Research · Summary
Back to sweepResearch sweep · deep · 2015 – 2026
AI in Weather and Climate Prediction
AI in weather and climate prediction across the 2015 to June 2026 machine-learning era, with historical context from mid-twentieth-century numerical weather prediction and Lorenz's chaos theory: the shift from physics-based NWP and statistical post-processing (MOS) to data-driven models (GraphCast, GenCast, Pangu-Weather, FourCastNet, Aurora, NeuralGCM, ECMWF AIFS), how forecasters at ECMWF, NOAA, and the Met Office have operationalised them, measured accuracy versus the IFS, and the predictability limits imposed by chaos, the Lorenz attractor, and the butterfly effect.
- Claude Opus 4.8
- financial
- frontier
- academic
- vc
- blogs
Synthesised 2026-06-26
AI in Weather and Climate Prediction: From Charney's ENIAC to the Diffusion Era
Overview
Weather forecasting has just lived through its biggest methodological break since ensemble prediction arrived in 1992. Between November 2023 and mid-2026, data-driven models trained on decades of reanalysis went from research curiosities to operational systems running alongside the physics-based engines they were built to challenge. The defining shift of the past 18 months is institutional: ECMWF made its AI Forecasting System operational on 25 February 2025, added a 51-member AI ensemble on 1 July 2025, and NOAA declared three AI-driven models operational in December 2025.
Sources: ECMWF (2025) (↗); ECMWF (2025) (↗); NOAA (2025) (↗)
The economic logic is brutal and simple. A 10-day GraphCast forecast costs a few dollars of GPU time against hundreds of millions of dollars of supercomputing for an equivalent IFS run, and ECMWF reports roughly a 1,000-fold reduction in energy use per forecast for AIFS. That cost collapse is reorganising a commercial market that had been stable for decades, in which public agencies bore the capital cost and private firms sold value-added analytics.
Sources: Science (2023) (↗); ECMWF (2025) (↗)
Why it matters now is that the headline accuracy claims and the operational reality have begun to separate. The models genuinely beat the IFS on standard medium-range metrics, but they underpredict the intensity and frequency of record-breaking extremes, the very events of greatest commercial and societal value. They also remain downstream of classical physics-based data assimilation for their initial conditions, so the "replacement" narrative is incomplete.
Sources: arXiv (2025) (↗); Science Advances (2026) (↗); Bloomberg (2025) (↗)
The deeper question sitting under all of it is whether these models change the predictability horizon that Edward Lorenz formalised in the 1960s, or merely reach it more cheaply. The evidence so far says the latter, and that the chaos ceiling of roughly two weeks for deterministic synoptic forecasting remains intact.
Sources: Geophysical Research Letters (2023) (↗); arXiv (2025) (↗)
Timeline
- ERA5 reanalysis released, the universal training corpus
- WeatherBench establishes a common ML benchmark
- FourCastNet and Pangu-Weather demonstrate big-tech ML on global weather
- GraphCast beats HRES across the board in Science
- WeatherBench 2 gives an independent scoreboard
- Selz and Craig show AI models miss the butterfly effect
- GenCast diffusion ensemble beats ECMWF ENS
- NeuralGCM bridges weather and climate
- Aurora foundation model published
- startup Series A wave (WindBorne, Brightband, Jua)
- ECMWF AIFS goes operational then adds AI ensemble
- NOAA budget and staff cuts begin
- conservation-constrained AIFS v1.1 ships
- NOAA declares three AI models operational
- Tomorrow.io raises 175m at over 1bn valuation
- physics still wins on record-breaking extremes
Key Findings
ERA5 is the substrate everything stands on, including the dependency that complicates the replacement story. Virtually every modern ML weather model trains on ECMWF's ERA5 reanalysis, a physically consistent, gap-free 0.25-degree hourly record from 1940 onwards produced by Hersbach and colleagues. The subtle catch is that ERA5 was itself generated by running the IFS with 4D-Var data assimilation, so these models learn a statistical approximation of the IFS's own state estimates and inherit its conservation inconsistencies.
