The November 13, 2026 paper introduces WenHai-E2E, which generates global eddy-resolving ocean forecasts directly from observations and reports better forecast accuracy than a numerical system. The abstract and publication metadata reviewed here identify no separate oceanographic discovery or specific field result beyond forecast performance. Full text was not assessed.
Technology: WenHai-E2E (end-to-end deep learning ocean forecast model)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Forecast accuracy against a numerical ocean forecasting system, without a separate discovery in the reviewed abstract.
- Still unknown: Admission review pending: primary full results needed to distinguish a specific scientific/application result from general method validation. Existing supported claims may be retained with bounded scope, but do not mark source checked until admission is resolved.
- AI role: Machine learning model produces the full eddy-resolving global ocean forecast from observations, replacing numerical initial conditions.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1126%2Fsciadv.aeh9058%22&resultType=core&format=json
Needs source check · editorial details
The source has not been fully checked or contains conflicting information. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
MIT researchers with Imperial College London designed a small synthetic set of deliberately unrealistic emissions scenarios to train climate emulators, which normally need six to eight full Earth system model runs that each take months. The method needs only one model run, and the emulator predicted unseen scenarios and local pollutant-control effects where traditional emulators failed. It was validated with a simple climate model rather than a full Earth system model, and the article reports no quantitative accuracy metrics.
Technology: Climate emulator trained on optimized synthetic scenarios
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: An emulator trained on a small set of synthetic scenarios predicted unseen scenarios and a local pollutant-control effect where traditional emulators failed. The source says it was validated with a simple climate model.
- Still unknown: The source says the method still needs testing on a full Earth system model scenario and on other climate model types; it does not say whether this has been done.
- AI role: An optimization algorithm generated the synthetic training scenarios; an AI emulator was trained on them to predict climate outcomes.
Evidence sources: https://news.mit.edu/2026/new-research-could-lead-faster-trustworthy-ai-models-climate-prediction-1112
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A 10-meter map classifies each land area into six forest or land-use types (including primary forest, plantations, and tree crops), with an initial time series from 2017–2025. Validation across more than 9,000 sites identifies the correct type in 90% of cases and distinguishes forest from non-forest in 95%.
Technology: Satellite-image classification models
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: The map was validated at more than 9,000 random sites, where two trained experts checked high-resolution imagery to determine land type: correct forest type in 90% of cases and forest vs non-forest in 95%.
- Still unknown: Per-class accuracy and regional breakdowns are not reported, the time series is early-release without separate validation, and results depend on the probability threshold chosen.
- AI role: A custom remote-sensing model trained on about 1.7 million labeled locations, using AlphaEarth embeddings, outputs probabilities for six forest and land-use classes.
Evidence sources: https://deepmind.google/blog/beyond-tree-cover-mapping-the-worlds-forest-types/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A diffusion model trained on ten CMIP6 climate models estimated event-specific temperature changes conditional on observed meteorology. It found >95% probability that anthropogenic emissions since the 2015 Paris Agreement intensified Europe’s summer temperature extremes in 2021–2025; for the June 2025 heatwave, regional mean temperature is estimated to have been 0.34 °C lower under 2015 cumulative emissions, with >99% probability.
Technology: Generative diffusion model for spatially explicit temperature fields
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A diffusion model trained on ten CMIP6 climate models estimated event-specific temperature changes conditional on observed meteorology.
- Still unknown: This event-specific attribution is conditioned on observed large-scale meteorology and trained on ten CMIP6 models; it is not a direct experiment that removed emissions. The paper emphasizes uncertainty from internal variability/model spread and reports robust regional-mean effects, while grid-cell-level changes are more limited. The June 2025 value is a median modeled counterfactual estimate.
- AI role: The model learned climate-model relationships between meteorological conditions, cumulative CO₂ emissions and daily temperature fields, then generated conditional counterfactual temperature distributions to attribute the incremental post-2015 emissions effect to recent European heat extremes.
Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026GL125079
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A model that learns from real-time geostationary satellite data and produces hourly forecasts, at resolutions down to 5 km for temperature and humidity (about five times finer than WeatherNext 2). For short-term precipitation, Google reports improvements of up to 60% over NASA's IMERG. These are ‘up to’ figures from Google's evaluations, with an independent ranking by Brightband.
Technology: WeatherNext 3
Family WeatherNext: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Precipitation forecasts were scored with CRPS against baselines: gains of up to 60% over IMERG, 30% over MRMS and 10% over rain gauges for early lead times; independent live evaluations by Brightband ranked the model.
- Still unknown: The 'up to' figures lack full test setup, longer-range gains are largest where forecasts are weak, and the atmosphere retains unpredictability.
- AI role: The model ingests live geostationary satellite data and historical analyses, trains on station observations and produces gridded fields, cyclone tracks and station-level forecasts.
Evidence sources: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
MAPL-EMIT used a vision transformer trained on 3.6 million synthetic plumes embedded in EMIT radiance data to map methane sources. Beyond synthetic tests, the paper abstract reports recovery of 84% of known annotated plume complexes across 1,084 EMIT granules. Comparisons with airborne data, landfills and controlled releases supported previously uncaptured sources. This is source-reported validation, not an independent emissions audit.
Technology: MAPL-EMIT (end-to-end vision transformer on EMIT hyperspectral radiance)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Synthetic tests plus known real plumes, coincident airborne observations, landfill comparisons and controlled-release data.
- Still unknown: Review scope: complete indexed abstract and metadata; full uncertainty budgets and individual source emissions were not assessed.
- AI role: An end-to-end vision transformer jointly estimated methane enhancements, plume outlines and source locations from hyperspectral radiance.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1073%2Fpnas.2612145123%22&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Using SINDy equation discovery on monthly observational indices from 1950–2022, researchers derived three linked low-order equations for Atlantic Multidecadal Variability, the North Atlantic Oscillation and North Atlantic precipitation. The model recovered a distinct roughly 20-year component and highlighted precipitation-linked terms as robust components of the learned dynamics; tests against the independent 1900–1939 record found skill for statistics of the variability indices and rainfall. The equations generate hypotheses about ocean-atmosphere feedbacks; they do not establish a single causal mechanism or guarantee forecasts.
Technology: Sparse Identification of Nonlinear Dynamics (SINDy) equation-discovery algorithm
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Using SINDy equation discovery on monthly observational indices from 1950–2022, researchers derived three linked low-order equations for Atlantic Multidecadal Variability, the North Atlantic Oscillation and North Atlantic precipitation.
- Still unknown: The learned equations are compact models/hypotheses from observational associations. The paper and Oxford source note that proposed mechanisms need further evaluation; reported hindcast skill is not prospective climate prediction validation.
- AI role: Searched candidate dynamical models against observational time series and selected sparse, interpretable coupled equations; the researchers analyzed robust equation terms and compared the resulting model against a historical holdout period.
Evidence sources: https://esd.copernicus.org/articles/17/1061/2026/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Three-day forecasts match what earlier models achieved at two days: about 100 km trajectory error and 11 knots intensity error, compared with ECMWF-ENS and HWRF. The work is with the National Hurricane Center, CIRA, and the Met Office. The comparison uses historical 2023–2024 cyclones; why results are good at 28 km remains an open question.
Technology: WeatherNext Cyclones
Family WeatherNext: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Historical cyclones from 2023-2024 were benchmarked deterministically and probabilistically against ECMWF-ENS (track) and HWRF (intensity): three-day errors matched earlier two-day forecasts; a 2025 storm is cited only as an anecdote.
