林轩杰, 江汉同, 李倩, 周玉玲, 臧怀娟, 任永生(通讯作者), 詹曙, 马文会. 基于机器学习的材料弹性性能预测及可视化分析(中国有色金属学会第十四届学术年会推荐论文)[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2023.08.10.003
引用本文: 林轩杰, 江汉同, 李倩, 周玉玲, 臧怀娟, 任永生(通讯作者), 詹曙, 马文会. 基于机器学习的材料弹性性能预测及可视化分析(中国有色金属学会第十四届学术年会推荐论文)[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2023.08.10.003
Prediction of Elastic Properties of Materials Based on Machine Learning and Visualization Analysis[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2023.08.10.003
Citation: Prediction of Elastic Properties of Materials Based on Machine Learning and Visualization Analysis[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2023.08.10.003

基于机器学习的材料弹性性能预测及可视化分析(中国有色金属学会第十四届学术年会推荐论文)

Prediction of Elastic Properties of Materials Based on Machine Learning and Visualization Analysis

  • 摘要: 在工程材料的应用中,弹性模量是重要的性能参数,找到特定弹性性能的材料是新材料合成领域的热点问题,如何快速且准确的预测弹性在工程上具有重要意义。通过实际实验测量大量材料的弹性性能并不现实。因此,通过计算机模拟筛选材料数据,选出候选材料,再通过实际实验进行验证,是一种理想的新材料发现方法。目前材料性能预测的主要计算方法是基于第一性原理的高通量计算,这类方法效率低下,难以高效地完成大批量的材料筛选任务。而基于材料统计学的机器学习预测方法,可通过大数据挖掘,快速预测材料性能,成为一种有可能替代高通量计算的方案。本文将特征选择方法和机器学习模型进行组合,从中选择最有效的弹性模量预测组合方案,并设计交互界面对输入特征和材料弹性性能的关系进行可视化分析。实验表明Pearson/RFE和GBDT的组合模型性能最好,同时通过可视化分析发现每原子能量、熔点、密度等特征对于预测结果的影响较大。这些重要的特征可以从特征-目标关系中初步预测弹性模量的范围,目标属性值也可反过来估计材料的重要特征。这些研究成果可应用于探究弹性的影响因素、预测大批量材料性能和可视化分析指导材料合成。

     

    Abstract: In the application of engineering materials,the elastic modulus is an important performance parameter. Finding the materials with specific elastic properties is a hot issue in the field of new materials synthesis. How to predict the elasticity quickly and accurately is of great significance in engineering. It is not practical to measure the elastic properties of a large number of materials by practical experiments. Therefore,screening material data through computer simulation,selecting candidate materials,and then verifying them through actual experiments is an ideal method for new material discovery. At present,the main calculation methods for material performance prediction are first-principles high-throughput calculation,which is inefficient and difficult to ? complete the task of high-volume material screening efficiently. The machine learning prediction method based on material statistics can quickly predict material properties through big data mining,which has become a possible alternative to high-throughput computing. In this paper,feature selection method and machine learning model are combined to select the most effective combination scheme of elastic modulus prediction,and interactive interface is designed to conduct visual analysis of the relationship between input features and elastic properties of materials. The experiment shows that the Pearson/RFE-GBDT combination model has the best performance. Meanwhile,through the visualization analysis,it is found that the energy of each atom,melting point,density and other characteristics have a great influence on the prediction results. These important characteristics can be used to preliminarily predict the range of elastic modulus from the feature-target relationship,and the value of target attributes can be used to estimate the important characteristics of the material in turn. These results can be applied to explore the influencing factors of elasticity,predict the properties of large quantities of materials and guide the synthesis of materials through visualization analysis.

     

/

返回文章
返回