Explainable AI for weather/climate, uncertainty quantification, and tests of physical consistency, balance, spectra and stability of AI models.
Search and filter this topic Plain list on GitHub
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.
Shows Pangu-Weather, FourCastNet and GraphCast fail to reproduce sub-synoptic and mesoscale phenomena, lacking physical consistency versus physics-based models.
Climate-invariant ML embeds physical knowledge into parameterization emulators so they generalize across climate regimes, maintaining offline accuracy in three atmospheric models.
Columbia University (LEAP, Lamont-Doherty)Google ResearchMassachusetts Institute of Technology (MIT)NVIDIAUniversity of California, Los Angeles (UCLA)
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.
Shows Pangu-Weather, FourCastNet and GraphCast fail to reproduce sub-synoptic and mesoscale phenomena, lacking physical consistency versus physics-based models.
Backpropagation through GraphCast optimizes initial conditions, cutting 10-day Pacific Northwest heatwave errors over 90% and extending predictability to about 23 days.
Compares uncertainty quantification methods for Pangu-Weather ensembles over Europe; ML post-hoc UQ gives skill up to 5 days beyond ECMWF ENS.
Finds an AI weather model fails to reproduce rapid early growth of small initial perturbations, implying falsely unlimited predictability unless perturbations are large.
Reservoir computers trained to enforce dynamical invariants gain longer, more stable forecasts with limited data on Lorenz 96 and a quasi-geostrophic atmosphere model.
Evidential deep learning gives one-model aleatoric and epistemic uncertainty for winter precipitation type and surface flux regression, matching ensemble accuracy; releases the MILES-GUESS package.
Classical dynamical experiments show Pangu-Weather reproduces Matsuno-Gill responses, polar lows and hurricane development, suggesting it encodes realistic atmospheric physics.
Analysis of Pangu-Weather shows its forecasts lack the fidelity and physical consistency of physics-based models, suggesting part of its deterministic skill advantage comes from these peculiarities.
Applies machine learning with uncertainty quantification to climate precipitation predictions over South America; the abstract is missing, so details are limited.
Recurrent neural operator with conformal prediction forecasts tipping points, including a stratocumulus cloud cover climate tipping point, using physics-constraint deviations.
ML-surrogate Sobol analysis of WRF physics parameters for southeast Australian heat extremes finds only three of 24 parameters matter.
Neural field compactly represents non-linear mutual-information dependencies across a 1000-member weather forecast ensemble, enabling interactive volume-rendered visualization with lower memory and compute.
Neuro-symbolic encoder-decoder translates climate questions into programs to direct deep learning simulations of AMOC tipping, aiming at explainability.
Architectural fixes to neural operators curb autoregressive error growth, giving lower long-horizon errors on Navier-Stokes, shallow water and a global weather forecasting system.
Trains reservoir-computing RNNs to enforce Lyapunov spectrum and fractal dimension, giving longer stable forecasts on Lorenz 96 and a quasi-geostrophic model.
Describes an AI2ES summer school teaching trustworthy and explainable AI for severe storms, tropical cyclones and space weather to 677 participants.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)NSF NCAR / UCARUniversity of WashingtonColorado State University (incl. CIRA)University of Oklahoma (AI2ES)
Introduces six neural network uncertainty methods and evaluation scores for environmental science, demonstrated on dewpoint profiles and Himawari-8 convection prediction.
Colorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
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.
Applies self-attention map analysis to the AtmoRep weather foundation model to interpret how meteorological variables drive its predictions; a short talk without quantitative results.
Gradient boosting and attention-LSTM models predict cirrus ice water content and crystal number (R² = 0.49), with SHAP attributions linking dust aerosols to ice crystal changes.
Tests explanation faithfulness on weather ML models, showing that reducing correlated features shrinks ranking uncertainty 2-10x and favors permutation importance.
Compares explanation methods on severe-weather and road-surface-temperature models, finding top-10 feature sets agree on about six features, and releases scikit-explain.
Compares uncertainty estimation methods for deep learning forecasts of meteorological time series, testing whether predicted uncertainty behaves as expected across horizons.