- Still unknown: Why the model is accurate at 28 km resolution is an open question; the 'extra day' is an average dependent on the chosen benchmarks, and the evaluation window is slightly unclear.
- AI role: A single model using Functional Generative Networks produced track, intensity and wind-structure ensembles of up to 1,000 members.
Evidence sources: https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Trained with a loss function that preserves energy across scales, it reaches a mean squared error equal to the Met Office global model and exceeds it on some metrics, with more realistic storm cores in Hurricane Ian and Storm Ciarán. The announcement gives no figures, covers only two extreme cases, and the system is not operational.
Technology: FastNet (spherical-harmonic loss)
Family FastNet: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: FastNet was compared with the Met Office global model on verification measures (RMSE and some other metrics, shown only in a figure) and qualitatively on Hurricane Ian and Storm Ciaran storm cores.
- Still unknown: The model is not operational, the text gives no numerical results, only two extreme cases are described, and comparable skill does not establish the same validation as the Global Model.
- AI role: A machine-learning weather model co-developed with the Alan Turing Institute, trained with a modified spherical-harmonic loss to keep energy across scales physically realistic.
Evidence sources: https://www.metoffice.gov.uk/about-us/news-and-media/media-centre/weather-and-climate-news/2026/new-research-shows-ai-can-produce-trustworthy-physically-realistic-weather-forecasts-
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A neural network trained on observations from five Arctic field campaigns reduced ERA5 errors in the Arctic sea-ice surface energy budget and removed ERA5’s wintertime near-surface warm bias of about 4 K. The corrected estimates imply about 50% weaker winter surface cooling and 80–120 cm seasonal thermodynamic ice growth, closer to SMOS–CryoSat thickness increases than ERA5’s 150–200 cm estimate.
Technology: Feed-forward neural network product SEBai
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: A neural network trained on observations from five Arctic field campaigns reduced ERA5 errors in the Arctic sea-ice surface energy budget and removed ERA5’s wintertime near-surface warm bias of about 4 K.
- Still unknown: Validation uses leave-one-campaign-out cross-validation and multiple field campaigns, but direct flux observations remain sparse in space and time. The climatology applies to ice-covered Arctic regions north of 70°N with sea-ice concentration above 80%; authors do not recommend extrapolation to marginal ice zones or coastal regions. The ice-growth values are thermodynamic estimates rather than direct pan-Arctic measurements.
- AI role: The network learned the relationship between ERA5 atmospheric states and observed fluxes from MOSAiC, SHEBA, AO2018, N-ICE2015 and ARTofMELT campaigns, then bias-corrected hourly Arctic sea-ice surface fluxes for 1994–2024 to derive a revised regional climatology and ice-growth estimates.
Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025GL121289
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
WeatherNext forecast a Category 5 landfall in Jamaica five days ahead with 80% probability, rising to nearly 100% at three days. According to the post, it was the best single model for track and intensity in the National Hurricane Center's 2025 verification report, which used it alongside its HAFS model. The post gives no numerical comparisons; rapid intensification remains very hard to predict.
Technology: WeatherNext
Family WeatherNext: profile · models page
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: The National Hurricane Center's 2025 annual verification report found WeatherNext the top-performing individual model for track and intensity; the Melissa forecast is presented as a single case.
- Still unknown: The page notes that predicting rapid intensification is very difficult and that Weather Lab is experimental; no dedicated limitations section.
- AI role: WeatherNext, an AI weather model with 50-member ensembles, forecast the Hurricane Melissa landfall alongside physics-based HAFS and was used by NHC forecasters.
Evidence sources: https://deepmind.google/blog/how-weathernext-helped-the-national-hurricane-center-better-predict-hurricane-melissas-historic-landfall-in-jamaica/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A stacked ensemble trained on 867 compiled degassing records and geospatial raster predictors generated a continuous 0.10° map of model-estimated CO₂ flux within an active-tectonics applicability mask. The authors report spatially coherent high-flux estimates along the Andes, Pacific Ring of Fire and other active orogens. This is a new model-derived spatial estimate for fault-system flux, rather than a direct global survey or measurement.
Technology: Stacked regression ensemble combining support-vector, k-nearest-neighbor, decision-tree, random-forest and gradient-boosting models
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A stacked ensemble trained on 867 compiled degassing records and geospatial raster predictors generated a continuous 0.10° map of model-estimated CO₂ flux within an active-tectonics applicability mask.
- Still unknown: Predictions are model estimates based on 867 compiled records with heterogeneous dates and measurement techniques, not direct measurements at every pixel. The mask is limited to areas within 40 km of mapped active faults or with PGA ≥0.10 g; the 0.10° raster smooths local geology. The abstract reports R²=0.677, while the full text gives stacking test R²=0.693 and validation R²=0.552; fit is moderate. Sampling gaps include central Africa, eastern Asia and North America. High-flux belts overlap known tectonic/geothermal regions, and SHAP associations do not establish causal controls.
- AI role: Five regression learners were tuned and combined in a stacking regressor; that model predicted flux pixel by pixel from global raster inputs and generated the paper’s continuous map of estimated fault-system flux. The mask excluded areas outside the model’s stated tectonically active applicability domain.
Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2025JH000715
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A CatBoost model trained on CMIP6 simulations used the observed 1971–2020 grid-cell warming pattern to identify tropical and polar regions that best constrain later global warming. Applying these spatial constraints reduced projection error variance by about 70%, compared with about 48% using global-mean warming alone; the results depend on model-based emergent relationships and are projections rather than direct observations of future climate.
Technology: CatBoost machine-learning emergent-constraint model
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A CatBoost model trained on CMIP6 simulations used the observed 1971–2020 grid-cell warming pattern to identify tropical and polar regions that best constrain later global warming.
- Still unknown: The study is a climate-model emergent-constraint analysis; uncertainty reduction was assessed with leave-one-model-out cross-validation, not by validating future outcomes. The reported probabilities are scenario-conditional projections.
- AI role: Learned grid-cell-scale relationships between historical warming fields and future global mean warming across CMIP6 simulations, then ranked grid cells by their influence on constrained projections.
Evidence sources: https://www.nature.com/articles/s41467-026-70205-9
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
AIGFS (retrained from GraphCast) uses up to 99.7% less computing than GFS (a 16-day forecast in about 40 minutes); AIGEFS has skill comparable to GEFS with 9% of the computing; and HGEFS (62 hybrid members) outperforms GEFS and AIGEFS on most metrics. Cyclone intensity worsened in AIGFS v1.0; the announcement gives no numerical scores.
Technology: AIGFS, AIGEFS, and HGEFS (GraphCast-derived)
Family GraphCast: profile · models page
Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: not reported
- Tested: The release describes initial or early testing and claims skill comparable to GEFS and better ensemble performance, but gives no metrics, verification datasets or time periods.
- Still unknown: The release states that AIGFS v1.0 degraded tropical cyclone intensity forecasts, that AIGEFS ensemble spread is still being improved, and that HGEFS hurricane intensity is being improved.
- AI role: AI global models: AIGFS forecasts with up to 99.7% less compute; AIGEFS is a 31-member AI ensemble with 9% of the GEFS computing; HGEFS combines 31 GEFS and 31 AIGEFS members.
