数据驱动的钢铁耐磨材料性能预测研究综述

Review on Data-driven Method for Property Prediction of Iron and Steel Wear-resistant Materials

数据驱动方法利用机器学习算法挖掘数据中隐藏的规则,是一种符合“第四范式”的研究方法。该研究方法的开展基于大量材料基础数据。通过对比国内外材料基础数据平台,分析利用现有数据平台已开展的研究,指出钢铁耐磨材料基础数据存在数据匮乏和缺乏统一采集标准两个问题。针对此,介绍符合材料基因组计划的数据采集标准,并给出钢铁耐磨材料专用数据平台的框架以及数据来源。分析钢铁耐磨材料性能的影响因素,讨论各种特征选择技术的特点。回顾在材料科学研究中成功应用的几种机器学习算法,分析每种算法的应用场景,讨论它们的优缺点,并对算法性能进行了比较。最后总结一些建议为特征提取和机器学习算法选择提供指导,并指出数据驱动方法在性能预测、发现新材料和自动化自主试验等方面具有良好的应用前景。

Data-driven method utilizes machine learning(ML)to mine hidden rules in data,conforming to the"fourth paradigm".A great deal of basic data is needed for this method.By comparing the domestic and aboard materials basic data platforms and analyzing researches based on these platforms,there are two problems:lack of data and lack of unified acquisition standard.In view of this,the data acquisition standard in line with materials genome initiative(MGI)is introduced.And the framework and sources of data platform specially for iron and steel wear-resistant materials are given.The factors affecting the properties of iron and steel wear-resistant materials are analyzed,and the characteristics of various feature selection techniques are discussed.Then several ML algorithms,applied in material science researches successfully,are reviewed.The application scenarios of each algorithm are analyzed,the relative merits of them are discussed,and their performances are compared.Finally,some suggestions are summarized to provide guidance on how to choose feature selection methods and ML algorithms.It is pointed out that the data-driven method has a good application prospect in property prediction,new material discovery and automatic experiment.