64 papers in the archive · tier-1 watchlist.
Enhanced DEVINE deep-learning model downscales terrain-forced near-surface winds and generalizes to unseen alpine and coastal terrain, with station MAE around 1.13 m/s.
Score-based data assimilation with a diffusion prior reconstructs Western North Pacific surface winds from sparse CYGNSS data, lowering tropical cyclone core intensity bias by 70-78%.
Observations & remote sensingChina Meteorological Administration (CMA)Météo-France (incl. CNRM)
Multi-fidelity ML parameterization trained on warmer-climate low-fidelity data generalises to warmer climates, with uncertainty that tracks prediction error.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)Météo-France (incl. CNRM)NVIDIA
Gradient boosting model trained on snowpack and weather data projects avalanche activity declining about 9% per decade since 1958, with future trends from downscaled simulations to 2100.
UNet with Sobel-penalty loss downscales AROME and ARPEGE to 300 m urban air temperature over Paris, with mean bias of about -0.3 to -0.5 °C against stations.
WassersteinGrad replaces pointwise SmoothGrad averaging with entropic Wasserstein barycenters of attribution maps, giving more robust explanations of autoregressive regional weather forecasts.
Explainability, uncertainty & physicsMétéo-France (incl. CNRM)
XGBoost-based ML observation operator assimilates TROPOMI solar-induced fluorescence into the IFS, improving leaf area index and gross primary productivity over cropland.
Evaluates statistical and ML decomposition methods on nudged simulations; they get the sign of circulation-driven summer warming right but underestimate its size, which is up to half of 1979–2023 warming.
Climate analysis & attributionETH ZurichMétéo-France (incl. CNRM)
STIPP jointly postprocesses station forecasts over the French Alps, producing hourly spatio-temporally consistent ensembles from six-hourly deterministic input with improved accuracy for several variables.
Compares empirical, physics, regression and ML methods for diagnosing heat-stress WBGT from NWP output; a neural network scores best for the red flag category (CSI 0.191).
Post-processing & calibrationMétéo-France (incl. CNRM)NSF NCAR / UCAR
Adds a learned triggering mechanism to a neural deep-convection parameterization in ARP-GEM, which runs stably online and generalizes to a warmer climate.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)
Benchmarks wavelet, gradient and other training losses and their combinations for deep statistical downscaling of wind speed and precipitation; no single loss dominates.
A GAN-based framework reconstructs multi-level tropical cyclone wind speed, including the 10 m level, from sparse dropsonde observations within seconds, tested on Hurricane Hunter data.
A machine-learning model trained on observed large-scale predictors attributes CMIP6 low-cloud biases about equally to large-scale conditions and to parameterizations.
Uses MLPs trained on ten years of observed surface turbulent fluxes to remove forcing errors when evaluating climate model heat fluxes, revealing a land-scheme bias in sensible heat.
Applies machine learning to analyze the Lagrangian evolution of low clouds in climate models, diagnosing cloud behaviour across model simulations.
A UNET maps daily sea-level pressure to European temperature anomalies in the MIROC6 large ensemble, beating analogs at daily and interannual scales.
Perspective on ML emulators that augment kilometre-scale regional climate simulations, identifying extremes, multivariate coherence and out-of-sample downscaling as open issues.
Downscaling & super-resolutionMet OfficeMétéo-France (incl. CNRM)
StyleGAN-2 generates extra members for the operational kilometre-scale AROME-EPS, so a few hundred GAN members match tail sampling of a much larger ensemble.
Probabilistic & ensemble forecastingECMWFMétéo-France (incl. CNRM)
Gradient boosting classifies deep convective cloud structures and retrieves ice water path from simulated C2OMODO microwave radiometry, reaching over 70% classification accuracy and R² up to 0.99.
Deep-learning segmentation of radar and satellite images tracks MCS life cycles in AROME-France, identifying three predictors for MCS nowcasting.
CNN on C-band dual-polarisation radar images detects severe hail at the ground, outperforming existing radar hail proxies on a 2018–2023 test set.
A neural network emulator of the deep convection scheme runs online in the ARP-GEM1 global model via OASIS coupling, with a 5-year simulation matching the physics-based reference.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)
Hybrid model enforces advection physics as a hard constraint on a U-Net for cloud cover nowcasting, beating EUMETSAT EXIM and a plain U-Net with one-third of the data.
CNN RCM-emulator with an asymmetric loss reproduces 12 km precipitation distributions and extremes, transferring across driving GCMs.
U-Net distributional regression post-processes AROME-EPS precipitation ensembles, matching quantile regression forests in CRPS and beating them on heavy precipitation.
LSTM post-processes AROME ensemble and satellite data to predict convection areas for air traffic, matching a probabilistic benchmark for moderate convection and beating it for total convection.
A U-Net detects tropical cyclone wind structure in kilometer-scale AROME forecasts with IoU near 0.8, outperforming an operational heuristic, especially for weak storms.
Machine learning is used to reduce parametrization errors in polar surface turbulent flux schemes; abstract unavailable, so classification relies on the title.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)
Two neural networks correct AROME wind forecasts at 1300 m, then downscale to 30 m with DEVINE, optimizing corrections jointly with the downscaling step.
Neural-network flux parametrizations trained on earlier Arctic campaigns reduce RMSE of sea-ice surface turbulent flux predictions by up to 70% versus a polar bulk algorithm on MOSAiC data.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Combines two CNNs with custom MSE losses to downscale 9-km WRF wind forecasts to 1 km over Alpine terrain, recovering topography-driven wind features.
