Towards Hybrid Earth System Modeling: A Living Review

October 13, 2025 · View on GitHub

This page presents an alphabetical review of emerging approaches that bring machine learning (ML) into Earth system modeling to simulate the full time evolution of climate variables in response to diverse forcings. These models hold promise for improving long-term projections of Earth's physical climate and biogeochemical cycles. Although new creative approaches continue to emerge, most designs fall into two broad categories:

  1. Hybridizing existing Earth system models (ESMs)
    Hybrid ESMs build on established ESM codebases, retaining key physics components (e.g., the dynamical core) while replacing or improving parameterizations of hard-to-model processes (e.g., storm formation) with ML. This often involves interfacing Fortran-based codebases with Python-based ML tools. For related technical resources, see this living review maintained by Julien Le Sommer and Alexis Barge.

  2. Developing data-driven climate models from scratch
    Climate emulators write prognostic equations directly in differentiable programming frameworks, incorporating explicit physical laws (e.g., conservation equations) only when needed. This is a longer-term endeavor, involving the progressive development and coupling of the atmosphere–ocean–land–cryosphere components.

If you notice any errors, omissions, or outdated information, please feel free to submit a pull request.

Author: Tom Beucler (UNIL); written in the context of AI4PEX and the WCRP Lighthouse Activities.

How to cite

Beucler, T. (2025). Towards Hybrid Earth System Modeling: A Living Review (v1.0). Zenodo. https://doi.org/10.5281/zenodo.16967529

Table of Contents


ACE

The Ai2 Climate Emulator (ACE) emulates NOAA's FV3GFS atmospheric model using spherical Fourier neural operators. ACE operates with six prognostic variables, can be forced through insolation and sea surface skin temperature, diagnoses radiative and energy fluxes at the atmosphere's boundaries, and runs on a single GPU. ACE2 improves upon ACE by enforcing global conservation of dry air mass and humidity, making it a hybrid climate model and improving climate stability and surface pressure representation. ACE2, which can be coupled to a slab ocean or a 3D ocean emulator, is trained and tested on historical climate reanalysis (1940-2020) and 100 km-resolution Unified Forecast System (UFS) simulations forced by historical sea surface temperatures and greenhouse gas concentrations.

Latest simulations in Duncan, J. P. C., Wu, E., Dheeshjith, S., Subel, A., Arcomano, T., Clark, S. K., ... & Bretherton, C. (2025). SamudrACE: Fast and Accurate Coupled Climate Modeling with 3D Ocean and Atmosphere Emulators. arXiv:2509.12490

See also:


CAMulator

CAMulator v1 is a machine-learned emulator of the Community Atmosphere Model v6 (CAM6) that predicts atmospheric states from sea surface temperatures and solar radiation. It conserves key physical quantities, captures major climate patterns like ENSO and NAO, and runs 350× faster than CAM6—enabling large-scale, physically grounded climate simulations. While it exhibits a cold bias in high-latitude winters outside its training range, CAMulator represents a major advance toward fast, realistic ML-based climate modeling.

Latest simulations in Chapman, W. E., Schreck, J. S., Sha, Y., Gagne II, D. J., Kimpara, D., Zanna, L., ... & Berner, J. (2025). CAMulator: Fast Emulation of the Community Atmosphere Model. arXiv preprint 2504.06007.

See also:


CBRAIN

Cloud Brain (CBRAIN) aims to break the convective parameterization deadlock in the Community Atmosphere Model (CAM) by training neural networks to emulate the total subgrid thermodynamic time tendencies. These tendencies represent the cumulative tendencies of prognostic thermodynamic variables (temperature and specific humidity) due to subgrid-scale processes such as convection, radiation, and turbulence.

Latest simulations in Lin, J., Yu, S., Peng, L., Beucler, T., Wong-Toi, E., Hu, Z., ... & Pritchard, M. S. (2024). Sampling Hybrid Climate Simulation at Scale to Reliably Improve Machine Learning Parameterization. arXiv preprint 2309.16177.

