~/aixsci
weekly updates

climate-weather/ai produced the result/Nature 2023 · v2

Neural networks trained on 39 years of weather data produce week-ahead global forecasts

Researchers trained four transformer networks on decades of past atmospheric data to predict the state of the global atmosphere. The networks replaced the physics simulation that normally generates a forecast, producing one in 1.4 seconds on a single graphics processor.

1. Obtain reanalysis, baseline forecast and cyclone track data2. Select variables, levels and splits; normalise fields3. Train four 3DEST forecast models at different lead times4. Produce deterministic forecasts by hierarchical temporal aggregation5. Run perturbed-initial-state ensemble forecasts6. Locate and track cyclone eyes in forecast fields7. Score forecasts and tracks against held-out data and baselines

spectrum · one line per step, placed by what the step does · bright lines used AI

Accurate medium-range global weather forecasting with 3D neural networks
Nature, 2023

doi:10.1038/s41586-023-06185-3 · record aix-00296 v2 · checked 2026-11-05

ai-resultrole of AI
AI was for
Forecasting
Model family
Transformer
Checked by
Held-out
Code
available

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes, never the evidence.

Introduction by AIxSci · plain language

What this research was about

A weather forecast begins with a snapshot of the atmosphere: temperature, pressure, humidity and wind, measured all over the globe and at many heights. The usual method is to feed that snapshot into equations describing how air moves and heats, then step the equations forward in time. This works, but it is costly. The equations must be solved on a fine grid over the whole planet, which takes a large supercomputer and considerable time. Small errors in the starting snapshot also grow as the forecast runs on, which is why forecasts beyond about a week become unreliable. Shortening the cost of a forecast, without losing accuracy, has been a long-standing aim.

An alternative is to skip the equations and learn the atmosphere's behaviour directly from its history. Decades of reconstructed past weather exist, in the form of reanalysis: a record that blends observations with model output to give a consistent global picture hour by hour. The researchers set out to train neural networks on that record so that, given one atmospheric state, a network returns the state a fixed number of hours later.

Where AI came in

The forecasts in this work are the networks' output. The team trained four separate networks from scratch on reanalysis fields covering 1979 to 2017, each learning to jump ahead by a fixed interval: one hour, three hours, six hours or twenty-four hours. Each network handled 69 variables, five upper-air quantities on 13 pressure levels plus four surface quantities, and treated them as a single three-dimensional volume rather than a stack of flat maps. The networks are transformers, an architecture that weighs how parts of an input relate to each other, here adapted so that position on the globe and height in the atmosphere are built into those comparisons.

To reach a longer forecast, the trained networks were called one after another, always using the largest step available, so a week-ahead forecast is a chain of jumps rather than a simulation. The same networks were also run from a hundred slightly disturbed starting states to give a spread of possible outcomes. What the networks stood in for is the physics-based forecast model itself. Steps either side of them stayed conventional: selecting and normalising the data, locating cyclone centres in the forecast fields by looking for pressure minima, and scoring the results against held-out years and existing forecasts.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.

The work

Technical · from the record

Pangu-Weather is a set of four three-dimensional Earth-specific transformer networks trained on 39 years of ERA5 reanalysis data (1979-2017) to map a global atmospheric state to the state at a fixed lead time of 1 h, 3 h, 6 h or 24 h, covering 69 variables at 0.25 degree resolution. Longer forecasts are built by chaining these models with a greedy hierarchical temporal aggregation scheme that uses the largest affordable lead time at each step. On the 2018 ERA5 test year the paper reports lower RMSE and higher ACC than the operational IFS of ECMWF and than FourCastNet for each tested variable, including an RMSE of 296.7 for a 5-day Z500 forecast against 333.7 for the operational IFS and 462.5 for FourCastNet, with an inference cost of 1.4 s on a single GPU. Tracking the eyes of 88 named 2018 tropical cyclones through forecast mean sea-level pressure minima gave 3-day and 5-day mean direct position errors of 120.29 km and 195.65 km, against 162.28 km and 272.10 km for ECMWF-HRES, a comparison the authors note is uneven because ECMWF-HRES was initialized from IFS initial conditions rather than reanalysis.

