ML for data assimilation and state estimation, learned observation operators, and end-to-end models that forecast directly from observations.
Search and filter this topic Plain list on GitHub
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.
China Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Tsinghua University (NowcastNet)University of California, Los Angeles (UCLA)
End-to-end model ingests raw observations to produce global and station forecasts, beating operational NWP on several variables using 8% of input data.
The Alan Turing InstituteECMWFGoogle DeepMindMicrosoft Research (AI for Science)University of Cambridge (Aardvark)
Adas is an AI data assimilation model that, combined with FengWu, forms an end-to-end AI forecast system with skill comparable to IFS.
Fudan 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.
Shanghai Artificial Intelligence LaboratoryTsinghua University (NowcastNet)
GraphDOP is an ECMWF graph model trained and initialised only from satellite and conventional observations, giving skilful forecasts up to five days ahead.
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.
China Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Tsinghua University (NowcastNet)University of California, Los Angeles (UCLA)
Score-based diffusion data assimilation at km scale assimilates 40 sparse weather stations, lowering held-out station RMSE by 10% versus HRRR.
AI-Var trains a neural network to minimise the variational DA cost function directly, producing accurate initial conditions without pre-existing analysis datasets in proof-of-concept tests.
FuXi-DA is a deep learning data assimilation framework that assimilates FY-4B AGRI satellite radiances, reducing analysis errors and improving forecasts.
China 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.
The Alan Turing InstituteECMWFGoogle DeepMindMicrosoft Research (AI for Science)University of Cambridge (Aardvark)
DiffDA adapts the pretrained GraphCast network as a diffusion backbone to assimilate sparse observations at 0.25 degree, losing at most 24 hours of lead time versus state-of-the-art analyses.
Adas is an AI data assimilation model that, combined with FengWu, forms an end-to-end AI forecast system with skill comparable to IFS.
Fudan 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.
Shanghai Artificial Intelligence LaboratoryTsinghua University (NowcastNet)
End-to-end variational Bayes neural scheme jointly learns a trainable solver and Gaussian posterior for data assimilation with uncertainty, tested on Lorenz-63 and river flow.
Score-based data assimilation with a memory-efficient score network is tested on a two-layer quasi-geostrophic atmosphere-like model, demonstrating scalability to high-dimensional geophysical systems.
U-Net predicts flow-dependent covariances for an Ensemble Kalman Filter in a quasi-geostrophic model, matching 3DVar and EnKF for ensemble sizes up to 80.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)NSF NCAR / UCARPrinceton University
VAE-based 3D-Var assimilates simulated temperature observations in a learned latent space, yielding flow-dependent background covariances across tropical and extratropical regions.
Uses denoising diffusion models to generate pseudo-ensembles for ensemble data assimilation, outperforming standard ensemble methods when the simulation model is imperfect.
A CNN estimates state-dependent forecast error covariance from one model run for hybrid data assimilation, matching ensemble Kalman filter performance on Lorenz96 with small ensembles.
Couples data assimilation with ML to optimise WRF land-surface states in the Heihe basin, improving temperature and humidity simulations and ET estimates.
Deep neural network operator surrogates with Bayesian inversion identify sources from sparse observations in an atmospheric dispersion exemplar, with uncertainty quantification under wind uncertainty.
Perspective reviewing machine learning combined with data assimilation for climate surrogate models and model error correction, with open challenges for operational use.
Latent diffusion model inpaints 3D atmospheric boundary-layer LES flow from observations covering under 1% of the volume, producing diverse physically plausible turbulent samples usable as LES initial conditions.
Particle filter with linear inverse model emulators of CMIP-class GCMs produces ensemble millennial temperature reconstructions and skillful one-year-ahead forecasts in a perfect-model test.
Introduces an operator-algebra framework for sequential data assimilation with kernel-based data-driven schemes, tested on Lorenz 96 and ENSO in a climate model.
Recurrent neural networks learn model bias from analysis increments in an idealized Lorenz-96 setup, outperforming linear regression and reducing extended-range forecast error growth.
