Climate Models & Frameworks — Cross-Cutting Comparison
This note compares the climate and Earth-system modeling frameworks a researcher actually chooses between when designing an experiment — full coupled CMIP6 ESMs vs convection-permitting regional models vs intermediate-complexity EMICs vs component-only models (ocean, land, sea-ice, ice-sheet, carbon-cycle) vs integrated assessment models vs the new ML-based neural ESMs — across every ClimateScience library note that touches simulation. Each family has a unified table mapping models against the axes that drive selection (spatial resolution, coupling complexity, computational cost, calibration target, openness). The closing decision tree picks the model class given the research question.
See also
- physical-climate-system
- climate-sensitivity-and-feedbacks
- carbon-cycle-and-greenhouse-gases
- atmospheric-chemistry-and-aerosols
- atmospheric-chemistry-and-radiative-transfer
- atmospheric-dynamics-deep
- ocean-biogeochemistry
- glaciology-and-cryosphere
- hydrology-and-water-cycle
- regional-climate-and-downscaling
- extreme-event-attribution
- paleoclimate
- paleoclimate-and-deep-time
- ipcc-scenarios-and-integrated-assessment
- solar-geoengineering-and-cdr
- ai-and-machine-learning-for-climate
- climate-mitigation-and-adaptation
- climate-impacts-and-adaptation
- climate-finance-and-investment
- carbon-accounting-and-mrv
1. The five axes every model choice answers
Climate-model selection looks like five orthogonal questions before any code runs.
| Axis | Cheap end | Expensive end | Why it matters |
|---|---|---|---|
| Spatial resolution | EMIC (~500 km) | LES / CRM (~100 m) | physics resolved vs parameterized |
| Coupling | atmosphere-only | atmosphere+ocean+land+sea-ice+ice-sheet+chemistry+veg dynamics+CO₂ | feedback completeness |
| Time horizon | hours (weather) | 10⁹ yr (deep-time paleoclimate) | spin-up + integration cost |
| Computational cost | EMIC: hr on desktop | high-res ESM: month on 10⁴ cores | wallclock per scenario |
| Calibration target | paleoclimate / pre-industrial / historical / future | which observations constrain tuning | propagates into projection skill |
The textbook ordering says “pick the simplest model that captures your feedbacks”. The practical reality: pick the model whose vetting cost (you understand it, the community trusts it, IPCC AR cites it) matches your timeline. A bespoke high-res run nobody else has done buys little if your reviewer trusts CESM2 outputs.
2. CMIP6 Earth-system models — the canonical coupled set
The Coupled Model Intercomparison Project Phase 6 (CMIP6) is the IPCC AR6 (2021) backbone — ~100 modeling groups submit standardized experiments (historical, SSP1-1.9 through SSP5-8.5, DECK, AMIP, RFMIP, GeoMIP, ScenarioMIP). The lineage continues with CMIP7 (in progress, IPCC AR7 deadline ~2028).
| Model | Group | Atm res | Ocn res | Coupling | Equilibrium climate sensitivity (ECS, °C) | Linked note |
|---|---|---|---|---|---|---|
| CESM2 | NCAR (US) | ~1° (CAM6) | 1° (POP2) | atm+ocn+land(CLM5)+ice(CICE6) + carbon + atm chem (CAM-chem) + WACCM(stratosphere) | 5.2 | physical-climate-system / climate-sensitivity-and-feedbacks |
| CESM2.2 / CESM3 | NCAR | ~0.25° HR variant | 0.1° HR variant | as CESM2 + improved aerosol | being released 2025-2026 | physical-climate-system |
| E3SM v2 / v3 | US DOE (LLNL+ANL+ORNL+PNNL+SNL) | regionally refined unstruct (MPAS-A) | unstruct (MPAS-O) | atm+ocn+ice+land+river+ice-sheet (MALI) | 4.0 | physical-climate-system |
