92 papers in the archive.
Proposes a process-additive neural network that attributes compensating tendency errors to individual physical processes in a single-column atmospheric model.
Explainability, uncertainty & physicsNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Cambridge (Aardvark)University of Oxford
Sequential Monte Carlo steers diffusion models to rare events, reaching 9x to 1413x speed-ups on a score-based climate emulator for 1e-3 to 1e-5 rarities.
Calibrated GraphCast, GenCast and FGN beat IFS for African rainfall by a median of about 5%, with IFS better only in dry high-elevation areas.
A U-Net maps 1980–2025 temperature evolution to 2030s–2090s local climatologies, keeping internal variability; forecasts are calibrated and skilful on withheld model families.
CNN classifier with explainable AI identifies which climate models and resolutions produce winter surface temperature fields, showing ocean eddy resolution drives model similarity.
Climate analysis & attributionMet OfficeMax Planck Institute for MeteorologyUniversity of Oxford
U-Net trained on unorthorectified EMIT hyperspectral imagery detects methane plumes onboard satellites, reaching IoU 16.9% versus 4.8% for matched-filter baseline.
Köppen-Geiger diagnostic of AIMIP AI climate emulators under +4K warming shows only hybrid NeuralGCM-HRD reorganizes climate zones consistently with thermodynamic scaling.
Equation discovery yields three-ODE low-order models of coupled AMV, NAO and precipitation, with a 20-year oscillation and out-of-sample decadal skill, suggesting new precipitation-mediated feedbacks.
Climate analysis & attributionMet OfficeUniversity of Oxford
Deep learning estimate of solar energy reduction during hurricanes. Abstract missing, so the classification is low confidence and based on the title.
Applications & impactsPrinceton UniversityUniversity of Oxford
An ML cloud microphysics scheme trained on 5 km ICON data runs stably online for decade-long simulations once physical constraints are enforced, matching the classical graupel scheme's climate.
Hybrid physics-ML & parameterizationECMWFUniversity of Oxford
ThousandWorlds offers about 1,700 multi-GCM exoplanet simulations for emulating 3D climate fields; Gaussian processes beat deep learning baselines.
Benchmarks, datasets & evaluationUniversity of Cambridge (Aardvark)University of Oxfordcode
Shows hybrid-ML climate emulators degrade under seasonal distribution shifts and proposes compositional, physically motivated decompositions to improve out-of-distribution robustness.
Climate emulationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oxford
Self-organising maps attribute 1980–2020 European heatwave trend mainly to greenhouse gases, with aerosol reductions adding amplification, and link blocking patterns to the trend.
Makes Julia Earth system model components differentiable via Enzyme.jl and Reactant.jl, enabling reverse-mode gradients and CPU/GPU/TPU portability across four ESM components.
Software, scaling & compressionArgonne National LaboratoryMassachusetts Institute of Technology (MIT)Potsdam Institute for Climate Impact Research (PIK)Technical University of Munich (TUM)University of Oxford
ACE2 atmospheric emulator is coupled to the NEMO ocean for 70-year runs, giving realistic fast air-sea coupling but unrealistically weak tropical Pacific ENSO-like variability.
Benchmarks U-Net, ConvLSTM and ClimaX under no-analog forcing shifts; ClimaX has lowest error but up to 8.44% precipitation degradation.
Applies the GenCast diffusion ensemble to seasonal forecasting with SST forcing; it matches SEAS5 in some regions and captures ENSO precipitation patterns.
Monotonic Mass Partition embeddings with physics-informed rate models give stable ML emulators of cloud collision-coalescence that best predict precipitation onset.
TCBench is a standardized benchmark for 1-5 day global tropical cyclone track and intensity forecasts, comparing AI models (AIFS, GenCast, Pangu) against TIGGE; AI tracks are skillful, intensity less so.
Benchmarks, datasets & evaluationColorado State University (incl. CIRA)Columbia University (LEAP, Lamont-Doherty)ETH ZurichUniversity of California, BerkeleyUniversity of Oxford
Intercomparison of statistical and machine learning methods for estimating forced climate response from single realizations; skilful in models but divergent when applied to observations.
Climate analysis & attributionColorado State University (incl. CIRA)ETH ZurichMax Planck Institute for MeteorologyNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)New York University (M2LInES)Stanford UniversityUniversity of California, Los Angeles (UCLA)University of Oxford
Using GraphCast forecasts, the authors trace a recurrent 2m temperature error over Ethiopia in ERA5 to optimal-interpolation analysis increments at 0600 UTC.
