Deep Learning for Physical Processes: Integrating Prior Scientific Knowledge:

Official Pytorch implementation of ICLR 2018 paper Deep Learning for Physical Processes: Integrating Prior Scientific Knowledge.

Guetting started:

Clone repository along with submodules: git clone --recursive https://github.com/emited/flow

Dataset

Download the data here.

Results

Note: By defalt, this implementation currently uses bilinear interpolation for warping. This scheme works well for modeling purely advective processes. For advective and diffusive processes, a gaussian warping scheme can be used (flow/modules/warps/GaussianWarpingScheme). The gaussian warping scheme will be integrated shortly into pytorch. Take a look at the pull request here for a status update. While waiting, it is possible to build pytorch from a forked version available here.