Scale-invariant Learning by Physics Inversion (SIP)

This repository contains the code for the NeurIPS 2022 paper Scale-invariant Learning by Physics Inversion. With the code published here, all experiments from the paper can be reproduced.

Requirements

Python 3.6 or higher including pytorch and phiflow.

Using the Code

The code is ordered by experiment.

Experiment	Figures	Source Code Directory
2D Optimization	1	optimization_trajectories
Inverting the exponential	2	exp
Experimental Characterization	5	sin_characterization
Poisson's equation	6a,b	poisson
Heat equation	6c,d	heat
Navier-Stokes equations	7	fluid

Inside the directories, you will find train_* and plot_* files. No external configuration is required, the settings are adjusted within the Python files. The train_* files train a neural network using the selected method and store checkpoints and learning curves in a subdirectory of ~/phi.

Once the networks are trained, the plot_* files can be used to visualize the results. You need to fill in the correct paths before running them.

Citing this Work

Please use the below citation:

@inproceedings{Holl2022Scale,
        title     = {Scale-invariant Learning by Physics Inversion},
        author    = {Philipp Holl and Vladlen Koltun and Nils Thuerey},
        booktitle = {Conference on Neural Information Processing Systems},
        year      = {2022},
}