| GFDL ESM4 / SPEAR | NOAA (US) | 1° (AM4) | 0.5° (MOM6) | atm+ocn(MOM6)+ice(SIS2)+land(LM4)+atm chem + biogeochem | 2.7 | atmospheric-chemistry-and-aerosols |
| HadGEM3-GC31 | Met Office (UK) | 60 km (N216) | 0.25° | atm(UM)+ocn(NEMO)+ice(CICE)+land(JULES); UKESM1 adds chem + ocn BGC | 5.3 (UKESM1) | physical-climate-system |
| UKESM1 | UK consortium (Met Office + NCAS + NERC) | 60 km | 0.25° | full Earth system (atm chem UKCA + MEDUSA ocn BGC + TRIFFID veg) | 5.3 | carbon-cycle-and-greenhouse-gases |
| MPI-ESM 1.2 | Max-Planck (Germany) | 1.9° (ECHAM6) | 0.4° (MPIOM) | atm+ocn+land(JSBACH)+ice + carbon | 3.0 | carbon-cycle-and-greenhouse-gases |
| MPI-ESM HR | Max-Planck | 0.5° | 0.4° | as MPI-ESM 1.2 high-res | 3.0 | physical-climate-system |
| IPSL-CM6A-LR | IPSL (France) | 2.5° (LMDZ6) | 1° (NEMO) | atm+ocn+land(ORCHIDEE)+ice+atm chem(REPROBUS/INCA) | 4.6 | atmospheric-chemistry-and-aerosols |
| ICON-A / ICON-ESM | DWD + MPI-M (Germany) | unstruct icosahedral 80-160 km; 5 km in ICON-Sapphire CRM mode | 0.4° | atm+ocn(ICON-O)+land(JSBACH) | ~3 | atmospheric-dynamics-deep |
| MIROC6 | JAMSTEC + NIES + U-Tokyo (Japan) | 1.4° | 1° (COCO) | atm+ocn+land(MATSIRO)+ice | 2.6 | physical-climate-system |
| MIROC-ES2L | as above | 2.8° | 1° | + carbon cycle + aerosol + chem | 2.7 | carbon-cycle-and-greenhouse-gases |
| NorESM2-LM / MM | Norway consortium | 2° / 1° | 1° (BLOM) | atm(CAM-OSLO aerosol)+ocn(BLOM)+ice(CICE)+land(CLM5) | 2.5 | atmospheric-chemistry-and-aerosols |
| AWI-CM-1-1-MR | AWI (Germany) | 1° (ECHAM6) | unstruct FESOM2 (varies) | atm+ocn(unstruct)+ice | 3.2 | physical-climate-system |
| CMCC-ESM2 | CMCC (Italy) | 1° (CAM5.3) | 0.25° (NEMO) | atm+ocn+land(CLM4.5)+ice | 3.6 | physical-climate-system |
| BCC-ESM1 | Beijing Climate Center (China) | T42 (~2.8°) | 1° (MOM4) | atm+ocn+land+ice+aerosol | 3.3 | physical-climate-system |
| CanESM5 | Environment Canada | T63 (~2.8°) | ~1° (NEMO) | atm+ocn+land(CLASS-CTEM)+atm chem (CMAM-chem) | 5.6 (high-ECS outlier) | climate-sensitivity-and-feedbacks |
| EC-Earth3 | European consortium | ~1° (IFS) | 0.25° (NEMO) | atm(IFS)+ocn(NEMO)+ice(LIM3)+land(HTESSEL) | 4.3 | physical-climate-system |
| ACCESS-ESM1-5 | Australia (CSIRO + BoM) | ~1° (UM) | 1° (MOM5) | atm+ocn+land(CABLE)+ice + carbon | 3.9 | carbon-cycle-and-greenhouse-gases |
| CNRM-ESM2-1 | Météo-France + CERFACS | T127 (~1.4°) | 1° (NEMO) | atm(ARPEGE-Climat)+ocn(NEMO)+ice(GELATO)+land(SURFEX-ISBA) | 4.8 | physical-climate-system |
| INM-CM5-0 | INM RAS (Russia) | 2° | 0.5° | atm+ocn+land+ice | 1.9 (low-ECS outlier) | climate-sensitivity-and-feedbacks |
| KIOST-ESM | KIOST (Korea) | T62 | ~1° | atm+ocn+land+ice | 3.4 | physical-climate-system |
| FIO-ESM | First Inst of Oceanography (China) | T42 | 1° | adds explicit wave model | 3.5 | physical-climate-system |
| MRI-ESM2-0 | Met Research Institute (Japan) | T127 | ~1° | atm+ocn+land+ice + atm chem + aerosol | 3.2 | atmospheric-chemistry-and-aerosols |
| SAM0-UNICON | Seoul National (Korea) | 1° | 1° | as CESM2 + UNICON convection | 4.0 | physical-climate-system |
A typical CMIP6 historical+SSP1-1.9 through SSP5-8.5 ensemble for one of these models is ~5-10 Mcore-hours on a Tier-1 HPC system. CMIP6 archive is on ESGF (Earth System Grid Federation, esgf-node.llnl.gov and mirrors) at >20 PB. Access via cdo, xarray, intake-esm, pangeo. The python-Pangeo + Zarr stack on AWS / GCP is the de-facto modern access path.