Sources: Journal of Advances in Modeling Earth Systems (2020) (↗); Science (2023) (↗)
The architectures diverged into four families, each with different inductive biases. FourCastNet used Adaptive Fourier Neural Operators on a Vision Transformer backbone; Pangu-Weather used 3D Earth-Specific Transformers and was the first to beat the deterministic IFS on tested variables; GraphCast used a multi-scale icosahedral mesh graph neural network; and NeuralGCM took the hybrid route, embedding learned parameterisations inside a differentiable dynamical core. Only the hybrid preserves the conservation properties that pure data-driven models lack.
Sources: arXiv (2022) (↗); Nature (2023) (↗); Science (2023) (↗); Nature (2024) (↗)
Operationalisation is real, not a demo, and it happened fast. ECMWF's AIFS went live in February 2025 and reported tropical cyclone track improvements of up to 20 percent. NOAA's December 2025 deployment included AIGFS, a 31-member AIGEFS, and a hybrid HGEFS, with AIGFS completing a 16-day forecast on 0.3 percent of the traditional GFS compute and AIGEFS extending lead time by 18 to 24 hours.
Sources: Earth.org (2025) (↗); NOAA (2025) (↗); CBS News (2025) (↗)
Diffusion solved the ensemble problem more credibly than deterministic models did. Price and colleagues' GenCast generates ensemble members by conditioning on the current state and injecting noise, outperforming ECMWF's 51-member ENS on 97.2 percent of verification targets while producing physically coherent power spectra. This matters because deterministic ML models trained on mean squared error produce oversmoothed fields at long lead times, spuriously improving RMSE while losing small-scale physical structure.
Sources: arXiv / Nature (2024) (2023) (↗); Nature (2024) (↗)
The strongest academic counter-evidence is that physics still wins where it matters most. Zhang, Fischer and colleagues found ECMWF HRES consistently outperforms GraphCast, Pangu-Weather and FuXi for record-breaking heat, cold and wind extremes, with AI models systematically underpredicting tail intensity. The April 2026 Science Advances paper sharpened this into a direct challenge to the headline accuracy narrative, and it sits uneasily against the vendor benchmarks.
Sources: arXiv (2025) (↗); Science Advances (2026) (↗)
The butterfly effect is sidestepped, not solved. Selz and Craig demonstrated that current AI models cannot simulate the butterfly effect: infinitesimal perturbations do not grow at the correct rate, which incorrectly implies unlimited atmospheric predictability. The 2025 follow-up testing whether ML could push past 30 days found only modest extensions consistent with better effective initial conditions, confirming the Lorenz horizon stands.
Sources: Geophysical Research Letters (2023) (↗); arXiv (2025) (↗); arXiv (2025) (↗)
Conservation laws are being bolted on after the fact, and it works. Sha and colleagues showed that adding global mass and energy conservation schemes to FuXi reduces forecast error and corrects biases such as excess light rain, and ECMWF's September 2025 AIFS update introduced output bounding layers to fix spurious negative precipitation present until the v1.1.0 release. These are not cosmetic fixes; they are the crux of whether ML can be trusted on extremes where the training distribution is thinnest.
Sources: Journal of Advances in Modeling Earth Systems (2025) (↗); arXiv (2025) (↗); Geoscientific Model Development (Copernicus) (2026) (↗)
The commercial market is consolidating around data moats, not model moats. Tomorrow.io closed 175 million dollars in February 2026 at over a billion-dollar valuation, earmarked for the DeepSky satellite constellation to address the data-density bottleneck. WindBorne's balloon-derived observations sell to NOAA, the US Air Force and Navy, and by June 2026 its WeatherMesh model claimed to beat the IFS on surface temperature five days out. The historical analogy favours data: ECMWF's superiority has always rested more on assimilation quality than on architecture.