A CNN detects clouds in camera photos to build a forward operator, enabling convective-scale assimilation of cloud cover into the ICON-D2 model; proof of concept over 6 days.
A neural-network shallow-water emulator of MPAS with tangent-linear and adjoint code enables a continuous 4D-Var assimilation system demonstrated on twin experiments.
Hybrid DA system uses a bilinear network as tangent-linear/adjoint and a CNN to adaptively blend 4DVar and EnKF analyses, tested on Lorenz-96 with lower cost.
Score-based data assimilation learns a generative prior over state trajectories and applies observations only at inference, enabling zero-shot observation scenarios for long dynamical-system trajectories.
Hybrid data assimilation pairs a stochastic reduced-order model with a learned nonlinear map to recover unobserved states in a precipitating quasi-geostrophic model. Idealized atmosphere testbed.
Hybrid network learns 4DVAR-style inversion directly from partial, noisy observations, using the variational mechanistic constraint for regularization and reduced computation.
Machine learning replaces the radiative transfer observation operator for AMSU-A radiance assimilation in NICAM-LETKF, performing near RTTOV but slightly worse with simpler bias handling.
Online neural-network model error correction is added to incremental 4D-Var in ECMWF's OOPS, showing online learning beats offline learning on a quasi-geostrophic model.
Couples POD-LSTM reduced-order models with an ensemble Kalman filter for geophysical flows, improving SST forecast error by about an order of magnitude.
Data-driven analog data assimilation with ML-emulated models selects the best GCM for observations, performing as well as using the true SPEEDY model.
Neural networks emulate 3D-Var analysis of surface and upper-air variables in a WRF setup over Rio de Janeiro, cutting assimilation CPU time 25–121 times.
Machine-learning regression maps AROME proxy variables to pseudo MTG lightning flash extent, giving a learned observation operator for storm-scale EnVar assimilation.
Kalman filter with a trained linear inverse model performs strongly coupled assimilation, cutting SST analysis error over 20% and extending forecast skill to 50 days.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)NSF NCAR / UCARUniversity of Washington
Convolutional networks learn IFS model bias from radio occultation temperatures, pretrained on ERA5 and transfer-learned on one year of IFS data, as an alternative to weak-constraint 4D-Var.
Replaces the forward model in variational data assimilation with a differentiable learned emulator, speeding geopotential height DA by four orders of magnitude.
Convolutional autoencoder reconstructs missing observation fields before EnSRF assimilation, cutting analysis error in a barotropic vorticity testbed with sparse data.
SVM detects shifts in the Lorenz-63 variable distribution to switch between Gaussian and mixed Gaussian-lognormal 3DVar, reducing analysis error.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Builds a pure data-driven 4D-Var framework (ML-4DVAR) using a bilinear neural network surrogate, tested on Lorenz-96 with better assimilation and faster computation than classical 4D-Var.
Deep learning enhances ensemble-based data assimilation for high-dimensional nonlinear dynamical systems; the abstract is missing and atmospheric application is unclear.
Fully connected network emulates the CRTM microwave radiative transfer model for ATMS brightness temperatures over ocean, with mean biases of hundredths of a kelvin for data assimilation.
A hybrid method uses ML regression inside 3DVAR to retrieve fine-scale boundary-layer winds from lidar and in-situ data, improving 2-D and 3-D velocity estimates by 42.3% on average.
Trains neural surrogates using adjoint information to speed up 4D-Var assimilation, showing better results than forward-only surrogates on Lorenz-63.
Uses LSTM networks to learn observation error covariances for data assimilation, outperforming DI01 and D05 tuning on Lorenz-63 and shallow-water twin experiments.
RNNs replace the forecast model, error covariance and adjoint in the DA cycle, with DA updating hidden states; the approach scales via localization and parallelization.
Bayesian neural networks estimate convective-scale model parameters from observations within data assimilation in a 1D shallow-water perfect-model setup, reducing initial state errors.
Neural network emulator of gravity wave drag yields tangent-linear and adjoint models that succeed in 4D-Var assimilation, easing code maintenance.