Evidence sources: https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 1/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Neural networks trained on earthquake catalogues from California, New Zealand, Italy, Japan and Greece forecast the number of aftershocks in the following 24 hours after magnitude 4+ earthquakes. Their quality is reported as comparable to the operational ETAS model, while producing results in seconds rather than hours or days. The abstract reports no numeric skill scores, so the comparison rests on the full paper.
Technology: Deep neural networks for spatiotemporal seismicity rate forecasting, benchmarked against ETAS
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Deep-learning seismicity forecasts were compared with an epidemic-type aftershock sequence benchmark and a persistence null hypothesis on Southern California and Italy test datasets; they matched ETAS at much lower processing time.
- Still unknown: The abstract says performance depends on training data that represent the properties of earthquake sequences.
- AI role: Deep neural networks learned short-term spatiotemporal seismicity rates from high-resolution catalogues.
Evidence sources: https://link.springer.com/article/10.1186/s40623-025-02241-6
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
An information-flow analysis of daily observations found statistically significant local surface-soil-moisture effects on precipitation over about 15.7% of global land (60°S–60°N) in the boreal warm season, with eight regional hotspots. Sensible-heat mediation dominated most hotspots, while evapotranspiration mediation dominated two African hotspots; remote moisture availability and boundary-layer-height variability helped explain regional differences. Only four of 16 assessed CMIP6 models reproduced the ERA5-derived relationship between soil-moisture variability and causal coupling.
Technology: Neural-network-inversion satellite soil-moisture product (NNSm), analyzed with Liang–Kleeman information-flow causal inference
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: An information-flow analysis of daily observations found statistically significant local surface-soil-moisture effects on precipitation over about 15.7% of global land (60°S–60°N) in the boreal warm season, with eight regional hotspots.
- Still unknown: The mapped effect is local surface-soil-moisture-to-precipitation information flow, statistically significant at the authors' p<0.1 threshold over about 15.7% of land; the analysis does not imply local soil moisture causes rainfall everywhere or exclude remote/upwind effects. The AI contribution is the neural-network-derived satellite soil-moisture product; the causal analysis is not itself an AI model. The spatial pattern was broadly checked with ERA5, while pathway attribution depends on ERA5-reanalysis variables and the assumptions/limitations of information-flow causal inference. Results concern boreal warm seasons and the analyzed 2003–2020 satellite period (ERA5 cross-check spans 1979–2021).
- AI role: The study's long-term daily satellite soil-moisture series was generated with a neural-network inversion algorithm, extending the observed record used to map global coupling. Researchers applied information-flow analysis to this AI-derived product and precipitation observations, then cross-checked the main spatial pattern using independent ERA5 reanalysis. AI provided the satellite soil-moisture input for a concrete global climate result; the causal-inference method itself is an information-theoretic statistical technique, not a learned model.
Evidence sources: https://www.nature.com/articles/s41467-025-63999-7
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 42 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A model combines optical, radar, 3D laser, and climate-simulation data into one representation for 10-meter squares, using 16 times less storage and with 24% lower average error than the compared models. The Google Earth Engine dataset has over 1.4 trillion embeddings per year and has been tried by more than 50 organizations. Benefits with LLM agents remain prospective.
Technology: AlphaEarth Foundations
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Compared against traditional methods and other AI mapping systems on land-use identification and surface-property estimation; average error 24% lower than the models tested, storage 16 times lower. Over 50 organizations tested the dataset; no sample sizes given.
- Still unknown: The page states no limitations; it says the team is still exploring time-based capabilities and that embeddings may be more useful combined with reasoning LLM agents.
- AI role: The model turns satellite and Earth observation data into compact embeddings; scientists and partners use them to create maps and classify ecosystems.
Evidence sources: https://deepmind.google/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/ · https://arxiv.org/abs/2507.22291
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
On an open-ended climate-analysis task, the EarthLink agent autonomously selected two Atlantic-internal predictors for the summer ATL3 index at an eight-month lead: the SON-to-November tendency of the equatorial east–west sea-level-pressure gradient and November western-equatorial-Atlantic sea level as a thermocline-tilt proxy. Rolling-origin hindcasts for 1995–2023 achieved a temporal correlation of about 0.46, above persistence (~0.13) and ENSO-only baselines. The paper proposes a Kelvin-wave/Bjerknes-feedback explanation for the association.
Technology: EarthLink multi-agent AI research system based on foundation language models and climate data/tools
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: On an open-ended climate-analysis task, the EarthLink agent autonomously selected two Atlantic-internal predictors for the summer ATL3 index at an eight-month lead: the SON-to-November tendency of the equatorial east–west sea-level-pressure gradient and November western-equatorial-Atlantic sea level as a thermocline-tilt proxy.
- Still unknown: This is an unrefereed arXiv preprint and a retrospective rolling-origin hindcast from one observational/reanalysis data setup. The physical pathway is an AI-generated, testable hypothesis that the authors explicitly say requires further validation; the source does not establish a causal mechanism or demonstrated prospective operational forecast. Skill is ~0.46 and falls short of the authors’ 0.5 target.
- AI role: Given an open-ended Atlantic Niño precursor problem, EarthLink autonomously planned analyses, generated code, tested predictors using regression models, reviewed literature and synthesized a physical mechanism. The paper states the proposed AI-generated hypothesis needs further validation.
Evidence sources: https://arxiv.org/abs/2507.17311
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Samudra, a UNet-style emulator trained on the ocean component of a state-of-the-art climate model (OM4 within CM4), reproduces sea surface height, velocities, temperature and salinity across the full ocean depth. It runs about 150 times faster than the original ocean model and remains stable for centuries. The authors note it struggles to reproduce the correct magnitude of forcing trends while staying stable, so it needs further work.
Technology: Samudra UNet-style ocean emulator
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Samudra was compared with the original climate model's ocean component: no drift relative to the truth, reproduction of depth structure and interannual variability, stability for centuries, and about 150 times faster.
- Still unknown: The authors state the emulator struggles to reproduce the correct magnitude of forcing trends while staying stable, and this needs further work.
- AI role: A ConvNeXt-UNet-style emulator trained on the ocean component of a climate model to predict sea surface height, velocities, temperature and salinity.
Evidence sources: https://arxiv.org/abs/2412.03795
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 52 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A single AI model replaces the entire forecasting chain: it starts from satellite and station observations and produces global and local forecasts in minutes on a desktop using about one-tenth of the data; the page says it beats the US GFS on many variables. The announcement gives no metrics; it is not operational and depends on ERA5. Published in Nature.
Technology: Aardvark Weather
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Forecasts were compared with the operational NWP baseline (GFS) and with NWS and an end-to-end system; the Nature paper reports the global forecasts beat the baseline for several variables and lead times.
- Still unknown: The source states it is not operational, depends on ERA5, and that hurricane, air quality and other applications are proposed rather than tested.
- AI role: A single model ingested satellite and station observations and produced global and local forecasts.
Evidence sources: https://www.eng.cam.ac.uk/node/7921 · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-025-08897-0&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 229 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A two-part framework (adaptive artifact removal plus a deep subdomain adaptation network) was trained on 1,627 Sentinel-2 images from six known methane super-emitters. Applied to an unannotated oil field in Algeria, it flagged 34 anomalous emission events over one year, attributed to three super-emitters linked to production and transmission infrastructure. DSAN achieved an average macro F1 of 0.86 on transfer tasks, versus 0.78 for the EfficientNet-V2L comparator; independent field verification is not reported in the abstract.