Calibrate-emulate-sample with a Gaussian process emulator quantifies gravity wave parameterization uncertainty, showing QBO parametric uncertainty comparable to internal variability under 2xCO2.
Applies layer-wise relevance propagation to echo state networks, showing they detect ENSO from sea surface temperature anomalies and explaining the learned features.
An ensemble-based algorithm targets data acquisition to regions and times that most reduce parametric uncertainty in a GCM convection scheme, within calibrate-emulate-sample.
Massachusetts Institute of Technology (MIT)California Institute of Technology (Caltech)
Shows XAI attributions for a climate-signal network change substantially with the chosen baseline, using large-ensemble SSP3-7.0 simulations, so baseline choice must match the science question.
Calibrate-Emulate-Sample with a Gaussian process emulator quantifies gravity wave parameterization uncertainty; QBO parametric uncertainty is comparable to 2xCO2 internal variability.
Distorted-physics experiments show a CNN El Niño predictor corrects ocean adjustment distortions well but struggles with changes in upwelling feedback strength.
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.
Benchmarks XAI attribution methods on a CNN that counts atmospheric rivers in climate simulations, exposing fidelity issues such as gradient shattering and sign ambiguity.
Calibrating a convection parameterization with seasonally averaged statistics cuts calibration error up to tenfold and narrows posterior spread by two to five times in an idealized GCM.
California Institute of Technology (Caltech)Massachusetts Institute of Technology (MIT)
An ensemble-based experimental design algorithm, using calibrate-emulate-sample, chooses high-resolution simulation regions that most reduce uncertainty in GCM convection parameters, with 10^2 rather than 10^5 model runs.
Massachusetts Institute of Technology (MIT)California Institute of Technology (Caltech)
Introduces a skill-decomposition method for ML week 3–4 forecasts, finding western US temperature patterns predictable mainly through Pacific ENSO-related SST variability.
Climate-invariant ML embeds physical knowledge into parameterization emulators so they generalize across climate regimes, maintaining offline accuracy in three atmospheric models.
Columbia University (LEAP, Lamont-Doherty)Google ResearchMassachusetts Institute of Technology (MIT)NVIDIAUniversity of California, Los Angeles (UCLA)
Proposes uncertainty and robustness-to-shift metrics for regression, evaluated with baselines on the Shifts weather prediction dataset under distributional shift.
XGBoost with SHAP predicts and explains NASA GEOS lightning-occurrence errors, linking them to convection and land-surface characteristics.
Neural network regression outputs a general probability distribution for heteroscedastic, asymmetric uncertainty, demonstrated on synthetic climate data.
Calibrate-emulate-sample with Gaussian process emulators and ensemble Kalman inversion quantifies convective parameter uncertainty in an idealized GCM at a fraction of MCMC cost.
Random forests show that climate sensitivity and aerosol forcing scaling dominate parametric uncertainty in 2020-2150 sea-level risk, with thermal expansion driving near-term hazards.
Philosophical argument that neural-network cloud parameterizations lacking physical process representation undermine the reliability of climate model predictions.
The NotWrong loss adds an abstention class so classifiers learn from confident samples, outperforming existing losses on several climate prediction tasks.
An abstention loss lets regression networks flag low-confidence samples during training, outperforming a standard uncertainty baseline on synthetic climate tasks.
An extended Lorenz-63 toy model reveals feedforward network instability, which training on Latin Hypercube samples removes, showing sampling design matters for stability.
Applies Gaussian process emulators and history matching to global parameter search in a climate model, framing uncertainty quantification as essential for model reproducibility.
Synthetic attribution benchmark for XAI in generic geoscience regression problems; not specifically about weather or climate.
Surveys physics-informed ML approaches with ten case studies for emulating, downscaling and forecasting weather and climate, reporting greater consistency and data efficiency.
California Institute of Technology (Caltech)Lawrence Berkeley National Laboratory / NERSCNVIDIAStanford UniversityUniversity of California, Berkeley
Proposes machine learning algorithms for uncertainty quantification in numerical weather prediction models; no abstract is available to confirm details.
Layer-wise relevance propagation applied to neural networks identifies atmospheric states associated with enhanced weather predictability, giving physical insight into predictability mechanisms.
