基于人工神经网络的混凝土类材料SHPB动态压缩性能预测  

Prediction of Dynamic Compressive Performance of Concrete-Like Materials Subjected to SHPB Based on Artificial Neural Network

作　　者：龙旭[1] 毛明晖 卢昶衡 苏天雄 贾冯睿 LONG Xu;MAO Minghui;LU Changheng;SU Tianxiong;JIA Fengrui(School of Mechanics,Civil Engineering and Architecture,Northwestern Polytechnical University,Xi’an 710072,China;School of Civil Engineering,Liaoning Petrochemical University,Fushun 113001,China;Yangtze Delta Region Institute of Tsinghua University,Jiaxing 314006,China)

机构地区：[1]西北工业大学力学与土木建筑学院,西安710072 [2]辽宁石油化工大学土木工程学院,抚顺113001 [3]浙江清华长三角研究院,嘉兴314006

出　　处：《南京航空航天大学学报》2021年第5期789-800,共12页Journal of Nanjing University of Aeronautics & Astronautics

基　　金：陕西省国际科技合作计划(2021KW-25)资助项目;航天科学技术基金(2021-HT-XG)资助项目。

摘　　要：针对混凝土类脆性材料高应变率下本构行为,结合ABAQUS有限元仿真与反向传播(Back propagation,BP)人工神经网络技术,对分离式霍普金森压杆(Split Hopkinson pressure bar,SHPB)实验过程中关键波形参数进行仿真和机器学习,建立了混凝土类材料SHPB高应变率下力学性能预测的机器学习模型,极大地提升了复杂脆性材料受冲击状态下变形行为与本构参数之间关联机制的计算效率。利用商业有限元软件ABAQUS的动态分析模块,通过在入射杆自由面设置4种不同的应力波,得到在不同应变率下材料应力-应变曲线,通过对比数值模拟结果和SHPB实验,验证了基于有限元分析的计算结果准确性。以20组ABAQUS仿真结果作为训练样本,其中入射波作为输入层,透射波和反射波作为输出层,建立相应的机器学习预测模型。研究结果表明:基于BP人工神经网络技术的机器学习预测模型具有良好的适用性,可代替量大且耗时的有限元仿真建模、分析及后处理流程,实现了高应变率下混凝土类材料应力-应变曲线形式本构行为的高效准确预测,同时可以预测给定训练样本以外更大应变率范围下材料应力-应变曲线。Regarding constitutive behavior of concrete-like brittle materials subjected to high strain rates,we combine ABAQUS finite element simulations and back propagation(BP)artificial neural network method to analyze the critical waveform parameters in the split Hopkinson pressure bar(SHPB)experiments,and propose the machine-learning based methodology for predicting mechanical properties of concrete-like materials under high strain rates.This model significantly improves the computational efficiency to reveal the correlation mechanisms between deformation behavior and constitutive parameters of complex brittle materials under impact loading.We adopt the dynamic analysis module of ABAQUS finite element software to apply four different stress waves on the free surface of the incident bar to obtain the stress-strain curve of materials under different strain rates.By comparing with SHPB experimental data,the accuracy of numerical predictions from finite element simulations is validated.20 sets of ABAQUS simulation results are exploited as training samples,in which the incident wave is used as the input layer while the transmitted and reflected waves are taken as the output layer.The results show that the machine-learning prediction model based on BP artificial neural network method owns satisfactory generality,and this proposed method could replace repetitive finite element modeling to considerably save the time for model creation,analysis and post-analysis process.It can accurately predict the constitutive behavior in the form of stress-strain curve for concrete-like materials under high strain rates,and can also predict the stress-strain responses under a wider range of strain rate beyond those provided from training samples.

关 键 词：分离式霍普金森压杆 机器学习 混凝土 动态力学性能 BP人工神经网络 

分 类 号：TU375[建筑科学—结构工程]