基于PINN的二维剪切流圆柱绕流场重构  

RECONSTRUCTION OF THE FLOW FIELD AROUND A CYLINDER IN A TWODIMENSIONAL SHEAR FLOW BASED ON PINN

作　　者：苑光耀 王俊淞 赵玄烈 耿敬[1,2] Yuan Guangyao;Wang Junsong;Zhao Xuanlie;Geng Jing(College of Shipbuilding Engineering,Harbin Engineering University,Harbin 150000,China;Sanya Nanhai Innovation and Development Base of Harbin Engineering University,Sanya 572024,Hainan,China)

机构地区：[1]哈尔滨工程大学船舶工程学院,哈尔滨150000 [2]哈尔滨工程大学三亚南海创新发展基地,海南三亚572024

出　　处：《力学学报》2025年第2期436-452,共17页Chinese Journal of Theoretical and Applied Mechanics

基　　金：海南省重点研发项目(ZDYF2023GXJS017);中央高校优秀青年团队项目(3072023JC0101)资助。

摘　　要：随着机器学习方法的发展,基于数据驱动的流场预测成为研究热点.相比于传统神经网络对大量训练数据的依赖以及可解释性较差等问题,物理信息神经网络(PINN)通过将物理约束直接嵌入神经网络损失函数,仅需少量训练数据即可实现流场预测.文章探究了基于PINN的低雷诺数条件下剪切流圆柱绕流场重构方法.基于开源CFD软件OpenFOAM生成的数值模拟数据构建训练集,构建引入剪切流边界条件的圆柱绕流PINN模型,利用流场采样点的速度和压强信息对PINN模型进行训练,以预测不同区域内的流速和压强分布.通过将PINN预测结果与数值模拟数据进行对比,评估了PINN在重构复杂流场中的表现.同时,研究了神经网络层数、网络节点数和监测点位置等参数变化对预测结果的影响,分析了参数调整对流场预测精度的优化作用.研究结果表明,PINN不仅能有效重构剪切流绕流场的流速和压强分布,其预测精度也与数值模拟结果较为吻合,验证了PINN在剪切流圆柱绕流场重构中的实用性和准确性.对比分析表明,监测点位置对预测结果的影响较为显著,合理选取监测点位置可使PINN预测精度提升1~2个数量级.With the advancement of machine learning techniques,data-driven flow field prediction has become a significant research focus in fluid mechanics.Compared with traditional neural networks,which often require large amounts of training data and suffer from poor interpretability,physics-informed neural networks(PINN)embed physical constraints directly into the loss function.This enables accurate flow field predictions with only minimal training data.This study investigates the reconstruction of shear flow around a circular cylinder under low Reynolds number conditions using a PINN-based approach.Numerical simulation data generated by the open-source computational fluid dynamics(CFD)software OpenFOAM were used to construct the training dataset.A PINN model incorporating shear flow boundary conditions for circular cylinder flow was developed,and velocity and pressure information from sampling points within the flow field were employed to train the model.The goal was to predict the velocity and pressure distributions across various regions of the flow field.To evaluate the performance of the PINN model,its predictions were compared with results from numerical simulations,demonstrating the model's ability to effectively reconstruct complex flow fields.Furthermore,this study investigated the impact of neural network parameters,such as the number of layers,the number of nodes per layer,and the spatial location of monitoring points,on the prediction accuracy.The mechanisms underlying the optimization of these parameters were analyzed in detail.The results indicate that the PINN model can accurately reconstruct the velocity and pressure distributions of shear flow around a circular cylinder,achieving a prediction accuracy that closely aligns with numerical simulation results.This validates the practical applicability and accuracy of PINN for reconstructing flow fields in low Reynolds number shear flows.Moreover,comparative analyses revealed that the location of monitoring points has a significant influence on the prediction

关 键 词：物理信息神经网络 NAVIER-STOKES方程 剪切流场重构 圆柱绕流 

分 类 号：O352[理学—流体力学] TP183[理学—力学]