基于深度学习的超高速碰撞碎片云模拟  

Application of Deep Learning in Debris Cloud Simulation

作　　者：文永 张浩[1] 李毅[2] 褚新坤 田志宇 张庆[1] Yong Wen;Hao Zhang;Yi Li;Xinkun Chu;Zhiyu Tian;Qing Zhang(Institute of Computer Application,CAEP,Mianyang,621900;China Aerodynamics Research and Development Center,Mianyang,621900)

机构地区：[1]中国工程物理研究院计算机应用研究所,绵阳621900 [2]中国空气动力研究与发展中心,绵阳621900

出　　处：《固体力学学报》2020年第5期455-469,共15页Chinese Journal of Solid Mechanics

基　　金：中国空气动力研究与发展中心超高速碰撞研究中心开放基金项目(HIRC201804)资助。

摘　　要：数据驱动的模型已经被广泛研究,并成功应用到了计算力学.基于深度学习技术,提出一种新的采用数据驱动的碎片云生成模型.此模型可以学习SPH数值模拟结果,然后在多种控制条件下快速生成碎片云.在模型训练前的数据预处理阶段,对SPH模拟结果进行空间网格划分和质量聚合,实现了改善数据分布规律、加速模型训练和提升模型泛化性的目的.以高速靶球撞击薄壁圆筒后的碎片云质量分布为例,模拟并测试了多种控制条件下深度学习模型计算结果的正确性和稳定性,以及计算速度的高效性.实验证明,深度学习模型可以从训练集学习碎片云的物理规律,然后在训练集控制参数范围内进行良好的推理及插值;并且可以在训练数据集控制参数范围外,进行小范围推理预测;同时深度学习模型的计算速度远快于SPH方法.通过深度学习方法建立碎片云模型,可能是一种在空间飞行器防护结构原型设计阶段,实现碎片云实时生成的潜在方案.Usually it can take up to several hours to solve the hypervelocity impact debris cloud distribution problem using the traditional numerical simulation methods. However, in industrial design, the work of prototype design stage needs fast iteration, so it is necessary to find a more efficient simulation method. Based on deep learning technology, the debris cloud generation model driven by data would be a potential method for the real-time simulation of hypervelocity impact. In this method, the conditional variational autoencoder(CVAE) model is used to learn the debris distribution law from existing SPH simulation results, and obtain a deep learning model for debris cloud generation. Firstly, data preprocessing is carried out before model training. The results of SPH simulation are divided into spatial grids, and the mass of debris is aggregated by grids. Data preprocessing could achieve three goals: smoothing data distribution, speeding up model training process, and improving model generalization capability. Secondly, the CVAE model is constructed based on the convolutional and deconvolutional neural network layer, and the projectile incident condition is used as the control condition of the CVAE model. Thirdly, the model training task is run. When the errors of training set and validation set are both small, the model is saved as a debris cloud generation model. The performance of the deep learning model is verified by an example of a high-speed ball impacting on a thin-walled cylinder. The accuracy, stability and efficiency of the model under various control conditions are simulated and tested. Experimental results show that the trained deep learning model can learn the mass distribution of debris cloud according to the training data set, and make a reasonable inference within the control parameter range of the training dataset. Besides, it can provide small-range prediction around the training set’s control parameter. On top of that, the computational speed of the deep learning model is much faster than the S

关 键 词：碎片云 条件变分自编码器 深度学习 人工智能 卷积神经网络 数据驱动 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论] TP18[自动化与计算机技术—计算机科学与技术]