Multi-Viscosity Physics-Informed Neural Networks for Generating Ultra High Resolution Flow Field Data
Sen Zhang, Xiao‐Wei Guo, Chao Li, Ran Zhao, Canqun Yang, Wei Wang, Yanxu Zhong
Abstract
To address the limited generalisation ability issue of physics-informed neural networks, we propose a multi viscosity physics-informed neural networks (μ-PINNs), along with two tailored training strategies. By using μ-PINNs, we train the model once for a specific scenario and obtain flow field data with varying fluid viscosity (μ). To validate μ-PINNs, we conduct experiments on three 2D fluid flow scenarios, comparing the results to those computed by OpenFOAM and providing their relative L2 error. The experiments demonstrate that μ-PINNs possess the capability of capturing the influence of viscosity on the output flow field data. Additionally, we compare the traditional scheme with the mixed-variable scheme. The memory usage and training time in mixed-variable scheme are 52.5% and 53.6% of that in traditional scheme, at the expense of lower accuracy. This comparison offers guidance for researchers in selecting an appropriate scheme. All code and data-sets are available on GitHub at https://github.com/Jensen1997/mu-PINNs.