A cGAN post-processing of the ECMWF ensemble gives East African rainfall skill comparable to IDR-calibrated FuXi and GraphCast, and was assessed operationally in Kenya and Ethiopia.
Active learning from Bayesian experimental design reduces parameter uncertainty in atmospheric parameterization calibration versus Latin hypercube sampling, though gains depend on observational constraint strength.
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
Bayesian neural network parameterisation in Lorenz 96 separates aleatoric and epistemic uncertainty, showing aleatoric dominates weather timescales while both matter for climate ensembles.
Kaggle-inspired ML subgrid emulators stay online-stable when coupled to a climate model with cloud microphysics, and some match state-of-the-art zonal-mean bias and global RMSE.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)NVIDIAUniversity of California, BerkeleyUniversity of Oxford
Random forest classifier tracks tropical cyclones from ERA5 aggregated predictors over the Eastern North Pacific and North Atlantic, a lighter alternative to deep learning trackers.
PseudospectralNet couples a quasi-geostrophic dynamical core with a UNet in spectral and grid space, improving short-term skill and long-term stability over pure ML models.
Hybrid physics-ML & parameterizationMassachusetts Institute of Technology (MIT)Potsdam Institute for Climate Impact Research (PIK)Technical University of Munich (TUM)University of Oxford
XGBoost with SHAP maps compound wind droughts and heat waves across China (2000-2022), finding Tmax and VPD as main drivers (R2 0.84 on test data).
Clustering-based selection of ECMWF ensemble members, guided by GSMaP observations, improves tropical cyclone rainfall forecasts over Vietnam in 70–90% of cases.
Shows recurrent networks emulate IFS radiation, gravity waves and moist physics with stable autoregressive runs, matching ResNets offline and running more efficiently.
Hybrid physics-ML & parameterizationECMWFUniversity of Oxford
PhyDA adds PDE-based physical regularization and a virtual reconstruction encoder to diffusion-based atmospheric data assimilation, improving physical plausibility on ERA5 over baselines.
Deep learning reconstructs AMOC strength from 1900 using climate model simulations, finding a 21st-century decline consistent with models and 26.5°N observations, unlike older SST-based indices.
Review of neural networks for predicting earthquakes, floods, hurricanes and wildfires, covering early warning, resource allocation and ethics; it offers no new method or quantitative results.
Random forest predicts deep convection triggering with SHAP clustering to characterize uncertainty, outperforming conventional triggers and suggesting use as a stochastic parameterization.
Position paper arguing that AI-driven equation discovery yields interpretable equations that extrapolate better than statistical emulators for climate projections and Earth system model development.
Review of neural lossy compression for Earth observation and Earth system model data, highlighting petabyte-scale climate model output as a key compression challenge.
Software, scaling & compressionIBM ResearchForschungszentrum Jülich / Jülich Supercomputing CentreUniversity of Oxford
Self-supervised geospatially-aware masked autoencoders reconstruct 3D cloud profiles from MSG/SEVIRI imagery and CloudSat radar, outperforming U-Nets.
Physics-informed ML cloud microphysics for ICON pairs a trigger classifier with a regressor under mass constraints, reaching combined offline R2 of 0.78.
MERCURY emulates multiple climate variables and wet-bulb globe temperature from global mean temperature using multi-resolution lifting, matching ESM spatial correlations and quantiles within about 5%.
Climate emulationColumbia University (LEAP, Lamont-Doherty)University of Oxford
Nature Climate Change perspective on machine learning in climate modelling and analysis; abstract absent, judged from title and venue.
Climate emulationAllen Institute for AI (Ai2, ACE)California Institute of Technology (Caltech)Colorado State University (incl. CIRA)Columbia University (LEAP, Lamont-Doherty)Google ResearchInriaLawrence Berkeley National Laboratory / NERSCNSF NCAR / UCARNVIDIANew York University (M2LInES)University of California, BerkeleyUniversity of Oxford
Pareto-front model hierarchies compare semi-empirical and deep learning models for cloud cover, shortwave radiation and precipitation parameterizations, condensing ML gains into a ten-parameter equation.
Hybrid physics-ML & parameterizationColorado State University (incl. CIRA)ECMWFNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)New York University (M2LInES)University of Oxford
Bidirectional vertical RNNs match feed-forward accuracy for sub-grid gravity wave, moist and unified parameterizations with fewer parameters and better offline autoregressive stability than ResNets.