Technology: Deep subdomain adaptation network (DSAN) with low-reflectance artifact detection (LRAD)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: The model was tested on 36 cross-region transfer tasks (macro F1 0.86 for 1-to-1, 0.89 for 5-to-1) and 33 of 369 positive detections were manually verified as true methane plumes.
- Still unknown: Real-world macro F1 was 0.56 because of 336 false positives, the dataset covers only three countries, label confidence is hard to quantify, the detection limit is roughly 5000 kg/h, and the abstract and body differ on the event count (34 vs 33).
- AI role: A ResNet-50 deep subdomain adaptation network classified plume presence in place of human inspection, after a band-threshold artifact-removal step.
Evidence sources: https://acp.copernicus.org/articles/25/4035/2025/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: at percentile 88 of works of the same type, year and subfield (OpenAlex work metrics, 9 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The WenHai data-driven global ocean forecast system outperforms a state-of-the-art eddy-resolving numerical model and an AI-based ocean forecast system for temperature and salinity profiles, sea surface temperature, sea level anomaly and near-surface currents at lead times from 1 day to at least 10 days. Skill beyond 10 days is not reported in the abstract, and the computational speed-up claimed in secondary coverage was not confirmed in the paper abstract.
Technology: WenHai deep neural network ocean forecasting model
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: WenHai forecasts were compared with a state-of-the-art eddy-resolving numerical system and an AI-based system for temperature, salinity, sea surface temperature, sea level anomaly and currents at 1 to at least 10 day leads.
- Still unknown: Skill beyond 10 days is not reported in the page read, and the authors describe ocean AI forecasting as remaining challenging.
- AI role: A data-driven deep neural network with bulk flux formulae, with ocean dynamics informing the architecture.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-025-57389-2&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 87 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The European center operationalized AIFS Single on February 25, 2025, alongside the physical IFS model. ECMWF reports gains of up to 20% on tropical-cyclone tracks and about 1,000 times lower energy use. It is deterministic, at 28 km versus 9 km for IFS.
Technology: AIFS (machine-learning model)
Family AIFS (ECMWF): profile · models page
Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: not reported
- Tested: Not reported
- Still unknown: The page does not describe the verification methods or datasets, the grid is 28 km versus 9 km for IFS, and ensemble forecasting is a future step.
- AI role: A machine learning model trained on past weather evolution projects the same initial conditions as the IFS model forward every six hours.
Evidence sources: https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A diffusion model takes atmospheric data from 25-kilometer to 2-kilometer scale and also predicts variables absent from the inputs, such as radar reflectivity. The paper appeared the same day in Communications Earth and Environment; the initial model was tuned on Taiwan data with the Central Weather Administration, and Taiwan's disaster-reduction center uses it for alerts. NVIDIA claims 500 times faster and 10,000 times more energy-efficient than high-resolution numerical forecasting on CPU: company figures, with no accuracy comparison on the page.
Technology: CorrDiff (diffusion model)
Family CorrDiff: profile · models page
Publication: Peer-reviewed paper · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: not reported
- Tested: Not reported
- Still unknown: The article gives no accuracy metrics or accuracy comparison and does not explicitly list limitations; the speed and energy figures are company claims.
- AI role: A diffusion model downscales atmospheric data from 25 km to 2 km and predicts variables missing from the input.
Evidence sources: https://blogs.nvidia.com/blog/earth-2-ai-high-resolution-forecasts/
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 274 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
GlaViTU combined convolutional and transformer models to map glacier outlines from satellite images. The original supplement applies it to the Aletsch complex between 1998 and 2023 and Hardangerjøkulen between 1996 and 2022, estimating area declines in both. The reported uncertainty bands are broad and overlap across dates. These are retrospective, model-derived geophysical estimates, not independent field measurements. Review covered the paper abstract, metadata, relevant main-text discussion and the first supplementary page.
Technology: GlaViTU (Glacier-VisionTransformer-U-Net)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Retrospective satellite-image mapping and regional glacier-area estimates at two periods; model-generalization and expert-outline comparisons.
- Still unknown: Supplementary area estimates have broad, overlapping 95% uncertainty bands. Only the relevant first supplementary page was reviewed; no field campaign or independent glacier-area estimate was performed in this review.
- AI role: GlaViTU mapped glacier outlines with convolutional-transformer models and inference-time bias optimization.
Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11695715/ · https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11695715/supplementaryFiles
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 61 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A 0.25° model that outperformed ECMWF's ENS ensemble on 97.2% of 1,320 variable-and-lead-time combinations (99.8% beyond 36 hours) and produces a 15-day ensemble in about 8 minutes on a TPU. Physical models remain essential because they provide training data and initial conditions.
Technology: GenCast (diffusion model)
Family GenCast: profile · models page
Publication: Publication status unspecified in annotation · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Not reported
- Test type: retrospective or benchmark
- Tested: Forecasts were compared with ECMWF ENS on held-out 2019 data (trained up to 2018) across 1,320 variable and lead-time combinations; GenCast was more accurate on 97.2% of targets.
- Still unknown: Forecasts are especially uncertain beyond a few days, the wind-power analysis is only a proof of principle, and physics-based models remain needed for training data and initial conditions.
- AI role: A diffusion model generates an ensemble of 50 or more 15-day weather trajectories from the current weather state.
Evidence sources: https://deepmind.google/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1072 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The new model, trained on 15,980 stations (up from 5,680), gives seven-day reliability comparable to the previous model's five-day forecast. Flood Hub covers over 100 countries and about 700 million people (up from 460 million). About 250,000 virtual points are unverified and intended for experts only; flash, urban, and coastal floods remain out of scope. The linked March 2024 Nature paper describes the predecessor model, also covered by the catalog's March entry; the November performance and coverage update is reported in this Google Research announcement.
Technology: Hydrological forecasting model
Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: retrospective or benchmark
- Tested: Gauged basins were checked with standard cross-validation, and ungauged locations were checked against Sentinel-1 radar inundation images; reliability was compared using Nash-Sutcliffe efficiency distributions across 3,079 basins.
- Still unknown: Ungauged validation covers only a fraction of locations, flash, urban and coastal floods are out of scope, and virtual-gauge forecasts are unverified.
- AI role: An LSTM network takes meteorological time series and static catchment attributes, adds input from DeepMind's medium-range forecast model, and outputs a probabilistic streamflow distribution.
Evidence sources: https://research.google/blog/a-flood-forecasting-ai-model-trained-and-evaluated-globally/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 580 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A partial-convolution U-Net trained on CMIP6 climate simulations and transferred to European station-based extreme indices reconstructed the poorly observed 1911 heatwave. The model estimates that in parts of southern France, more than 40% of September days exceeded the 90th percentile of maximum temperature, extending the heatwave signal into fall. The reconstructed pattern is similar to 20th Century Reanalysis Version 3 and aligns spatially with higher French mortality in 1911; the HadEX3 gridded record does not show the same pronounced heat.
Technology: Climate Reconstruction AI (CRAI), a U-Net with partial convolutions trained on CMIP6 historical simulations and transferred to the HadEX-CAM European extreme-index dataset
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A partial-convolution U-Net trained on CMIP6 climate simulations and transferred to European station-based extreme indices reconstructed the poorly observed 1911 heatwave.