See also:


CliMA

The Climate Modeling Alliance (CliMA) is developing a new Earth system model that includes atmosphere, ocean, and land components, leveraging GPUs and advances in machine learning and data assimilation. The models are first pretrained with high-resolution simulations (e.g., for atmospheric turbulence and convection) or site-level field data (for the land model) and are then fine-tuned with global observational data. A key aspect of the project is the use of a common derivative-free ensemble Kalman methodology to train the models with climate statistics, rather than focusing only on weather state increments. This ensures that the models reproduce long-term climate statistics well. The models incorporate both physics-based dynamical equations and machine learning components to represent processes such as snow or turbulence and convection in the atmosphere.

Selected Publications (Full List Available Here)


ClimSim

ClimSim, the first benchmark dataset for hybrid ML-physics climate emulation, includes simulation data from the Energy Exascale Earth System Model Multi-scale Modeling Framework (E3SM-MMF). E3SM-MMF embeds GPU-accelerated cloud-resolving models within each grid cell of E3SM and uses explicit scalar momentum transport to ensure the quality of subgrid-scale fluxes. ClimSim provides billions of multivariate input/output vector pairs, capturing the aggregate effect of cloud-resolving models on E3SM's macro-scale state. ClimSim also inspired a Kaggle competition and includes an end-to-end workflow for developing hybrid ML-physics simulators.

Latest simulations in Hu, Z., Subramaniam, A., Kuang, Z., Lin, J., Yu, S., Hannah, W. M., ... & Pritchard, M. S. (2024). Stable Machine-Learning Parameterization of Subgrid Processes with Real Geography and Full-physics Emulation. arXiv preprint 2407.00124.

See also:


Corrective ML

Building on early efforts to enhance subgrid-scale physics through machine learning with near-global storm-resolving aquaplanet simulations, AI2 has developed a series of data-driven solutions to improve the (thermo)dynamics of FV3-GFS, the atmospheric component of the Unified Forecast System (UFS). The latest efforts focused on learning apparent dynamic tendencies to nudge temperature and humidity toward a reference state derived from a global storm-resolving GFDL X-SHiELD simulation, informally called "Corrective ML."

Latest simulations in Watt‐Meyer, O., Brenowitz, N. D., Clark, S. K., Henn, B., Kwa, A., McGibbon, J., ... & Bretherton, C. S. (2024). Neural network parameterization of subgrid‐scale physics from a realistic geography global storm‐resolving simulation. Journal of Advances in Modeling Earth Systems, 16(2), e2023MS003668.

See also:


DLESyM

The Deep Learning Earth SYstem Model (DLESyM) couples a deep learning weather prediction model with a deep learning prognostic sea surface temperature (SST) model. The resulting model is parsimonious, using an order of magnitude fewer prognostic variables per grid point (9 for the atmosphere and only SST for the ocean), partly due to its use of the Hierarchical Equal Area isoLatitude Pixelization (HEALPix) spatial grid.

Latest simulations in Cresswell-Clay, N., Liu, B., Durran, D., Liu, B., Durran, D. R., Liu, Z., Espinosa, Z. I., Moreno, R. A., & Karlbauer, M. (2025). A deep learning Earth system model for efficient simulation of the observed climate. AGU Advances, 6, e2025AV001706.


Hybrid ARP-GEM

Hybrid ARP-GEM1 combines the dynamical core of the new global atmospheric model ARP-GEM1 (Global, Efficient, and Multiscale version of ARPEGE version 1) with neural network-based parameterizations. It employs the Python interface of the Message Passing Interface-based “field-exchange” method OASIS3, enabling neural network integration on heterogenous High-Performance Computing (HPC) architectures. Initial prototypes emulate deep learning parameterization, and Hybrid ARP-GEM1's modular design enables the coupling of diverse data-driven parameterizations in the near term.

Latest simulations in Balogh, B., Saint-Martin, D., & Geoffroy, O. (2024). Online test of a neural network deep convection parameterization in ARP-GEM1. arXiv preprint 2410.21920


Hybrid CAM

Hybrid versions of the Community Atmosphere Model (CAM) preserve CAM's dynamical core while replacing uncertain CAM parameterizations with ML surrogates. This includes the emulation of moist physics and warm microphysics.

