Source-linked AI summary
GraphCast: Learning skillful medium-range global weather forecasting
Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, Peter Battaglia
TL;DR
Medium-range forecasting needs accurate predictions, but traditional numerical weather prediction does not directly use historical data to improve its models. GraphCast trains on reanalysis data to produce global 10-day forecasts, achieving higher skill than HRES on most evaluated targets while retaining efficiency and supporting severe-event prediction. Its deterministic formulation leaves uncertainty representation as an important limitation.
Problem
Traditional numerical weather prediction does not directly use historical weather data to improve forecast models, despite the importance of medium-range forecasts.
Method
GraphCast is a machine-learning weather model trained on reanalysis data that autoregressively predicts global weather states on a 0.25° grid.
Results
GraphCast outperformed HRES on 90.3% of 1380 evaluated targets, while optimally blurred GraphCast exceeded optimally blurred HRES on 88.0%.
Takeaways & Limitations
GraphCast’s skill and efficiency make machine-learning forecasting competitive with traditional methods and support severe-event applications including tropical cyclones, atmospheric rivers, and extreme temperatures.
Takeaways & Limitations
GraphCast produces deterministic forecasts whose MSE-trained spatial blurring may limit value for applications requiring uncertainty distributions at longer lead times.
Abstract
from arXiv · showhide
Global medium-range weather forecasting is critical to decision-making across many social and economic domains. Traditional numerical weather prediction uses increased compute resources to improve forecast accuracy, but cannot directly use historical weather data to improve the underlying model. We introduce a machine learning-based method called "GraphCast", which can be trained directly from reanalysis data. It predicts hundreds of weather variables, over 10 days at 0.25 degree resolution globally, in under one minute. We show that GraphCast significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets, and its forecasts support better severe event prediction, including tropical cyclones, atmospheric rivers, and extreme temperatures. GraphCast is a key advance in accurate and efficient weather forecasting, and helps realize the promise of machine learning for modeling complex dynamical systems.
Introduction
Traditional numerical weather prediction remains highly accurate but does not directly use growing historical weather archives to improve forecast models. GraphCast addresses this gap with a data-trained graph neural network that autoregressively predicts global weather states and is evaluated against operational and machine-learning baselines.
- Motivation: Traditional numerical weather prediction solves governing weather equations on supercomputers, but increasing historical data does not directly improve its underlying models.Improving NWP instead relies on expert-designed models, algorithms, and approximations.
- Motivation: Machine learning-based weather prediction trains forecast models directly from historical data and may improve efficiency through deep-learning hardware.In medium-range forecasting, however, NWP systems such as ECMWF HRES remained the most accurate systems.
- GraphCast: GraphCast predicts global weather states on a 0.25° grid, using the two most recent states to forecast the next state six hours ahead.Each state contains surface and atmospheric variables, and forecasts are generated by feeding predictions back into the model autoregressively.
- GraphCast: GraphCast uses an encode-process-decode graph neural network that maps grid data to a multi-mesh, performs learned message-passing, and decodes features back to the grid.The processor uses 16 unshared GNN layers, while the multi-mesh provides spatially homogeneous global resolution through an icosahedral hierarchy.
- Training and evaluation: The model was trained on 39 years of ERA5 reanalysis data with an autoregressive, vertically weighted mean squared error objective.The number of autoregressive training steps increased from 1 to 12, corresponding to six hours through three days.
- Training and evaluation: Evaluation compares GraphCast with HRES and machine-learning baselines using RMSE and anomaly correlation across selected variables, levels, and lead times.The study evaluates 69 of GraphCast’s 227 variable-level combinations against HRES and reports additional analyses in the supplement.
Forecast verification results
GraphCast showed higher forecast skill than HRES across most evaluated targets over 10-day forecasts, with advantages across lead times and weather variables. Its relative performance remained strong after accounting for forecast blurring and exceeded that of Pangu-Weather on nearly all shared targets.
- GraphCast produced forecasts at 0.25° horizontal resolution and 13 vertical levels.
- 7%–14% skill score improvement was observed for GraphCast over HRES on z500 across lead times.z500 represents geopotential at 500 hPa and is used as a meteorologically important headline field.
- 90.3% of 1380 evaluated targets favored GraphCast over HRES by RMSE skill, with significant superiority on 89.9%.The evaluation covered variables and pressure levels across 10-day forecasts.
