climate-weather/ai produced the result/Nature 2024 · v2
Weather and climate model built from physics equations plus a trained neural network
Researchers built NeuralGCM, an atmospheric model that pairs a conventional solver for large-scale air flow with a neural network standing in for small-scale processes. The network was trained on decades of archived weather data and then used for forecasts and climate runs.
spectrum · one line per step, placed by what the step does · bright lines used AI
Neural general circulation models for weather and climate
Nature, 2024
doi:10.1038/s41586-024-07744-y · record aix-00297 v2 · checked 2026-11-05
- AI was for
- Forecasting, Simulation surrogate
- Model family
- Multilayer perceptron
- Checked by
- Benchmark732 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
The atmosphere is modelled by dividing the globe into a grid and stepping the equations of fluid motion forward in time. That works well for large features such as jet streams and weather fronts. It fails for anything smaller than a grid box: individual clouds, raindrops forming, the way sunlight is absorbed, swirls of air too fine to resolve. These still matter enormously to the forecast, so modellers write approximate formulas for them, known as parameterisations. Choosing and tuning those formulas rests heavily on expert judgement, and they are a long-standing source of error in both weather forecasts and climate projections.
The researchers set out to keep the physics solver but replace those hand-written small-scale formulas with a component learned from data. Because their solver is differentiable, meaning the effect of changing any internal setting can be traced mathematically through the whole simulation, the learned part could be fitted by running the model and comparing its output with the real atmosphere.
Where AI came in
A neural network, a fully connected one with the same weights applied to every vertical column of air, takes that column's conditions, their horizontal gradients, incoming sunlight, sea-ice cover and sea surface temperature, and outputs the rates of change due to unresolved processes: cloud formation, radiation, precipitation and fine-scale air motion. Related networks also correct the steps that translate between the archived data's coordinates and the solver's own. So the AI stands in for the hand-written parameterisations and the expert tuning behind them.
Training used ERA5, a reconstruction of past weather assembled by the European Centre for Medium-Range Weather Forecasts. The model was run forward and its errors pushed back through hundreds of simulation steps to adjust the network, with the stretches it had to predict lengthened from six hours to five days. The trained model then produced forecasts out to fifteen days and climate simulations lasting up to forty years.
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
NeuralGCM is an atmospheric model that couples a differentiable pseudo-spectral dynamical core, implemented in JAX, with a neural network that predicts the tendencies from unresolved processes in each vertical atmospheric column. The neural components, including learned encoder and decoder corrections, were trained end-to-end on ERA5 rollouts of 6 hours to 5 days, and the model was then run for 1- to 15-day weather forecasts and for multi-year and 40-year prescribed-SST climate simulations. Evaluated on WeatherBench2 variables over 732 forecasts initialized in 2020, NeuralGCM-0.7° and GraphCast gave the lowest RMSE at 1-3 day lead times depending on variable, and the stochastic NeuralGCM-ENS at 1.4° had lower error than ECMWF-ENS across almost all variables, lead times and vertical levels for ensemble-mean RMSE, RMSB and CRPS. In climate settings, 35 of 37 two-year runs and 22 of 37 forty-year runs completed without instability; the 1.4° model produced 83 tropical cyclones against 86 in ERA5 for the corresponding period, and the authors report that the model's response diverges from expectations for a +4 K sea surface temperature increase.
How AI was used
A fully connected neural network with residual connections and weights shared across atmospheric columns takes the column's prognostic variables, their horizontal gradients, incident solar radiation, sea-ice concentration and sea surface temperature as standardized inputs, and outputs scaled tendencies representing unresolved processes such as cloud formation, radiation, precipitation and subgrid dynamics. The same column-based architecture supplies learned corrections to the encoder and decoder that interpolate between ERA5 pressure levels and the solver's sigma coordinates. Because the dynamical core is differentiable, the network parameters were fitted by back-propagating through hundreds of simulation steps, with rollout length increased from 6 hours to 5 days during training. Deterministic models were trained with a combination of three losses in a spherical harmonics basis (a wavenumber-filtered mean squared error, a spectrum-matching term and a bias term), followed by fine-tuning of the decoder on 24-hour rollouts. Stochastic models took additional Gaussian random fields with learned spatial and temporal correlation as inputs to the encoder and learned physics, and were trained with a CRPS-based loss using two ensemble members per forecast. Learned physics tendencies were held constant across multiple ODE solver steps, typically 30 minutes of simulation time. Forecasts and climate runs were then produced by integrating the coupled model forward, with non-learned post-processing used for cyclone tracking and water-budget diagnostics.
The shape of the work
Structural · the record, drawn
Assemble ERA5 training and initialization data
Obtaining raw data, whether by measurement, download or retrieval.
NeuralGCM is trained on forecasting up to 5-day weather trajectories sampled from ERA5.where the paper describes this · verbatim
Encode reanalysis into model state with learned encoder
Encoding data into features, descriptors, embeddings or graphs. The AI stood in for conventional algorithm.
We additionally introduce learned corrections to both encoder and decoder stepswhere the paper describes this · verbatim
Train learned physics end-to-end through the differentiable solver
Fitting model parameters, including fine-tuning an existing model. The AI stood in for expert judgement.
we advance the model multiple time steps before employing stochastic gradient descent to minimize discrepancies between model predictions and reanalysiswhere the paper describes this · verbatim
Run medium-range weather forecasts
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
the dynamical core and learned physics (Fig. 1b) then compute tendencies, which are integrated in time using an implicit–explicit ordinary differential equation solverwhere the paper describes this · verbatim
Run multi-year and decadal climate simulations with prescribed SST
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
For extended climate simulations, we prescribe historical sea surface temperature (SST) and sea-ice concentration.where the paper describes this · verbatim
Diagnose derived fields and track tropical cyclones
Extracting understanding from model behaviour.
we use the tropical cyclone tracker TempestExtremes, as described in Supplementary Information section I.2where the paper describes this · verbatim
Evaluate against reanalysis, operational and baseline models
Testing outputs against ground truth.
average over all 732 forecasts made at noon and midnight UTC in the year 2020, which was held-out from training datawhere the paper describes this · verbatim
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.
The neural network learned-physics module is a constituent part of the model whose forecasts and climate simulations are the paper's result; the reported skill cannot be separated from it.
a GCM that combines a differentiable solver for atmospheric dynamics with machine-learning componentswhere the paper describes this · verbatim
Our evaluation set-up focuses on quantifying accuracy and physical consistency, following WeatherBench2.where the paper describes this · verbatim
both publicly available on GitHub at https://github.com/google-research/dinosaur and https://github.com/google-research/neuralgcmwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Version of NeuralGCM (deterministic; 2.8°, 1.4°, 0.7°)Which version of the model was used is not stated.
- Version of NeuralGCM-ENS (stochastic, 1.4°)Which version of the model was used is not stated.
- Version of GraphCastWhich 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-00297, 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). Weather and climate model built from physics equations plus a trained neural network. Record aix-00297, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00297/