Sources: PR Newswire (2026) (↗); CTech / Calcalist (2026) (↗); TechCrunch (2026) (↗); Business Wire / WindBorne Systems (2024) (↗)
NeuralGCM is the only credible weather-to-climate bridge so far. Kochkov and colleagues fused a differentiable dynamical core with learned physics, matching GraphCast on medium-range skill while running stable multi-decade climate simulations. The January 2026 extension to precipitation suggests the hybrid path is maturing, but no model has demonstrated robust physical responses substantially outside the training envelope.
Sources: Nature (2024) (↗); Science Advances (2026) (↗); Google Research (2026) (↗)
Evidence & Data
The verification numbers are consistent across the deep-learning wave. GraphCast beat ECMWF's HRES on 90 percent of 1,380 verification targets and produced a 10-day forecast in under a minute. GenCast outperformed ENS on 97.2 percent of probabilistic targets. Aurora, a 1.3-billion-parameter foundation model pre-trained on over a million hours of geophysical data, beat existing numerical and AI models on 91 percent of targets and extended to air quality, ocean waves and hurricanes.
Sources: Science (2023) (↗); arXiv / Nature (2024) (2023) (↗); Nature (2025) (↗)
Compute and energy figures anchor the cost argument. AIFS reports roughly a 1,000-fold reduction in energy per forecast, and AIGFS runs a 16-day forecast on 0.3 percent of GFS compute. Bloomberg's September 2025 reporting put ECMWF's commercial licensees above 800, up almost 20 percent in 2024, with nearly half owned by energy companies.
Sources: ECMWF (2025) (↗); NOAA (2025) (↗); Bloomberg (2025) (↗)
Market sizing is wide and unreliable. Grand View Research put AI-based climate modelling at 343 million dollars in 2024, GM Insights at 266 million, and Transpire Insight projected 7.2 billion dollars by 2033 at a 26.4 percent CAGR. PwC's State of Climate Tech 2024 found AI climate ventures raised 6 billion dollars in the first three quarters of 2024, 14.6 percent of climate tech total, up from 7.5 percent in 2023, even as overall climate tech funding fell 29 percent to 56 billion dollars.
Sources: Grand View Research (2025) (↗); GM Insights (2024) (↗); Transpire Insight (2026) (↗); PwC (2024) (↗)
The independent verification picture is mixed. WeatherBench 2 provides a live scoreboard against 0.25-degree ground truth, and the npj evaluation of five global AI models across Eastern Asia and the Western Pacific is the kind of third-party assessment that vendor releases lack. McKinsey's September 2025 resilience report cited 27 billion-dollar US weather and climate disasters in 2024 as the demand catalyst for a claimed 1 trillion dollar private-capital opportunity by 2030.
Sources: Journal of Advances in Modeling Earth Systems (2024) (↗); npj Climate and Atmospheric Science (2024) (↗); McKinsey & Company (2025) (↗)
Signals & Tensions
The accuracy headline contradicts the extremes evidence, and the lanes know it. Vendors and ECMWF report broad outperformance, while academic and independent work shows physics-based models still win on record-breaking extremes. Bloomberg's own April 2025 piece on looking beyond conventional metrics signals that even the financial press has noticed RMSE flatters the smoothing problem.
Sources: Science Advances (2026) (↗); Bloomberg (2025) (↗); Geophysical Research Letters (2024) (↗)
The replacement narrative is oversold while the dependency is underreported. Independent writers and the Open-Meteo practitioner commentary stress that every ML model relies on classical physics-based assimilation upstream, and that GraphCast initialised on GFS rather than ERA5 produces systematic inconsistencies. ML-driven data assimilation is the live frontier that would close this gap.
Sources: Open-Meteo (Substack) (2024) (↗); arXiv (2025) (↗)
The NOAA cuts create a circular dependency nobody has resolved. Roughly 20 percent of NOAA's 13,000 staff were cut in early 2025 with proposed 27 percent budget reductions for 2026, degrading the radiosonde and satellite inputs that private AI models depend on. Companies including WindBorne, Tomorrow.io and Sofar are filling gaps, raising the governance question of proprietary dependency on a public good.
