Papers with a high importance score: watchlist labs, top venues, open code, operational use, citations, and a significance judgment by Claude. See how the score works. RSS.
Operational AI ensemble WeatherNext Cyclones predicts tropical cyclone track, intensity and size to 15 days, gaining about 1 day of lead time over leading operational models.
Extremes & hazardsColorado State University (incl. CIRA)Google DeepMindGoogle ResearchMet OfficeNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Updates ECMWF's AIFS with bounding layers for physical consistency, better precipitation and new variables; operational at ECMWF since February 2025.
AERIS is a 1.3 to 80B parameter Swin diffusion transformer reaching 10.2 ExaFLOPS on Aurora, stable to 90-day forecasts and outperforming IFS ENS.
Software, scaling & compressionAllen Institute for AI (Ai2, ACE)Argonne National LaboratoryUniversity of California, Los Angeles (UCLA)University of Chicago
SamudrACE is a coupled atmosphere-ocean ML emulator giving stable centuries-long runs at 1 degree with realistic ENSO variability absent in uncoupled mode.
Climate emulationAllen Institute for AI (Ai2, ACE)Columbia University (LEAP, Lamont-Doherty)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)New York University (M2LInES)Princeton University
ACE2 produces skilful winter seasonal forecasts from SST-persisted initialisation, reaching NAO correlation 0.47 over 1993–2016, comparable to a physics-based model.
Sub-seasonal to seasonalAllen Institute for AI (Ai2, ACE)Met Office
AIFS-CRPS trains an ML ensemble model on a CRPS-based loss, outperforming the IFS ensemble for most medium-range variables and competing at subseasonal lead times.
ArchesWeather is a transformer deterministic model, and ArchesWeatherGen a flow-matching ensemble that beats IFS ENS and NeuralGCM on WeatherBench headline variables at academic compute cost.
A NeuralGCM-based hybrid model trained on satellite precipitation cuts climate-model precipitation biases and beats the ECMWF ensemble at mid-range forecasts.
ACE2 is a 450M-parameter emulator that conserves dry air and moisture, runs about 1500 simulated years per day, and reproduces El Niño responses over 80 years.
Climate emulationAllen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
FourCastNet trained without Category 3-5 cyclones cannot forecast Category 5 storms, showing AI weather models fail to extrapolate to unseen extremes.
Extremes & hazardsNew York University (M2LInES)University of Chicago
Combining RCM dynamical downscaling with diffusion refinement lets large CMIP6 ensembles be downscaled cheaply, giving better uncertainty bounds than BCSD and smaller dynamical ensembles.
Downscaling & super-resolutionCalifornia Institute of Technology (Caltech)Google Research
Shows weather-trained ML models forecast skillfully across pre-industrial, present-day and +2.9K climates, though two models show global-mean drift biases toward their training climate.
Graph-based stretched-grid model refines a global AI model to 2.5 km over the Nordics, beating the MEPS ensemble mean for 2 m temperature but underestimating extremes.
Regional & km-scale modelsECMWFMET Norway (Norwegian Meteorological Institute)
StormCast, a generative diffusion emulator of the 3 km HRRR convection-allowing model, gives competitive 1-6 hour radar reflectivity skill with realistic cold pools and power spectra.
FuXi Weather is an ML system that assimilates multi-satellite observations and produces 10-day global forecasts, outperforming ECMWF HRES in observation-sparse regions like central Africa.
Data assimilationChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Tsinghua University (NowcastNet)University of California, Los Angeles (UCLA)
A 1.1-billion-parameter SFNO ensemble with bred vectors and perturbed parameters passes spectral tests and is calibrated, setting up ML huge ensembles of 1,000+ members.
Probabilistic & ensemble forecastingLawrence Berkeley National Laboratory / NERSCNVIDIAUniversity of California, Berkeley
A U-Net emulator of embedded cloud-resolving models gives a stable 5-year hybrid climate run with realistic cloud condensate and a 0.96 mm/day precipitation RMSE.
Shows Pangu-Weather, FourCastNet and GraphCast fail to reproduce sub-synoptic and mesoscale phenomena, lacking physical consistency versus physics-based models.
AIFS is ECMWF's operational GNN-transformer forecast system, trained on ERA5 and operational analyses, with skilled upper-air, surface and tropical cyclone forecasts run publicly.
Aurora is a large Earth-system foundation model trained on over a million hours of data that beats operational systems for air quality, waves, cyclone tracks and high-resolution weather.
Foundation modelsThe Alan Turing InstituteMicrosoft Research (AI for Science)Silurian AIUniversity of Cambridge (Aardvark)
FuXi-ENS, a VAE-based ensemble trained with CRPS and KL loss at 0.25°, beats ECMWF ENS in CRPS on 98.1% of 360 variable and lead time combinations.
Probabilistic & ensemble forecastingChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
End-to-end model ingests raw observations to produce global and station forecasts, beating operational NWP on several variables using 8% of input data.
Data assimilationThe Alan Turing InstituteECMWFGoogle DeepMindMicrosoft Research (AI for Science)University of Cambridge (Aardvark)
First data-driven global model at 0.09 degree resolution, initialized by transfer from a low-resolution model, outperforming IFS-HRES in 2022 hindcasts.
