AW-PINN for solving 3D flow over buildings

Adoptive weights Physics-Informed Neural networks for solving 3D turbulent flow in steady-state condition. flow over a big cities in Japan. two types of optimization is used simultaneously , Adam and L-BFGS. A kind of novel adaptive method for weight balance is applied for making the trainer more robust. L2 regularization in Adam optimization is considered.

The adoptive weight balance method is still neive and needs evolution.

This case solves 3D flow over a city.the benchmark is Case E in Aij institute:

_Guidebook for CFD Predictions of Urban Wind Environment Architectural Institute of Japan.

follow this site for data of the case E : https://www.aij.or.jp/jpn/publish/cfdguide/index_e.htm

Step 1 - How to run this example

• provide data and boundary condition points file in csv format.
• share your data in your gdrive
• load your drive in colbab
• open the repository github link with google colab
• Run All

Step 2 - consideration

• Epochs number is up to your case
• it's recommended to run it with GPU

Step 3 - Libraries

• torch
• pandas
• numpy
• matplotlib
• sklearn _optional

you can install these libraries in one line easily. just copy this single in in your console:

• pip

  python -m pip install torch pandas numpy matplotlib sklearn

Contact

Amirreza Rezayan - [email protected]

Project Link: https://github.com/arezayan/Adaptive-weights-PINN-caseE