- 96.9% of the remaining 1280 targets significantly favored GraphCast when the 50 hPa level was excluded, rising to 99.7% of 1180 when 50 and 100 hPa were excluded.HRES’s relative advantages were disproportionately localized in the stratosphere, where training loss weight was lowest.
- 88.0% of 1380 targets favored optimally blurred GraphCast over optimally blurred HRES.This test assessed whether GraphCast’s advantage persisted when both systems were allowed to blur forecasts.
- GraphCast outperformed Pangu-Weather on 99.2% of the 252 shared targets.
Severe event forecasting results
GraphCast supports severe-event forecasting across tropical cyclones, atmospheric rivers, and extreme temperatures, despite not being specifically trained for these downstream tasks. Its advantages over HRES vary by event type and lead time.
- Severe event forecasting results: GraphCast forecasts were evaluated for tropical cyclones, atmospheric rivers, and extreme temperatures as important downstream applications.These events were not specific training targets.
- Tropical cyclones: GraphCast had lower median tropical-cyclone track error than HRES over 2018–2021.Paired errors showed GraphCast was significantly better from 18 hours to 4.75 days.
- Figure 3: Figure 3 compares cyclone median track error, cyclone paired error difference, atmospheric-river RMSE, and extreme-heat precision-recall across lead times.Error bars show bootstrapped 95% confidence intervals for cyclone medians and 95% confidence intervals for atmospheric-river RMSE.
- Extreme temperatures: GraphCast precision-recall curves exceeded HRES at 5- and 10-day lead times for extreme-heat classification.HRES performed better at the 12-hour lead time; results were consistent across several thresholds, variables, and extreme cold forecasting.
Effect of training data recency
Training GraphCast on more recent data improved its 2021 z500 skill scores, while the older model remained competitive with HRES. The comparison used variants trained through different years.
- Experimental setup: Four GraphCast variants used ERA5 data ending in 2017, 2018, 2019, or 2020 and were evaluated on 2021 test data.All variants began training with data from 1979.
- Results: Training through before 2021 further improved GraphCast’s z500 skill scores, although the variant trained through before 2018 remained competitive with HRES.The authors speculate that recent weather trends may be captured through retraining.
- Figure 4: Figure 4 plots 2021-test z500 RMSE skill scores relative to GraphCast trained through before 2018 across lead times.Colored lines represent training cutoffs from 2018 to 2021; the dashed line marks 3.5 days and the black line represents HRES.
Conclusions
The conclusions position GraphCast as an efficient, data-trained weather forecasting approach with downstream value, while identifying resolution, compute, uncertainty, and dependence on NWP-derived data as important boundaries.
- Conclusions: GraphCast’s forecast skill and efficiency compared to HRES indicate that MLWP methods are competitive with traditional weather forecasting methods.Its severe-event performance also demonstrates robustness and potential downstream value.
- Scope and constraints: The current GraphCast version uses 36.7 million parameters and 0.25° resolution because of memory, data-resolution, and engineering constraints.The authors view it as a model family that could scale with greater compute and higher-resolution data.
- Uncertainty: GraphCast’s deterministic forecasts do not explicitly represent uncertainty as ENS does, while MSE training can spatially blur predictions.The authors identify explicit uncertainty modeling as a crucial next step.
- Relationship to NWP: Data-driven MLWP depends critically on large quantities of high-quality data assimilated through NWP, so GraphCast is not presented as a replacement for traditional forecasting methods.Traditional methods have been rigorously tested and offer features not yet explored by GraphCast.
- Broader directions: Learned simulators trained on rich real-world data may support forecasting research in climate, ecology, energy, agriculture, and other complex dynamical systems.These are described as directions beyond weather forecasting.
Data and Materials Availability
GraphCast’s code and trained weights are publicly available, and the study uses publicly available ECMWF, NOAA, and related weather datasets.
- Data and Materials Availability: GraphCast’s code and trained weights are publicly available on GitHub.The repository is identified as github.com/deepmind/graphcast.
- Data and Materials Availability: The study uses publicly available ERA5, HRES, TIGGE, and IBTrACS Version 4 data under the stated access and licensing arrangements.ECMWF products are governed by CC BY 4.0.
Supplementary materials
The supplementary materials include Figures 5–53 and Tables 3–5.
- Figures 5–53 and Tables 3–5 are provided in the supplementary materials.
1. Datasets
The paper constructs training, evaluation, baseline, and tropical-cyclone datasets from ECMWF archives and IBTrACS, aligning GraphCast and HRES data for comparison.
