单向复合材料均质化弹性性能预测的集成学习模型:基于SHAP方法的可解释性分析  

Ensemble machine learning for predicting the homogenized elastic properties of unidirectional composites:A SHAP-based interpretability analysis

作　　者：王文照 赵云妹 李岩 Wenzhao Wang;Yunmei Zhao;Yan Li(School of Aerospace Engineering and Applied Mechanics,Tongji University,Shanghai 200092,China)

机构地区：[1]School of Aerospace Engineering and Applied Mechanics,Tongji University,Shanghai 200092,China

出　　处：《Acta Mechanica Sinica》2024年第3期109-122,共14页力学学报（英文版）

基　　金：supported by the National Key Research and Development Program of China (Grant No.2022YFB4602000);Liaoning Open Competition Science and Technology Major Project (Grant No.2022JH1/10400043);the National Natural Science Foundation of China (Grant No.12302184);Shanghai Pujiang Talent Program (Grant No.22PJ1413800).

摘　　要：本研究提出了一种可解释的集成学习(EML)数据驱动方法,用于快速预测单向纤维增强复合材料的均质化弹性性能.我们选择了三种机器学习模型――随机森林(RF)、极端梯度提升机(XGBoost)和轻量梯度提升机(LGBM)来构建EML模型.基于实验并建立微观纤维分布的概率统计模型,结合有限元仿真,建立了反映纤维随机分布特征的代表性体元(RVE)模型,并创建相关数据集.通过准确性、效率、可解释性和泛化性等指标对EML模型的性能进行了全面评估,研究结果表明:(1)EML模型有效提升了基机器学习模型的预测精度(R2=0.962,MS E=5.41);(2)我们使用基于合作博弈理论的SHAP可解释性方法对预测结果进行了分析,其中,全局解释发现纤维的体积含量是影响复合材料均质化弹性性能的决定性变量,而局部解释阐明了输入特征对预测结果的关键影响机制;(3)EML模型在实验数据上表现出良好的泛化能力,计算结果与实验数据吻合度高,实现了复合材料均质化弹性性能的高效精准预测.This study aims to develop an interpretable ensemble machine learning(EML)method for predicting the homogenized elastic properties of unidirectional fiber-reinforced composite.Three machine learning models—Random Forest(RF),eXtreme Gradient Boosting Machine(XGBoost),and Light Gradient Boosting Machine(LGBM)are selected for constructing the EML model.The ground-truth dataset is created via knowledge-based finite element(FE)simulations with representative volume element(RVE)models of composites considering randomly dispersed fibers.We thoroughly evaluate the EML model’s performance using the metrics of accuracy,efficiency,interpretability,and generalization.The obtained results indicate that:(1)the EML model outperforms the base ML model with high precision(R2=0.962,MSE=5.41);(2)the SHapley Additive explanations(SHAP)based on cooperative game theory is used to interpret the predictions,with the global interpretations identifying fiber volume content as the most influential variable on the composite and the local interpretations examining the key influencing mechanisms of each feature;(3)the EML model demonstrates good generalization ability on experimental data,and it can accurately forecast the homogenized properties of the composite with genuine fibers.

关 键 词：Composite materials HOMOGENIZATION Ensemble learning INTERPRETATION Model generalization 

分 类 号：TP33[自动化与计算机技术—计算机系统结构]