Compares resolvent and tendency ML error corrections within data assimilation on a two-scale Lorenz model; tendency correction wins in assimilation and enables online learning.
Auto-differentiable ensemble Kalman filters jointly learn surrogate dynamics and assimilate states, outperforming EM and particle-filter baselines on Lorenz-96.
An autoencoder latent-space ensemble Kalman filter with a neural surrogate reduces data assimilation cost and beats ETKF-Q on an augmented Lorenz 96 system.
Equivariance-preserving U-STN networks improve data-driven Z500 forecasts by 45% over U-Nets, and coupling them with ensemble Kalman data assimilation gives stable cycles under high observation noise.
Patched autoencoder-based constructed analog ensembles outperform unpatched versions on a 1D toy model, offering a route to scale to global domains.
Nonlinear physics-informed autoencoder surrogates extend multifidelity ensemble Kalman filtering, outperforming linear surrogates on Lorenz 96 data assimilation.
A learned observation-to-state mapping improves variational data assimilation, tested on Lorenz96 and 2D turbulence with better forecast quality.
Perspective framing data assimilation and machine learning as equivalent inverse methods in a Bayesian network, e.g. linking 4D-Var to recurrent neural networks, to learn earth system models from observations.
Multilayer perceptron-based method for hybrid data assimilation of atmospheric states (no abstract; classification based on title).
Neural networks reproduce weak-constraint 4D-Var model error corrections in the operational ECMWF IFS and could extend correction to the whole atmospheric column.
Hybrid model combines a knowledge-based quasi-geostrophic model with deep learning, learned iteratively within 4D-Var, to correct model error and improve short- to mid-range forecasts.
Deep recurrent architecture learns data assimilation with approximate prior and posterior densities, matching EnKF performance on the Lorenz-95 toy system with Gaussian errors.
A conditional sampling method uses short partial observations to generate training data that reduces model error and improves machine learning forecasts of turbulent and climate-like systems.
Ensemble Kalman filter estimates model error from sparse noisy observations to train a neural parametrisation, yielding hybrid forecasts with better skill on Lorenz and a coupled ocean-atmosphere model.
Lawrence Berkeley National Laboratory / NERSCUniversity of Reading
End-to-end neural variational data assimilation, with learned model and solver, outperforms classical gradient descent on Lorenz-63 and Lorenz-96 reconstructions.
Combines random feature map forecasting with ensemble Kalman filter data assimilation, learning sequentially from noisy observations to give long forecast skill, reliable ensembles and closure for multiscale chaotic systems.
Reservoir computing trained on LETKF analyses predicts the Lorenz 96 system under sparse, noisy observations and imperfect models, outperforming LETKF when the model is imperfect.
LSTM-learned nudging gain for data assimilation outperforms EKF and EnKF on Lorenz 96 with sparse observations, acting as a stabilizer at lower computational cost.
An artificial neural network estimates IASI cloud-top pressure and fraction closer to optimal values, cutting retrieval time and improving 1D-Var cloudy-sky assimilation with near-neutral forecast impact.
Neural network emulates WRF 3D-Var data assimilation over Rio de Janeiro using airport station data, matching analyses to about 0.3 K and running 71 times faster.
Quantum-mechanics-inspired, kernel-based data assimilation using Koopman operators, applied to Indo-Pacific SST and Niño 3.4 probabilistic forecasts. Novel but early-stage.
Unifies data assimilation, machine learning and expectation-maximization to jointly estimate model, state trajectory and error statistics for noisy observations of low-order chaotic geophysical models.
Iterating an ensemble Kalman filter with a neural network emulates the chaotic Lorenz-96 model from sparse noisy observations, giving forecast skill up to two Lyapunov times.
Support vector machines and neural networks detect non-Gaussian error distributions in the Lorenz-63 model, informing relaxation of Gaussian assumptions in data assimilation.
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.
A neural-network tree-ring width proxy system model outperforms linear and VS-Lite models for paleoclimate data assimilation, ranking best among the four compared.
