基于PINN方法的不可压流场求解

Solving Incompressible Flow Fields Based on PINN Method

将残差网络(Res Net)结构引入物理信息神经网络(PINN),提出一种基于Res Net的PINN方法(Res Net-PINN)。采用该方法对二维不可压圆柱绕流尾迹流场进行重建和预测,结果表明:Res Net-PINN能更准确地重建绕流尾迹的非定常变化规律;在圆柱绕流尾迹的短期预测方面,Res Net-PINN的预测精度和收敛速度相比全连接PINN均能提升3倍左右。Res Net能提高PINN对非定常流场的求解和预测能力,该研究可为采用机器学习方法求解更复杂流动问题提供参考。

The residual network(ResNet)structure is introduced into the physics-informed neural network(PINN),and a PINN method(ResNet-PINN)is proposed.It is applied to the reconstruction and prediction of a two-dimensional incompressible wake flow around a cylinder.The results show that the ResNet-PINN can more accurately reconstruct the unsteady variation laws of the wake flow.In terms of short-term prediction of the wake flow,both the prediction accuracy and the training speed of the ResNet-PINN can be improved by about 3 times compared with the fully connected neural network PINN(FCNN-PINN).Therefore,the residual network can improve the ability of PINN to solve and predict unsteady flow fields.This study provides a basis for machine learning methods to solve more complex flow problems.