Notable ECS spread: CanESM5 (5.6 °C) and HadGEM3/UKESM1 (5.3 °C) sit high; INM-CM5 (1.9 °C) sits low. The IPCC AR6 “very likely range” of ECS = 2.0–5.0 °C (best 3.0 °C) deliberately narrower than the model spread because emergent-constraint analyses tightened the range — see climate-sensitivity-and-feedbacks.
3. Convection-permitting and regional models
Below ~10 km grid spacing, deep convection is partially resolved rather than parameterized. This regime is where most regional climate-change adaptation work happens.
| Model | Group | Range | Native res | Where used | Linked note |
|---|---|---|---|---|---|
| WRF (Weather Research & Forecasting) | NCAR (US) | regional, ~50 km–100 m | 1 km common; 100 m LES | dominant US regional (NCAR, EPA, DOE, universities) | regional-climate-and-downscaling |
| WRF-Chem | NCAR / NOAA | regional + atm chem | 1 km | air quality + climate feedback | atmospheric-chemistry-and-aerosols |
| RegCM4 / RegCM5 | ICTP (Italy) | regional | 10–25 km | CORDEX (Coordinated Regional Climate Downscaling) workhorse | regional-climate-and-downscaling |
| COSMO-CLM | Climate Limited-area Modelling Community | regional | 2.8 km common | EU + Germany regional | regional-climate-and-downscaling |
| REMO | Helmholtz-Zentrum Geesthacht / HZG | regional | 12–25 km | EU CORDEX | regional-climate-and-downscaling |
| ALADIN-Climate | Météo-France + ALADIN consortium | regional | 12 km common; HARMONIE 2.5 km | EU operational + climate | regional-climate-and-downscaling |
| HadRM3 / HadREM3 | Met Office (UK) | regional | 12 km / 2.2 km | UK climate projections (UKCP18, UKCP-Local) | regional-climate-and-downscaling |
| RACMO2 | KNMI (Netherlands) | regional + ice sheet | 5.5–11 km | Antarctica + Greenland surface mass balance | glaciology-and-cryosphere |
| MAR (Modèle Atmosphérique Régional) | U-Liège (Belgium) | polar regional | 7.5–25 km | Greenland + Antarctic SMB; ice-sheet projections | glaciology-and-cryosphere |
| NICAM (Nonhydrostatic Icosahedral) | RIKEN + JAMSTEC + U-Tokyo | global cloud-resolving | 870 m / 3.5 km | global CRM Aqua-planet + DYAMOND-class | atmospheric-dynamics-deep |
| ICON-Sapphire | MPI-M (Germany) | global CRM | 5 km / 2.5 km | DYAMOND + nextGEMS Earth Virtualization Engines | atmospheric-dynamics-deep |
| X-SHiELD | NOAA-GFDL | global CRM | 3.25 km | DYAMOND class | atmospheric-dynamics-deep |
| SCREAM (Simple Cloud-Resolving E3SM Atm Model) | DOE (US) | global CRM | 3 km | E3SM CRM-mode, GPU-native (Frontier) | atmospheric-dynamics-deep |
CORDEX (Coordinated Regional Climate Downscaling Experiment) is the regional analog of CMIP — standardized regional simulations for 14 CORDEX domains (Africa, Europe, North America, etc.). EURO-CORDEX is the most active subset.
4. EMICs — Earth-system Models of Intermediate Complexity
Below CMIP6 ESMs lies a class of models that trade resolution for speed, enabling deep-time paleoclimate (10⁵–10⁸ yr), large ensembles for uncertainty quantification, and IAM coupling.
| Model | Group | What’s simplified | Range | Where used | Linked note |
|---|---|---|---|---|---|
| UVic ESCM | U-Victoria (Canada) | 1.8°×3.6° energy-moisture-balance atm; full ocn | up to 10⁵ yr | climate-carbon-cycle feedback, deep-time, MIP6 (CMIP6 contribution) | carbon-cycle-and-greenhouse-gases |
| CLIMBER-2 / CLIMBER-X | PIK (Potsdam, Germany) | sectoral statistical-dynamical atm; full ocn; ice sheet | up to 10⁶ yr | tipping points, glacial cycles, AMOC collapse | paleoclimate-and-deep-time |
| Bern3D | U-Bern (Switzerland) | 2D zonally-averaged atm; 3D ocn | up to 10⁵ yr | carbon cycle, ocean BGC, paleoclimate | carbon-cycle-and-greenhouse-gases |
| GENIE / cGENIE | U-Bristol + UEA (UK) | reduced-cmplx atm (EMBM); GOLDSTEIN ocn | up to 10⁸ yr | deep-time paleoclimate, PETM, snowball Earth | paleoclimate-and-deep-time |
| LOVECLIM | UCLouvain (Belgium) | quasi-geostrophic atm (ECBilt); CLIO ocn | up to 10⁵ yr | Holocene, last deglaciation | paleoclimate |
| CESM (low-res FullPhys-CESM) | NCAR | 4° + 3° ocn | up to 10⁴ yr | bridges EMIC to full ESM | paleoclimate-and-deep-time |
| MAGICC | Climate Resource (Australia) | global-mean energy balance + reduced ocn + carbon cycle | seconds-to-minutes per scenario | reduced-complexity for IAM use; emulates CMIP6 spread | ipcc-scenarios-and-integrated-assessment |
| FaIR | UK consortium (Smith et al.) | impulse response | seconds per scenario | reduced-complexity for AR6 + IAM coupling | ipcc-scenarios-and-integrated-assessment |
| OSCAR | LSCE / IIASA | linearized box | seconds per scenario | AR6 Probabilistic Earth-System Emulator | ipcc-scenarios-and-integrated-assessment |
| Hector | PNNL | 1D box | seconds per scenario | IAM-coupled (GCAM partner) | ipcc-scenarios-and-integrated-assessment |
The MAGICC + FaIR + OSCAR + Hector class is what IAMs actually couple to — full CMIP6 ESMs are too slow for IAM-loop iteration. AR6 ChapAtlas chapter constructs probabilistic projections from MAGICC + FaIR calibrated to CMIP6 spread.
5. Specialized component models
A modular view — these are the components that combine into CMIP6 ESMs above, but are also run standalone for component-specific science.
5.1 Ocean-only
| Model | Group | Native res | Where used | Linked note |
|---|---|---|---|---|
| MOM6 | NOAA-GFDL + Princeton + NCAR | 0.1° (HR), 0.25° (MR), 1° (LR) | de-facto US ocean model; CESM, GFDL, NCAR | ocean-biogeochemistry |
| NEMO | NEMO Consortium (CNRS + Met Office + INGV + Mercator Ocean) | 0.25° (ORCA025) → 1/12° | dominant EU ocean (UKESM1, IPSL, EC-Earth) | ocean-biogeochemistry |
| MITgcm | MIT | 1/3° → eddy-resolving | ECCO ocean state estimate; Lagrangian analysis | ocean-biogeochemistry |
| POP2 | NCAR (legacy) | 1° | pre-MOM6 CESM ocean | ocean-biogeochemistry |
| MIROC OGCM (COCO) | JAMSTEC | 1° | MIROC ocean component | ocean-biogeochemistry |
| FESOM2 | AWI (Germany) | unstruct (varies) | AWI-CM coupled, eddy-permitting in key regions | ocean-biogeochemistry |
| MPAS-O | LANL + DOE | unstruct | E3SM ocean component | ocean-biogeochemistry |
5.2 Atmosphere-only
| Model | Group | Native res | Where used | Linked note |
|---|---|---|---|---|
| CAM6 / CAM7 | NCAR | ~1° | CESM atmosphere; CAM-chem variant + WACCM stratosphere | atmospheric-chemistry-and-radiative-transfer |
| IFS (Integrated Forecasting System) | ECMWF | 9 km HRES (TCo1279) → 0.1° AIFS | ECMWF medium-range + EC-Earth coupled | atmospheric-dynamics-deep |
| MPAS-A | NCAR / LANL | unstruct (varies) | E3SM + research; mesh-refinement | atmospheric-dynamics-deep |
| FV3 | NOAA-GFDL | 13 km (UFS) | NOAA UFS operational; SHIELD HR | atmospheric-dynamics-deep |
| COSMO | DWD + COSMO Consortium | 2.2 km operational | DWD operational + COSMO-CLM | atmospheric-dynamics-deep |
| UM (Unified Model) | Met Office | 10 km operational | HadGEM atmosphere + GloSea seasonal | atmospheric-dynamics-deep |
| ARPEGE-Climat | Météo-France | T127–T255 | CNRM-CM/ESM atmosphere | atmospheric-dynamics-deep |
| ECHAM6 | MPI-M | T63/T127 | MPI-ESM atmosphere | atmospheric-dynamics-deep |
| LMDZ | LMD (France) | 2.5° | IPSL atmosphere | atmospheric-dynamics-deep |
5.3 Land-surface
| Model | Group | Where used | Linked note |
|---|---|---|---|
| CLM5 | NCAR | CESM + NorESM + CMCC; gold-standard for plant functional types + hydrology + biogeochem | hydrology-and-water-cycle |
| JULES | Met Office + CEH | HadGEM + UKESM | hydrology-and-water-cycle |
| ISBA (SURFEX-ISBA) | Météo-France | CNRM | hydrology-and-water-cycle |
| ORCHIDEE | IPSL + LSCE | IPSL ESM | carbon-cycle-and-greenhouse-gases |
| JSBACH | MPI-M | MPI-ESM | carbon-cycle-and-greenhouse-gases |
| Noah-MP | NCAR + NOAA | WRF + NOAA UFS | hydrology-and-water-cycle |
| CABLE | CSIRO (Australia) | ACCESS | carbon-cycle-and-greenhouse-gases |
| CLASS-CTEM | Environment Canada | CanESM | carbon-cycle-and-greenhouse-gases |
5.4 Sea-ice
| Model | Group | Where used | Linked note |
|---|---|---|---|
| CICE6 | LANL + NCAR + Met Office | CESM + UKESM + NorESM + CanESM + EC-Earth | glaciology-and-cryosphere |
| LIM3 | UCLouvain | NEMO-based (IPSL-CM, EC-Earth) | glaciology-and-cryosphere |
| SIS2 | NOAA-GFDL | GFDL ESM4 | glaciology-and-cryosphere |
| GELATO | CNRM | CNRM-ESM2 | glaciology-and-cryosphere |
5.5 Ice-sheet
| Model | Group | Where used | Linked note |
|---|---|---|---|
| PISM (Parallel Ice Sheet Model) | U-Alaska + PIK | research workhorse; Antarctic + Greenland projections | glaciology-and-cryosphere |
| ISSM (Ice-Sheet System Model) | JPL / NASA | adjoint capability for inverse problems, satellite assimilation | glaciology-and-cryosphere |
| BISICLES | LBNL + Bristol | AMR (adaptive mesh refinement) for shelves + grounding lines | glaciology-and-cryosphere |
| Elmer/Ice | CSC Finland + IGE Grenoble | finite-element, full Stokes; thermomechanical | glaciology-and-cryosphere |
| CISM | LANL + NCAR | CESM-coupled ice-sheet | glaciology-and-cryosphere |
| MALI (MPAS-Albany Land Ice) | DOE (LANL+SNL+ORNL) | E3SM-coupled; unstruct grids | glaciology-and-cryosphere |
| ÚaMITgcm | various | grounded-ice + ice-shelf MITgcm-coupled | glaciology-and-cryosphere |
ISMIP6 (Ice Sheet MIP for CMIP6) is the standardized intercomparison; produced AR6 ice-sheet sea-level contributions.
5.6 Carbon cycle
| Model | Group | What it does | Linked note |
|---|---|---|---|
| CASA | NASA + Stanford | terrestrial primary productivity from satellite | carbon-cycle-and-greenhouse-gases |
| CTE (CarbonTracker Europe) | Wageningen | atmospheric inverse for CO₂ + CH₄ | carbon-cycle-and-greenhouse-gases |
| CarbonTracker | NOAA + CIRES | atmospheric inverse for CO₂ flux (operational since 2007) | carbon-cycle-and-greenhouse-gases |
| TM5 (Transport Model 5) | KNMI + Wageningen | atmospheric transport for inversions | carbon-cycle-and-greenhouse-gases |
| TM5-MP | KNMI + JRC | massively-parallel TM5 | carbon-cycle-and-greenhouse-gases |
| GEOS-Chem | Harvard + community | global atm chem + carbon | atmospheric-chemistry-and-aerosols |
6. Integrated Assessment Models (IAMs)
IAMs couple climate (a reduced-complexity emulator like MAGICC or FaIR) with economy + energy + land-use + technology. The IPCC AR6 WG3 used 7 IAM frameworks to build the SSP-RCP scenario space.
| Model | Group | Type | Where used | Linked note |
|---|---|---|---|---|
| GCAM (Global Change Analysis Model) | PNNL + Joint Global Change Research Institute | partial-equilibrium energy-economic + land + water | AR6, US DOE policy | ipcc-scenarios-and-integrated-assessment |
| IMAGE | PBL Netherlands Environmental Assessment | partial-equilibrium energy-land-economy | AR6, OECD work | ipcc-scenarios-and-integrated-assessment |
| MESSAGE-GLOBIOM | IIASA | linear-programming energy + GLOBIOM land | AR6, SDP scenarios | ipcc-scenarios-and-integrated-assessment |
| REMIND-MAgPIE | PIK | general-equilibrium energy + MAgPIE land | AR6, EU-Green-Deal-class | ipcc-scenarios-and-integrated-assessment |
| WITCH | RFF-CMCC (Italy) | game-theoretic optimal-growth + energy + climate | climate policy + game theory | climate-finance-and-investment |
| AIM/CGE | NIES + Mizuho (Japan) | computable-general-equilibrium | AR6, Japan / Asia scenarios | ipcc-scenarios-and-integrated-assessment |
| POLES | JRC (EU) + Enerdata | partial-equilibrium energy | EU climate policy + JRC | ipcc-scenarios-and-integrated-assessment |
| DICE / RICE (Nordhaus) | Yale | cost-benefit growth model | benchmark for SCC (Nordhaus, Nobel 2018) | ipcc-scenarios-and-integrated-assessment |
| FUND | Anthoff + Tol | enumerated impact-cost regions | SCC sensitivity | ipcc-scenarios-and-integrated-assessment |
| PAGE | Hope (Cambridge) | probabilistic damage + SCC | UK Stern review + sensitivity | ipcc-scenarios-and-integrated-assessment |
| GEM-E3 | EU consortium | EU-policy CGE | EU 2050 long-term strategy | ipcc-scenarios-and-integrated-assessment |
7. ML-based and neural Earth-system models (2022+)
The post-2022 wave of ML weather + climate models trained on ERA5 reanalysis + CMIP. Speed gain is 10⁴-10⁶× vs traditional NWP/ESMs; skill is now competitive with conventional ECMWF IFS at medium range for many variables. Climate-range projection skill (multi-decadal, response to forcing) is still an open question — most demonstrations stop short of multi-decadal forcing-response.
| Model | Group | Range | Resolution | Architecture | Where it wins | Linked note |
|---|---|---|---|---|---|---|
| Pangu-Weather | Huawei Cloud (2022) | 7 d forecast | 0.25° | 3D Earth-Specific Transformer | first ML to beat IFS on RMSE for several variables (Bi et al. 2023 Nature) | ai-and-machine-learning-for-climate |
| GraphCast | Google DeepMind (2023) | 10 d forecast | 0.25° | graph neural network on icosahedral mesh | beats IFS on 90% of 1380 variable-lead-time pairs (Lam et al. 2023 Science) | ai-and-machine-learning-for-climate |
| AIFS (AI Forecasting System) | ECMWF (2024) | 10 d operational | 0.25° | graph neural | first operational ML forecast at ECMWF, deployed Feb 2025 | ai-and-machine-learning-for-climate |
| Aurora | Microsoft (2024) | 5 d weather + air quality + waves + tropical cyclone | 0.1° (waves) | foundation-model transformer | first multi-domain ESM emulator (Bodnar et al. 2024) | ai-and-machine-learning-for-climate |
| NeuralGCM | Google (2024) | hours-to-decade hybrid | 1.4° / 0.7° / 2.8° | diff-physics + ML | beats CMIP6 ensemble on key climate diagnostics with physics constraint (Kochkov et al. 2024 Nature) | ai-and-machine-learning-for-climate |
| ClimaX | Microsoft Research (2023) | climate + weather | flexible | foundation transformer | downstream-task fine-tuning baseline | ai-and-machine-learning-for-climate |
| FourCastNet | NVIDIA + Caltech + LBNL (2022) | 7 d forecast | 0.25° | adaptive Fourier neural operator | first ML to operate at IFS scale | ai-and-machine-learning-for-climate |
| Stormer | NVIDIA + others (2024) | medium range | 0.25° | transformer | weather forecast at IFS-class | ai-and-machine-learning-for-climate |
| NVIDIA Earth-2 | NVIDIA | full platform | flexible | platform for digital-twin climate | NVIDIA Omniverse + Modulus + FourCastNet + CorrDiff downscaling | ai-and-machine-learning-for-climate |
| Ai2 Climate Emulator (ACE) | Allen Institute for AI | climate-range emulator | C96 (~100 km) | spherical harmonic neural operator | trained to emulate CMIP6 forced response (Watt-Meyer 2023) | ai-and-machine-learning-for-climate |
| GenCast | Google DeepMind (2024) | probabilistic ensemble | 0.25° | generative diffusion | beats IFS ensemble on probabilistic skill (Price et al. 2024 Nature) | ai-and-machine-learning-for-climate |
| CorrDiff (NVIDIA) | NVIDIA | downscaling | km-scale | conditional diffusion | super-resolution from km-scale models | regional-climate-and-downscaling |
| Prithvi-WxC | NASA + IBM (2024) | foundation model | 0.25° | transformer | NASA’s open foundation weather-climate model | ai-and-machine-learning-for-climate |
The key open question (still being resolved through 2025-2026) is whether these neural models capture climate-change response rather than just short-range weather. Hybrid (NeuralGCM, Ai2 ACE) and physics-constrained approaches look most promising for decade-to-century projection. Pure-data emulators trained only on ERA5 (a 1940–present record) cannot extrapolate to future forcings they have never seen.
8. Unified five-axis cross-class summary
A compact view of where each class lives. “Time horizon” is the longest a researcher would run it in production; “Cost” is rough wallclock for one historical+SSP scenario at standard resolution.
| Class | Resolution | Coupling | Time horizon | Cost (compute) | Calibration target | Openness |
|---|---|---|---|---|---|---|
| CMIP6 ESM (CESM2, UKESM1, …) | ~1° / ~0.25° | full ESM | 1850–2100 + perturbed | 5–10 Mcore-hr | historical SST + atmosphere | mostly open (LANL, NCAR, MPI, GFDL); some national (HadGEM3) |
| Convection-permitting global (NICAM, SCREAM, ICON-Sapphire, X-SHiELD) | 2–5 km | atm only (mostly) | months | 100 Mcore-hr | DYAMOND benchmark | open |
| Regional CRM (WRF) | 100 m–4 km | regional | 1 month–10 yr | 10–100k core-hr | reanalysis lateral boundary | open |
| Regional CORDEX (RegCM, REMO, COSMO-CLM) | 10–25 km | regional | 1990–2100 | 100k–1M core-hr | reanalysis lateral boundary | open |
| EMIC (CLIMBER, Bern3D, GENIE, UVic) | 5° / box | atm + ocn + carbon | up to 10⁸ yr | hours-days | paleoclimate records | open |
| Reduced-complexity (MAGICC, FaIR, OSCAR, Hector) | global mean | energy-balance + carbon | seconds-minutes | desktop | CMIP6 ensemble + obs | open |
| Component (MOM6, NEMO, CLM5, CICE6, PISM, ISSM) | varies | one component | varies | 0.1–5 Mcore-hr | reanalysis or proxy | open |
| IAM (GCAM, IMAGE, MESSAGE, REMIND, WITCH) | regional (32 regions) | energy+land+economy + emulator | 2020–2100 | minutes-hours per scenario | historical economic + emission | mostly open; some IIASA/PNNL specific |
| Neural ESM (Pangu, GraphCast, AIFS, GenCast, NeuralGCM, Aurora) | 0.25° / 0.1° | atm (mostly) | days for weather; emerging climate-range (Ai2 ACE) | inference: seconds | ERA5 reanalysis + CMIP6 | open (most); proprietary weights some |
9. Recent shifts (2022–2026)
- Storm-resolving global modeling is now possible — DYAMOND-class runs (3 km global) on Frontier, Aurora, LUMI HPCs. SCREAM, ICON-Sapphire, X-SHiELD all crossed the threshold.
- AI weather models beat IFS — Pangu-Weather (2022), GraphCast (2023), AIFS (2024 operational at ECMWF). Climate-range skill is the next frontier.
- Neural ESM emulators — Ai2 ACE, NeuralGCM. Trade fidelity for orders-of-magnitude speed.
- Probabilistic ML ensembles — GenCast (Google 2024) replaces 51-member IFS ensemble at lower cost.
- CMIP7 in preparation — IPCC AR7 reports due 2027-2029; CMIP7 design finalized 2024-25, runs 2025-2027.
- Digital twin Earth — Destination Earth (EU ECMWF + ESA, launched 2022, operational 2024), NVIDIA Earth-2 platform. Each combines storm-resolving simulation + ML downscaling + observational nudging.
- AR6 emergent constraints tightened ECS — model-spread 1.5–6 °C; AR6 assessed range 2.0–5.0 °C, best 3.0 °C — see climate-sensitivity-and-feedbacks.
- CDR / geoengineering — GeoMIP under CMIP6/7 increases SRM + CDR scenario coverage — solar-geoengineering-and-cdr.
Adjacent
- Sensitivity feedbacks — climate-sensitivity-and-feedbacks for ECS / TCR diagnostics from these models.
- Reanalysis (data side) — ERA5 (ECMWF), MERRA-2 (NASA), JRA-3Q (JMA), NCEP/NCAR — primary observational base for ML training and model evaluation.
- Carbon-cycle observations — carbon-cycle-and-greenhouse-gases.
- Glaciology component — glaciology-and-cryosphere.
- Ocean biogeochemistry — ocean-biogeochemistry.
- Hydrology — hydrology-and-water-cycle.
- Paleoclimate forcing + targets — paleoclimate and paleoclimate-and-deep-time.
