NeuralGCM: A Hybrid Atmospheric General Circulation Model with a Differentiable Dynamical Core
New1. Overview
NeuralGCM was developed jointly by Google Research, Google DeepMind, the European Centre for Medium-Range Weather Forecasts, the Massachusetts Institute of Technology, and Harvard University. It is the first hybrid atmospheric general circulation model in which a differentiable dynamical core and a neural parameterization are fully coupled and trained end to end. The repository lives at neuralgcm/neuralgcm, with the earlier address google-research/neuralgcm now a 301 redirect. It was created on 5 February 2024, last pushed on 11 November 2026, and uses an Apache-2.0 license. At the time of verification for this entry it carries roughly 1037 stars, 145 forks, and 70 open issues, with main as the default branch.
Two formal publications correspond to the project. The first is the arXiv preprint 2311.07222, published on 13 November 2023 and updated to version 3 on 8 March 2024, with the comment noting 92 pages, 54 figures, Nature (2024). The second is the Nature article "Neural general circulation models for weather and climate", published online on 22 July 2024, volume 632, pages 1060 to 1066, DOI 10.1038/s41586-024-07744-y. The author list is substantial, with Dmitry Kochkov and Janni Yuval as co-first authors and Stephan Hoyer among the corresponding authors, spanning Google Research in Mountain View, Google DeepMind in London, ECMWF in Reading, MIT's Department of Earth, Atmospheric and Planetary Sciences, and Harvard's School of Engineering and Applied Sciences.
The project marks an important turn in the AI weather modeling lineage. Pure machine learning models such as GraphCast and Pangu had already matched or beaten traditional numerical models on deterministic forecast metrics, but the paper identifies three structural gaps in that approach: no calibrated uncertainty estimates; a mean squared error objective that rewards averaging over uncertainty and therefore produces blurry multi-day forecasts; and misrepresented derived quantities such as geostrophic wind. Traditional GCMs, meanwhile, rely on manual tuning and retain persistent errors, biases, and climate drift.
NeuralGCM's proposition is to combine both. The dynamical core handles resolved large-scale fluid motion and thermodynamics, while the learned module handles unresolved processes such as cloud formation, radiation, and precipitation. The idea is not new, but earlier hybrid models trained their machine learning components offline, independently of their interaction with the governing equations, and then inserted them into an existing GCM. That lack of coupling during training can cause instability and climate drift, and such models had largely been confined to idealized aquaplanet experiments. NeuralGCM differs in performing end-to-end online training.
Within this site, NeuralGCM represents the deep coupling of physics and machine learning. The GraphCast, Pangu-Weather, and FourCastNet entries collected here belong to the same forecasting direction but follow a different technical route: those are purely data-driven AI weather models, while NeuralGCM retains the conservation equations and the numerical solver.
2. How It Works
The architecture has two parts, shown in full in Figure 1.
The first is the differentiable dynamical core, implemented in JAX with automatic differentiation. It solves the hydrostatic primitive equations with moisture using horizontal pseudo-spectral discretization and vertical sigma coordinates, and it evolves seven prognostic variables: the vorticity and divergence of horizontal wind, temperature, surface pressure, and three water species, namely specific humidity together with specific ice and liquid cloud water content.
The second is the learned physics module. It follows the single-column approach of GCMs, using information from one atmospheric column alone to predict the effect of unresolved processes within that column. A fully connected neural network with residual connections implements this, with weights shared across all columns. Inputs include the prognostic variables in the column, total incident solar radiation, sea ice concentration, and sea surface temperature. The paper adds that horizontal gradients of the prognostic variables are also supplied, and that this was found to improve performance. All inputs are standardized to zero mean and unit variance using statistics precomputed at model initialization, and outputs are prognostic variable tendencies scaled by the fixed unconditional standard deviation of the target field.