How AI was used

ERA5 hourly reanalysis was reduced to 5 upper-air variables on 13 pressure levels plus 4 surface variables and 3 constant masks, normalised per two-dimensional field using means and standard deviations computed over 1979-2017, and split into training (1979-2017), validation (2019) and test (2018) sets. Four separate 3D Earth-specific transformer networks, each about 64 million parameters, were trained from scratch with a weighted mean-absolute-error loss and the Adam optimizer for 100 epochs on 192 Tesla-V100 GPUs to predict the atmospheric state at lead times of 1 h, 3 h, 6 h and 24 h from a single input state. The architecture embeds upper-air and surface variables into a joint 3D volume by patch embedding, propagates it through an 8-layer encoder and 8-layer decoder of shifted-window transformer blocks with an Earth-specific positional bias indexed by absolute pressure-level and latitude coordinates, and restores resolution by patch recovery. At inference, forecasts for arbitrary lead times are produced by iteratively calling these models under a greedy hierarchical temporal aggregation rule. The same models were run on 100 initial states perturbed with three octaves of Perlin noise to form an ensemble, and their forecast fields were post-processed by a non-learned algorithm that locates local mean sea-level pressure minima to track tropical cyclone eyes.

The shape of the work

Structural · the record, drawn

ACQUISITIONPREPARATIONTRAININGSIMULATIONSIMULATIONINTERPRETATIONVALIDATION1234567AIAIAIObtainreanalysis,baseline forecas…Select variables,levels andsplits; normalis…Train four 3DESTforecast modelsat different lea…Producedeterministicforecasts by hie…Runperturbed-initial-stateensemble forecas…Locate and trackcyclone eyes inforecast fieldsScore forecastsand tracksagainst held-out…↤ simulation↤ simulation↤ simulation
AI stepNo AI↤ what the AI stood in for

Obtain reanalysis, baseline forecast and cyclone track data

Obtaining raw data, whether by measurement, download or retrieval.

we downloaded a subset of the ERA5 dataset (around 60 TB) from https://cds.climate.copernicus.eu/where the paper describes this · verbatim
in the paper

Select variables, levels and splits; normalise fields

Cleaning, filtering, normalising or labelling data already obtained.

We followed WeatherBench to choose 13 out of 37 pressure levelswhere the paper describes this · verbatim
in the paper

Train four 3DEST forecast models at different lead times

Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.

The four individual models were trained for 100 epochs using the Adam optimizer.where the paper describes this · verbatim
in the paper

Produce deterministic forecasts by hierarchical temporal aggregation

Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.

we introduced hierarchical temporal aggregation, a greedy algorithm that always calls for the deep network with the largest affordable lead timewhere the paper describes this · verbatim
in the paper

Run perturbed-initial-state ensemble forecasts

Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.

we obtained a 100-member ensemble forecast by simply averaging the forecast resultswhere the paper describes this · verbatim
in the paper

Locate and track cyclone eyes in forecast fields

Extracting understanding from model behaviour.

We followed a classical algorithm that locates the local minimum of MSLP to track the eye of tropical cyclones.where the paper describes this · verbatim
in the paper

Score forecasts and tracks against held-out data and baselines

Testing outputs against ground truth.

validated them on 2019 data and tested them on 2018 datawhere the paper describes this · verbatim
in the paper

What the record says

Technical · every part carries its own basis

+ in the paper~ our reading− not reported

How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.

~Role of AI
AI produced the resultour reading

The forecasts the paper reports are produced by the trained deep networks themselves; every reported result (deterministic forecast skill, ensemble mean, cyclone tracks) derives from model output.

+What the AI was for
Forecastingin the paper
We further designed a three-dimensional (3D) Earth-specific transformer (3DEST) architecture to inject Earth-specific priors into the deep networks.where the paper describes this · verbatim
+Model families
Transformerin the paper
+How it was taught
Supervisedin the paper
+How results were checked
Held-outin the paper
validated them on 2019 data and tested them on 2018 datawhere the paper describes this · verbatim
+Code · weights · data
code availableweights availabledata availablein the paper
We released the trained models, inference code and the pseudocode of details to the public at a GitHub repositorywhere the paper describes this · verbatim
+Compute
Training: approximately 16 days per model for 100 epochs on a cluster of 192 NVIDIA Tesla-V100 GPUs, batch size 192 (1 sample per GPU). Inference: 1.4 s on a single Tesla-V100 GPU.in the paper

What this paper did not report

Technical · absence is published deliberately

Reported as not stated: 2 items
  • How many were testedThe paper gives no count of what was tested.
  • Version of Pangu-Weather (3D Earth-specific transformer, 3DEST)Which version of the model was used is not stated.

About this article

This paper was chosen for the collection rather than drawn at random, so it is left out of the statistics this site gives about the field. Record aix-00296, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error

Cite this record: AIxSci (2026). Neural networks trained on 39 years of weather data produce week-ahead global forecasts. Record aix-00296, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00296/