- The datasets combine subsets of ECMWF archives with IBTrACS, distinguishing source archives from constructed datasets.
- ERA5: ERA5 represents global weather from 1959 onward at 0.25° resolution and hourly intervals across hundreds of variables.
- ERA5: The GraphCast ERA5 dataset covers 1979–2022, 37 pressure levels, and six-hour intervals at 00z, 06z, 12z, and 18z.
- HRES: HRES forecasts were spatially regridded to 0.25° and temporally downsampled to six-hour intervals for evaluation.
- HRES-fc0: HRES-fc0 uses the initial state of HRES forecasts at 00z, 06z, 12z, and 18z as evaluation ground truth.
- Tropical cyclones: Tropical-cyclone ground truth comes from IBTrACS tracks recorded every six hours with positions, categories, and meteorological features.
2. Notation and problem statement
The forecasting problem is to learn an efficient approximation of weather dynamics from partial observations and past context, then generate multi-step trajectories autoregressively.
- The true weather state evolves through a discrete-time dynamics function that generates the next state from the current state.
- Because observations incompletely represent the true state, the model receives the current state plus one or more past states.
- Predictions are fed back into the model to produce a full forecast trajectory over the chosen horizon.
- Forecast skill is assessed by comparing predicted trajectories with corresponding ground-truth trajectories using an objective function.
- The paper uses six-hour forecast steps and a maximum 10-day horizon, corresponding to 40 steps.
- ERA5 serves as ground truth for GraphCast, while HRES skill is evaluated against HRES-fc0.
3. GraphCast model
GraphCast is a graph-neural-network learned simulator that maps weather-grid inputs through a multi-scale icosahedral mesh and decodes forecasts back to the global grid. It generates 0.25° 10-day forecasts in under 60 seconds on one TPU.
- Autoregressive forecasting: GraphCast is a one-step learned simulator that predicts the next weather state from two consecutive input states.
- Architecture: Its encode-process-decode architecture uses GNNs to map grid features onto a multi-mesh, pass messages, and return grid forecasts.
- Multi-mesh: The multi-mesh enables long-range interactions with few message-passing steps while maintaining broadly homogeneous global spatial resolution.
- Multi-mesh: The mesh is built by refining an icosahedron six times, producing 40,962 nodes and multiscale edges for local and long-range communication.
- Efficiency: Under 60 seconds is sufficient for GraphCast to generate a 0.25°-resolution, 10-day forecast on one Cloud TPU v4 device.
- Efficiency: ECMWF’s IFS generates a 0.1°-resolution 10-day forecast in about an hour using an 11,664-core cluster.
- Inputs and outputs: The model predicts 227 variable-and-level combinations at each grid point, using additional static and external variables as context.
4. Training details
GraphCast training used causally separated ERA5 data, a weighted multi-step objective, autoregressive curriculum learning, and distributed memory-saving techniques.
- Data split: The development set covered 1979–2017, with training on 1979–2015 and validation on 2016–2017 before freezing model choices.
- Objective: The training objective minimized weighted mean squared error between GraphCast predictions and ERA5 targets across forecast steps, variables, levels, locations, and batches.
- Autoregressive training: GraphCast used autoregressive training by feeding predicted states back as inputs, with the final model trained for 12 steps corresponding to three days.
- Curriculum training: Training used three phases that progressed from one to 12 autoregressive steps while varying learning-rate schedules.
- Implementation: Training took about four weeks on 32 TPU devices, using batch parallelism, bfloat16 activations, and gradient checkpointing to reduce memory usage.
5. Verification methods
Verification used causally held-out data and carefully aligned ground truths, initialization windows, validity times, and skill metrics for GraphCast–HRES comparisons.
- Evaluation design: Four GraphCast versions were trained on progressively extended periods and evaluated on later years, preserving causal separation between training and test data.
- Ground truth: HRES was evaluated against an HRES-fc0 ground truth because verifying it against ERA5 would introduce non-zero error at the initial forecast step.
- Initialization alignment: GraphCast was evaluated from 06z/18z initializations so its ERA5 inputs had comparable lookahead to HRES inputs, while targets incorporated a 3h lookahead for both systems.
- Lead-time coverage: Like-for-like HRES comparisons extend to 3.5 days; from 4 days onward, HRES uses 00z/12z initializations while GraphCast remains at 06z/18z.
- Metrics: Skill was quantified using latitude-weighted RMSE and ACC, with normalized differences used for skill scores.
