GraphDOP (Graph-based Direct Observation Prediction) was proposed by ECMWF. It is an end-to-end observation-driven (AI-DOP) weather forecasting model based on graph neural networks (GNNs). The model uses only Earth-system observations (polar-orbiting/geostationary satellite brightness temperatures, occultation bending angles, scatterometer backscatter, radar altimetry, radiosonde and conventional surface observations, and so on) as inputs and training targets. It does not use any physics-based reanalysis field and can produce medium-range forecasts beyond 5 days.
Paper:GraphDOP: Towards skilful data-driven medium-range weather forecasts learnt and initialised directly from observations
GraphDOP uses an encoder-processor-decoder architecture: the GNN encoder maps observations in the input window to an O96 (approximately 1°) latent grid according to spatial proximity, the Transformer processor advances the atmospheric state in latent space, and the GNN decoder maps the latent grid back to target observation locations and outputs forecasts channel by channel. The training objective is channel-weighted mean squared error (WMSE). This repository is a minimal reproduction based on the paper and is integrated with the OneScience data loading and training workflow.
Use Cases
Scenario
Description
Observation-driven medium-range weather forecasting research
Learn atmospheric-state representations directly from observations and forecast future windows.
Graph + Transformer latent-space model research
Reproducible encoder-processor-decoder architecture and WMSE objective.
Local quick validation
Use synthetic data to check data loading, training, inference, and result scripts.
ModelScope/OneCode execution
Download the model package, install dependencies, and run the scripts directly.
Multi-card training
Launch multi-process training with torchrun.
Usage
1. OneCode Usage
Use the OneCode online environment for intelligent one-click AI4S programming:
The OneScience community provides ERA5 data for training (the current repository contains complete data slices subject to data-file size limits). Download it with the command below and confirm that the data path in conf/config.yaml is correct:
For a quick workflow validation, run the synthetic data script first:
python scripts/fake_data.py
Note: scripts/fake_data.py generates [T, C, H, W] data from the model input/output windows and grid_shape. Because ERA5Datapipe supports only regular grids, this project approximates the irregular Level-1 observations in the paper with six observation-grid channels.
The weight/ folder is reserved for model weights. Pretrained weights are not provided by default; users may train the model using the paper configuration. The paper model (1024 latent channels, an O96 latent grid, and 70k steps on 64×H100 GPUs) has no publicly released weights.
Inference
Inference reads data/checkpoints/model_bak.pth：
python scripts/inference.py
Prediction results are written frame by frame to:
result/output/
Evaluation and Visualization
python scripts/result.py
Outputs include：
result/rmse.npy
result/acc.npy
result/loss.png
Forecast comparison plots for the specified date and variables
Official Source and Reproduction Notes
The paper is an ECMWF AI-DOP preprint, and no official implementation is publicly available (the source is based on PyTorch Geometric and builds dynamic observation graphs per batch). This repository's model/graphdop.py is a pure PyTorch minimal reproduction that preserves the paper's encoder-processor-decoder GNN structure and WMSE objective.
Differences from the paper (due to limitations of the OneScience gridded data pipeline): the paper consumes irregular raw Level-1 observations, with each observation connected to its nearest latent-grid neighbors using a dynamically constructed graph. This reproduction uses ERA5 gridded h5 channels as a placeholder observation field and fixes the latent grid to a regular 8-neighbor graph. The edge features (azimuth and Haversine distance) match the paper. Training augmentations such as random observation dropout (25% for satellites/50% for conventional observations) are not implemented.
conf/config.yaml uses a small configuration (32×32 grid, 8×8 latent grid, latent_dim=64) for connectivity validation by default; paper-level reproduction requires an O96 latent grid, latent_dim=1024, and a larger data scale.
The following details are not disclosed in the paper and are assumptions in this reproduction:per-channel weights w_{c,i} (currently all 1 by default), graph-construction details, and the specific implementation of windowed attention in the processor.
This repository is an independent GraphDOP reproduction (the model code is an original minimal implementation), with the architecture design based on the paper by Alexe et al. (2024).
Please cite:Alexe, M., E. Boucher, P. Lean, E. Pinnington, P. Laloyaux, A. McNally et al. GraphDOP: Towards skilful data-driven medium-range weather forecasts learnt and initialised directly from observations. arXiv:2412.15687, 2024.