Latest simulations in Chen, J., Zhang, M., Zhang, T., Lin, W., & Xue, W. (2025). Stable simulation of the community atmosphere model using machine‐learning physical parameterization trained with experience replay. Journal of Advances in Modeling Earth Systems, 17(6), e2024MS004722.

See also:


Hybrid HadGEM

HadGEM3-GC5.0 is the latest global coupled configuration of the UK Met Office Hadley Centre model, combining the Global Atmosphere and Land v9.0 and the Global Ocean and Sea Ice v9 components. In an ML-enhanced configuration, a machine-learned cloud fraction/condensate scheme is coupled to GC5.0, replacing the model’s native prognostic cloud fraction/condensate parameterization.

Latest simulation in Morcrette, C., Cave, T., Reid, H., da Silva Rodrigues, J., Deveney, T., Kreusser, L., ... & Budd, C. (2025). Scale‐aware parameterization of cloud fraction and condensate for a global atmospheric model machine‐learned from coarse‐grained kilometer‐scale simulations. Journal of Advances in Modeling Earth Systems, 17(4), e2024MS004651.


Hybrid Land Surface Modeling

Attempts to augment global land surface models with ML components that learn hard-to-parameterize processes are emerging, from data-driven fire schemes coupled to carbon fluxes and vegetation to emulators of surface fluxes (e.g., evaporation). ML models are trained on observations and high-resolution products, often with physical constraints to preserve the host model's water/energy/carbon budgets.

See:


Hybrid SAM

Using the hypohydrostatic configuration of the System for Atmospheric Modeling (SAM), quasi-global aquaplanet simulations can represent convection and large-scale circulation simultaneously at horizontal resolutions as coarse as 12 km. This provides an ideal testbed for machine learning parameterization approaches, such as leveraging non-local information across grid columns to model subgrid momentum fluxes, employing reduced-precision computations, and making parameterizations scale-aware. Numerical stability is ensured through a per-process flux prediction framework and a short integration timestep, which allow the use of tailored SAM prognostic equations and precipitation diagnostics.

Latest simulations in Yuval, J., & O’Gorman, P. A. (2023). Neural‐network parameterization of subgrid momentum transport in the atmosphere. Journal of Advances in Modeling Earth Systems, 15(4), e2023MS003606.

See also:


Hybrid SPEEDY

Hybrid SPEEDY builds on the combined hybrid-parallel prediction (CHyPP) framework, coupling reservoir computing with the Simplified Parameterization, primitive‐Equation Dynamics (SPEEDY) atmospheric model. Beyond the standard SPEEDY atmospheric state (3.75°×3.75°, 8 levels), it prognoses 6-h accumulated precipitation, sea-surface temperature (SST), and 0-300 m upper-ocean heat content via a hybridized recurrent neural network. Hybrid SPPEDY captures variability from intraseasonal (MJO, convectively coupled waves) to interannual (ENSO) scales.

Latest simulations in Patel, D., Arcomano, T., Hunt, B., Szunyogh, I., & Ott, E. (2025). Prediction beyond the medium range with an atmosphere-ocean model that combines physics-based modeling and machine learning. Journal of Advances in Modeling Earth Systems, 17, e2024MS004480.

See also:


Hybrid WRF

Hybrid versions of the Weather Research & Forecasting Model (WRF) preserve WRF's dynamical core while replacing uncertain WRF parameterizations with ML surrogates. This includes the emulation of radiation, convection, cloud microphysics, and select chemical formation processes.

See:

Radiation

Convection

Cloud microphysics

Chemistry


ICON-MLe

The Machine Learning–enhanced (MLe) ICOsahedral Non-hydrostatic (ICON) climate model builds on ICON-A at approximately 80 km resolution (R2B5), primarily using the ECHAM physics package. Suboptimal parameterizations, such as those for cloud cover and convection, are progressively replaced with improved data-driven versions trained on high-fidelity model outputs (e.g., DYAMOND, ClimSim) and observations. This results in a flexible, hybrid AI-climate version of ICON.