- Still unknown: The September 1911 values are AI reconstructions rather than newly recovered thermometer readings; the paper states the source observations are sparse or missing. CRAI is trained on CMIP6 simulations and applied to European station-based extreme-index fields, so model-prior dependence remains. The article reports a similar spatial pattern in 20CRv3 and correlation with 1911 mortality, but these support the reconstructed pattern rather than directly validate each missing grid cell. The reported >40% refers to predicted percentage of days above the local 90th percentile for maximum temperature.
- AI role: The deep-learning reconstruction inferred missing monthly temperature-extreme indices from spatial context and learned climate-model patterns. This enabled a detailed estimate for the poorly observed September 1911 period; the model output exposes hot extremes in southern France that are not clearly represented in the standard HadEX3 map and supports comparison with independent 20th Century Reanalysis and historical mortality patterns.
Evidence sources: https://www.nature.com/articles/s41467-024-53464-2
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 36 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A convolutional neural network trained on idealized Earth-system-model simulations classified eruption occurrence and broad latitude band from global surface-temperature anomalies in the first post-eruption summer. Applied to reanalysis and observational climate records, the method identified major historic events and assigned a northern-extratropical footprint to the unidentified 1809 eruption. The same analysis found that some smaller events leave weaker, less distinguishable temperature signals.
Technology: Convolutional neural network with explainable-AI relevance mapping
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A convolutional neural network trained on idealized Earth-system-model simulations classified eruption occurrence and broad latitude band from global surface-temperature anomalies in the first post-eruption summer.
- Still unknown: The model is trained on idealized simulations with only three broad eruption latitude classes and four sulfur-injection strengths. The 1809 geographic inference is an AI classification of a climate fingerprint, not direct observation of the eruption source; predictions show sensitivity to normalization and differing observational datasets. It should be interpreted as a clue, not a definitive eruption-location determination.
- AI role: The CNN classified volcanic presence/location using post-eruption temperature patterns; explainability analysis localized climate regions driving those classifications and helped infer the 1809 event’s northern-extratropical footprint from observational/reanalysis data.
Evidence sources: https://www.nature.com/articles/s43247-024-01617-y
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: at percentile 41 of works of the same type, year and subfield (OpenAlex work metrics, 2 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
At 1.4° resolution, the model outperformed ECMWF-ENS in 95% of cases for forecasts from 2 to 15 days; at 2.8°, its mean 1980–2020 temperature error was 0.25 °C versus 0.75 °C for atmosphere-only models. It is more than 3,500 times faster than X-SHiELD on a TPU. It models only the atmosphere and is not yet a complete climate model.
Technology: NeuralGCM (differentiable dynamical core and neural networks)
Family NeuralGCM: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Evaluated with the WeatherBench 2 benchmark against ECMWF-ENS (better in 95% of 2-15 day cases at 1.4 degrees) and against AMIP climate baselines for 1980-2020 temperature (0.25 vs 0.75 degrees C error).
- Still unknown: The model covers only the atmosphere and is not yet a complete climate model; oceans and the carbon cycle are future goals.
- AI role: A neural network learns small-scale physics such as cloud formation, replacing hand-built parameterizations, coupled to a physics-based fluid dynamics solver.
Evidence sources: https://research.google/blog/fast-accurate-climate-modeling-with-neuralgcm/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 780 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A 1.3-billion-parameter model for high-resolution weather and air-pollution forecasting: it matches or exceeds CAMS on 74% of targets and matches or beats GraphCast on 94%. The roughly 5,000-fold speedup over the IFS system is an estimate; GraphCast remains slightly better up to five days.
Technology: Aurora (foundation model)
Family Aurora: profile · models page
Publication: Publication status unspecified in annotation · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Not reported
- Test type: retrospective or benchmark
- Tested: Forecasts were compared with CAMS for air pollution (Jun-Nov 2022, matched or beat 74% of targets) and with GraphCast (matched or beat 94% of targets) over 2022 evaluation periods.
- Still unknown: GraphCast was slightly better on specific humidity up to five days, and the page has no dedicated limitations section.
- AI role: A 1.3-billion-parameter foundation model, pretrained on over a million hours of simulations, forecasts weather and air pollution at 0.1 degrees.
Evidence sources: https://www.microsoft.com/en-us/research/blog/introducing-aurora-the-first-large-scale-foundation-model-of-the-atmosphere/
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A vision transformer applied to multispectral Sentinel-2 data detects methane plumes as small as 0.01 km2, corresponding to sources of about 200 to 300 kg CH4 per hour, and is reported to improve on the prior state of the art by an order of magnitude. Detections were compared with airborne methane measurement campaigns; the claims rest on that comparison and the abstract gives no formal detection-rate metrics.
Technology: Vision transformer for multispectral methane plume detection (Sentinel-2)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Detections were compared with airborne methane measurement campaigns, and the method is reported to detect plumes down to 0.01 km2 (about 200 to 300 kg CH4 per hour); the authors describe this as supporting evidence rather than a definitive validation.
- Still unknown: Not reported; the page lists no explicit limitations and gives no formal detection-rate metrics.
- AI role: Vision transformer applied to multispectral Sentinel-2 data to detect methane plumes, reported to improve on prior methods by an order of magnitude.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-024-47754-y&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 115 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Using satellite observations of four Kīlauea volcanic outgassing episodes in 2008 and 2018, an ML counterfactual analysis found that volcanic aerosols increased cloud cover and reflected sunlight over tropical marine shallow-convective clouds, even as rainfall strengthened. Cloud-cover adjustment contributed about 65% of aerosol–cloud cooling under mixed meteorological conditions and 84–87% in humid conditions; under humid, stable conditions, a 30% increase in droplet number could correspond to up to a 50% relative increase in cloud cover. The result identifies a tropical regime in which aerosol cloud expansion is strong and its response varies with humidity and stability.
Technology: Random-forest machine-learning surrogates of MODIS cloud properties, radiative fluxes, and precipitation, trained on 20 years of satellite observations and meteorology while excluding volcanic years 2008 and 2018
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Using satellite observations of four Kīlauea volcanic outgassing episodes in 2008 and 2018, an ML counterfactual analysis found that volcanic aerosols increased cloud cover and reflected sunlight over tropical marine shallow-convective clouds, even as rainfall strengthened.
- Still unknown: This is an observational ML-counterfactual analysis of four Kīlauea plume months over tropical marine shallow-convective clouds, not a controlled aerosol intervention or a direct test of deliberate cloud seeding. The source reports strong cloud-cover responses and radiative estimates, especially under humid/stable conditions, but results depend on satellite retrievals and the learned unperturbed baseline. The authors’ proposed mechanism for increased cloud cover alongside stronger rainfall is explicitly a hypothesis requiring further testing; do not treat the marine-cloud-brightening scenario as experimentally demonstrated.
- AI role: The random-forest surrogates estimated unperturbed cloud properties and radiative fluxes for given meteorological conditions. Comparing those counterfactuals with observations in four volcanic plume episodes isolated aerosol-associated signals from meteorological co-variability and enabled estimates of cloud-cover, rainfall, and radiative-forcing responses across distinct tropical meteorological regimes.
Evidence sources: https://www.nature.com/articles/s41561-024-01427-z
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 63 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The model produces 256 members at 2° resolution every three minutes on a TPU and matches or exceeds the operational GEFS ensemble (31 members) on rank histograms, RMSE and CRPS. For the Lisbon heatwave (July 2022), with 16,384 members, none of GEFS’s 31 members was as warm as observed. One extreme case; depends on physics-based forecasts as seeds; published in Science Advances.
