ABSTRACT

In this work, we construct a data-driven model to address the computing performance problem of the moving particle semi-implicit method by combining the physics intuition of the method with a machine-learning algorithm. A fully connected artificial neural network is implemented to solve the pressure Poisson equation, which is reformulated as a regression problem. We design context-based feature vectors for particle-based on the Poisson equation. The neural network maintains the original particle method’s accuracy and stability, while drastically accelerates the pressure calculation. It is very suitable for GPU parallelization, edge computing scenarios and real-time simulations.

摘要

在这项工作中，我们构建了一个数据驱动模型，通过将方法的物理直觉与机器学习算法相结合来解决运动粒子半隐式方法的计算性能问题。实施完全连接的人工神经网络来求解压力泊松方程，将其重新表述为回归问题。我们基于泊松方程为粒子设计基于上下文的特征向量。神经网络保持了原始粒子法的准确性和稳定性，同时大大加快了压力计算。非常适合GPU并行化、边缘计算场景和实时仿真。