search

openclimatefix / graph_weather

PyTorch implementation of Ryan Keisler's 2022 "Forecasting Global Weather with Graph Neural Networks" paper (https://arxiv.org/abs/2202.07575)
MIT License
194 stars 50 forks source link

[Paper] Graph-based Neural Weather Prediction for Limited Area Modeling #78

Open jacobbieker opened 11 months ago

jacobbieker commented 11 months ago

Arxiv/Blog/Paper Link

https://www.climatechange.ai/papers/neurips2023/17

Detailed Description

Interesting work using graph weather models for limited area models

Code is available here: https://github.com/joeloskarsson/neural-lam

Context

This somewhat does the more fine-grained local area forecasts with graph weather models. Its only limited-area-models, so not quite the same, but considering we want to have our global weather models here also be able to have higher-resolution areas that they output to/take in data from. So would be quite interesting to look through, especially the graph creation, since it is also in PyTorch.

jacobbieker commented 11 months ago

A lot of these LAM has higher update rates than the global models that they use for boundary conditions (i.e. HRRR is every 15 minutes, while GFS is hourly). One approach for training a model that outputs to both would be mask the non-LAM model area's loss for all timesteps that are not covered by the global model. So the LAM might still get 4 updates per hour, but the global areas would only get 1 per hour.