6. Comparison with previous machine learning baselines
Against the strong 0.25° Pangu-Weather baseline, GraphCast was evaluated from matched 00z/12z inputs and outperformed it on nearly all targets.
- Evaluation setup: GraphCast and Pangu-Weather were compared using the same 00z/12z initializations, which provided both models with the same +9 hours of lookahead.
- Overall comparison: 99.2% of targets favored GraphCast over Pangu-Weather.
- Surface variables: For surface variables, GraphCast’s error was around 10–20% lower during the first several days and around 7–10% lower at longer lead times.
- Exceptions: Pangu-Weather exceeded GraphCast on only two of 252 metrics: z500 at 6 and 12 hours, where GraphCast had 1.7% higher average RMSE.
7. Additional forecast verification results
Additional verification analyses examine GraphCast’s skill across variables, regions, resolutions, training horizons, and post-processing choices. They identify where GraphCast retains advantages, where performance is weaker, and how blurring and autoregressive training affect results.
- Biases: GraphCast’s biases were smaller than HRES’s for most variables up to 6 days but generally exceeded HRES’s at longer lead times.For 2m temperature, GraphCast’s biases began exceeding HRES’s at 4 days.
- Biases: GraphCast and HRES bias errors were uncorrelated or weakly correlated at short lead times, with correlations reaching 0.6 at 10 days.The correlation coefficient generally increased with lead time.
- Regional and vertical variation: GraphCast’s skill relative to HRES decreased with increasing surface elevation at short lead times, but this effect was less noticeable after 5 to 10 days.The short-lead advantage often fell below HRES’s at sufficiently high elevations.
- Training horizon and blurring: Longer autoregressive training horizons reduced the need for optimal blurring within the trained horizon, but still produced lower RMSE after blurring.Models trained with fewer autoregressive steps traded longer-lead accuracy for shorter-lead accuracy.
8. Additional severe event forecasting results
GraphCast’s severe-event evaluation covers cyclone tracking and examines robustness across event categories, tracker behavior, and statistical treatment. The results report comparable cyclone detection with HRES and stronger tracking performance in several settings.
- Scope: GraphCast’s severe-event evaluation spans cyclone tracking, atmospheric rivers, and extreme-temperature classification, despite not being specifically trained for these downstream tasks.These tasks probe spatial-temporal structure, nonlinear derivation of quantities, and characterization of extreme events.
- Cyclone tracking protocol: Cyclone-center error is computed as geodesic distance from the predicted center to the true center.
- Cyclone tracking protocol: Median track error is used because cyclone-track errors are non-Gaussian with many outliers, including tracker-related errors.The authors report that the median is less affected by tracking errors and better represents the overall population than the mean.
- Results: GraphCast significantly outperforms HRES in mean track error for lead times between 2 and 5 days.Using the final tracker, the mean results are similar to the median results, with significance assessed through paired analysis.
- Results: GraphCast and HRES drop a similar number of cyclones on the 2018–2021 test set, supporting a fair comparison.This check addresses the risk that blurred ML forecasts could cause more missed cyclones and distort positional-error comparisons.
- Results: GraphCast performs equally or better than HRES across cyclone categories, with significant advantages for category 2 and especially category 5 storms.The same pattern is reported when mean rather than median performance is measured.
9. Forecast visualizations
The paper presents representative GraphCast forecasts for seven weather variables. Each visualization uses a 2018 initialization and shows predictions at 2-, 6-, and 10-day lead times.
- Visualization set: Seven variables—2t, 10u, msl, z500, t850, v500, and q700—are visualized using representative predictions selected by median 2018 performance.
- Visualization examples: The 2t visualization uses a forecast initialized at 2018-05-05 12:00 UTC and displays 2-, 6-, and 10-day lead times.
- Visualization examples: The 10u visualization uses a forecast initialized at 2018-12-22 00:00 UTC and displays 2-, 6-, and 10-day lead times.
- Visualization examples: The msl visualization uses a forecast initialized at 2018-03-03 12:00 UTC and displays 2-, 6-, and 10-day lead times.
- Visualization examples: The z500, t850, v500, and q700 visualizations likewise show 2-, 6-, and 10-day lead times from their respective 2018 initializations.The initializations are 2018-11-23 12:00 UTC, 2018-11-12 12:00 UTC, 2018-03-30 12:00 UTC, and 2018-11-19 12:00 UTC, respectively.