Position paper outlining an agenda for combining physics and machine learning in Earth system science, covering equation discovery, physics constraints, parameterizations and emulation.
Tutorial-style guidance on evaluating, tuning and interpreting neural networks for meteorological satellite images, covering receptive fields, relevance propagation and image-to-image translation.
Philosophical framework arguing data-driven machine learning climate models can provide scientific understanding, illustrated with an atmospheric research case study.
Derives analytic derivatives of kernel methods for interpretability and illustrates them on Earth system data cubes, though the core contribution is generic ML.
Practical guidance on evaluating, tuning and interpreting meteorological neural networks, covering effective receptive fields, relevant performance measures, and layer-wise relevance propagation for image classification and translation.
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.
Applies backwards optimization and layerwise relevance propagation to neural networks, recovering scientifically meaningful patterns in climate variability data.
Synthesizes ML interpretation methods such as saliency maps and permutation importance, demonstrated on tornado, hail and convective-mode applications.
NSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Architectural and loss-based constraints make neural emulators of convection conserve physical laws to machine precision without degrading accuracy.
Columbia University (LEAP, Lamont-Doherty)Technical University of Munich (TUM)
Applies machine learning to global sensitivity analysis of Earth system models; abstract unavailable, so scope is inferred from the title.
Tests backward optimization and layerwise relevance propagation on a neural network for the Madden-Julian Oscillation, recovering known spatial structure and suggesting seasonality shifts.
Interactive visualization system helps users inspect recurrent convolutional network forecasts of temperature, pressure and wind trained on 25 years of climate data.
Machine learning learns WRF precipitation errors from past observations to correct systematic bias and identify which physics schemes drive forecast uncertainty.
Five recommendations for standardized, independent evaluation of ML-based Earth system models, stressing physical consistency for climate projections without historical analogues.
Benchmarks, datasets & evaluationColorado State University (incl. CIRA)Lawrence Berkeley National Laboratory / NERSCNSF NCAR / UCAROak Ridge National LaboratoryUniversity of California, Berkeley
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
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.
Object-based probabilistic deep learning on GOES-16 infrared imagery predicts convection initiation up to 1 hour ahead, beating logistic regression, with explanations of moisture and cloud drivers.
Feature-importance methods for echo state networks identify climate pathways, applied to reanalysis data around the 1991 Pinatubo eruction as a proxy for stratospheric aerosol injection.
CNN-based backward optimization attributes 1900-2014 global temperature change to forcings without assuming additivity, with uncertainty estimates compared to regularized optimal fingerprinting.
Compares deep learning architectures for European precipitation, including extremes; a two-level U-Net matches deeper versions at lower cost, with layer-wise relevance attribution.
MEDIDA uses sparse equation discovery with random Fourier features to learn Earth system model errors from sparse analysis increments in a quasi-geostrophic test case.
Online re-training of two layers of a gravity-wave CNN via ensemble Kalman inversion yields realistic QBOs in a 1D model when offline small-data training fails.
A CNN trained on daily precipitation from climate models detects emerging anthropogenic signals in observed rainfall since the mid-2010s.
Climate analysis & attributionNanjing University of Information Science & Technology (NUIST)University of Washington
Neural networks with explainable AI identify decadal variability in tropically driven subseasonal precipitation skill in CESM2 and confirm the relationships in ERA5 reanalysis.
Sub-seasonal to seasonalNASANSF NCAR / UCARColorado State University (incl. CIRA)
Surface wind asymmetry descriptors from 1-km tropical cyclone simulations, classified by machine learning, improve intensification prediction skill by 11% to about 75%.
Uses machine-learning emulators to select climate model parameter configurations, showing that conventional tuning underestimates warming uncertainty while matching present-day observations.
Downscales CanESM5 temperature and precipitation at two Iranian stations with feedforward and LSTM networks, adding LUBE prediction intervals; LSTM gains about 55% higher NSE.
Neural network with SHAP projects US fire PM2.5 emissions to rise 4–75% by mid-century under SSP5-8.5, driven by warming and drying.
Applies a Fourier neural operator with a Clausius–Clapeyron constraint for zero-shot super-resolution of WRF near-surface heat fields over the Great Lakes, reaching 1-km resolution.
