基于二级网格和卷积神经网络的流体模拟方法

A 2-LEVEL-GRID-BASED AND CNN-BASED FLUID SIMULATION METHOD

基于欧拉法的流体模拟是计算机图形学的重要方向之一,但其计算复杂度较高,直接求解精确的流体运动状态会带来较大的性能开销。提出基于二级网格和卷积神经网络(CNN)的流体模拟方法,在提高模拟性能的同时基本保持流体的形态。二级网格的流体空间结构抽象,能够有效减少流体计算的规模,达到加速模拟的效果。在此基础上,使用CNN来减少由于计算量减少而引入的误差,通过对欧拉法精确计算结果进行机器学习,得到较为精确的流体形态。实验结果表明,该方法在提高计算效率的同时能一定程度上保持流体形态细节。

Euler-method-based fluid simulation is an important research direction of computer graphics. However, its computation complexity is high, and directly solving accurate fluid motion state will bring great performance overhead. This paper proposes a 2-level-based and CNN-based fluid simulation method, and it can improve the simulation performance while basically keeping the fluid morphology. The fluid space structure of the 2-level grid is abstract, which can effectively reduce the scale of fluid calculation and achieve the effect of speeding up the simulation. On this basis, a convolutional neural network was used to reduce the error caused by the reduction of calculation amount. And a more accurate fluid morphology was obtained through the machine learning of the accurate calculation results of Euler method. The experimental results show that this method can improve the computational efficiency while keeping the details of fluid morphology to a certain extent.