结合卷积神经网络的复杂介质动力学建模与加速方法研究

With the development of virtual reality and the increasing demand for the physical authenticity of virtual scenes, physical-based simulation has become a hot topic in computer graphics research. However, existing physical-based simulation methods still cannot model complex media well, such as non-Newtonian fluids, hyperelastic-plastic materials and granular flows. Besides, existing numerical methods have poor uniformity and usually cannot guarantee the accuracy and efficiency at the same time. To solve these problems, we intend to unify the modeling of any complex materials including non-Newtonian fluids, hyperelastic-plastic materials and granular flows based on the nonlocal theory. In order to guarantee both the accuracy and efficiency, we propose to apply the coarsening technology and multi-GPU parallel acceleration to improve the efficiency and apply the convolutional neural network to improve the accuracy. The purpose of this project is to establish a highly efficient and accurate simulation method as well as a unified mechanical theory suitable for the modeling of arbitrary complex material.

随着虚拟现实相关技术的发展以及人们对虚拟场景物理真实性需求的增长，基于物理的仿真动画技术已经成为计算机图形学研究的一大热点。然而现有物理仿真方法在建模非牛顿流体、超线性弹塑性材料及颗粒流等复杂介质时依然存在力学理论建模能力不足、现有数值计算方法统一性差以及物理仿真精度与效率无法同时兼顾等矛盾。针对这些不足，本课题拟基于非局部理论对包括非线性流体、超线性弹塑性材料以及颗粒流在内的任意复杂介质进行统一建模，同时为了保证精度和效率，本课题拟结合卷积神经网络对复杂介质的本构模型进行训练，并利用粗糙化技术及多GPU并行加速来提升物理仿真的效率，旨在建立适用于统一建模任意复杂介质动力学的高效、高精度物理仿真理论与方法。

随着虚拟现实相关技术的发展以及人们对虚拟场景物理真实性需求的增长，基于物理的仿真动画技术已经成为计算机图形学研究的一大热点。然而现有物理仿真方法在建模非牛顿流体、超线性弹塑性材料及颗粒流等复杂介质时依然存在力学理论建模能力不足、现有数值计算方法统一性差以及物理仿真精度与效率无法同时兼顾等矛盾。. 针对这些不足，本课题基于非局部理论对包括非线性流体、超线性弹塑性材料以及颗粒流在内的任意复杂介质进行统一建模，同时为了保证精度和效率，本课题结合AI技术及数值优化法方法对复杂介质的本构模型进行建模，并利用粗糙化技术及多GPU并行加速来提升物理仿真的效率，建立了适用于统一建模任意复杂介质动力学的高效、高精度物理仿真理论与方法。搭建了开源物理仿真引擎PeriDyno。