Between data coordinates and model coordinates sits a pair of encoder and decoder components. They perform linear interpolation between pressure levels and sigma levels and additionally introduce learned corrections to both steps, using the same column-based network architecture as the learned physics module. The paper highlights one specific role of the encoder: it eliminates gravity waves from initialization shock, which would otherwise contaminate forecasts.
Time advancement uses an implicit-explicit ODE solver, whose step size is limited by the CFL condition on dynamics and can be small relative to the timescale of atmospheric change. The paper gives a concrete cost figure: evaluating learned physics is approximately 1.5 times as expensive as one dynamical core time step. Following normal GCM practice, learned physics tendencies are therefore held constant over multiple ODE time steps, typically corresponding to 30 minutes of simulation time. The appendix gives finer settings, with state updated iteratively every 4 to 30 minutes and learned physics tendencies recomputed every 60 minutes for the lowest resolution 2.8 degree model and every 30 minutes for the 1.4 and 0.7 degree models.
Training proceeds end to end. The model advances multiple time steps before stochastic gradient descent minimizes the discrepancy between predictions and conservatively regridded ERA5 data. Rollout length grows gradually from 6 hours to 5 days, which the paper calls critical because the models are not accurate for multi-day prediction early in training. Backpropagation through hundreds of simulation steps lets the neural networks account for interactions between learned physics and the dynamical core.
Two loss formulations are used. Deterministic models combine three terms. The first is a standard MSE computed in a spherical harmonics basis, with contributions from higher total wavenumbers progressively filtered out at longer lead times to address the double penalty problem, which would otherwise penalize placing high-wavenumber features in the wrong location beyond the predictability horizon. The second applies squared loss to the total wavenumber spectrum of the prognostic variables so the spectrum matches the training data. Both are evaluated on sigma and pressure levels. The third adds MSE on the batch-averaged mean amplitude of each spherical harmonic coefficient to discourage bias. Combined, these losses allow models trained on 3-day rollouts to remain stable during years-to-decades long simulations. Just before final evaluation, the decoder alone is fine-tuned on short 24-hour rollouts.
Stochastic models instead use the Continuous Ranked Probability Score, which balances a mean absolute error term encouraging accuracy against a similar term encouraging ensemble spread, summing CRPS in grid space and CRPS in the spherical harmonic basis below a maximum cutoff wavenumber for each variable, over rollout lengths from 6 hours to 5 days. Randomness enters by injecting Gaussian random fields with learned spatial and temporal correlation into the learned encoder and the learned physics module.
3. Scope and Assumptions
Four families of prerequisites apply.
Data prerequisites: training and evaluation use the public ERA5 reanalysis, originally downloaded from the Copernicus Climate Data Store and available in Zarr format at gs://gcp-public-data-arco-era5/ar/full_37-1h-0p25deg-chunk-1.zarr-v3. The evaluation set comprises 732 forecasts at noon and midnight UTC during 2020, a year held out from training data for all machine learning models.
Hardware prerequisites: the documentation explicitly recommends running on a machine with an attached GPU or TPU, otherwise performance is very slow. All training in the paper was done on Google Cloud TPUs with data parallelism, running one example per TPU device and distributing batches across multiple TPU cores, with additional model parallelism across spatial dimensions for the 0.7 degree and ensemble models.
Coordinate and variable prerequisites: NeuralGCM currently models the atmosphere only, with no ocean or land surface. Sea surface temperature and sea ice concentration must be supplied by the user as forcing data for climate simulation. Internal state uses non-dimensional units and simulation time rather than SI units and numpy.datetime64, with dedicated conversion utilities.
Interface prerequisites: the user-facing API centers on the PressureLevelModel class, offering from_checkpoint construction, encode and decode state conversion, advance and unroll for stepping and rolling forward, and methods to convert to and from xarray.Dataset. Because Xarray is not compatible with JAX, all conversions must go through these explicit interfaces.
