LithiumDA/L_inf-PINN

[NeurIPS 2022] Minimizing L_inf Physics-Informed Loss for PINN with adversarial training

L_inf PINN

This is the official implementation for Is $L^2$ Physics-Informed Loss Always Suitable for Training Physics-Informed Neural Network?, which proposes a novel PINN training algorithm to minimize the $L^{\infty}$ loss in a similar spirit to adversarial training.

Dependencies

The required packages are listed in requirements.txt, which can be installed by running pip install -r requirements.txt.

Getting started

To reproduce our result on 250-dimensional HJB Equation on a single GPU, run python run.py.

Multi-GPU training is also supported, e.g.,

python run.py hjb.gpu_cnt=2 \
      hjb.train.batch.domain_size=25 \
      hjb.train.batch.boundary_size=25
# shrink the batch size when there are mutiple GPUs

Training scripts for other experiments are provided in the scripts directory. For example, to train vanilla PINN on 100-dimensional HJB Equation, run bash scripts/100-HJB-PINN.sh.

Contact

May you have any questions on our work or implementation, feel free to reach out to shandal@cs.cmu.edu!

Citation

If you find this repository useful, please consider giving a star ⭐ and cite our paper.

@inproceedings{wang2022is,
      title={Is {$L^2$} Physics-Informed Loss Always Suitable for Training Physics-Informed Neural Network?}, 
      author={Chuwei Wang and Shanda Li and Di He and Liwei Wang},
      booktitle={Advances in Neural Information Processing Systems},
      year={2022},
}