Technology: SEEDS (diffusion model)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: The ensemble was compared with the operational GEFS (31 members) using rank histograms, RMSE and CRPS, and with a 16,384-member ensemble for the Lisbon heatwave of July 2022.
- Still unknown: The heatwave comparisons are single examples, the model depends on physics-based seed forecasts, and the page has no explicit limitations section.
- AI role: A denoising diffusion model generates large weather-forecast ensembles from one or two operational forecasts.
Evidence sources: https://research.google/blog/generative-ai-to-quantify-uncertainty-in-weather-forecasting/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 172 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A model trained on 5,680 gauges improved average global forecast reliability from zero to five days: forecasts at days 4–5 were similar to or better than day-zero forecasts from GloFAS v4. Flood Hub covers more than 80 countries. F1 comparison across 2,092 basins for two-year return-period events; flash and urban floods remain to be covered.
Technology: Hydrological forecasting model
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: F1 scores were compared with GloFAS v4 for 2-year return-period events across 2,092 watersheds (2014-2023); reliability at 4-5 day lead times was similar to or better than GloFAS at day zero.
- Still unknown: Coverage must expand to more locations, flash and urban floods are not yet covered, and the gain is average-level rather than uniform.
- AI role: Two sequential LSTMs with a mixture density output produce probabilistic streamflow forecasts, trained on 5,680 gauges.
Evidence sources: https://research.google/blog/using-ai-to-expand-global-access-to-reliable-flood-forecasts/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 580 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Kernel ridge regression reconstructed present and future air–sea CO₂-flux variability in five Earth-system models. The analysis found dissolved inorganic carbon (DIC) and alkalinity to be the main drivers, while DIC’s relative importance generally declines in warming scenarios—especially in the equatorial Pacific, Southern Ocean and North Atlantic—as anthropogenic carbon uptake flattens vertical DIC gradients. This result is conditional on the models and scenarios analyzed.
Technology: Kernel ridge regression and machine-learning clustering/classification
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Kernel ridge regression reconstructed present and future air–sea CO₂-flux variability in five Earth-system models.
- Still unknown: This is an analysis of outputs from five Earth-system models and four historical/future scenarios, not a measurement of future ocean conditions. The interpretation that anthropogenic-carbon-driven gradient flattening contributes to reduced DIC importance is the authors’ proposed mechanism; relative predictor importance does not alone establish causal attribution.
- AI role: Kernel ridge regression reconstructed CO₂-flux variability from Earth-system-model variables; machine-learning analysis compared predictor importance among spatial clusters and climate scenarios to identify the changing roles of DIC and alkalinity.
Evidence sources: https://www.nature.com/articles/s43247-024-01257-2
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: at percentile 88 of works of the same type, year and subfield (OpenAlex work metrics, 13 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
It outperforms ECMWF’s HRES system on over 90% of 1,380 variables and test forecast lead times, and produces a 10-day forecast in under a minute on a single TPU. Comparison is limited to deterministic HRES; training uses ERA5 data.
Technology: GraphCast (graph neural network)
Family GraphCast: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: 10-day forecasts of hundreds of variables at 0.25° in under a minute; outperforms the most accurate operational deterministic systems on 90% of 1,380 verification targets; supports better tropical-cyclone, atmospheric-river and extreme-temperature prediction.
- Still unknown: Comparison with ensemble systems is not described in the abstract.
- AI role: A machine-learning model trained directly on reanalysis data.
Evidence sources: https://doi.org/10.1126/science.adi2336
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2507 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The model forecasts precipitation, temperature, wind and dew point up to 24 hours ahead at 1–4 km resolution, learning from about 1,000 stations, and outperforms HRRR and ENS in several regions. It operates in Google Search and the app in the United States and 27 European countries, with precipitation forecasts up to 12 hours. Preprint, no numerical values in the text and one case study.
Technology: MetNet-3
Family MetNet: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: MetNet-3 was benchmarked against HRRR and ENS over the contiguous US for forecasts up to 24 hours, scored with CRPS; a single January 2019 case study is also shown.
- Still unknown: The post has no limitations section; it notes operational coverage limited to the contiguous US and parts of Europe, hyperlocal wind results depend on test-station input, and the case study is a single example.
- AI role: A neural model forecasts precipitation, temperature, wind and dew point up to 24 hours ahead, learning from about 1,000 stations.
Evidence sources: https://arxiv.org/abs/2306.06079 · https://research.google/blog/metnet-3-a-state-of-the-art-neural-weather-model-available-in-google-products/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Applying THOR to one eddy-permitting CM4 climate simulation, researchers detected a projected dynamical-regime shift where the Antarctic Circumpolar Current crosses the Pacific–Antarctic Ridge under the SSP5-8.5 scenario. Neural-network explanations and transport analysis attributed the modeled change to stronger wind stress shifting the current northward into less topographically constrained waters, where the flow intensifies. This is a model-based scenario result, not an observed future change.
Technology: THOR (Tracking global Heating with Ocean Regimes), using NEMI clustering, an ensemble of 50 multilayer perceptrons, SHAP and layer-wise relevance propagation
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Applying THOR to one eddy-permitting CM4 climate simulation, researchers detected a projected dynamical-regime shift where the Antarctic Circumpolar Current crosses the Pacific–Antarctic Ridge under the SSP5-8.5 scenario.
- Still unknown: This is an arXiv preprint and computational climate-model analysis using CM4/MOM6, one model and the SSP5-8.5 scenario. The paper reports neural-network validation accuracy around 65–70%; it describes the model output and projected shift, not an observed real-world regime transition.
- AI role: The clustering and neural-network ensemble identified and tracked local ocean dynamical regimes across model scenarios; SHAP and layer-wise relevance propagation highlighted wind-stress curl and bathymetry for follow-up physical analysis.
Evidence sources: https://arxiv.org/abs/2310.13916
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A convolutional neural network trained on 80 CESM2 large-ensemble members detected an emerging climate-change signal in observed daily precipitation fields. Across four satellite, merged-observation and reanalysis products, the model’s annual estimated global-mean temperature exceeded the historical internal-variability range from the mid-2010s onward. The signal was driven mainly by precipitation variability on timescales shorter than 10 days, with hotspots including the tropical eastern Pacific and mid-latitude storm tracks; rain-gauge checks over the eastern United States were consistent with the increase in high-frequency variability.
Technology: Deep detection (DD) convolutional neural network
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A convolutional neural network trained on 80 CESM2 large-ensemble members detected an emerging climate-change signal in observed daily precipitation fields.
- Still unknown: This is an observational detection and attribution result using a CNN trained on climate-model simulations; it is not a direct causal experiment or a forecast of individual storms. The long-term mean precipitation change remained indistinguishable from natural variability in the analysis. Precipitation products disagree in some estimates, especially where direct ocean measurements are sparse, and the reported hotspots depend on the selected datasets and attribution methods. The paper reports an emerging signal beginning in the mid-2010s, not a uniform increase in all regions or at every timescale.
- AI role: The CNN was trained on paired daily precipitation anomaly maps and annual global-mean surface-air-temperature values from 80 CESM2 Large Ensemble members. Applied to observed precipitation products, it learned spatial and temporal patterns that predicted the warming metric and exposed an emergent signal that conventional linear detection did not resolve. Explainability analyses (occlusion sensitivity, with SHAP and integrated-gradients checks) localized the strongest contributions to high-frequency precipitation variability.
