基于深度学习的超高速碰撞碎片云生成模型

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。

Application of Deep Learning in Debris Cloud Simulation

With the rapidly development of deep learning technology, various models based on data-driven have been widely studied and used in Computational Solid Mechanics and computational fluid dynamics. Based on the deep learning method, this project proposes a data-driven model for debris cloud generation. By the combination of Conditional Variation-Auto-Encoder model and numerical simulation results of SPH method, a deep learning model for simulating hypervelocity impact debris cloud was constructed. Before training deep learning model, some data preprocessing steps are needed to improve the data distribution law, such as spatial grid division and quality aggregation, which are conducive to improving the training speed and generalization performance. Experimental results show that the deep learning model can make a good prediction and interpolation in the range of training data set and also showed this ability just near training data. It provided to be a potential way to realize the comprehensive utilization of existing experimental data and numerical simulation results. And also maybe a potential way to improve the accuracy of debris cloud model.