A review of satellite-based objective tropical cyclone analysis from 2019–2022, surveying AI methods and calling for better Best Track training data.
Extremes & hazardsMétéo-France (incl. CNRM)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Espresso is a global deep-learning rainfall estimator from geostationary imagery, calibrated to GPM, now operational at Météo-France with 30-minute updates.
A spectrally normalized residual GAN emulates 10 m wind and 2 m temperature from the kilometre-scale AROME-EPS, matching distributions and spectra and exposing positional-bias limits.
Reviews seasonal and multi-annual tropical cyclone forecasting, including emerging ML/AI methods, ENSO-neutral predictability and landfall skill gaps relative to user needs.
Sub-seasonal to seasonalBureau of Meteorology (Australia)China Meteorological Administration (CMA)Colorado State University (incl. CIRA)Japan Meteorological Agency (JMA)Met OfficeMétéo-France (incl. CNRM)NSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Argues machine learning in climate model parameterization still requires expert judgment, so it is no more objective than conventional tuning.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)
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
DRAIN, a U-Net on GPM passive microwave brightness temperatures, retrieves rain rates with performance similar to or slightly better than the official GPROF algorithm.
DEVINE, a CNN trained on ARPS simulations, downscales AROME winds to 30 m over the Alps and cuts the mean wind-speed bias by 27% at 61 stations.
Machine-learning regression maps AROME proxy variables to pseudo MTG lightning flash extent, giving a learned observation operator for storm-scale EnVar assimilation.
A neural-network RCM emulator hybrid-downscales daily 12 km near-surface temperature from large-scale predictors, reproducing RCM variability with training in under 2 hours and inference in under a minute.
Proposes ML parameterizations with parameters as inputs, calibrated online against long-term statistics, which significantly reduces long-term biases in two simple dynamical systems.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)
DeepFire neural emulator of fire spread turns high-resolution weather forecasts into next-day fire danger maps and ensembles for 13 Corsican fires.
Compares linear regression and neural networks for filling gaps in micrometeorological station data in complex terrain, finding significant wind differences only in synoptic flow.
A CNN detects bow echoes in French kilometre-scale model output with 86% hit rate and 39% false alarms, and forecasters found the products useful.
Proposes tunable neural-network climate parameterizations calibrated online against long-term statistics, reducing long-term biases of the ML model in a simple dynamical system.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)
An extended Lorenz-63 toy model reveals feedforward network instability, which training on Latin Hypercube samples removes, showing sampling design matters for stability.
Explainability, uncertainty & physicsMétéo-France (incl. CNRM)
An ECMWF and ESA workshop report reviews the status and opportunities of machine learning for Earth system observation and prediction.
Observations & remote sensingECMWFMétéo-France (incl. CNRM)University of Reading
Calibrates cloud geometry parameters for 3D radiation using ML surrogate emulators, reducing top-of-atmosphere and surface flux errors from about 10 to 5 W/m2.
Hybrid physics-ML & parameterizationThe Alan Turing InstituteECMWFMétéo-France (incl. CNRM)
Wasserstein GAN with a CNN generator learns a 64-dimensional latent space of synthetic PLASIM climate states, reproducing leading EOF statistics and geostrophic balance.
Gaussian process calibration tunes boundary-layer parameterization parameters against large-eddy simulations in single-column mode, before global model tuning.
Hybrid physics-ML & parameterizationThe Alan Turing InstituteMétéo-France (incl. CNRM)
Emulator-based calibration of an atmospheric model's parameterizations against large-eddy simulations matches hand-tuned global performance while keeping process-level behaviour physically sound.
Hybrid physics-ML & parameterizationThe Alan Turing InstituteMétéo-France (incl. CNRM)
CNNs on Meteosat imagery for cloud cover nowcasting; U-Net beats persistence and the AROME physical model on meteorological metrics.
End-to-end neural variational data assimilation, with learned model and solver, outperforms classical gradient descent on Lorenz-63 and Lorenz-96 reconstructions.
A U-Net with a learned predictor-weighting layer post-processes ARPEGE cloud cover forecasts over Europe, outperforming random forests and logistic quantile regression.
POD combined with Kriging or Random Forest emulates Arome-NWC rainfall nowcasts over France at 1-minute resolution with much lower computational cost.
Operational quantile regression forests and ensemble copula coupling improve station temperature and gridded hourly rainfall forecasts in France, better capturing cold and heat waves than raw ensembles.
Random forests classify continuous versus intermittent precipitation textures in AROME high-resolution forecasts, outperforming other classifiers and transferring to gridded observations for verification.
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
Machine-learning regression diagnoses horizontal visibility from mesoscale NWP forecasts; abstract not available, so the scope is inferred from the title.
ML-tuned blending of multiple ensembles gives probabilistic thunderstorm forecasts from 9 hours to 4 days, clearly better than raw ensemble output.
Neural-network soil moisture retrievals from SMOS are assimilated into ECMWF's land model, improving Southern Hemisphere and some Northern Hemisphere forecasts through atmospheric feedback.
Quantile regression forests calibrate Météo-France ensemble temperature and wind forecasts, outperforming EMOS and giving better-calibrated, sharper probabilistic forecasts from 3 to 54 h.