Evidence sources: https://www.nature.com/articles/s41586-023-06474-x
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 115 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
In a trial with American Airlines and Breakthrough Energy, pilots flew 70 flights over six months using AI forecasts to avoid forming contrails, reducing them by 54% compared with flights without the forecasts. Flights that avoided contrails used 2% more fuel; across an airline’s flights, the impact could fall to 0.3% if only a small share need rerouting.
Technology: AI contrail-formation forecasts and satellite imagery
Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: direct test
- Tested: Google, American Airlines and Breakthrough Energy flew 70 pilot flights over six months using AI forecasts to avoid contrail-forming altitudes; satellite imagery reported a 54% contrail reduction.
- Still unknown: The article does not state a control group, the baseline for the 54% figure, statistical methods or peer review; fuel impact (2% more fuel, 0.3% airline-wide) is an estimate.
- AI role: AI contrail forecasts guided pilots to avoid altitudes likely to form contrails.
Evidence sources: https://blog.google/technology/ai/ai-airlines-contrails-climate-change/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
According to the authors, it is the first AI model to outperform ECMWF’s IFS at medium range: five-day Z500 RMSE of 296.7 versus 333.7, a global 24-hour forecast in 1.4 seconds on a GPU, and mean cyclone-track error of 120 km at three days versus 162 km. Source reviewed: Huawei Cloud blog, not the Nature paper (July 6, 2023); it underestimates extremes.
Technology: Pangu-Weather
Family Pangu-Weather: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Trained on 39 years of global data; stronger deterministic results on reanalysis data than ECMWF's operational IFS in all tested variables; also evaluated on extreme weather and ensemble forecasts.
- Still unknown: Comparisons are initialized with reanalysis data, as the abstract states; performance from operational analyses is not given there.
- AI role: Three-dimensional deep networks with Earth-specific priors and hierarchical temporal aggregation.
Evidence sources: https://doi.org/10.1038/s41586-023-06185-3
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2232 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A model combining neural networks and physical constraints forecasts precipitation up to three hours ahead; 62 meteorologists preferred it in 71% of 2,400 evaluated events. Tsinghua reports it was already deployed on the National Meteorological Center’s operational SWAN 3.0 platform. Expert preference does not measure reduced damage or guarantee accuracy for every event.
Technology: NowcastNet
Publication: Peer-reviewed paper · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: prospective or independent check
- Tested: 62 forecasters from 23 meteorological stations ran retrospective and prospective checks on about 2,400 extreme-precipitation events and ranked NowcastNet first in 71% of cases against leading methods.
- Still unknown: No limitations of NowcastNet are stated; the abstract reports no verification scores beyond expert preference, and the page says the team plans to extend the approach to other physics problems.
- AI role: A neural network combined with physics-based precipitation evolution forecasts extreme precipitation up to three hours ahead from radar data.
Evidence sources: https://www.thss.tsinghua.edu.cn/info/1023/2133.htm · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06184-4&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 576 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A model using geostationary imagery tracks fire perimeters with updates every 10–15 minutes and, according to the page, is more accurate than similar products and the previous model based on NOAA GOES-16/17 data; shown in Search and Maps for large fires in the United States, Mexico, much of Canada and Australia. Resolution is 2 km; clouds can hide fires; figures appear only in an image.
Technology: Model on geostationary satellite imagery
Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: retrospective or benchmark
- Tested: Fire perimeters were compared with official fire scars using precision, recall and distance metrics; a comparison with an earlier GOES-16/17 model and a 2022 US test set unavailable during training are also reported.
- Still unknown: The post states limitations: smoke and clouds hide fires, resolution is about 2 km and coarser away from nadir, and coverage is limited to large fires in the listed regions; no paper is cited.
- AI role: A model using geostationary satellite imagery tracks fire perimeters with updates every 10-15 minutes.
Evidence sources: https://research.google/blog/real-time-tracking-of-wildfire-boundaries-using-satellite-imagery/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Using MODIS satellite observations and a random-forest counterfactual trained on meteorology and non-eruption years, researchers estimated cloud responses to the 2014 Holuhraun volcanic aerosol plume. They found cloud fraction increased by about 10% across the studied region; this change contributed more to the estimated aerosol cooling than the smaller radiative effect from brighter clouds with smaller droplets. The study supplies an observational constraint on aerosol–cloud climate forcing.
Technology: Random-forest ML-MODIS surrogate trained on meteorological reanalysis and MODIS satellite data from 2001–2020, excluding the 2014 eruption year
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Using MODIS satellite observations and a random-forest counterfactual trained on meteorology and non-eruption years, researchers estimated cloud responses to the 2014 Holuhraun volcanic aerosol plume.
- Still unknown: The result concerns one volcanic natural experiment, Holuhraun in 2014, and satellite observations of a North Atlantic/Northern European region. The cloud response is estimated by comparing observations with an ML counterfactual and is not a randomized intervention; it depends on model and MODIS retrieval assumptions. The paper reports about a 10% cloud-cover increase and attributes the leading radiative contribution to cloud fraction, while the droplet-size effect is also present and the liquid-water-path contribution is weak.
- AI role: The random forest estimated the cloud properties expected under observed meteorological conditions without volcanic aerosol, creating a counterfactual baseline. Comparing these predictions with 2014 MODIS observations let the authors isolate cloud changes associated with the Holuhraun aerosol perturbation from meteorological co-variability and estimate the relative radiative contributions of cloud-fraction, liquid-water-path, and droplet-size changes.
Evidence sources: https://www.nature.com/articles/s41561-022-00991-6
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 144 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The preprint presents global forecasts at 0.25° resolution and a one-week forecast in under two seconds. Comparison with IFS is favorable at short lead times for large-scale variables and some finer-scale variables; this does not mean general superiority or operational use.
Technology: Adaptive Fourier Neural Operators
Family FourCastNet: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Forecasts at 0.25 degrees are compared with ECMWF IFS, matching it at short lead times for large-scale variables and outperforming it for some fine-scale variables such as precipitation.
- Still unknown: Not reported in the arXiv abstract read.
- AI role: Adaptive Fourier neural operator produced global weather forecasts.
Evidence sources: https://arxiv.org/abs/2202.11214
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Developed with the UK Met Office, the model forecasts precipitation up to two hours ahead; over 50 expert meteorologists chose it as their first choice in 89% of cases compared with the most widely used methods. The authors cite long-range accuracy and rare, intense events as limitations.
Technology: Deep generative radar model (DGMR)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Over 50 Met Office meteorologists chose the model first in 89% of cases, and it was compared with PySTEPS and UNet on UK and US radar data.
- Still unknown: No method is without limitations; long-range accuracy and rare intense events need improvement.
- AI role: Generative model forecast precipitation up to two hours ahead.
Evidence sources: https://deepmind.google/blog/nowcasting-the-next-hour-of-rain/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-021-03854-z&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1232 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Using machine learning on data from over 300 satellites and 11,100 sensors (over 59 trillion bytes), emissions from oil and gas in major countries were estimated at about twice UNFCCC reports (about 1 billion tonnes more), while rice emissions in India were nearly three times the official 2016 inventory. No peer review; the page gives both September 14 and September 16, 2021.
Technology: Machine learning on satellite and sensor data
Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: not reported
- Tested: The page gives no methodology for checking the estimates; it says more than 50 outside organisations reviewed methodology and validated findings, without describing results.
- Still unknown: The page notes granularity is still developing and that estimates are hedged and constantly evolving; no peer review is mentioned.