CNN recognizes El Nino and La Nina phases with 86.4% validation accuracy, and weight analysis identifies the few active kernels driving the classification.
Physically constrained neural network with a 2D latent space predicts boundary-layer turbulent fluxes from LES data, outperforming the eddy-diffusivity mass-flux scheme.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)
Explainable neural networks trained on Earth system model output quantify climate distinguishability after stratospheric aerosol injection, showing slower change under SAI.
Climate analysis & attributionColorado State University (incl. CIRA)
Deep ensembles improve uncertainty quantification of statistical downscaling under climate change, tested in a pseudo-reality experiment.
Causal discovery selects inputs for neural convection and radiation parameterizations that replace superparameterization, giving stable climate simulations matching the high-resolution model.
An explainable ANN corrects ECMWF extended-range European summer temperature exceedance probabilities and traces errors to western Pacific SST teleconnections, improving skill.
Post-processing & calibrationKNMI (Royal Netherlands Meteorological Institute)
Deep neural network operator surrogates with Bayesian inversion identify sources from sparse observations in an atmospheric dispersion exemplar, with uncertainty quantification under wind uncertainty.
Physics-informed neural networks predict regional ENSO variables and aim to add physical interpretability; the abstract is unavailable, so results cannot be verified.
Perspective reviewing machine learning combined with data assimilation for climate surrogate models and model error correction, with open challenges for operational use.
Adversarial generators find Earth system model configurations that trigger AMOC collapse, with a discriminator predicting collapse accurately without the oracle.
Hybrid neural and neuro-symbolic approach uses a surrogate ocean model to predict Atlantic Meridional Overturning Circulation collapse with high accuracy, adding explainability.
XGBoost and random forest models trained on 25 years of Southern Ocean surface observations cut ERA5 shortwave mean bias to near zero and reduce RMSE by about 25%.
Post-processing & calibrationBureau of Meteorology (Australia)
Intercompares deep learning perfect-prognosis temperature downscaling models for CORDEX North America and uses XAI to expose extrapolation behaviour beyond standard validation.
A rough-set feature-interaction method feeds a generalized linear model for interpretable station weather prediction at one Kenyan farm station.
Explainable neural network on 100 MPI Grand Ensemble members identifies spring tripolar North Atlantic SST patterns as precursors of European summer heatwaves.
Climate analysis & attributionMax Planck Institute for MeteorologyColorado State University (incl. CIRA)
Logistic regression and neural networks show stratospheric aerosol injection signals in regional temperature emerge within a decade, while precipitation signals are harder to detect.
Climate analysis & attributionColorado State University (incl. CIRA)Princeton University
Interpretable deep learning ENSO model uses global ocean heat signatures to reach 22-month effective forecast length, with skill maintained using SST alone.
A physics-informed CloudMask loss improves deep networks that reconstruct 3D multilayer cloud masks from 2D satellite imagery over binary cross-entropy.
Sinh-arcsinh-normal output distributions give neural networks asymmetric, heteroscedastic uncertainty, demonstrated on tropical cyclone intensity forecast errors.
Probabilistic & ensemble forecastingColorado State University (incl. CIRA)
CNNs predict daily maximum temperature from soil moisture and geopotential, and partial dependence reveals nonlinear soil moisture-temperature coupling across mid-latitude regions.
Climate analysis & attributionStanford UniversityColorado State University (incl. CIRA)
Introduces a distributionally robust supervised fingerprint method using anchor regression for climate change detection and attribution that stays robust under interventions on other drivers.
Neural network post-processing of Swiss surface forecasts enforces thermodynamic equations, yielding physically consistent temperature and humidity without losing accuracy, especially with scarce data.
A neural network trained on CMIP6 global temperature and precipitation means predicts the year, and Alopex-derived patterns reveal how climate change signals evolve.
Bayesian physics-informed neural networks model wildfire fronts under changing wind, adding data assimilation and uncertainty quantification to level-set fire spread.
A Bayesian neural-network closure with memory, sampled by Hamiltonian Monte Carlo, gives skillful Lorenz-96 forecasts with calibrated uncertainty from noisy, sparse data.
