基于多任务图神经网络的非定常流体预测  

Unsteady Flow Prediction Based on Multi-task Graph Neural Network

作　　者：周恒安 成乐 施克权 欧洺余 朱宏娜[1] ZHOU Heng-an;CHENG Le;SHI Ke-quan;OU Ming-yu;ZHU Hong-na(School of Physical Science and Technology,Southwest Jiaotong University,Chengdu 610031,China;School of Information Science and Technology,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学物理科学与技术学院,成都610031 [2]西南交通大学信息科学与技术学院,成都610031

出　　处：《科学技术与工程》2024年第27期11733-11740,共8页Science Technology and Engineering

摘　　要：在高维非线性流体模拟中,传统计算流体力学(computational fluid dynamics,CFD)求解器存在计算成本与难度较高的局限性。为提高计算效率,提出一种结合多任务学习与图神经网络的流体模拟方法,旨在高效、快速预测非定常流体动力过程。首先采用非结构化网格对流体计算域的空间分布建模,结合图神经网络提取其多维度空间特征,并通过消息传递机制的聚合更新特性来模拟流体的时空变化规律。考虑流体中不同物理场参数之间存在相关性,采用多任务学习策略对多个物理场参数并行学习和预测,以提高模型的准确性与泛化性。构建数据集并开展验证,对比了不同图神经网络的预测精度,结果表明,相较于图卷积神经网络,本文提出的模型表现出更好的预测性能。预测100步时,均方误差降低了7.2%,预测200步时相对下降29.9%。本文方法在计算效率上也有显著提升,与常规CFD求解器相比,预测速度提升了一到两个数量级,为实时预测提供支撑。Traditional computational fluid dynamics(CFD)solver encounters significant computational costs and struggles with simulating fluids due to the inherently high-dimensional and nonlinear properties.To improve the computational efficiency,a framework combining multi-task learning and graph neural networks(GNN)was proposed here,which could predict unsteady flow quickly and efficiently.Firstly,the fluid computational domain's spatial distribution was modeled by unstructured grids.Then,GNN was employed to extract multidimensional spatial features from this distribution.Finally,the aggregation and updating properties of the message passing mechanism were used to simulate the spatio-temporal variation patterns of the fluid.Considering that there were some correlations between different physical fields,a multi-task learning strategy was adopted to learn the variations of multiple physical field parameters in parallel,which could improve the accuracy and generalization of the model.Validation experiments were carried out on a simulation dataset and compared with graph convolutional neural network.The results show that our model has the best prediction results,with a relative decrease of 7.2%in mean square error(MSE)at 100 steps and 29.9%at 200 steps.This model also has a large improvement in computational efficiency,and its prediction speed is improved by one to two orders of magnitude compared with the conventional CFD solver,which provides a good solution on real-time prediction.

关 键 词：图神经网络 多任务学习 深度学习 非定常流体 流体模拟 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] O351.3[自动化与计算机技术—控制科学与工程]