Hybrid physics-ML & parameterizationECMWFUniversity of Oxford
Machine learning shows near-surface zonal wind alone predicts ENSO 11 to 21 months ahead as well as SST, via an Indian Ocean atmospheric bridge.
Score-based diffusion data assimilation at km scale assimilates 40 sparse weather stations, lowering held-out station RMSE by 10% versus HRRR.
A definition and metric separate fixable model error from unfixable chaos in autoregressive ML weather models, and a regularization penalty improves RMSE and spread-skill.
Explainability, uncertainty & physicsThe Alan Turing InstituteUniversity of Cambridge (Aardvark)University of Oxford
A conditional GAN postprocesses ECMWF hourly East African rainfall, improving the diurnal cycle and extremes up to the 99.9th percentile.
Climinator uses a mediator-advocate LLM framework to fact-check climate change claims against IPCC and literature, with reported high accuracy.
A position paper argues for combining stochastic parameterization with machine learning, reviewing early studies and outlining open challenges for data-driven stochastic subgrid schemes.
Hybrid physics-ML & parameterizationUniversity of Cambridge (Aardvark)University of Oxford
Bayesian Gaussian-process emulator obeying energy-balance equations emulates global and spatial surface temperature across scenarios, with interpretable parameters and uncertainty quantification.
A hybrid RBF network downscales sparse station data into high-resolution precipitation, temperature and humidity maps for Bangladesh, outperforming MLR and kriging.
A Wasserstein GAN downscales tropical cyclone rainfall from about 100 km to 10 km, reproducing spatial structure and spectra with mean bias within 5% and extrapolating to extreme storms.
GPT-4 grounded with IPCC AR6 content via retrieval; expert-scored answers show the hybrid chatClimate more accurate than GPT-4 alone.
A deep learning atmospheric river tracker is applied to a high-resolution large-ensemble climate dataset and checked for consistency against established algorithms over the US West Coast.
Extremes & hazardsETH ZurichNSF NCAR / UCARNVIDIAUniversity of California, BerkeleyUniversity of Oxford
LLM agent retrieves emissions data from ClimateWatch and web search to give more precise answers on climate change topics.
LLMs & agentsColumbia University (LEAP, Lamont-Doherty)ETH ZurichUniversity of Oxford
Applies unpaired UNIT image-to-image translation, combined with quantile mapping, to bias-correct HadGEM3 monsoon simulations toward ERA5, preserving cross-variable correlations and spatial structure.
Linear regression, random forests and neural networks correct ECMWF IFS 2 m temperature and 10 m wind errors, improving RMSE by 10–15% with uncertainty estimates.
Surrogate-based statistical framework constrains UKESM1 aerosol parameters against MODIS observations on a cloud platform, using 3-hourly outputs for tighter constraints.
Climate analysis & attributionMet OfficeUniversity of Oxford
Randomly wired dense layers in ClimateBench emulators improve performance in 14 of 24 cases, up to 30.4% for MLPs, with no slower prediction.
Causal machine learning on satellite observations estimates aerosol-cloud interaction with uncertainty bounds, finding only one of three climate models reproduces the cooling trend.
Pyrocast pipeline with a 148-event pyroCb database and CNN or random-forest models predicts pyroCb six hours ahead with AUC 0.90.
Extremes & hazardsColumbia University (LEAP, Lamont-Doherty)University of Cambridge (Aardvark)University of Oxford
A neural network trained on ERA5 and MODIS skin temperature quality-checks IFS physiographic upgrades, improving land-surface temperature prediction by 0.37 K for updated lakes.
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
Random forest and boosting models estimate pre-industrial cloudiness, suggesting aerosols increased global cloud fraction by 1.27% with about -0.42 W/m2 cooling.
A CNN gives skilful probabilistic MJO forecasts whose uncertainty tracks instantaneous predictability, using predicted standard deviation and Monte Carlo dropout.
RICE-N integrated assessment model with multi-agent reinforcement learning simulates climate-economy negotiations to test cooperation frameworks for climate mitigation agreements.
Scalable sensitivity bounds with uncertainty-aware deep models for continuous-treatment causal effects, applied to how human emissions affect cloud properties in 15 years of satellite data.
Climate analysis & attributionETH ZurichUniversity of Oxford
Introduces a nonparametric regression with a dependence-based regulariser that enforces physical consistency in Earth data models, with a consistency-versus-accuracy trade-off diagram.