Global medium-range NWPChina Meteorological Administration (CMA)Nanjing University of Information Science & Technology (NUIST)Shanghai Artificial Intelligence Laboratory
Lagged ensembles give a parameter-free probabilistic comparison showing GraphCast and Pangu tie on CRPS, and that multi-step losses hurt probabilistic skill.
Probabilistic & ensemble forecastingLawrence Berkeley National Laboratory / NERSCNVIDIAUniversity of California, Berkeley
GenCast, a diffusion ensemble trained on reanalysis, beats ECMWF ENS on 97.4% of 1320 targets, with better tropical cyclone and wind power prediction, in 8 minutes.
Probabilistic & ensemble forecastingGoogle DeepMindGoogle Research
Adas is an AI data assimilation model that, combined with FengWu, forms an end-to-end AI forecast system with skill comparable to IFS.
Data assimilationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Shanghai Artificial Intelligence Laboratory
Couples the FengWu forecast model with 4D-Var using autodifferentiation instead of adjoints, producing cyclic analyses from simulated observations.
Data assimilationShanghai Artificial Intelligence LaboratoryTsinghua University (NowcastNet)
Machine learning S2S model outperforms ECMWF's S2S system for precipitation and OLR up to 42 days, extending MJO prediction skill from 30 to 36 days.
Sub-seasonal to seasonalChina Meteorological Administration (CMA)Fudan University (FuXi)
NeuralGCM couples a differentiable dynamical solver with ML components, matching ML and ECMWF ensemble skill for 1-15 day forecasts and stable multi-decade climate runs.
Hybrid physics-ML & parameterizationECMWFGoogle DeepMindGoogle ResearchMassachusetts Institute of Technology (MIT)
ACE is a 200M-parameter autoregressive emulator that stays stable for 100 years, reproduces a 100 km reference model's climate, and runs about 100x faster.
Climate emulationAllen Institute for AI (Ai2, ACE)NVIDIAUniversity of California, Berkeley
Two-step UNet plus diffusion residual model downscales 25 km reanalysis to 2 km over Taiwan, recovering spectra and sharpening fronts and typhoon structures.
WeatherBench 2 provides an open evaluation framework, training and ground-truth data, and baselines for data-driven global 1-14 day forecasts.
Benchmarks, datasets & evaluationECMWFGoogle DeepMindGoogle Research
Compares PanguWeather with ECMWF IFS and finds comparable skill for global and extreme metrics, with forecast smoothness and bias drift as drawbacks.
NowcastNet unifies advection physics with learned conditional generation for extreme precipitation nowcasts up to 3 hours, ranked first by meteorologists in 71% of cases.
NowcastingChina Meteorological Administration (CMA)Tsinghua University (NowcastNet)University of California, Berkeley
Releases EUPPBench, a time-aligned forecast-observation dataset with evaluation methods and a first benchmark of 2 m temperature postprocessing.
Benchmarks, datasets & evaluationDeutscher Wetterdienst (DWD)ECMWFKarlsruhe Institute of Technology (KIT)Met OfficeMétéo-France (incl. CNRM)MeteoSwisscode
SEEDS uses diffusion models to emulate operational weather ensembles, generating thousands of realistic members cheaply with improved reliability and extreme-event probabilities after bias correction.
FuXi is a cascaded ML system giving 15-day global forecasts at 0.25 degrees, the first ML model to match the ECMWF ensemble mean at 15 days.
MetNet-3 learns from sparse and dense sensors to forecast precipitation, wind, temperature and dew point up to 24 hours ahead, beating HRRR and ENS over CONUS and running in Google Search.
FengWu is a multi-modal transformer forecasting system that extends skillful forecasts to 10.75 days, beating GraphCast on 80% of 880 predictands.
Global medium-range NWPInstitute of Atmospheric Physics, Chinese Academy of SciencesNanjing University of Information Science & Technology (NUIST)Shanghai Artificial Intelligence Laboratory
3D-Geoformer, a self-attention transformer, predicts Niño3.4 SST anomalies 18 months ahead with high correlation skill for ENSO.
Sub-seasonal to seasonalNanjing University of Information Science & Technology (NUIST)
ClimaX is a transformer foundation model pretrained on CMIP6 that fine-tunes to many weather and climate tasks, outperforming data-driven baselines.
Foundation modelsMicrosoft Research (AI for Science)University of California, Los Angeles (UCLA)code
GraphCast, a graph neural network trained on ERA5, makes 10-day global forecasts at 0.25 degrees in under a minute, beating operational systems on 90% of targets.
3D Earth-Specific Transformer trained on ERA5 is the first AI model to beat IFS across variables and lead times from 1 hour to one week.
Introduces ClimateBench, a CMIP-based benchmark with baseline ML emulators predicting temperature and precipitation responses to emissions scenarios.
Benchmarks, datasets & evaluationETH ZurichMET Norway (Norwegian Meteorological Institute)University of Oxford
FourCastNet uses adaptive Fourier neural operators to produce global medium-range forecasts about 80,000 times faster than NWP, scaling to thousands of GPUs.
