, fracture strength and hardness,based on 360 data samples from the Japan National Institute of Material Science(NIMS) database. Three methods, Multi Task Lasso,feature importance and Multiple correlation, were applied to rank features. The "Area method" was proposed to balance the model complexity(number of features) and prediction accuracy. The linear regression, with selected features, of the entire data leads to four analytic formulas, which predict very accurately the four mechanical properties of steels. Furthermore, nine atomic features, such as atomic radius, valence electron number, and Pauling electronegativity, of the involved elements in the steels were replaced the elements in the multi-objective machine learning and the results show that the atomic features are able to accurately predict the four mechanical properties of steels as well.
作者
魏清华
熊杰
孙升
张统一
WEI QingHua;XIONG Jie;SUN Sheng;ZHANG TongYi(Materials Genome Institute,Shanghai University,Shanghai 200444,China;Department of Mechanical Engineering,The Hong Kong Polytechnic University,Hong Kong 999077,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2021年第6期722-736,共15页
Scientia Sinica(Technologica)
基金
国家重点研发计划(编号:2018YFB0704400)
国家自然科学基金(批准号:12072179)
高等学校学科创新引智计划(编号:D16002)
香港理工大学(编号:1-ZE8R和G-YBDH)资助项目。