Explainability, uncertainty & physicsETH ZurichMax Planck Institute for BiogeochemistrySwiss Data Science Center (SDSC)University of Oxford
GAN and VAE-GAN downscale low-resolution forecasts to radar-scale precipitation maps with CRPS matching state-of-the-art pointwise post-processing, including heavy rain.
A probabilistic physically informed recurrent network learns stochastic parameterizations in the Lorenz 96 atmosphere model, outperforming autoregressive and GAN baselines on temporal patterns.
Hybrid physics-ML & parameterizationThe Alan Turing InstituteUniversity of Cambridge (Aardvark)University of Oxford
CorrectorGAN jointly bias-corrects and super-resolves global ensemble precipitation forecasts over the US against radar, approaching an operational regional model in seconds.
Downscaling & super-resolutionThe Alan Turing InstituteUniversity of Oxford
Neural network emulates the M7 aerosol microphysics scheme in ECHAM-HAM with 77.1% mean R² and up to 64× GPU speed-up, with physical mass constraints.
Information-bottleneck method learns causal macrovariables from high-dimensional data and recovers both known El Niño variants in real climate data.
Non-linear causal machine learning reveals how aerosol effects on cloud droplet radius vary across meteorological regimes in the Southeast Pacific.
ESEm, an open-source Gaussian-process emulation and calibration toolkit, reduces parametric uncertainty in a GCM and explores CMIP6 scenario uncertainty.
Neural network emulator of gravity wave drag yields tangent-linear and adjoint models that succeed in 4D-Var assimilation, easing code maintenance.
Bayesian Gaussian-process deconditioning method for statistical downscaling of atmospheric fields, with a proven minimax-optimal rate and gains over existing methods.
A DOE perspective argues for AI-enabled stochastic parameterizations that account for atmospheric state, replacing hand-tuned noise in Earth system models.
Hybrid physics-ML & parameterizationECMWFNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oxford
Neural-network emulators replace bin-microphysics warm-rain processes in a GCM, reproducing its climate behaviour at the control run's computational cost.
Hybrid physics-ML & parameterizationNSF NCAR / UCARUniversity of Oxford
Neural emulators of the non-orographic gravity wave drag scheme run stably to seasonal timescales, improve medium-range forecasts and run ten times faster on GPUs than the CPU scheme.
Hybrid physics-ML & parameterizationECMWFUniversity of Oxford
RainBench is a multimodal global benchmark combining simulated satellite data, ERA5 and IMERG precipitation, with the PyRain library and medium-range baselines.
Overview of neural networks for climate change prediction, discussing opportunities and challenges without new methods or quantitative results.
Machine-learned preconditioners for the semi-implicit linear solver in a shallow-water model are competitive with conventional preconditioners and generalize beyond the training range.
Hybrid physics-ML & parameterizationECMWFNSF NCAR / UCARUniversity of Oxford
Review of big data and machine learning for climate change modeling using satellite, sensor and climate model data, covering challenges and outlook without new methods.
Position review arguing climate emulation with machine learning differs from weather forecasting, since climate is a boundary-condition problem with steady-state emulation opportunities.
Deep convolutional networks detect two anthropogenic aerosol-driven perturbations in satellite cloud imagery to assess their prevalence and climatic effects.
Perspective arguing that machine learning should mine existing climate data and simulations for teleconnections and warnings, not only improve Earth system models.
Cumulo is a global dataset pairing one year of 1km MODIS imagery with CloudSat cloud labels, with a flow-based baseline for cloud classification.
Benchmarks, datasets & evaluationThe Alan Turing InstituteUniversity of Oxford
Nonlinear causal discovery method estimates interdependency networks from large generic time series datasets.
Climate analysis & attributionThe Alan Turing InstitutePotsdam Institute for Climate Impact Research (PIK)University of Oxford
Trains a GAN as a stochastic sub-grid parameterization for the Lorenz '96 model; some configurations beat a bespoke scheme at weather and climate timescales.
Hybrid physics-ML & parameterizationNSF NCAR / UCARUniversity of Oxford
Statistical emulators built from perturbed-physics ensembles iteratively tune weather@home parameters, cutting northwestern US warm biases by 74%.
ML learns timestep-by-timestep error corrections to a physical model; tested on Lorenz-96, it gives stable runs, better initialised skill and improved long-term statistics.