Sources: Undark (2025) (↗); Inside Climate News (2025) (↗); Inside Climate News (2026) (↗)
Foundation-model framing blurs the purpose-built distinction. Aurora's open release and Microsoft's Azure integration position weather as one capability within a general AI cloud, while NeuralGCM and AIFS are purpose-built. The VC lane notes that the most consequential capital is flowing inside hyperscaler R&D budgets, not standalone weather-AI startups.
Sources: Microsoft On the Issues (2025) (↗); arXiv (2025) (↗)
Climate non-stationarity is the quiet structural risk. Robustness studies found AI models trained on present-day ERA5 hold skill in pre-industrial and +2.9 K climates but drift back towards the training distribution in warmer scenarios. Insurers, whose pricing depends on tail statistics, face a harder validation bar than energy traders, which explains their slower adoption.
Sources: arXiv (2024) (↗); arXiv (2025) (↗)
Open Questions
Can ML-based data assimilation remove the dependency on physics-based analysis, and would that finally make the systems end-to-end data-driven? Brightband's bet on training from raw observations is the clearest commercial test of this.
Sources: TechCrunch (2024) (↗)
Do hybrid architectures such as NeuralGCM recover physically correct error-growth statistics, or do they only stabilise the climate run while still missing the butterfly effect at small scales? The conservation work suggests progress, but no model has demonstrated correct upscale error propagation.
Sources: Nature (2024) (↗); Geophysical Research Letters (2023) (↗)
How much of the headline outperformance survives independent verification on operational, real-time data rather than retrospective ERA5 hindcasts? WeatherBench 2 and SAFE-style stratified assessments are early attempts, but the vendor-versus-academic gap on extremes is unresolved.
Sources: arXiv (2025) (↗); Journal of Advances in Modeling Earth Systems (2024) (↗)
Will the data moat or the model moat win commercially? The WindBorne, Brightband and Jua theses are mutually exclusive enough that the next funding cycle will adjudicate.
Sources: TechCrunch (2026) (↗); Ananda Impact Ventures (2025) (↗)
Does ECMWF's decision to make AIFS freely downloadable in near-real-time collapse the value-added window that private resellers occupied, or expand the total market enough to offset it?
Sources: ECMWF (2026) (↗); Bloomberg (2025) (↗)
Can ML extend usefully into subseasonal and seasonal ranges where slow boundary conditions dominate, or is medium-range the genuine ceiling for the current generation?
Sources: arXiv (2025) (↗); Nature (2024) (↗)
What happens to forecast quality if NOAA's observing backbone degrades far enough that the public training and initialisation data the whole field depends on starts to thin? The models cheapened the forecast; nobody cheapened the observations.
