60 papers in the archive.
Argues AI weather and climate foundation models need physical evaluation beyond skill scores, proposing five priorities and recommendations such as shared physics-based reporting standards.
Explainability, uncertainty & physicsNASAUniversity of Reading
Recreates the 1944 D-Day forecasts with ECMWF's AIFS-ENS, showing the ML ensemble moves the storm too fast and gives overconfident landing conditions.
Probabilistic & ensemble forecastingECMWFUniversity of Reading
Perspective arguing AI is merging weather forecasting and climate prediction into unified Earth-system models, while flagging trust, energy and equity challenges.
Foundation modelsColumbia University (LEAP, Lamont-Doherty)Max Planck Institute for BiogeochemistryPotsdam Institute for Climate Impact Research (PIK)Technical University of Munich (TUM)University of Reading
Feature-selection workflow picks 3 of 28 Himawari channels plus radar for lightning identification (CSI 0.433) and nowcasting (CSI 0.284), beating full-channel models.
NowcastingNanjing University of Information Science & Technology (NUIST)University of Reading
GraphCast inherits and amplifies ERA5's premature afternoon peak and weak nocturnal boreal-summer rainfall, while a 5-km Met Office model improves night-time precipitation.
Global medium-range NWPMet OfficePrinceton UniversityUniversity of Reading
Review arguing machine learning and data assimilation are converging, with ML accelerating DA and DA supplying uncertainty quantification and physical constraints.
Data assimilationBureau of Meteorology (Australia)Colorado State University (incl. CIRA)ECMWFMax Planck Institute for MeteorologyNASANOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)University of Reading
Position paper arguing that statistical ML cannot and should not replace physics-based simulation for weather forecasting.
Position paper argues AI weather forecasts need adapted verification methods, identifying where and when they are useful beyond aggregate MSE skill.
Benchmarks, datasets & evaluationECMWFMet OfficeUniversity of Cambridge (Aardvark)University of Reading
DAG-VAE embeds a causal graph in a VAE latent space to disentangle Pacific and Indian Ocean influences on East African short rains, validated against SST-replacement experiments.
Trains a U-Net to detect Arctic atmospheric rivers in CANARI large-ensemble simulations, removing threshold tuning and matching a rule-based detector's threshold sensitivity.
Extremes & hazardsUniversity of Cambridge (Aardvark)University of Reading
RDSRNet recursively decouples structure and detail to super-resolve radar and satellite meteorological data, claiming gains across both data types.
Observations & remote sensingNanjing University of Information Science & Technology (NUIST)University of Reading
TriPhysGAN-Attn is a physics-guided GAN for radar echo nowcasting that splits advection, initiation and deformation into branches, improving accuracy under severe convective weather.
NowcastingNanjing University of Information Science & Technology (NUIST)University of Reading
Physics-encoded GAN super-resolves tropical cyclone satellite image sequences 4x, using a vorticity-based module and temporal discriminator for motion consistency.
Observations & remote sensingChina Meteorological Administration (CMA)Nanjing University of Information Science & Technology (NUIST)University of Reading
Seven MLWP models match IFS midlatitude cyclone track skill to 10 days but underestimate cyclone central pressure by over 5 hPa and peak winds at all leads.
Global medium-range NWPUniversity of Cambridge (Aardvark)University of Reading
Gaussian process emulator predicts fast surface temperature responses to greenhouse gas and aerosol forcings at a small fraction of GCM cost, with sensitivity analysis.
Climate emulationKarlsruhe Institute of Technology (KIT)University of OxfordUniversity of Reading
Targeted RMM-VAE weather regimes explain 34% of Western European summer warming and link a southerly-flow regime to most heatwave days.
Python interface pyPDAF to the Fortran PDAF data assimilation framework, demonstrated for coupled atmosphere-ocean assimilation in the MAOOAM model.