Assimilating SAPHIR microwave radiances into WRF with an ANN radiative transfer surrogate improves 72-hour track forecasts of cyclone Vardah.
Integrates neural network and case-based reasoning into a precipitation data assimilation system to improve weather forecasts; results reported as effective.
Proposes machine-learning-based adaptive covariance localization for ensemble Kalman filters, tuning the localization radius in time and space to counter sampling errors.
MACIN combines a CNN cloud detector with data assimilation to nowcast direct irradiance up to 20 minutes ahead, improving on persistence beyond one minute.
A physics-constrained deep learning radiative transfer model over oceans recovers correct Jacobians, enabling radiance sensitivities for satellite data assimilation.
Observations & remote sensingNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
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.
Latent-space digital twin of JULES-INFERNO for global wildfire prediction, with latent data assimilation, running 500 times faster than the original model.
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
Combines polynomial chaos expansion and Gaussian process regression to estimate and forecast wind fields from ADS-B aircraft-derived data, with confidence intervals.
Neural networks learn FV3-GFS systematic error from data assimilation analysis increments and apply online corrections, improving 10-day forecast skill at all lead times.
Hybrid physics-ML & parameterizationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Columbia University (LEAP, Lamont-Doherty)
Machine-learning quality control for ABI water vapor radiances blends seven bands to reject surface-contaminated pixels, improving Hurricane Harvey track forecasts.
Observations & remote sensingNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
Regression trees, MLPs, CNNs and LSTMs estimate local Lyapunov exponents from trajectories of Lorenz-63 and Rössler systems, a feasibility study motivated by weather predictability.
Paleoclimate data assimilation with ANN proxy system models produces a millennial, spatially continuous Northern Hemisphere temperature reconstruction from tree rings.
A CNN remaps FY-3D MWRI brightness temperatures to NWP grids, reducing observation minus background bias and spread in GRAPES 4D-Var.
Observations & remote sensingChina Meteorological Administration (CMA)
Compressed sensing learns a stochastic parameterization of fine-scale processes in the Lorenz-96 atmospheric model, combined with ensemble Kalman filtering and compared with polynomial regression.
A deep neural network fuses NUCAPS, ABI and RTMA data to cut low-level temperature and moisture profile RMSE by over 30% versus ERA5.
Observations & remote sensingNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Institute of Atmospheric Physics, Chinese Academy of Sciences
Discusses whether deep learning could fully replace numerical weather models and data assimilation, reviewing ML concepts and their fit for weather data.
Global medium-range NWPForschungszentrum Jülich / Jülich Supercomputing Centre
Compares CDF-matching and neural network retrievals of ASCAT soil moisture for NWP assimilation, finding inversion better suited than forward modelling.
Reviews how artificial intelligence is applied across numerical weather prediction, including data assimilation, error correction, and end-to-end neural forecasting models.
A CNN called GREMLIN fuses GOES-R ABI and GLM data into synthetic radar reflectivity, proposed to feed convective-scale data assimilation and latent heating initialization.
Observations & remote sensingColorado State University (incl. CIRA)
Data assimilation corrects neural-network subgrid closures during online runs, improving long-term prediction of chaotic Lorenz 96 and Kraichnan turbulence dynamics.
Recurrent network predicts global spatiotemporal uncertainty of 10 m ocean wind reanalyses, trained at a few locations and generalizing globally, for ensemble generation and assimilation.
A CNN converts GOES-R radiances and lightning into synthetic radar reflectivity, enabling convective radar assimilation with much better skill than pixel-wise methods.
Observations & remote sensingColorado State University (incl. CIRA)
Uses a Gaussian process to decide when to release sensors from a weather balloon, targeting better stratospheric analyses and forecasts, tested on real flights.
Review arguing modern AI can augment or replace parts of satellite remote sensing and NWP processing chains, from observations to forecasts.
Observations & remote sensingColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Neural network combined with 1D-variational retrieval of FY-3C microwave temperature and humidity profiles during Hurricane Matthew, improving WRF data-assimilation track forecasts.
Observations & remote sensingNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
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
