Page 40 - 《爆炸与冲击》2026年第3期

第 46 卷 第 3 期 爆 炸 与 冲 击 Vol. 46, No. 3
2026 年 3 月 EXPLOSION AND SHOCK WAVES Mar., 2026

DOI：10.11883/bzycj-2025-0339

基于人工神经网络的金属材料本构模型

在显式有限元中的实现 *

1
2
康正东 ，王少喆 ，苏步云 ，康佳鑫 ，邱 吉 ，树学峰 1
1
1
1
（1. 太原理工大学航空航天学院，山西 太原 030024；
2. 山西建筑工程集团有限公司，山西 太原 030002）
摘要： 以 CoCrFeNiMn 高熵合金为研究对象，首先开展了不同温度与应变率下的压缩实验，获得了应力-应变数
据；随后基于实验结果建立了修正的 Johnson-Cook 本构模型，并用于有限元仿真生成机器学习训练数据。在此基础上
构建人工神经网络（artificial neural network, ANN）模型，对材料流动应力进行学习与预测。为实现神经网络在有限元框
架中的高效应用，开发了基于 FORTRAN 的自动代码生成工具，将训练完成的 ANN 模型嵌入到 Abaqus/Explicit 计算平
台中。结果表明，该方法预测精度高，相对误差低于 1%，且计算效率优于传统本构模型。基于数据驱动的神经网络方
法可有效替代传统本构模型在有限元中的应用，为金属材料的数值建模与模拟提供了一条有效路径。
关键词： 机器学习；人工神经网络；本构模型；CoCrFeNiMn 高熵合金；有限元方法；数值实现
中图分类号： O347.3; TP181 国标学科代码： 13015 文献标志码： A

Implementation of metallic material constitutive models based on artificial
neural networks in explicit finite element analysis

1
1
1
1
2
KANG Zhengdong , WANG Shaozhe , SU Buyun , KANG Jiaxin , QIU Ji , SHU Xuefeng 1
（1. College of Aeronautics and Astronautics, Taiyuan University of Technology, Taiyuan 030024, Shanxi, China;
2. Shanxi Construction Engineering Group Co., Ltd., Taiyuan 030002, Shanxi, China）
Abstract: Machine learning techniques have been increasingly applied to the prediction of material behavior and have
demonstrated clear advantages over conventional constitutive modeling approaches. The objective of this study was to develop
an accurate and computationally efficient data-driven constitutive description for metallic materials under coupled temperature
and strain-rate loading conditions. A CoCrFeNiMn high-entropy alloy was selected as the representative material
system.Compression experiments were performed over a wide range of temperatures and strain rates to obtain true stress-strain
data. Based on the experimental results, a modified Johnson-Cook constitutive model was calibrated to describe strain
hardening, strain-rate sensitivity, and thermal softening effects. The calibrated model was then implemented in finite element
simulations to generate a large, physically consistent dataset spanning broad thermo-mechanical conditions. This simulation-
assisted data generation strategy expanded the training domain while ensuring continuity and stability of the dataset. Using the
generated data, an artificial neural network (ANN) model was constructed to learn the nonlinear relationship between strain,
strain rate, temperature, and flow stress. The network architecture and training strategy were optimized to improve prediction
accuracy and generalization performance. To enable efficient application of the trained ANN within an explicit finite element
framework, an automatic FORTRAN code generation tool was developed. The trained ANN parameters were converted into a

* 收稿日期： 2025-10-11；修回日期： 2025-12-24
基金项目： 国家自然科学基金（13202477，12272256）
第一作者： 康正东（2000－ ），男，硕士研究生，18726217603@163.com
通信作者： 邱 吉（1992－ ），男，博士，副教授，qiuji@tyut.edu.cn
树学峰（1964－ ），男，博士，教授，shuxuefeng@tyut.edu.cn

031403-1