Apply for a GPU community grant: Academic project

#1
by Shuxian - opened
Owner

PhyTopo-Net is an open academic application for reconstructing high-resolution coastal typhoon wind fields from ERA5 reanalysis data. The project accompanies an ongoing research manuscript on WRF-supervised, topography-aware reconstruction of typhoon winds over complex coastal terrain.

The model learns mesoscale wind and pressure structures from high-resolution WRF simulations and transfers them to historical ERA5 typhoon events. The public Space allows researchers and students to upload hourly ERA5 surface fields, reconstruct 3-km U10, V10 and mean sea-level pressure fields, inspect time-dependent wind maps, and download the reconstructed NetCDF product. It also provides a documented engineering post-processing option for supported observation stations. The released workflow has been applied to a catalogue of 117 paired historical typhoons to investigate upper-tail winds, sub-grid coastal contrasts and changes in strong-wind exceedance frequency.

This is an open-science and non-commercial project. The Space is intended to support manuscript reproducibility, peer review, teaching and reuse by the coastal meteorology and wind-hazard communities. The source code, model configuration, example inputs and downloadable outputs are made publicly accessible.

GPU support is needed because the final inference workflow uses overlapping 64 × 64 patches, Hann-weighted blending and multiple offset grids to reduce patch-boundary artefacts and obtain spatially smooth full-domain fields. A six- or 24-hour event window therefore requires multiple model evaluations. The current shared ZeroGPU quota is sufficient for only a few short demonstrations and prevents reliable reproduction by reviewers and other users.

We would be grateful for a community GPU grant for a T4- or A10G-class GPU. The application processes bounded user-selected time windows, does not perform training on the Space, and can be configured to sleep when inactive. The grant would make the published reconstruction workflow continuously reproducible without changing the scientific inference procedure.

Thank you for considering this academic open-science project.

Hello Hugging Face team, @hysts

I would like to kindly follow up on this Community GPU Grant application, which has been open for approximately 30 days.

Our PhyTopo-Net Space is already publicly available on ZeroGPU, but its shared GPU quota is insufficient for reproducible scientific inference over multi-hour typhoon events.

Unlike short interactive demonstrations, the full reconstruction workflow requires repeated inference over overlapping spatial patches. A dedicated T4 GPU would be sufficient for our research and educational use case, and we are happy to configure automatic sleep when the Space is inactive to minimize resource consumption.

Could you please let us know whether the application is still under consideration or whether any additional information is needed?

Thank you very much for your consideration!

Hi @Shuxian , thanks for following up.
Community GPU Grants are currently limited to ZeroGPU, and this Space is already running on ZeroGPU. So there isn't any additional hardware grant we can provide for this Space.
Thanks for understanding!

Sign up or log in to comment