Training capability is the project's clearest current limitation. The documentation states plainly that the team would like to support training, modifying, and fine-tuning NeuralGCM models, but with the present codebase based on Haiku and Gin configuration this is much trickier than it needs to be, and that in May 2024 the modeling code was being refactored to improve usability. The neuralgcm/legacy/ directory preserves pseudo-code and reference implementations of the paper's training procedure, including reference_code/ with experiment.py, metrics.py, stochastic_losses.py, a reader, and four paper configurations named deterministic_0_7_deg.gin, deterministic_1_4_deg.gin, deterministic_2_8_deg.gin, and stochastic_1_4_deg.gin.
4. Getting Started
The shortest path runs inference only. Installation takes one line with pip install neuralgcm, which pulls dependencies automatically, including JAX and the Dinosaur dynamical core package. Pretrained weights are hosted on Google Cloud Storage at gs://neuralgcm/models/ and are all distributed under a Creative Commons Attribution-ShareAlike 4.0 International license.
The documentation gives specific selection guidance. For weather forecasts, the 1.4 degree stochastic model is generally recommended because it performs best on most metrics, especially for 5 day and longer forecasts. If shorter lead times matter most, the 0.7 degree deterministic model is considerably slower but somewhat more accurate over the first few days. For seasonal to climate timescales, the higher resolution models are not stable: the unmodified 1.4 degree stochastic model can be relied on for roughly 6 months and the 1.4 degree deterministic model for roughly 2 years, with stability further enhanced by fixing mean log surface pressure. For multi-decadal climate simulation, the 2.8 degree stochastic models offer the bonus of predicting precipitation and evaporation separately. All of these cases require the user to supply their own sea surface temperature and sea ice concentration forcing, because NeuralGCM currently models the atmosphere only. Where precipitation is needed, either the 2.8 degree stochastic precipitation model or the 2.8 degree stochastic evaporation model can be used. Both predict precipitation and evaporation, and they differ in how they are trained. The precipitation model produces realistic cumulative precipitation at temporal resolutions of 6 hours or greater but struggles at finer sub-6-hour resolutions, while the evaporation model can produce slightly negative precipitation rates. These two models are slightly less accurate than the 1.4 degree stochastic model for weather, but they are reliably stable for 20-year atmosphere-only climate simulations.
The six available weight paths are v1/deterministic_0_7_deg.pkl, v1/deterministic_1_4_deg.pkl, v1/deterministic_2_8_deg.pkl, v1/stochastic_1_4_deg.pkl, v1_precip/stochastic_precip_2_8_deg.pkl, and v1_precip/stochastic_evap_2_8_deg.pkl. The first four correspond to the Nature paper and the last two to the precipitation-focused preprint, arXiv 2412.11973, released on 16 December 2024.
The repository also ships lightweight test data in neuralgcm/data/, an ERA5 tl31 file and a tl63 stochastic mini checkpoint, usable together with the neuralgcm.demo.load_data and load_checkpoint_tl63_stochastic helpers to exercise the pipeline without pulling large datasets from cloud storage.
One organizational evolution is worth noting. The main packages are base/ for core operators and parameterizations, classical/ for the primitive equations, coupled/ for coupling interfaces, hybrid/ for the hybrid model itself, ml/ for learned updates, and legacy/ for the older full implementation and paper reference code. The documentation also describes an experimental API built on coordax for coordinate-aware data representation and nnx for stateful model components, providing a Field data model, a Model subclassing interface, an immutable InferenceModel, and a VectorizedModel for batch and ensemble simulation, with Apache Beam compatibility; its tutorials have moved to a separate terrax repository. The documentation homepage states that it focuses mainly on the inference API, which the team believes is the most immediately useful part for third parties and the part they can commit to supporting in roughly its current form.