Optimizes initial conditions of the FuXi subseasonal model to improve prediction of the 2022 Pakistan floods and identifies moisture and circulation processes behind the gain.
Sub-seasonal to seasonalChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)University of Reading
Reviews data assimilation for land surface model parameter estimation and outlines how machine learning could reduce carbon-cycle and water uncertainty.
Data assimilationColumbia University (LEAP, Lamont-Doherty)ECMWFMax Planck Institute for BiogeochemistryNASANSF NCAR / UCARUniversity of Reading
AeroGP, a Gaussian process emulator trained on NorESM ensembles, reproduces aerosol-driven surface temperature changes at a fraction of the cost.
Climate emulationMET Norway (Norwegian Meteorological Institute)University of Reading
A targeted VAE identifies probabilistic circulation regimes that are more informative of Moroccan extreme precipitation while keeping subseasonal predictability.
Climate analysis & attributionThe Alan Turing InstituteECMWFUniversity of Reading
Argues for Tweedie deviance loss over RMSE for precipitation models; shows gains in diffusion downscaling and ConvLSTM nowcasting, with up to 16% MAE improvement.
A ViT-LSTM encoding stratospheric winds and OLR outperforms ECMWF subseasonal hindcasts beyond week 4 for North Atlantic-European weather regimes.
Sub-seasonal to seasonalThe Alan Turing InstituteUniversity of Reading
ANN corrects ERA5 upper-tropospheric relative humidity using IAGOS aircraft data, cutting MAE from 15.7% to 6.6% and improving contrail predictions.
Post-processing & calibrationForschungszentrum Jülich / Jülich Supercomputing CentreUniversity of Reading
Ridge regression on atmospheric and oceanic predictors outperforms GCM-based S2S precipitation forecasts for Senegal's monsoon season, offering a cheaper complement to dynamical systems.
Sub-seasonal to seasonalForschungszentrum Jülich / Jülich Supercomputing CentreUniversity of Reading
Comparison of GAN, WGAN-GP, U-Net diffusion and diffusion-GAN models trained on ERA5 shows each generates UK windstorm populations with different biases for extreme intensities.
FIDN, a DenseNet-based model, predicts final wildfire burnt area from meteorological and environmental drivers for 300+ western US fires, with 82% lower MSE than cellular automata.
A residual CNN analysis step matches the best tuned ensemble Kalman filter on Lorenz 96 using a single forecast state, identifying unstable-subspace perturbations without an ensemble.
Interpretable ML framework attributes gross primary productivity differences among Earth system models to climate drivers versus land process representation.
Learns the analysis map of filters such as the ensemble Kalman filter via variational inference, reducing bias in gain, inflation and localization estimates for nonlinear systems.
Data assimilationCalifornia Institute of Technology (Caltech)University of Reading
Random forest, XGBoost and MLP bias-correct and downscale Met Office UKV urban temperatures over London to 100 m, cutting UHI profile error from 0.64°C to 0.15°C.
Downscaling & super-resolutionBureau of Meteorology (Australia)Met OfficeUniversity of Reading
Compares the deep-learning TCDetect cyclone detector with the TRACK tracker and IBTrACS on ERA-Interim, finding good recall for hurricane-strength storms but many false positives.
RMM-VAE is a variational autoencoder that finds Mediterranean weather regimes targeted to Moroccan precipitation, more predictive of extremes than linear and ML baselines.
Case study of Storm Ciarán finds FourCastNet, Pangu, and GraphCast capture synoptic structure but underestimate peak winds and miss fine warning-relevant features.
Interpretable machine learning attributes differences in Earth system model gross primary productivity to atmospheric drivers versus land process representation across models.
Review of AI methods for predicting climate extremes such as heatwaves, droughts and heavy rain, covering hybrid AI-climate-model approaches, challenges and future perspectives.
Extremes & hazardsKNMI (Royal Netherlands Meteorological Institute)University of Reading
Unsupervised Markov-state analysis of 500 hPa heights identifies Northern Hemisphere winter blocking and teleconnection-like regimes with stronger dynamical grounding than EOFs.