Couples POD-LSTM reduced-order models with an ensemble Kalman filter for geophysical flows, improving SST forecast error by about an order of magnitude.
SHAP-explained SVM, RF, CNN and LSTM forecast climatic water balance 1-3 months ahead over Bangladesh; SVM performs best with antecedent CWB inputs.
Sub-seasonal to seasonalNanjing University of Information Science & Technology (NUIST)
Shows transfer-entropy causal graphs can be spurious in climate and ecohydrology data and proposes a subsample ensemble approach for more robust causality analysis.
Uses a VAE with knockoff interventions to estimate causal links under hidden confounding, applied to aerosol-cloud-climate time series with better accuracy than a deconfounding benchmark.
Gaussian processes with non-stationary Gibbs kernels reconstruct Upper Indus Basin precipitation patterns with uncertainty, outperforming stationary and deep GP baselines.
A neural network with Deep Taylor Decomposition attributes simulated soil moisture changes to forcing agents and compares the signals with SMAP observations.
Deep learning predicts next-day wildfire danger in the Eastern Mediterranean, beating the Fire Weather Index, with explainable AI attributing importance to wetness variables.
A CNN gives skilful probabilistic MJO forecasts whose uncertainty tracks instantaneous predictability, using predicted standard deviation and Monte Carlo dropout.
Uncertainty-aware regression network in CESM2 finds low-uncertainty initial states for decadal sea surface temperature predictability linked to Atlantic and Pacific variability.
Climate analysis & attributionColorado State University (incl. CIRA)
CNN predicts ENSO nine months ahead with about 0.82 correlation; contribution maps identify three precursors with asymmetric roles in El Nino and La Nina.
Deep feedforward networks with LIME explanations assess building-level hurricane wind and surge risk using updated forecasts for a Louisiana case.
Trains a shallow neural network to identify which climate model produced Arctic temperature maps, then applies it to reanalysis, using attribution to reveal model-specific regional patterns.
Climate analysis & attributionColorado State University (incl. CIRA)
Explainable neural networks using layer-wise relevance propagation combine multivariate climate fields to detect forced climate change earlier than single fields.
Climate analysis & attributionColorado State University (incl. CIRA)
Deep learning precipitation forecasting that jointly models input, target and model uncertainty via error propagation and weight sampling, giving tighter predictive uncertainty than prior joint approaches.
ENSO-MC multichannel deep network forecasts spatial SST anomalies for eastern and central Pacific El Niño up to one year ahead; saliency maps locate precursors.
Neural network with explainability predicts decadal global warming slowdowns from ocean heat content in CESM2 large ensemble, linking them to Interdecadal Pacific Oscillation-like patterns and testing on observations.
Climate analysis & attributionColorado State University (incl. CIRA)
Shallow neural networks predict the real-time MJO index out to about 18 days in October–March, outperforming linear models and explained via attribution techniques.
Sub-seasonal to seasonalColorado State University (incl. CIRA)
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.
Variational encoder-decoder compresses superparameterized aquaplanet convection into five latent nodes that separate convective regimes, offering an interpretable subgrid parameterization.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)
End-to-end differentiable training of subgrid energy-backscatter parameterizations for 2D quasi-geostrophic turbulence yields models more stable and accurate than a priori-trained schemes.
Explainable ML with Shapley values on satellite and reanalysis data shows liquid water path rises nonlinearly with droplet number, with sensitivity dependent on precipitation and SST.
Climate analysis & attributionKarlsruhe Institute of Technology (KIT)
Random forests with explainable AI on ERA5 identify drivers of subseasonal high summer temperatures in Europe, with useful skill for warm anomalies at 15-day lead.
Sub-seasonal to seasonalKNMI (Royal Netherlands Meteorological Institute)Potsdam Institute for Climate Impact Research (PIK)
Saliency-map explainable deep learning on global SST reveals ENSO-related information beyond indices and improves river flow prediction with uncertainty estimates.
Bayesian neural networks map GPM microwave imager data to precipitation type with calibrated uncertainty and decompose satellite retrieval errors.
Neural network cloud cover parameterizations trained on ICON storm-resolving simulations generalise across regional and global data, with SHAP used to explain their feature use.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)Max Planck Institute for Meteorology
AI sequence models with statistical inference reveal interpretable convective structure patterns preceding tropical cyclone rapid intensification, aiding forecaster insight.
