搜索

x
中国物理学会期刊

基于MO-PINN二维圆柱绕流流场重构与时序预测研究

Flow field reconstruction and time series prediction of two-dimensional flow past a circular cylinder using MO-PINN

PDF
导出引用
  • 物理信息神经网络(Physics-Informed Neural Network,PINN)通过嵌入物理方程,在稀疏数据条件下实现流场精准重构,适用于流体问题求解。针对PINN在流场重构中存在物理量耦合干扰严重、优化效率低,时序预测能力不足的问题,本文提出一种基于多通道输出的物理信息神经网络模型(Multi-output PINN,MO-PINN)。MO-PINN模型在输出层为各类物理量搭建独立输出通道,并分别开展参数优化,有效实现多物理量解耦。流场输入数据经由共享特征层提取后,通过多通道输出分别映射得到对应物理量结果,同步完成流场高精度重构与时序预测。以典型二维圆柱绕流为验证算例,同传统PINN进行性能对比。数值结果表明:MO-PINN流场重构精度提升42.9%,时序预测精度提升26.9%,复杂流动预测性能显著提升。

     

    Physics-informed neural networks (PINN) is applied to solve the fluid dynamics problems. It reconstructs regional or whole flow fields under sparse data conditions by embedding governing physical equations. However, conventional PINN models usually employ a single shared output structure for multiple physical variables, which may introduce cross-variable coupling errors among pressure and velocity components. This design can introduce cross-variable optimization interference. In addition, their time series extrapolation capability is often limited when dealing with unsteady flow fields. To address these limitations, a multi-output physics-informed neural network (MO-PINN) is developed here, for the reconstruction and time series prediction of unsteady flow fields. In the proposed MO-PINN framework, separate output channels are assigned to different physical variables, allowing each channel to be optimized independently. Such a multi-channel architecture effectively alleviates the coupling among physical variables, thereby enhancing the overall prediction accuracy. Specifically, the spatiotemporal coordinates are first fed into shared feature layers to extract the common latent representation of the flow field. Then, independent output channels are constructed for pressure and velocity components. Each output channel maps the shared features to a specific physical variable and is optimized separately. By reducing coupling interference among different physical quantities, the proposed architecture improves the capability of flow field reconstruction and prediction. The proposed MO-PINN is applied to the flow field reconstruction and time series prediction of two-dimensional flow past a circular cylinder. The numerical flow field data are obtained from computational fluid dynamics simulations and used to evaluate the reconstruction accuracy and temporal prediction performance of the model. Comparative analyses are conducted between the PINN and the proposed MO-PINN. The results show that MO-PINN can more accurately reconstruct the spatial distribution of pressure and velocity fields and better capture the unsteady wake characteristics behind the circular cylinder. In time series prediction, MO-PINN also shows improved prediction stability and temporal continuity. Quantitative results demonstrate that, relative to the baseline PINN, the reported accuracy gains are 42.9% for flow-field reconstruction and 26.9% for time-series prediction, indicating a significant improvement in prediction performance. These results demonstrate that the proposed multi-output architecture can effectively reduce coupling interference among physical variables and provide a feasible method for improving unsteady flow field reconstruction and prediction using PINN.

     

    目录

    /

    返回文章
    返回