5. Comparison with Alternatives
Against pure machine learning weather models, the dividing line is whether a dynamical core is retained. The paper offers a memorable code-size comparison: GraphCast has roughly 5417 lines of core model code against roughly 376578 lines for NOAA's FV3 atmospheric model, a factor of about seventy. Pure machine learning approaches already lead on deterministic 1 to 10 day forecasting, but the paper's analysis identifies three structural limitations: uncalibrated uncertainty, blurriness caused by the mean squared error objective, and distorted derived quantities. NeuralGCM avoids these by keeping the dynamical core. The vertical structure of geostrophic wind is depicted more accurately than in GraphCast across various rollouts, though ECMWF-HRES still aligns slightly more closely with ERA5, and within NeuralGCM the vertical structure of the geostrophic wind degrades only slightly in the first few days with no noticeable change thereafter, particularly beyond day 5.
Against ECMWF, the paper separates deterministic HRES from ensemble ENS. Over the first 1 to 3 days, NeuralGCM-0.7 degree and GraphCast achieve the best RMSE, with slight variation across variables. At longer lead times RMSE rises rapidly because of chaotic divergence and becomes less informative, so the paper turns to RMSB, the root mean square of persistent error, which indicates longer-term behavior; NeuralGCM again compares favorably, with notably less bias for specific humidity in the tropics. For ensembles, the 1.4 degree NeuralGCM-ENS shows lower error than ECMWF-ENS across almost all variables, lead times, and vertical levels for ensemble-mean RMSE, RMSB, and CRPS, with similar spatial patterns of skill, and a spread-skill ratio of approximately one, which is a necessary condition for calibrated forecasts. The paper frames NeuralGCM as the first machine learning model to make accurate ensemble weather forecasts.
Against global cloud resolving models such as GFDL's X-SHiELD, the comparison falls in the climate section. The annual bias in precipitable water gives an RMSE of 1.09 mm for NeuralGCM, 1.74 mm for X-SHiELD, and 1.36 mm for climatology, and the temperature bias in the upper and lower troposphere is also lower for NeuralGCM. For tropical cyclones, NeuralGCM produces 83 at the coarse 1.4 degree resolution against 86 in ERA5 for the corresponding period, while X-SHiELD regridded to 1.4 degrees yields only 40.
Against the later precipitation-optimized variant, the original NeuralGCM directly predicts tendencies for an atmospheric column and therefore cannot distinguish precipitation from evaporation, a limitation the paper acknowledges and that the arXiv 2412.11973 preprint was written to address.
In the broader framing of this site, NeuralGCM shares a methodological thread with the GNoME and MACE entries: handing parameterizations that are hard to write or not differentiable to trainable components so that data tunes the parameterization rather than replacing the equations. The difference is that NeuralGCM strictly preserves the conservative structure, which is what allows it to extrapolate beyond the training distribution in lead time and in climate state.
6. Benchmarks and Evidence
The evaluation protocol follows WeatherBench2. All forecasts are regridded conservatively to a 1.5 degree grid and averaged over 732 forecasts initialized at noon and midnight UTC during 2020. NeuralGCM, GraphCast, and Pangu use ERA5 as ground truth, while the ECMWF ENS and HRES are referenced to the ECMWF operational analysis, that is HRES at zero lead time, to avoid penalizing operational forecasts for biases that differ from ERA5.
Accuracy metrics comprise RMSE, bias, CRPS, and spread-skill ratio. Concrete figures in the paper include a global mean temperature at 850 hPa ensemble-mean RMSE of 0.16 K against 0.45 K for climatology; an annual precipitable water bias RMSE of 1.09 mm against 1.36 mm for climatology and 1.74 mm for X-SHiELD; and a tropical cyclone count of 83 against 86 in ERA5 and 40 for X-SHiELD.
Stability is the decisive test for a hybrid model, and the paper gives figures at two durations. For two-year simulations with NeuralGCM-1.4 degree, 37 initial conditions spaced every 10 days through 2019 were used, and 35 completed the full two years without instability. For 40-year AMIP-like simulations with NeuralGCM-2.8 degree, 37 initial conditions spaced every 10 days during 1980 were run, of which 22 remained stable for the entire 40-year period; the analysis focuses on those 22 and compares them with 22 AMIP experiments from CMIP6 Phase 6. All 22 NeuralGCM runs have smaller spatially averaged bias over 1981 to 2014 than the CMIP6 AMIP runs, and this remains true even after removing the global temperature bias from the CMIP6 runs.
