Journal of Computer Applications
Next Articles
Received:
Revised:
Online:
Published:
蒋权1,2,黄文清2,苟志勇2
通讯作者:
基金资助:
Abstract: In recent years, Graph Neural Network (GNN) have been widely applied to complex fluid system predictions due to the ability to effectively handle structured grids and strong compositional generalization capabilities. However, from a Lagrangian mesh-free perspective, unpredictable output issues have arisen when fluid particle information, transformed through translation, rotation, or reflection, was fed into GNN. To address this problem, a prediction method based on the Equivariant Graph Neural Network-based Simulator (EGNS) was proposed. Geometric vectors were first converted into relative equivariant quantities. Then, equivariant message passing was employed at each step to ensure that the entire neural network maintained equivariance, aligning network outputs with spatial transformations of the input equivariant variables. Finally, the optimized EGNS model was trained using particle trajectories simulated by the Smoothed Particle Hydrodynamics (SPH) method. Experimental results on public fluid simulation datasets demonstrated that EGNS exhibited strong predictive performance. Compared to the Graph Neural Network-based Simulator (GNS), EGNS achieved greater accuracy in predicting fluid particle movement, velocity, and typical details, improving prediction precision by 16%.
Key words: equivariant graph neural network, equivariance, smoothed particle hydrodynamics, fluid particle, flow prediction
摘要: 摘 要: 近年来,图神经网络(GNN)因能较好解决结构网格的问题,且具有较强的组合泛化能力,被越来越多地应用于复杂的流体系统预测中。然而,在拉格朗日无网格视角下,经过平移、旋转或翻转变换的流体粒子信息输入GNN会出现不可预测的输出问题。为了解决该问题,提出了基于等变图神经网络模拟(EGNS)的方法。首先将几何向量转换成相对的等变量,然后通过每一步具有等变性的消息传递使整个神经网络具有等变性,网络输出与输入等变量的空间变换保持一致,最后在光滑粒子动力学(SPH)方法模拟的粒子轨迹里训练得到较优的EGNS模型。在公开流体仿真数据集上的实验结果表明,EGNS具有良好预测效果。相较于图神经网络模拟(GNS)的方法,EGNS在流体粒子运动形态、速度及典型细节的表现力上更加准确,预测精度提高了16%。
关键词: 等变图神经网络, 等变性, 光滑粒子动力学, 流体粒子, 流动预测
CLC Number:
O35
TP183
蒋权 黄文清 苟志勇. 基于等变图神经网络的拉格朗日粒子流模拟[J]. 《计算机应用》唯一官方网站, DOI: 10.11772/j.issn.1001-9081.2024111601.
0 / Recommend
Add to citation manager EndNote|Ris|BibTeX
URL: https://www.joca.cn/EN/10.11772/j.issn.1001-9081.2024111601