Mapping microstructure to shock-induced temperature fields using deep learning

Chunyu Li, Juan Carlos Verduzco, Brian H. Lee, Robert J. Appleton & Alejandro Strachan

npj Computational Materials 9: 178 (2023).

编辑概述: 深度学习预测材料的冲击温度场

材料对冲击载荷的响应对行星科学、航空航天工程和高能材料非常重要。热激发过程如化学反应和相变在能量集中处会发生显著的加速。这是由冲击波与材料微观结构相互作用产生的结果，并受复杂的耦合过程控制，而其中的过程控制机制尚未被完全了解。这些过程大多发生在温度、压力和应变速率的极端条件下，而且其中各种能量局部集中和微观结构特征存在于不同长度和时间尺度。因此，现有的模型都无法在没有强近似假设的情况下预测冲击诱导的热点形成。分子动力学（MD）已被广泛用于研究激波诱导的热点形成，包括孔隙率的坍塌、剪切、摩擦和局部塑性变形，但是MD方法需要巨大的计算成本。同时，深度学习已经被用于模拟材料在冲击载荷下的中尺度热机械响应，其精度与基于物理的模拟相当，但只需要的计算成本相对而言非常小。在本工作中，来自普渡大学材料工程学院的Alejandro Strachan教授等人，基于UNet网络结构设计了冲击诱导温度网络（MISTnet）， 实现了材料中冲击温度场的预测。作者通过MD方法对数个百万原子规模的系统进行冲击模拟采集数据，通过将系统内原子坐标和原子速度以及网格化的局部密度场作为输入，将网格化的局部温度场作为输出，训练的神经网络能够将初始的、冲击前的微观结构映射到冲击后的温度场，其计算成本比MD模拟的计算成本小108倍，且准确性超过了现有的模型。在泛化测试中，模型仍然准确地预测了热点的形状，虽然温度被高估了，但却准确地捕捉到了孔隙大小和方向的趋势。该工作可以减少现有方法的经验主义，并为将微观结构与冲击载荷下材料的响应联系起来提供了有效的手段。

Editorial Summary: Deep learning predicts shock-induced temperature fields

Material response to shock loading is important to planetary science, aerospace engineering, and energetic materials. Thermal excitation processes such as chemical reactions and phase transitions are significantly accelerated at energy localization. This results from the interaction of shock waves with the material's microstructure and is controlled by complex coupling processes. The process control mechanisms are not fully understood. Most of these processes occur under extreme conditions of temperature, pressure, and strain rate, and various local energy concentrations and microstructural features exist at different length and time scales. Therefore none of the existing models are able to predict shock-induced hotspot formation without strong approximation assumptions. Molecular dynamics (MD) has now been widely used to study shock-induced hotspot formation, including collapse of porosity, shear, friction, and local plastic deformation, but MD methods require huge computational costs. Deep learning has been used to simulate the mesoscale thermomechanical response of materials under shock loading with an accuracy comparable to physics-based simulations, but at a relatively small computational cost. In this work, Prof. Alejandro Strachan et al. from the School of Materials Engineering, Purdue University, designed the Microstructure-Informed Shock-induced Temperature net (MISTnet) based on the UNet network structure, in order to predict the temperature field caused by shock loading. The authors used the MD method to perform shock loading simulations on several million-atom-scale systems to collect data. They used the atomic coordinates and atomic velocities in the system as well as the gridded local density field as input, and used the gridded local temperature field as the output. The trained neural network is able to map the initial, pre-shock microstructure to the post-shock temperature field with a computational cost that is 108 times smaller than that of MD simulations and an accuracy that exceeds existing models. In generalization tests, the model still accurately predicted the shape of hot spots, and although temperatures were overestimated, it accurately captured trends in pore size and orientation. This study can reduce the empiricism of existing methods and provides an effective means to relate microstructure to material response under shock loading.