Sources: Undark (2025) (↗); Inside Climate News (2026) (↗)
![[sources-ai-in-weather-and-climate-prediction-across-the-20]]
Sources
Summary: ↑ Back to summary
Financial Press
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| f1 | AI Drives Weather Data Demand Surge for Europe's Top Forecaster | Bloomberg | 2025-09 | Documents a nearly 20% rise in ECMWF commercial data licences in 2024 to over 800 firms, with energy traders and renewable producers the dominant buyers, directly quantifying AI-driven commercial demand. |
| f2 | Energy Traders Turn to AI to Forecast the Weather Forecast | Bloomberg | 2026-03 | Shows how European energy traders have moved beyond using AI weather models to using ML tools that predict the forecast itself, illustrating a second-order commercial application layer. |
| f3 | DeepMind's Latest AI Weather Model Targets Energy Traders | Bloomberg | 2025-11 | Bloomberg coverage of Google DeepMind's WeatherNext 2 model explicitly designed with tools for energy traders, marking the commercialisation of leading AI weather research. |
| f4 | AI Models Struggle at Forecasting Extreme Weather Events | Bloomberg | 2025-09 | Financial-press scepticism piece documenting AI models' limitations on the low-frequency, high-impact events of greatest commercial and safety relevance. |
| f5 | How AI Could Make Weather Forecasting Faster, Cheaper, More Accurate | Bloomberg | 2025-05 | Bloomberg explainer synthesising the business case for AI weather forecasting across industries that absorb billions in losses from extreme events. |
| f6 | AI Weather Forecasting Looks Beyond Conventional Metrics | Bloomberg | 2025-04 | Documents the frontier of AI forecasting moving beyond traditional meteorological metrics to non-meteorological factors relevant to enterprise decision-making. |
| f7 | Tomorrow.io Announces $175M Financing to Deploy DeepSky, The World's First AI-Native Weather Satellite Constellation | PR Newswire | 2026-02 | Primary source for Tomorrow.io's $175 million round at over $1 billion valuation, the largest disclosed weather-tech financing of the period, with Amazon and BNSF as named enterprise clients. |
| f8 | Tomorrow.io raises $175 million at over $1 billion valuation to build AI-driven weather satellite network | CTech / Calcalist | 2026-02 | Provides revenue detail (~$100 million ARR) and the CEO's commentary on the US shift toward privatisation of weather infrastructure, contextualising the investment thesis. |
| f9 | Tomorrow.io raises $175M to deploy its AI-native weather satellite constellation | SiliconAngle | 2026-02 | Confirms operational detail including 13 satellites deployed, 60-minute global revisit, and AWS as infrastructure partner, substantiating commercial traction claims. |
| f10 | ECMWF's AI forecasts become operational | ECMWF | 2025-02 | Primary institutional announcement confirming AIFS went fully operational on 25 February 2025, with approximately 1,000-fold energy reduction versus IFS, and naming energy pricing, insurance, and shipping as key commercial beneficiaries. |
| f11 | ECMWF's ensemble AI forecasts become operational | ECMWF | 2025-07 | Marks AIFS ENS going operational on 1 July 2025 with 51-member ensemble capability, enabling probabilistic commercial products previously unavailable from a fully AI system. |
| f12 | Another year of rapid development in machine-learning-based forecasting | ECMWF Annual Report 2024 | 2024 | Official ECMWF account of AIFS development through 2024, documenting resolution improvements, ensemble launches, and the acknowledged underestimation of extreme rainfall as an ongoing limitation. |
| f13 | Europe's AI Weather Forecasting Model Up to 20% More Accurate | Earth.org | 2025-07 | Synthesises ECMWF director Pappenberger's comments to the Financial Times on predictability horizons alongside the AIFS 20% accuracy gain for tropical cyclone tracks, bridging institutional and press coverage. |
| f14 | In NOAA Cuts Fallout, Private Companies Fill Weather Data Gaps | Undark | 2025-08 | Documents how NOAA budget and staffing cuts are creating commercial opportunities and governance risks simultaneously, with WindBorne, Tomorrow.io, and Saildrone named as private data suppliers. |
| f15 | How Massive Cuts to NOAA Could Impact Everything From Weather Apps to Agriculture to National Security | Inside Climate News | 2025-05 | Detailed analysis of the proposed 27% NOAA budget cut and 74% reduction to the Office of Atmospheric Research, with implications for AI model training data and long-term forecast accuracy. |