Lightweight deep learning detects tropical cyclone presence during simulations for online data reduction, reaching 92% recall on held-out ERA-Interim years.
Neural-network gas optics (RRTMGP-NN 2.0) is implemented in ECMWF's IFS radiation scheme, matching RRTMGP accuracy while speeding up radiation.
Hybrid physics-ML & parameterizationECMWFUniversity of Reading
Position paper on Earth Virtualization Engines that pair high-resolution physics-based climate models with machine learning and federated access to exabyte-scale climate data.
Software, scaling & compressionCSCS - Swiss National Supercomputing CentreETH ZurichMeteoSwissMassachusetts Institute of Technology (MIT)Max Planck Institute for MeteorologyNSF NCAR / UCARNVIDIAUniversity of Reading
SeaUnet, a fully convolutional network trained on CMIP TC tracks, gives skillful seasonal forecasts of Western North Pacific typhoon track density up to four months ahead.
Sub-seasonal to seasonalChina Meteorological Administration (CMA)Tsinghua University (NowcastNet)University of Reading
Proposes a framework to evaluate XAI methods for climate-science networks on robustness, faithfulness, randomization, complexity and localization, ranking methods on a temperature-map task.
Implements a neural-network gas optics scheme in ECMWF's IFS, speeding radiation by about one third with a similar model climate in 1-year coupled runs.
Hybrid physics-ML & parameterizationECMWFUniversity of Reading
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.
A neural-network-driven stochastic cloud population model parameterizes cumulus convection in WRF, reproducing realistic precipitation statistics, MJO propagation and diurnal cycle in regional tests.
Neural networks correct the ecRad 1D radiation scheme for 3D cloud effects, recovering 3D signal with 20-30% error at about 1% extra runtime versus the 5x costlier SPARTACUS.
Hybrid physics-ML & parameterizationECMWFUniversity of Reading
A neural network emulates an ensemble of urban land surface models and, coupled to WRF, is more accurate and cheaper than the reference TEB scheme.
Hybrid physics-ML & parameterizationECMWFUniversity of Reading
A lightweight deep learning filter detects tropical cyclones in running simulations with 92% recall and 36% precision, so models save only relevant output.
Online sequential-learning combinations of NWP and statistical extended-range forecasts improve probabilistic skill by 5-15% over standard multimodel methods up to lead week 5.
Post-processing & calibrationColumbia University (LEAP, Lamont-Doherty)University of Reading
DGMR, a deep generative model of radar, produces skillful 5-90 minute precipitation nowcasts that Met Office experts ranked first in 88% of cases.
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
Copula-generated synthetic data augments training of a neural emulator of longwave radiation, reducing mean absolute error by up to 62% in a toy setting.
Hybrid physics-ML & parameterizationECMWFUniversity of Reading
Machine-learning-constrained autoconversion and accretion parameterizations, fitted to Azores cloud-probe data, cut uncertainty to about 15% and 5% and reveal drizzle number's role.
Hybrid physics-ML & parameterizationColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of ReadingUniversity of Washington
Machine learning maps short-term climate model responses to long-term temperature change under forcing scenarios, enabling cheaper climate projections.
Combines quasipotential landscape theory and manifold learning to locate multistable climate states in an intermediate-complexity model, finding a third stable intermediate state.
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.
Data assimilationLawrence Berkeley National Laboratory / NERSCUniversity of Reading
Machine learning with PCA, ANN and quantile mapping downscales GCM output to 30 m sub-daily temperature, humidity and wind in complex terrain, improving on GCM accuracy.
Uses machine learning on Darwin C-band radar data to derive transition functions for a stochastic convective-stratiform cloud model aimed at global climate model parameterization.
Hybrid physics-ML & parameterizationBureau of Meteorology (Australia)University of Reading
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.
