openclimatefix/graph_weather

Graph-based weather forecasting models. Originally, PyTorch implementation of Ryan Keisler's 2022 "Forecasting Global Weather with Graph Neural Networks" paper (https://arxiv.org/abs/2202.07575)

What it solves

Graph Weather provides a modular framework for weather forecasting and data assimilation using graph neural networks (GNNs). It allows researchers and developers to implement and experiment with various graph-based AI models to predict weather patterns more efficiently and accurately than traditional numerical weather prediction (NWP) methods.

How it works

The library implements several state-of-the-art graph-based forecasting models, including DeepMind's GenCast and FGN, Microsoft's Aurora, and FengWu-GHR. It handles the generation of graphs internally based on latitude and longitude coordinates, allowing users to pass node features directly into the models. It also includes a dedicated assimilation model that can map irregular observations (with latitude, longitude, and height) to a desired output grid.

Who it’s for

This project is designed for AI researchers, meteorologists, and developers working on machine learning for weather and climate science.

Highlights

  • Multiple Model Architectures: Implements a variety of forecasting models like GenCast (diffusion-based), Aurora, and WeatherMesh-3 (Neighborhood Attention).
  • Modular Design: Components can be swapped to facilitate experimentation with different graph-based approaches.
  • Data Assimilation: Includes tools to process irregular observations into structured grids.
  • PyTorch-based: Built on PyTorch for flexible training and inference.

Related

  • Project
    TorchSpatiotemporal/tsl

    A PyTorch-based library for neural spatiotemporal forecasting that provides a unified framework for data preprocessing, model prototyping, and Graph Neural Network implementation.

  • Project
    openclimatefix/metnet

    A PyTorch implementation of Google Research's MetNet models for short-term and global precipitation forecasting using satellite and atmospheric data.

  • Project
    sktime/pytorch-forecasting

    A PyTorch-based library for time series forecasting that provides a high-level API and state-of-the-art deep learning architectures like Temporal Fusion Transformers and N-BEATS.

  • Project
    awslabs/gluonts

    A Python package for probabilistic time series modeling that uses deep learning and PyTorch to generate forecasts with uncertainty intervals.

  • Project
    iMoonLab/DeepHypergraph

    A PyTorch-based deep learning library for Graph and Hypergraph Neural Networks that supports high-order message passing and Auto-ML tuning.