基于机器学习的轻质低膨胀幕墙玻璃组分设计研究

Composition Design of Light and Low Expansion Glass Curtain Walls Based on Machine Learning

随着机器学习技术的不断发展和玻璃材料数据的逐步积累,基于数据驱动的组分设计方法已成为玻璃新材料开发的一种有力手段。本文采用随机森林回归算法构建了包含56种氧化物的玻璃组分与性能预测模型,并采用SHAP分析等方法进行了可解释性研究,实现了在高维组分空间对线膨胀系数、密度及弹性模量的准确预测。利用该预测模型在Si-Al-B-Ca-Mg-Na六元氧化物组分空间中对约118万个玻璃配方进行快速预测,并对优选的4组硼硅酸盐玻璃样品进行测试。结果表明,样品的线膨胀系数分布在(52.00~58.00)×10^(-7)℃^(-1),密度分布在2.34~2.39 g/cm^(3),弹性模量分布在67.00~74.00 GPa,与模型预测结果相符,且优于相关规范要求。

The data-driving method for glass composition design has become a desirable way to create novel glass materials,thanks to the ongoing development of machine learning algorithms and the progressive collection of glass materials.In the present work,the machine learning approach based on random forest regression algorithm was used to develop prediction model between the composition and performance of glass materials which contains 56 different oxides.In addition,the interpretability was studied by SHAP analysis.The accurate prediction of linear expansion coefficient,density,and elastic modulus in high-dimensional composition space was realized.The obtained models were utilized to forecast about 1.18 million Si-Al-B-Ca-Mg-Na six-component oxide glass composition swiftly.Four preferred borosilicate glass samples were tested.The results show that linear expansion coefficient of sample ranges from 52.00×10^(-7)℃^(-1) to 58.00×10^(-7)℃^(-1),the density ranges from 2.34 g/cm^(3) to 2.39 g/cm ^(3),and the elastic modulus ranges from 67.00 GPa to 74.00 GPa,respectively.These values are consistent with the predicted values and outperform the applicable standards.