Adversarial-learning-based closed-loop training strategy for PINN-based fluid simulators that generalize

LIU, Ran; XIAO, Ziruo; ZENG, Lingqi; ZHANG, Rushan (ordered alphabetically by last name)

Install

cd AL_PINN

conda env create -f environment.yml

Train

python train_ours.py

Interactive interface

python interface.py --explicit_weights ./pretrain_weights/ours.pth

Usage:
Use mouse to drag the objects
Press 'x' to increase flow velocity
Press 'y' to decrease flow velocity
Note: The interface is only tested on Windows 11

Evaluation tools

evaluation/inference.py : given an input dataset and a PINN solver weight, obtain the inference results
evaluation/visualization.py : given inference results in .npy, visualize the results and save as .gif and .png

Evaluation metrics

run.py : runs the evaluation for the overfitted results and the inference results. The inputs are the folder of those dataset.
metrics/scripts/overfit.sh : commands for overfitting the model for the evaluation
metrics/scripts/basic.sh : basic commands for the evaluation
metrics/scripts/ablation.sh : commands for ablation study
metrics/scripts/training_tracking.sh : commands for tracking the training performance