| f16 | NOAA Defends Cuts to Research and Climate Monitoring at Budget Hearing | Inside Climate News | 2026-04 | Most recent (April 2026) account of NOAA's FY2027 budget defence, with administrator Neil Jacobs' responses and congressional pushback, quantifying the proposed further cuts. |
| f17 | The Next Phase of Aurora: Open and Collaborative AI for Weather and Climate Forecasting | Microsoft On the Issues | 2025-11 | Microsoft's commitment to Aurora as an open-source foundation model, with integration into Azure weather and climate services, representing a major hyperscaler's commercial strategy for AI weather. |
| f18 | An update to ECMWF's machine-learned weather forecast model AIFS | arXiv (ECMWF authors) | 2025-09 | Technical paper documenting AIFS 1.1.0 release in August 2025, including the physical-consistency bounding layers added to correct precipitation forecasts, addressing a key commercial limitation. |
| f19 | AI Drives Weather Data Demand Surge for Europe's Top Forecaster (Insurance Journal reprint) | Insurance Journal / Bloomberg | 2025-10 | Insurance-sector reprint of Bloomberg's ECMWF data-demand story, confirming the insurance industry as a key commercial adopter of AI weather data alongside energy traders. |
| f20 | A.I. Is Quietly Powering a Revolution in Weather Prediction | Yale Environment 360 | 2025 | Authoritative practitioner synthesis including ECMWF's Peter Dueben on operational reliability, and the 1,000-fold energy reduction claim for AIFS versus IFS, relevant to computing cost economics. |
| f21 | Weather forecasting in a changing climate: the rise of AI and Machine learning? | ScienceDirect (peer-reviewed) | 2026-05 | May 2026 peer-reviewed synthesis documenting AIFS operational performance against IFS, physical-interpretability advances, and the democratisation argument for smaller meteorological services. |
| f22 | Mitsui & Co. Global Strategic Studies Institute Monthly Report December 2025 | Mitsui Global Strategic Studies Institute | 2025-12 | Business-intelligence report covering Google WeatherNext, NOAA staffing cuts (~20% of 13,000 employees, $1.3 billion budget reduction), and the US shift toward weather privatisation from an institutional investor perspective. |
| f23 | Tomorrow.io Raises $35M to Expand AI Weather Intelligence Infrastructure | Ventureburn | 2026-05 | Documents Tomorrow.io's May 2026 Series F extension of $35 million, confirming continued investor conviction and agentic AI expansion for operational enterprise decision-making. |
| f24 | Artificial intelligence could dramatically improve weather forecasting | Sustainability by Numbers | 2025-10 | Cites Nobel laureate economist Michael Kremer's estimate that AI weather forecasts generate more than $100 in farmer returns per dollar of government investment, providing an ROI framework for development-finance audiences. |
| f25 | The AI Revolution in Weather Forecasting Is Here | Eos (American Geophysical Union) | 2025-10 | Documents 375% average annual growth in AI weather-forecasting publications from 2011 to 2022 and maps the five most-studied forecast variables to major economic sectors including energy, providing a quantitative research-intensity baseline. |
Frontier Lab & Model News
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| t1 | Learning skillful medium-range global weather forecasting | Science | 2023-11 | Peer-reviewed publication of GraphCast by Google DeepMind, demonstrating superior accuracy on 90% of verification targets versus ECMWF HRES and establishing the graph neural network approach as a benchmark for the field. |
| t2 | GenCast: Diffusion-based ensemble forecasting for medium-range weather | Nature / arXiv | 2024-12 | Introduces GenCast, a diffusion-model ensemble forecaster that outperforms ECMWF ENS on 97.2% of verification targets, directly addressing probabilistic forecasting and ensemble spread without MSE-driven blurring. |
| t3 | A foundation model for the Earth system | Nature | 2025-05 | Peer-reviewed publication of Microsoft Aurora, a 1.3-billion-parameter model trained on over one million hours of geophysical data, covering weather, air quality, ocean waves, and cyclone tracks in a single foundation-model framework. |
| t4 | Neural general circulation models for weather and climate | Nature | 2024-08 | Nature paper on NeuralGCM, the Google hybrid model that embeds ML parameterisations inside a classical atmospheric dynamical core, bridging medium-range weather forecasting and multi-decade climate simulation. |