Latest simulations in Grundner, A., Beucler, T., Savre, J., Lauer, A., Schlund, M., & Eyring, V. (2025). Reduced Cloud Cover Errors in a Hybrid AI-Climate Model Through Equation Discovery And Automatic Tuning. arXiv preprint 2505.04358.

See also:


LUCIE

The Lightweight Uncoupled ClImate Emulator (LUCIE) is a fully data-driven, Spherical Fourier Neural Operator (SFNO)-based emulator of the ERA5 meteorological reanalysis. By prognosing only 5 variables on 8 σ-levels, LUCIE allows large ensembles via weight perturbation, which approximately reproduce the main global climate's modes of variability, including long-term responses to CO2 forcing.

Latest simulations in Guan, H., Arcomano, T., Chattopadhyay, A., & Maulik, R. (2025). LUCIE-3D: A three-dimensional climate emulator for forced responses. arXiv preprint 2509.02061.

See also:


MOM6

The Modular Ocean Model version 6 (MOM6) is a major testbed for hybrid ocean modeling, where ML augments or replaces oceanic parameterizations to correct systematic biases. This includes bias-correcting sea ice concentration, parameterizing mesoscale eddy fluxes and vertical mixing, all of which improve the emergent statistics of the coupled ML-dynamical ocean simulation.

Latest simulations in Gregory, W., Bushuk, M., Zhang, Y. F., Adcroft, A., Zanna, L., McHugh, C., & Jia, L. (2025). Advancing global sea ice prediction capabilities using a fully-coupled climate model with integrated machine learning. arXiv preprint:2505.18328.

See also:


NCAM

The Neural Community Atmosphere Model (NCAM) uses a real geography setup and combines residual and convolutional strategies, with memory (two time steps are sufficient). It replaces only the moist physics components of CAM, including deep and shallow convection and latent heating from microphysics, but does not replace radiation.

Latest simulations in Han, Y., Zhang, G. J., & Wang, Y. (2023). An ensemble of neural networks for moist physics processes, its generalizability and stable integration. Journal of Advances in Modeling Earth Systems, 15(10), e2022MS003508.

See also:


NeuralGCM

The Neural General Circulation Model (NeuralGCM) is based on a differentiable pseudo-spectral dynamical core implemented in JAX. Processes not represented by the core are learned in an end-to-end manner using a single-column parameterization that optimizes medium-range weather forecasting. NeuralGCM enables stable, multi-decadal simulations of climate variability under prescribed sea surface temperatures and is being updated to accurately simulate observed global precipitation fields.

Latest simulations in Kochkov, D., Yuval, J., Langmore, I., Norgaard, P., Smith, J., Mooers, G., ... & Hoyer, S. (2024). Neural general circulation models for weather and climate. Nature, 1-7.

See also:


Ola

The Ocean-linked-atmosphere (Ola) model is a 0.25°-resolution model coupling two autoregressive spherical Fourier neural operators for the atmosphere and the ocean.

Latest simulations in Wang, C., Pritchard, M. S., Brenowitz, N., Cohen, Y., Bonev, B., Kurth, T., ... & Pathak, J. (2024). Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model. arXiv preprint 2406.08632.


Samudra

Trained on coupled simulations using the OM4 model, Samudra is a fully data-driven, global 3D ocean emulator that autoregressively predicts sea-surface height, horizontal velocity, potential temperature, and salinity. Samudra's architecture is based on the ConvNeXt U-Net architecture and reproduces interannual variability signatures such as ENSO while accelerating the parent model by a factor ~150. The atmosphere-coupled SamudrACE variant simulates ~800 model-years per day at 1 degree horizontal, 6-hourly atmospheric, and 5-daily oceanic resolution.

Latest simulations in Dheeshjith, S., Subel, A., Adcroft, A., Busecke, J., Fernandez‐Granda, C., Gupta, S., & Zanna, L. (2025). Samudra: An AI global ocean emulator for climate. Geophysical Research Letters, 52(10), e2024GL114318.

See also: