Compare · 32 models
Weather, climate and Earth observation models you can run online
These models forecast global weather, emulate multi-year climate variability, analyse remote sensing imagery, and predict solar events for Earth and space scientists.
What weather, climate and earth observation models do
Earth system and atmospheric science tasks require modelling physical dynamics across vastly different temporal and spatial regimes. These span planetary-scale circulation, multi-decadal ocean-atmosphere coupling, regional storm development, downscaled surface precipitation, and satellite observation analysis. Machine learning approaches address these tasks by replacing or complementing classical numerical solvers with deep neural networks, diffusion architectures, graph neural networks, and vision transformers.
Across weather forecasting and climate emulation, models are trained on gridded reanalysis fields such as ERA5, numerical simulation runs, or high-resolution observational archives. These frameworks forecast prognostic variables forward in time, downscale coarse global simulations to kilometre scales, or simulate coupled ocean-atmosphere interactions. Parallel Earth-observation models process optical, radar, and hyperspectral satellite imagery to segment environmental hazards, track greenhouse gas emissions, or generate general-purpose visual representations.
Compared
- AIFS Single 1.0ecmwfA data-driven weather forecasting model by ECMWF for deterministic global predictions.Check termsOn demand
- Indus SDE v0.2nasa-impactA 125M-parameter masked language model for extracting scientific concepts from Earth and atmospheric science texts.Check termsOn demand
- ACE2-ERA5allenaiA global atmospheric emulator for subseasonal to decadal climate variability and multi-year projections.See its pageNot yet
- AIFS ENS 1.0ecmwfA probabilistic data-driven weather forecasting model for upper-air, surface variables and cyclone tracks.See its pageNot yet
- AIFS ENS 2.0ecmwfA data-driven medium-range probabilistic weather forecasting model generating global ensemble forecasts.See its pageNot yet
- AIFS Single 0.2.1ecmwfA data-driven medium-range global weather forecasting model developed by ECMWF.See its pageNot yet
- AIFS Single 1.1ecmwfA data-driven medium-range weather forecasting model producing 6-hour interval atmospheric and surface predictions.See its pageNot yet
- AIFS Single 2.0ecmwfA data-driven medium-range weather model by ECMWF providing deterministic 15-day global atmospheric and wave forecasts.See its pageNot yet
- AuroramicrosoftAn Earth system foundation model for global weather forecasting, atmospheric chemistry, and cyclone tracking.See its pageNot yet
- cBottlenvidiaA cascaded diffusion model for generating kilometer-scale Earth atmospheric states on HEALPix geometry.See its pageNot yet
- FourCastNet 3nvidiaA geometric deep learning model for rapid global probabilistic weather and subseasonal-to-seasonal forecasting.See its pageNot yet
- HiRO-ACEallenaiClimate simulation framework generating 3 km precipitation fields from 100 km inputs using diffusion downscaling.See its pageNot yet
- ML4Floodsisp-uv-esOptical satellite image segmentation model for global flood extent mapping from Sentinel-2 and Landsat data.See its pageNot yet
- MosaicAMLab-UvAA 214M-parameter transformer for fast, probabilistic global weather forecasting with preserved spectral fidelity.See its pageNot yet
- OlmoEarth v1 BaseallenaiA remote sensing foundation model for analysing Sentinel-1, Sentinel-2, and Landsat imagery and time series.See its pageNot yet
- OlmoEarth v1 LargeallenaiA 308M-parameter Vision Transformer foundation model for satellite imagery and Earth observation tasks.See its pageNot yet
- OlmoEarth v1 NanoallenaiA lightweight vision transformer for remote sensing analysis across Sentinel and Landsat satellite imagery.See its pageNot yet
- OlmoEarth v1 TinyallenaiA 6.2M-parameter vision transformer foundation model for analysing satellite imagery and image time series.See its pageNot yet
- OlmoEarth v1.1 BaseallenaiA remote sensing foundation model for analysing Sentinel-1, Sentinel-2, and Landsat satellite imagery.See its pageNot yet
- OlmoEarth v1.1 NanoallenaiA 5.5-million-parameter vision transformer for remote sensing analysis on satellite imagery and time series.See its pageNot yet
- OlmoEarth v1.1 TinyallenaiA 12M-parameter foundation model for remote-sensing tasks across Sentinel-1, Sentinel-2, and Landsat imagery.See its pageNot yet
- OlmoEarth v1.2 BaseallenaiAn Earth-observation foundation model for analysing Sentinel-1, Sentinel-2, and Landsat satellite imagery.See its pageNot yet
- OlmoEarth v1.2 NanoallenaiA compact remote-sensing foundation model extracting features from Sentinel and Landsat satellite imagery.See its pageNot yet
- OlmoEarth v1.2 SmallallenaiA 36M-parameter vision transformer extracting representations from satellite imagery for Earth and climate science.See its pageNot yet
- OlmoEarth v1.2 TinyallenaiA compact 12-million-parameter vision model for Earth observation and remote-sensing analysis.See its pageNot yet
- SamudrACE-E3SMv3allenaiA stochastic coupled atmosphere–ocean emulator for simulating long-term preindustrial climate variability.See its pageNot yet
- StarCOPisp-uv-esA segmentation model for detecting methane plumes in airborne and satellite hyperspectral remote sensing imagery.See its pageNot yet
- StormCast V1nvidiaA mesoscale neural model forecasting 99 atmospheric state variables at kilometre scale using generative diffusion.See its pageNot yet
- SuperIXisp-uv-esEnhances Sentinel-2 satellite imagery to finer resolutions and evaluates super-resolution algorithms for researchers.See its pageNot yet
- Surya 1.0nasa-ibm-ai4scienceA 366M-parameter foundation model trained on solar observatory data for space weather forecasting.See its pageNot yet
- Surya Solar Flaresnasa-ibm-ai4scienceA solar physics model that predicts solar flares from full-disk magnetogram and EUV time series.See its pageNot yet
- Surya Solar Windnasa-ibm-ai4scienceForecasts solar wind plasma and interplanetary magnetic fields at the L1 point for solar physicists.See its pageNot yet
Plus 5 task-specific fine-tunes of the models above, each with its own page in the catalogue.
How to choose
For deterministic global weather prediction, the ECMWF Artificial Intelligence Forecasting System (AIFS) Single family provides 6-hour interval forecast trajectories. AIFS Single 0.2.1 and 1.0 were foundational releases, with AIFS Single 1.1 correcting spurious rainfall artefacts caused by earlier soil moisture loss weighting. AIFS Single 2.0 expands vertical coverage to 10 hPa, integrates an 11-variable ocean wave component, and diagnoses vertical velocity directly. If a unified Earth-system foundation backbone is needed instead, Aurora models atmospheric dynamics, chemistry, ocean waves, and tropical cyclone tracking within a single network.
For probabilistic global forecasting, FourCastNet 3 produces a 60-day forecast at 0.25-degree resolution in under four minutes on a single GPU using spherical convolutions to preserve kinetic energy spectra. Mosaic generates 24-member 10-day forecasts in under 12 seconds on a single H100 GPU using block-sparse attention at 1.5-degree resolution without spectral damping. For operational ensemble forecasting, AIFS ENS 1.0 provides predictions trained on continuous ranked probability scores, while AIFS ENS 2.0 generates 51-member, 15-day forecasts with stratospheric coverage to 10 hPa and wave modelling, though it requires multi-node hardware split across four 120 GB GH200 GPUs per instance.
For regional, mesoscale, or kilometre-scale simulation, StormCast V1 uses a UNet-based diffusion framework to forecast 99 atmospheric variables hourly at convection-allowing scales conditioned on ERA5 and HRRR data. cBottle uses cascaded diffusion on HEALPix geometry to synthesise kilometre-scale states combining ERA5 and ICON simulations. For long-term climate modelling, ACE2-ERA5 provides multi-decadal atmospheric simulation under prescribed sea surface temperatures and carbon dioxide levels while strictly conserving dry air mass and moisture. SamudrACE-E3SMv3 instead simulates coupled atmosphere-ocean variability under preindustrial control forcing, and HiRO-ACE downscales 100 km coarse inputs to 3 km surface precipitation rates using diffusion.
For satellite imagery and text analysis, OlmoEarth provides foundation vision transformers processing Sentinel-1 radar and Sentinel-2 and Landsat optical imagery. Selection across OlmoEarth depends on compute constraints: Nano checkpoints (1.4M to 5.5M parameters) and Tiny checkpoints (6.2M to 12M parameters) suit embedded or high-throughput tasks, while Small (36M), Base (89M to 114M across v1, v1.1, and v1.2), and Large (308M) offer greater capacity for fine-tuning. For specific remote sensing tasks, ML4Floods handles flood segmentation from optical Sentinel-2 and Landsat scenes, StarCOP isolates methane emission plumes in EMIT and AVIRIS hyperspectral data, and SuperIX evaluates Sentinel-2 super-resolution. Outside terrestrial systems, Surya 1.0 forecasts solar flares and space weather from native-resolution solar observatory imagery, while Indus SDE v0.2 extracts scientific entities from geoscience literature.
Questions
Can I run these models without a GPU or installing anything?
2 of the 32 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models generate probabilistic ensemble forecasts, and what hardware do they require?
FourCastNet 3, Mosaic, AIFS ENS 1.0, and AIFS ENS 2.0 generate probabilistic forecasts. FourCastNet 3 runs a 60-day ensemble forecast on a single NVIDIA GPU in under four minutes. Mosaic runs a 24-member 10-day forecast in under 12 seconds on a single H100 GPU, needing roughly 9 GB of memory per rollout. In contrast, AIFS ENS 1.0 requires 24 GB to 38 GB of GPU memory, and AIFS ENS 2.0 splits each ensemble instance across four 120 GB GH200 GPUs.
How do ACE2-ERA5 and SamudrACE-E3SMv3 differ for climate research?
ACE2-ERA5 is trained on ERA5 reanalysis from 1940 to 2022 to simulate global atmospheric dynamics over days to centuries given specified greenhouse gas and sea surface temperature forcings. SamudrACE-E3SMv3 is a stochastic emulator of the DOE E3SMv3 model designed to capture coupled atmosphere–ocean interactions, but is restricted to preindustrial control forcing conditions. Neither model is suitable for operational weather forecasting.
What are the input differences between the OlmoEarth variants?
All OlmoEarth checkpoints (across the v1, v1.1, and v1.2 series) share the same sensor compatibility: they accept single images or multi-temporal image series from Sentinel-1 radar, Sentinel-2 optical, and Landsat optical sensors. They differ in model parameter sizes—ranging from 1.4 million parameters in OlmoEarth v1 Nano up to 308 million parameters in OlmoEarth v1 Large—to balance inference throughput with representational capacity.
Can ML4Floods map flood extent under all weather conditions?
No. ML4Floods is built on multioutput Unet architectures designed specifically for optical imagery from Sentinel-2 and Landsat. Because it relies on optical and shortwave infrared bands, it cannot map ground surface inundation beneath opaque cloud cover.
What distinguishes AIFS Single 2.0 from AIFS Single 1.1 and 1.0?
AIFS Single 2.0 extends vertical coverage into the stratosphere up to 10 hPa, whereas versions 1.0 and 1.1 terminate at 50 hPa. Version 2.0 also integrates an 11-variable marine wave component for wave predictions and treats vertical velocity as a diagnostic field rather than stepping it forward prognostically.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.