- AI methods overview — ai-and-machine-learning-for-climate.
- IPCC scenario framework — ipcc-scenarios-and-integrated-assessment.
- Energy + carbon markets — _index.
When to pick what
What's the question?
├─ Multi-decade global climate change projection
│ ├─ AR6 reference / official scenario → CMIP6 ESM ensemble (already on ESGF; analyze in Pangeo)
│ ├─ Fast scenario sweep / emulation → MAGICC + FaIR (probabilistic)
│ └─ Need ECS uncertainty → CMIP6 ensemble + emergent constraints
├─ Deep-time paleoclimate (>10 kyr, e.g. glacial cycles, PETM, snowball Earth)
│ ├─ Glacial cycles → CLIMBER-X or LOVECLIM
│ ├─ PETM / Cretaceous / Permian → cGENIE
│ └─ Last deglaciation → LOVECLIM, CESM low-res, or transient TraCE-21k
├─ Regional climate adaptation
│ ├─ Standard CORDEX downscaling → RegCM, REMO, COSMO-CLM, WRF
│ ├─ Polar / ice sheet SMB → MAR or RACMO2
│ └─ Hyper-local hazard → WRF at 1 km or CorrDiff ML downscaling
├─ Storm-permitting global climate
│ └─ ICON-Sapphire / SCREAM / NICAM / X-SHiELD (DYAMOND class) — needs Frontier/Aurora/LUMI access
├─ Component-only science
│ ├─ Ocean (eddies, AMOC) → MOM6, NEMO, MITgcm, FESOM2
│ ├─ Atmosphere (radiation, chemistry) → CAM6, IFS, CAM-chem, GEOS-Chem, WACCM
│ ├─ Land (carbon, hydrology) → CLM5, JULES, ORCHIDEE, Noah-MP
│ ├─ Sea-ice → CICE6
│ └─ Ice-sheet → PISM (research), ISSM (inverse), BISICLES (shelves), Elmer/Ice (full Stokes), MALI/CISM (coupled)
├─ Carbon-cycle / flux attribution
│ └─ CarbonTracker / CTE / TM5 for inversions; CASA for satellite-driven NPP; bottom-up via JULES/ORCHIDEE
├─ Climate-economy / policy / SCC
│ ├─ Long-term IPCC scenario → MESSAGE, GCAM, REMIND, IMAGE, AIM
│ ├─ SCC sensitivity → DICE, FUND, PAGE
│ └─ EU policy CGE → GEM-E3, POLES
├─ Operational weather forecast (1-10 d)
│ ├─ Traditional → IFS, GFS, ICON, UKM, GEM
│ ├─ ML deterministic → AIFS, Pangu-Weather, GraphCast, FourCastNet
│ └─ ML probabilistic → GenCast, AIFS-ENS
├─ Sub-seasonal / seasonal (S2S)
│ ├─ Traditional → SEAS5 (ECMWF), CFSv2 (NOAA), GloSea (Met Office)
│ └─ ML emerging → CorrDiff, ClimaX-S2S
├─ Climate-range ML emulator
│ ├─ Hybrid (physics + ML, multi-decadal) → NeuralGCM
│ ├─ Pure data-driven climate response → Ai2 ACE
│ └─ Multi-domain foundation model → Aurora, Prithvi-WxC
├─ Extreme event attribution
│ ├─ Pseudo-global-warming → WRF-PGW
│ ├─ Counterfactual ensemble → CMIP6 LE + storyline approach (NCAR, KNMI, WWA)
│ └─ Operational → World Weather Attribution + ClimaMeter
├─ Solar geoengineering / CDR
│ └─ GeoMIP / G6solar within CMIP6/7 framework — see [[Sciences/ClimateScience/solar-geoengineering-and-cdr]]
└─ Earth observation + data assimilation
→ ECCO (ocean), ICESat-2 + CryoSat-2 + Operation IceBridge (cryo), Sentinel + Copernicus, GRACE-FO (mass), OCO-2/3 + MERLIN + GOSAT-3 (carbon)
The biggest practical lesson 2010–2026 is that model class matters more than model identity. The within-class CMIP6 spread (e.g. CESM2 vs MPI-ESM vs UKESM1) is smaller than the cross-class difference (CMIP6 ESM vs EMIC vs ML emulator). Pick the class that matches your question’s resolution, time horizon, and feedback needs — then within that class pick the model whose community + documentation + reanalysis-comparison best supports your study. Default to multi-model ensembles for any policy-relevant projection; default to single-model multi-realization ensembles (Large Ensembles like CESM2-LE, MPI-GE, CanESM5-LE) for internal-variability questions.