Layer-wise relevance propagation attributes regional temperature signals to aerosol versus greenhouse forcing in single-forcing large ensembles.
Climate analysis & attributionColorado State University (incl. CIRA)
Explainable neural networks with layerwise relevance propagation identify North Atlantic and North Pacific SST regions that drive decadal predictability of continental temperature in an Earth-system model.
Climate analysis & attributionColorado State University (incl. CIRA)
Neural network identifies subseasonal forecasts of opportunity over the North Atlantic at 22-day lead, with layer-wise relevance propagation revealing tropical predictors.
Sub-seasonal to seasonalColorado State University (incl. CIRA)
Applies machine learning feature analysis to E3SM climate model output to explain its disparity with observed climate change; abstract unavailable.
Shows shallow-network emulators of radiative transfer, built for an older GCM, stay stable and realistic when transplanted into a newer state-of-the-art GCM under AMIP-like runs.
Hybrid physics-ML & parameterizationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Interpretable neural networks forecast MJO convection several weeks ahead, using layer-wise relevance propagation to explain sources of predictability compared with dynamical models.
Sub-seasonal to seasonalColorado State University (incl. CIRA)
Conditional inference trees link modelled weather and snowpack variables to avalanche problem types in Glacier National Park, Canada.
Gaussian process calibration tunes boundary-layer parameterization parameters against large-eddy simulations in single-column mode, before global model tuning.
Hybrid physics-ML & parameterizationMétéo-France (incl. CNRM)The Alan Turing Institute
Spatiotemporal graph neural networks forecast ENSO at 1 and 3 months lead, outperforming state-of-the-art systems and modeling teleconnections more interpretably.
A probabilistic deep network trained on climate simulations beats a dynamical seasonal system for precipitation and temperature, with saliency analysis of ENSO-driven predictability.
Combines quasipotential landscape theory and manifold learning to locate multistable climate states in an intermediate-complexity model, finding a third stable intermediate state.
Deep learning framework monitors physically relevant predictors and their relationships to aid human forecasters in tropical cyclone intensity change prediction.
Recurrent-convolutional networks with self-attention forecast daily weather at 18 European cities, with occlusion analysis explaining which features and cities matter.
DISTANA with active tuning infers hidden causal factors, recovering Earth's land-sea mask from temperature dynamics alone while improving temperature prediction.
A neural network trained to predict the year from climate model temperature and precipitation maps, with relevance visualization, reveals indicator patterns of forced climate change.
Climate analysis & attributionColorado State University (incl. CIRA)
CNNs predict next-hour tornado occurrence (AUC 0.94) and detect fronts in reanalysis, with interpretation methods explaining tornadogenesis failure.
Introduces tools to interpret neural convection parameterizations and diagnose gravity-wave instability when coupled to climate models, using SPCAM and cloud-resolving training data.
Hybrid physics-ML & parameterizationAllen Institute for AI (Ai2, ACE)Columbia University (LEAP, Lamont-Doherty)University of Washington
Enforces physical constraints in neural convection emulators via loss or architecture, and rescales data so models generalize to unseen climates.
Support vector machines and neural networks detect non-Gaussian error distributions in the Lorenz-63 model, informing relaxation of Gaussian assumptions in data assimilation.
A neural network trained to predict year from temperature maps detects forced climate patterns without explicitly separating signal from noise.
Climate analysis & attributionColorado State University (incl. CIRA)
A CNN outperforms logistic regression at predicting severe hail from convection-allowing NWP fields, and interpretation reveals learned storm and environment features consistent with hail physics.
Bayesian deep learning gives daily solar irradiance forecasts from weather forecasts with confidence estimates; confident outputs show 5% max error versus 22.8% for non-confident ones.
Quantile regression neural networks give case-specific posterior uncertainty for Bayesian retrievals, applied to MODIS cloud top pressure with consistent uncertainty; code released.
Discrete-continuous Bayesian deep learning with lognormal likelihood gives calibrated aleatoric and epistemic uncertainty for DeepSD precipitation downscaling, especially at extremes.