Software, scaling & compressionLawrence Berkeley National Laboratory / NERSCNVIDIA
FourCastNet uses adaptive Fourier neural operators to forecast global 0.25 degree weather in under 2 seconds, matching IFS at short leads and beating it for precipitation.
Global medium-range NWPCalifornia Institute of Technology (Caltech)Lawrence Berkeley National Laboratory / NERSCNVIDIA
Machine-learned nudging tendencies correct a 200-km FV3GFS model toward 3-km X-SHiELD, gaining 18 hours of 3-7 day skill and 30% better precipitation.
Hybrid physics-ML & parameterizationAllen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Washington
Climate-invariant ML embeds physical knowledge into parameterization emulators so they generalize across climate regimes, maintaining offline accuracy in three atmospheric models.
Explainability, uncertainty & physicsColumbia University (LEAP, Lamont-Doherty)Google ResearchMassachusetts Institute of Technology (MIT)NVIDIAUniversity of California, Los Angeles (UCLA)
Machine learning that accounts for model uncertainty detects a physically interpretable anthropogenic signal in extreme precipitation across all global observational datasets.
Climate analysis & attributionUniversity of California, Los Angeles (UCLA)
DGMR, a deep generative model of radar, produces skillful 5-90 minute precipitation nowcasts that Met Office experts ranked first in 88% of cases.
A CNN-based cubed-sphere ensemble of 32 deep-learning weather models gives 2-6 week forecasts that are only modestly worse than the ECMWF S2S ensemble over land.
Sub-seasonal to seasonalMicrosoft Research (AI for Science)University of Washington
Review of statistical postprocessing methods for NWP forecasts and their challenges in big-data settings, with little explicit machine learning focus.
Post-processing & calibrationDeutscher Wetterdienst (DWD)ECMWFKarlsruhe Institute of Technology (KIT)KNMI (Royal Netherlands Meteorological Institute)MET Norway (Norwegian Meteorological Institute)Met OfficeMétéo-France (incl. CNRM)MeteoSwiss
MetNet uses axial self-attention on radar and satellite inputs to forecast precipitation up to 8 hours at 1 km, beating NWP for 7-8 hour lead times over the US.
CNN on a cubed-sphere grid with multi-step loss gives stable global forecasts for weeks, beating persistence, climatology and coarse NWP at short-to-medium range.
Global medium-range NWPMicrosoft Research (AI for Science)University of Washington
WeatherBench introduces an ERA5-derived benchmark with standard metrics and baselines (linear regression, deep learning, physical models) for data-driven medium-range global weather forecasting.
Benchmarks, datasets & evaluationECMWFTechnical University of Munich (TUM)University of Washingtoncode
Machine-learning-based detection indicates the climate change signal is identifiable from a single day of global weather; abstract unavailable, so classification is tentative.
Climate analysis & attributionETH ZurichSwiss Data Science Center (SDSC)
WeatherBench provides an ERA5-derived benchmark with metrics and baselines for data-driven global medium-range forecasting, enabling direct model comparison.
Benchmarks, datasets & evaluationECMWFUniversity of WashingtonTechnical University of Munich (TUM)code
A U-Net treats 1-hour, 1 km precipitation nowcasting as image-to-image translation and outperforms optical flow, persistence and HRRR.
Deep learning model trained on climate model simulations produces multi-year ENSO forecasts, a landmark result for long-lead ENSO prediction.
Sub-seasonal to seasonalNanjing University of Information Science & Technology (NUIST)
Position paper surveying how machine learning can help climate change mitigation and adaptation across sectors such as energy and disaster management; not focused on weather or climate science.
Applications & impactsETH ZurichGoogle DeepMindGoogle ResearchLawrence Berkeley National Laboratory / NERSCMassachusetts Institute of Technology (MIT)New York University (M2LInES)Stanford UniversityUniversity of California, Berkeley
Nature review on deep learning for data-driven Earth system science, covering weather, climate and process understanding; abstract absent, judged from title and venue.
Climate analysis & attributionLawrence Berkeley National Laboratory / NERSCMax Planck Institute for BiogeochemistryMax Planck Institute for Meteorology
Discusses challenges and design choices for deep-learning global weather forecasting, illustrated with a toy global neural forecast model.
A neural network trained on cloud-resolving output replaces subgrid convection in a global climate model, giving stable multi-year runs that capture precipitation extremes and approximately conserve energy.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)
Neural networks for ensemble post-processing outperform benchmark distributional regression for 2-m temperature at German stations, using embeddings and auxiliary predictors.
Post-processing & calibrationKarlsruhe Institute of Technology (KIT)
A blueprint proposes Earth system models that learn parameterizations from observations and high-resolution simulations via data assimilation and machine learning, illustrated on a toy system.
Hybrid physics-ML & parameterizationCalifornia Institute of Technology (Caltech)NASA
Neural network emulates CAM longwave radiation 50-80 times faster with near-identical 10-year climate simulations, an early ML physics emulation.
Hybrid physics-ML & parameterizationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
