| 苟志勇,蒋权,黄文清.并行重采样物理信息神经网络:用于低雷诺数下不可压缩流体流动分析[J].计算力学学报,2025,42(4):693~698 |
|
| 
码上扫一扫！ |
| 并行重采样物理信息神经网络:用于低雷诺数下不可压缩流体流动分析 |
| Parallel resampling physical information neural network for fluid dynamics analysis at low Reynolds number |
| 投稿时间：2024-04-02 修订日期：2024-05-26 |
| DOI：10.7511/jslx20240402003 |
| 中文关键词: 物理信息神经网络 重采样 并行计算 网格搜索 PDE智能求解 |
| 英文关键词:PINNs resampling parallel computing grid search PDE intelligent solution |
| 基金项目:广西科技计划(桂科AD21220002);2023中央支持地方高校改革发展资金-智能信息技术现代产业学院(301021501);广西民族大学引进人才科研项目(2020KJQD24)资助. |
|
| 摘要点击次数: 833 |
| 全文下载次数: 802 |
| 中文摘要: |
| 流场重建和流场参数识别是流体力学研究中的关键任务。当使用基于物理信息神经网络PINNs(Physics-Informed Neural Networks)来解决流体流动问题时,传统的PINNs方法在模拟不可压缩流体时遭遇若干挑战,如预测精度不足、采样点布置和超参数选择难题。这些挑战使得PINNs的训练过程缓慢且难以收敛并陷入局部最优解。本文改进PINNs的主要思想为,首先,使用传统PINNs采样方式对求解域进行采样,并对其训练过程中残差较大的位置进行重采样;其次,运用并行神经网络结构以提高预测精度;最后,通过网格搜索技术来确定最优超参数。即并行重采样物理信息神经网络PRS-PINNs(Parallel Resampling PINNs)。通过两个经典的计算流体动力学问题,即低雷诺数下的Kovasznay流和圆柱绕流对PRS-PINNs进行了测试。实验结果显示,PRS-PINNs的预测结果与解析解高度一致,利用流场数据也能较为准确地推算出未知参数,为计算流体动力学问题的求解提供了一种新方法。 |
| 英文摘要: |
| Flow field reconstruction and flow field parameter identification are key tasks in fluid mechanics research.When using Physics-Informed Neural Networks(PINNs) to solve fluid flow problems,the traditional PINNs approach encounters several challenges in modeling incompressible fluids,such as the placement of sampling points,insufficient prediction accuracy,and hyper-parameter selection.These challenges make the training process of PINNs slow and difficult to converge or fall into local optimal solutions.The main ideas of this improved PINNs are:first,the traditional PINNS sampling method is used to sample the solution domain,and the position with a large residual error in the training process is resampled;second,a parallel neural network structure is utilized to improve the prediction accuracy;and finally,the optimal hyperparameters are determined by a grid search technique that is,Parallel Resampling Physical Information Neural Networks(PRS-PINN).In the article,PRS-PINN are tested by two classical computational fluid dynamics problems Kovasznay flow at low Reynolds number and cylindrical winding flow.The experimental results show that the prediction results of the PRS-PINNs are highly consistent with the analytical solutions,and the unknown parameters can also be derived more accurately using the flow field data,which provides a novel method for computational fluid dynamics solutions. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
|