- AI role: Machine learning on satellite and sensor data estimated global emissions.
Evidence sources: https://climatetrace.org/news/climate-trace-releases-first-comprehensive-independent
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 1/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The model outperforms ECMWF’s SEAS5 dynamical system in seasonal forecasts of summer sea ice, especially for extreme events, and is over 2,000 times faster (under ten seconds on a GPU). SEAS5 is better at a one-month lead, and the September comparison covers nine years; published in Nature Communications.
Technology: IceNet (probabilistic deep learning)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: IceNet is compared with the SEAS5 dynamical model in seasonal summer sea-ice forecasts, especially extreme events.
- Still unknown: Not reported in the abstract read.
- AI role: Probabilistic deep-learning model forecast the next six months of sea-ice concentration.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=PMCID:PMC8390499&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Non-negative matrix factorization (NMF) was applied to stream-chemistry time series from three watersheds to infer solute endmembers and their changing contributions without predefining endmember chemistry. Combined with geological interpretation and independent geochemical evidence, the model separated shallow acid-rain-affected flow from deeper pyrite-weathering contributions. The analysis found that acid rain can reduce a watershed’s long-timescale CO₂-sequestration capacity and that omitting sulfide oxidation/acid-rain reactions overestimates weathering-driven CO₂ drawdown.
Technology: Non-negative matrix factorization (NMF) machine-learning blind-source-separation model applied to stream chemistry
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Non-negative matrix factorization (NMF) was applied to stream-chemistry time series from three watersheds to infer solute endmembers and their changing contributions without predefining endmember chemistry.
- Still unknown: This is an inference from historical stream chemistry and watershed-specific geology, not a direct measurement of global CO₂ sequestration or a causal intervention. The NMF endmembers were interpreted as subsurface flowpaths using geological knowledge and supporting geochemical evidence; the inferred acid-rain sulfate export at Shale Hills includes a 19–31-year lag estimate. The result applies to three studied watersheds and the weathering timescales considered. Endmember counts and interpretations depend on the model setup and watershed reaction-front assumptions.
- AI role: The NMF model simultaneously inferred endmember chemical signatures and per-sample mixing proportions from solute ratios without preset endmember compositions. Researchers interpreted those patterns using watershed geology, sulfur isotopes, weathering profiles and calcite saturation, then estimated mineral-weathering rates and CO₂ sequestration. The ML decomposition enabled source separation where conventional endmember assumptions were uncertain; it did not directly measure a global carbon flux.
Evidence sources: https://hess.copernicus.org/articles/25/3397/2021/index.html
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 35 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Using explainable AI on single-forcing CESM1 large ensembles, the study identified the North Atlantic, Southern Ocean and Southeast Asia as key regional temperature patterns for dating the climate simulations, particularly before the early 21st century. When applied to reanalysis observations, the neural network trained on an ensemble with industrial aerosols fixed at 1920 levels predicted the actual year more closely than networks trained on the all-forcing or greenhouse-gas-fixed ensembles. The authors interpret this as evidence that simulated regional temperature signals are sensitive to aerosol forcing.
Technology: Artificial neural network trained on CESM1 single-forcing and all-forcing large ensembles, interpreted using layer-wise relevance propagation and an uncertainty mask
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Using explainable AI on single-forcing CESM1 large ensembles, the study identified the North Atlantic, Southern Ocean and Southeast Asia as key regional temperature patterns for dating the climate simulations, particularly before the early 21st century.
- Still unknown: This is a model-based attribution analysis from one climate model (CESM1) and historical/RCP8.5 simulations, with reanalysis used for evaluation; it is not a direct experiment or causal attribution of every observed regional trend. The relevance maps identify patterns useful to the trained network and do not by themselves prove a physical mechanism. The authors discuss other forcings, feedbacks and internal variability, and the net influence of aerosols remains uncertain.
- AI role: The ANN learned to predict year from global near-surface-temperature maps in simulations with different combinations of greenhouse-gas and industrial-aerosol forcing. Layer-wise relevance propagation identified regional patterns driving those predictions, and the trained networks were then tested on 20th Century Reanalysis observations. This analysis exposed the model’s aerosol-sensitive regional signal and the higher year-prediction correlation of the network trained with aerosols held at 1920 levels.
Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021MS002464
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 77 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Analysis of high-resolution images located and measured the crowns of over 1.8 billion trees across 1.3 million km² of West Africa. Validation used manual annotations and field data; the study counts trees but does not yet measure their stored carbon.
Technology: Convolutional neural network
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Tree detections checked against manually annotated data and field data from the region; the abstract itself does not describe the validation method.
- Still unknown: The study counts trees but does not yet measure stored carbon; the method can mistake stones for trees and gets harder with more detail.
- AI role: Deep learning (fully convolutional network trained on about 90,000 hand-marked trees) on submetre satellite imagery mapped the crowns of over 1.8 billion trees.
Evidence sources: https://www.nasa.gov/centers-and-facilities/goddard/nasa-supercomputing-study-breaks-ground-for-tree-mapping-carbon-research/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-020-2824-5&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 496 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Relevance-vector-machine sparse regression applied to high-resolution idealized ocean simulations produced compact momentum and temperature-flux closures for coarse grids. In the authors’ simulated setup, the inferred momentum terms include energy transfers absent from then-current closures; the parameterizations improved selected statistics when implemented in an idealized 30-km ocean model. This is a model-derived closure result, not an observation of a previously unknown ocean process.
Technology: Relevance vector machine (RVM) sparse Bayesian regression; physically constrained convolutional neural network used as a comparison
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Relevance-vector-machine sparse regression applied to high-resolution idealized ocean simulations produced compact momentum and temperature-flux closures for coarse grids.
- Still unknown: Derived from idealized, eddy-resolving double-gyre simulations and tested in an idealized coarse-resolution shallow-water model, not validated against observational ocean data or a global coupled climate model. The inferred tensor structure relates to earlier Reynolds-stress and deformation-based theories; the paper’s stated novelty is the additional data-derived terms/energy pathways and resulting parameterizations, not discovery of a new observed ocean phenomenon.
- AI role: RVM selected compact closed-form terms from candidate functions of simulated velocity, vorticity and deformation fields to infer subgrid momentum and temperature forcing. The authors report the expressions captured energy pathways omitted from existing parameterizations. A physically constrained CNN was also trained and evaluated, but the interpretable closure result comes from the RVM.
Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2020GL088376
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 288 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The model forecasts precipitation at 1 km resolution and 2-minute intervals up to 8 hours ahead from radar and satellite, without physical equations. According to the page, it outperforms NOAA’s operational HRRR model up to about 7–8 hours, in seconds rather than about an hour. US coverage only, author-only comparisons, a preprint, and no stated operational use.
Technology: MetNet (neural network)
Family MetNet: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Evaluated on a precipitation-rate forecasting benchmark against NWP (HRRR) and an optical-flow baseline, scored by F1 at 1.0 mm/h; outperforms NWP at forecasts up to about 7 to 8 hours.
- Still unknown: Demonstrated only for the continental US; the authors call it a small stepping stone; comparisons are author-run.
- AI role: Deep neural network with axial self-attention forecasting precipitation probabilities from radar and satellite inputs, without physical equations.
Evidence sources: https://research.google/blog/a-neural-weather-model-for-eight-hour-precipitation-forecasting/ · https://arxiv.org/abs/2003.12140
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.