Page 49 - 《爆炸与冲击》2026年第5期

第 46 卷 第 5 期 爆 炸 与 冲 击 Vol. 46, No. 5
2026 年 5 月 EXPLOSION AND SHOCK WAVES May, 2026

DOI：10.11883/bzycj-2025-0324

单晶金属中微孔洞生长过程的深度学习预测方法 *

苏 浩 ，赵雷洋 ，丛龙跃 ，陈 聪 ，关添元 ，刘 岩 1,3
1,2
1,3
1,3
1,2
4
（1. 清华大学航天航空学院，北京 100084；
2. 北京宇航系统工程研究所，北京 100076；
3. 清华大学空间高效能推进技术及应用全国重点实验室，北京 100084；
4. 北京理工大学（珠海）智能制造技术研究中心，广东 珠海 519088）
摘要： 针对单晶金属中微孔洞生长过程的预测问题，建立了一种基于 U-Net 和 Transformer 的深度神经网络模型：
基于包含初始椭球双孔洞的单晶铜原子模型的分子动力学模拟结果构建数据集；提出了一种基于背景网格的数据预
处理方法，在数据集中对模拟结果进行局部统计。算例结果表明，上述深度学习方法能够对单晶金属中微孔洞生长过
程中的整体物理量和局部细节信息进行准确预测。
关键词： 深度学习；微孔洞生长；分子动力学；U-Net；Transformer
中图分类号： O341 国标学科代码： 13015 文献标志码： A

A deep learning prediction method for growth of
micro voids in single-crystal metal

4
1,3
1,2
1,3
1,2
SU Hao , ZHAO Leiyang , CONG Longyue , CHEN Cong , GUAN Tianyuan , LIU Yan 1,3
（1. School of Aerospace Engineering, Tsinghua University, Beijing 100084, China;
2. Beijing Institute of Astronautical Systems Engineering, Beijing 100076, China;
3. State Key Laboratory of Advanced Space Propulsion, Tsinghua University, Beijing 100084, China;
4. Research Center of Intelligent Manufacturing Technology, Beijing Institute of Technology (Zhuhai),
Zhuhai 519088, Guangdong, China）

Abstract: A novel deep neural network was proposed to predict the growth of micro voids in single-crystal metal based on U-
Net and Transformer in this paper. The dataset was constructed through molecular dynamics (MD) simulation results of a
single-crystal copper atom model with initial double ellipsoidal voids. A data preprocessing scheme based on background mesh
was proposed to perform local statistics on the simulation results. The information obtained from simulation results, such as
void morphology, dislocation distribution, and von Mises effective stress, was converted into local statistics on the background
mesh. These statistics were then converted into pixel matrix format as the input of the deep neural network. Multiple data
samples can be generated from the results of one single MD simulation, which significantly reduces the computational
resources required for dataset generation. The samples encompass typical stages of the void growth, which enables the network
to capture key features and to facilitate data augmentation conveniently. The deep neural network model consists of four parts:
U-Net composed of down-sampling and up-sampling networks, a generation model, a Query network model, and a regression
prediction network. The model input includes both physical information and positional information. The output is the predicted
physical information for the next time step. The loss function is a superposition of loss functions for each predicted variable.

* 收稿日期： 2025-09-29；修回日期： 2026-03-17
基金项目： 国家自然科学基金（12572228）
第一作者： 苏 浩（1994－ ），男，博士，工程师，suhao9406@163.com
通信作者： 刘 岩（1978－ ），男，博士，长聘副教授，yan-liu@tsinghua.edu.cn

051423-1