Generalization to unseen climate is tested directly. AMIP simulations with increased sea surface temperature show that NeuralGCM exhibits some robust features of the warming response for modest increases of 1 K and 2 K, while at a more substantial 4 K increase the response diverges from expectations, and these increased-SST AMIP simulations also exhibit climate drift, underscoring the limitations.
For unseen data, the paper compares NeuralGCM-0.7 degree and GraphCast trained through 2017 over five years of forecasts beyond the training period from 2018 to 2022, finding that NeuralGCM shows no clear trend of increasing error when initialized further into the future from the training data, unlike GraphCast. Extending the test beyond five years, the authors trained a NeuralGCM-2.8 degree model on data before 2000 and tested its skill over 21 unseen years.
Compute is the most persuasive part of the work. The paper states that compared with traditional GCMs of similar skill, NeuralGCM is compute efficient and low complexity, running at 8 to 40 times coarser horizontal resolution than ECMWF IFS and global cloud resolving models, which enables 3 to 5 orders of magnitude savings in compute. The concrete comparison is that NeuralGCM-1.4 degree simulates 70000 simulation days in 24 hours on a single TPU, while X-SHiELD manages 19 simulated days on 13824 CPU cores. Appendix G6 gives finer figures. Training time and hardware are 1 day on 16 TPU v4 for the 2.8 degree deterministic model, 1 week on 16 TPU v4 for the 1.4 degree deterministic model, 3 weeks on 256 TPU v4 for the 0.7 degree deterministic model, and 10 days on 128 TPU v5e for the ensemble model. Parameter counts are 14.5 M, 18.3 M, 31.1 M, and 11.5 M respectively. Inference time on one TPU v4 core for a 10-day forecast is 2.5 seconds, 12.6 seconds, 119 seconds, and 12.4 seconds, with inference speeds of 350000, 69000, 7300, and 70000 simulation days per day and peak inference memory of 255 MB, 1100 MB, 4177 MB, and 1011 MB.
Training hyperparameters are equally explicit. All models use Adam with beta1 of 0.9, beta2 of 0.95, and epsilon of 1e-6. The learning rate rises linearly to its peak over the first 2000 steps, stays constant until step 15000, and then decays exponentially at rate 0.5 with a decay time of 10000 steps. Peak learning rates are 0.002 for the 2.8 and 1.4 degree models and 0.001 for the 0.7 degree and ensemble models, with training steps of 38000, 26000, 25000, and 43000 respectively.
7. Role in Physics-AI Workflows
On the operational weather side, NeuralGCM demonstrates a route previously considered unavailable to pure machine learning approaches: ensemble forecasting. Ensembles require not only accurate mean fields but also spread matched to error, with a spread-skill ratio near one. The way NeuralGCM achieves this is instructive. It does not predict precipitation directly but diagnoses precipitation minus evaporation, because advection is handled by the dynamical core and the machine learning parameterization models only local processes within atmospheric columns, so local sources and sinks can be separated from tendencies due to horizontal transport. This division of labor, delegating the diagnosable to physics and the non-diagnosable to learning, is the most transferable methodological point in the work.
On the climate side, it provides a capability pure machine learning models had struggled to reach: multi-decadal integration under prescribed historical sea surface temperature that reproduces historical warming trends, the Hadley circulation, monsoons, and the frequency and trajectories of cyclones. The paper explicitly frames this as strong evidence for a disputed hypothesis, namely that learning to predict short-term weather is an effective way to tune parameterizations for climate, since models trained on 72-hour forecasts are capable of realistic multi-year simulation.
From a cross-domain perspective, the discussion carries an unusually broad claim. Models based on physical laws and empirical relationships are ubiquitous in science, and the paper argues that the differentiable hybrid modeling approach of NeuralGCM has the potential to transform simulation for a wide range of applications, naming materials discovery, protein folding, and multiphysics engineering design. Those three directions correspond on this site to the materials discovery entry, the protein structure prediction entry, and the multiphysics coupling collection. The comparison underscores that NeuralGCM's real contribution is the hybrid modeling paradigm itself, with weather forecasting being only its first successful case.
There is also a methodological value worth noting. The paper observes that the training strategies which promoted a more realistic spectrum and the ability to produce ensemble forecasts could equally be applied in purely machine learning approaches, which points to a way of improving data-driven models without discarding physics.
8. Limitations and Failure Modes
The failure modes disclosed by the paper and documentation map onto concrete constraints.
Long-integration instability is endemic to hybrid models, and NeuralGCM reduces rather than eliminates it. Two years of simulation keeps 35 of 37 initial conditions stable, while 40 years keeps 22 of 37, and that latter failure rate of roughly forty percent underlines that running once and running for decades are different things.
Climate drift appears in the warming scenario. The response is acceptable for 1 K and 2 K sea surface temperature increases but diverges clearly at 4 K, with accompanying drift, which means extrapolative capability cannot be inferred from performance under historical forcing alone.
The trade-off between resolution and stability is a hard constraint. The documentation states plainly that for seasonal to climate timescales the higher resolution NeuralGCM models are not stable, so climate simulation must fall back to 2.8 degrees, which directly limits characterization of fine-scale climate phenomena.
Inability to separate precipitation from evaporation is a structural limitation of the original model, which directly predicts column tendencies. The documentation notes that the precipitation model struggles below 6-hour temporal resolution while the evaporation model can produce slightly negative precipitation rates.
The cost threshold for ensemble forecasting is real, with 1.4 degree ensemble inference requiring 12.4 seconds and about 1011 MB of memory, so member count scales cost linearly. The documentation lists large-scale ensemble forecasting among the new possibilities this efficiency enables, but that still requires proportional hardware investment.
Training support is the most practical obstacle. The documentation states explicitly that with the current Haiku plus Gin codebase, third parties training or fine-tuning models would find it much trickier than necessary, and the model-building and training pseudo-code left in neuralgcm/legacy/reference_code/ serves as paper reproduction material rather than an engineering interface.
Pure atmosphere limitation: there is no coupling to ocean or land surface, and no chemical composition module for greenhouse gases or aerosols. The paper lists coupling to other Earth system components as a next reformulation step for climate projection.
Hidden cost in code volume and complexity cuts both ways. The reduction by two orders of magnitude relative to FV3 is a positive, but the global spectral method for the dynamical core and the fully connected networks for learned physics do not remove the implementation and numerical analysis barriers. The paper notes that substantial headroom remains to pursue higher accuracy using advanced numerical methods and machine learning architectures.
9. Version and Source Notes
Repository address change: the code availability section of both the arXiv version 3 from March 2024 and the Nature paper writes https://github.com/google-research/neuralgcm and https://github.com/google-research/dinosaur. Both addresses are now 301 redirects, and the actual repositories are neuralgcm/neuralgcm and neuralgcm/dinosaur. Citations should use the new addresses and note the redirect.
Two publication tracks: arXiv 2311.07222, first released on 13 November 2023 with version 3 on 8 March 2024 at 92 pages and 54 figures, and the Nature article in volume 632, pages 1060 to 1066, published online on 22 July 2024 with DOI 10.1038/s41586-024-07744-y, overlap heavily but carry different version dates, so a citation must name which one is meant.
Layered licensing: the README distinguishes code under the Apache License 2.0 from trained model weights under Creative Commons Attribution-ShareAlike 4.0 International, and the checkpoints documentation likewise states that all checkpoints are distributed by Google under CC BY-SA 4.0. This differs from arrangements such as OpenFold parameters under CC BY 4.0, since the share-alike clause carries its own obligations.
Numerical conventions must not be mixed: ensemble and deterministic time series are not directly comparable, and the 1.4 degree ensemble model and the 1.4 degree deterministic model are different products with different stability limits, roughly 6 months and roughly 2 years respectively. The 2.8 degree precipitation model and the 2.8 degree evaporation model each output both precipitation and evaporation, differing only in how they were trained.
The denominator behind stability statistics: both the 35 out of 37 for two years and the 22 out of 37 for 40 years count initial conditions spaced every 10 days and count only runs that completed without instability. The paper additionally notes that the 40-year analysis focuses on the 22 stable runs and compares them with 22 CMIP6 AMIP experiments.
The hardware premise of the compute comparison: the 70000 simulation days in 24 hours on a single TPU against 19 simulated days on 13824 CPU cores is a single-TPU versus 13824-CPU-core figure, drawn from the same source as the inference speeds reported per single TPU v4 core in Appendix G6, and must not be restated as a single-GPU claim.
Resolution naming: 0.7, 1.4, and 2.8 degrees are horizontal grid spacings, with the vertical direction uniformly using 32 equidistant sigma levels.
Release timing: the repository was created on 5 February 2024, later than both the November 2023 arXiv first version and the July 2024 Nature first version, so the timing of public code and of publication does not coincide.
10. References
- NeuralGCM paper, "Neural General Circulation Models for Weather and Climate", arXiv 2311.07222, published 13 November 2023, version 3 updated 8 March 2024, 92 pages and 54 figures.
- NeuralGCM formal paper, "Neural general circulation models for weather and climate", Nature, volume 632, pages 1060 to 1066, 22 July 2024, DOI 10.1038/s41586-024-07744-y.
- NeuralGCM official repository: https://github.com/neuralgcm/neuralgcm , Apache-2.0, with google-research/neuralgcm as a 301 redirect.
- NeuralGCM official documentation: https://neuralgcm.readthedocs.io/ , covering installation, inference, the checkpoint list, and the experimental API.
- Pretrained weight hosting page on Google Cloud Storage:
gs://neuralgcm/models/, all six checkpoints distributed under CC BY-SA 4.0. - Dinosaur repository for the differentiable dynamical core: https://github.com/neuralgcm/dinosaur , Apache-2.0.
- Precipitation-optimized paper, "Neural general circulation models optimized to predict satellite-based precipitation observations", arXiv 2412.11973, 16 December 2024.
- WeatherBench2 evaluation project: https://github.com/google-research/weatherbench2 , Apache-2.0, with the evaluation protocol and data guide in its documentation.
- ERA5 reanalysis dataset paper, "The ERA5 global reanalysis", Quarterly Journal of the Royal Meteorological Society, 2020.
- GraphCast paper, "Learning skillful medium-range global weather forecasting", Science volume 382, pages 1416 to 1421, 2023.
- Pangu-Weather paper, "Accurate medium-range global weather forecasting with 3D neural networks", Nature volume 619, pages 533 to 538, 2023.
- WeatherBench 2 preprint, "WeatherBench 2: A benchmark for the next generation of data-driven global weather models", arXiv 2308.15560.
- Koopman perspective article, "Are general circulation models obsolete?", PNAS, 2022.
- GFDL X-SHiELD global cloud resolving model data, GA DoE open data bucket
gs://ai2cm-public-requester-pays/C3072-to-C384-res-diagnostics. - RAOBCORE V1.9 radiosonde observation correction reanalysis dataset, used as reference for tropical temperature trends.
- CMIP6 data catalog: https://catalog.pangeo.io/browse/master/climate/ .
- NeuralGCM official mailing list for release announcements: https://groups.google.com/g/neuralgcm-announce .
- Numerical method reference for the Dinosaur dynamical core, covering the pseudo-spectral discretization of the primitive equations and the hydrostatic balance framework in appendices B and C of the paper.