| t5 | AIFS Single 1.1.0: an update to ECMWF's machine-learned weather forecast model AIFS | Geoscientific Model Development (Copernicus) | 2026-05 | Peer-reviewed technical description of ECMWF's operational AIFS, including physical consistency constraints, updated training schedule, and verification against IFS on the full operational forecast portfolio. |
| t6 | ECMWF's AI forecasts become operational | ECMWF | 2025-02 | Official ECMWF announcement that AIFS moved to operational status on 25 February 2025, running alongside the physics-based IFS with tropical cyclone track gains of up to 20% and ~1,000-fold energy reduction. |
| t7 | AIFS Machine Learning data | ECMWF | 2026-05 | ECMWF data page documenting the operational timeline of AIFS Single and AIFS ENS versions, including the May 2026 v2 upgrade, providing primary source evidence for operational deployment status. |
| t8 | Another year of rapid development in machine-learning-based forecasting | ECMWF Annual Report | 2024-12 | Official ECMWF review documenting 2024 upgrades to AIFS: resolution increase from 1 degree to 0.25 degree, forecast frequency from 2 to 4 times daily, and lead time extension to 15 days. |
| t9 | AIFS: a new ECMWF forecasting system | ECMWF Newsletter | 2024-01 | Describes AIFS architecture (GNN encoder-decoder with sliding window transformer processor), training data, and context of ECMWF's decision to build an in-house ML model alongside proprietary physics-based IFS. |
| t10 | AIFS -- ECMWF's data-driven forecasting system | arXiv | 2024-06 | Primary technical paper by the ECMWF team on AIFS architecture, training methodology on ERA5 and operational IFS analyses, and initial verification results versus IFS. |
| t11 | An update to ECMWF's machine-learned weather forecast model AIFS | arXiv / Geoscientific Model Development | 2025-09 | Describes AIFS 1.1.0 operational release, the introduction of physical consistency bounding layers, expanded variable set, and the Anemoi open framework now adopted by multiple ECMWF member states. |
| t12 | GraphCast: AI model for faster and more accurate global weather forecasting | Google DeepMind | 2023-11 | Official Google DeepMind blog post accompanying the Science paper, providing public documentation of GraphCast's architecture, ERA5 training, Hurricane Lee forecasting case study, and ECMWF collaboration. |
| t13 | Accurate medium-range global weather forecasting with 3D neural networks | Nature | 2023-09 | Peer-reviewed publication of Pangu-Weather by Huawei, the first AI model demonstrated to outperform ECMWF IFS on all tested weather variables, establishing the 3D Earth-Specific Transformer as a competitive architecture. |
| t14 | FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators | arXiv | 2022-02 | NVIDIA paper introducing FourCastNet, the first deep-learning global weather model to reach 0.25-degree resolution and comparable short-range skill to IFS, using Adaptive Fourier Neural Operators on ERA5 training data. |
| t15 | FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators | arXiv / PASC Conference | 2022-08 | Extended NVIDIA FourCastNet conference paper demonstrating five orders of magnitude inference speedup versus NWP and enabling large ensemble generation for extreme event probability estimation. |
| t16 | Introducing Aurora: The first large-scale foundation model of the atmosphere | Microsoft Research | 2024-06 | Microsoft Research announcement of Aurora as a 1.3-billion-parameter foundation model using a 3D Swin Transformer with Perceiver encoders, establishing the pretrain-then-fine-tune paradigm for atmospheric prediction. |
| t17 | NOAA deploys new generation of AI-driven global weather models | NOAA | 2025-12 | Official NOAA press release announcing operational deployment of AIGFS, AIGEFS, and HGEFS in December 2025, with AIGFS completing a 16-day forecast using 0.3% of GFS computing resources. |
| t18 | NOAA says its new AI-driven weather models improve forecast speed and accuracy | CBS News | 2025-12 | Reports NOAA estimates of 91-99% compute reduction for AI models and a 18-24 hour extension of effective forecast skill from AIGEFS versus the traditional GEFS, with NOAA confirmation that AI models augment rather than replace physics-based systems. |
| t19 | Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect? | Geophysical Research Letters | 2023-10 | Peer-reviewed study by Selz and Craig demonstrating that current AI models cannot simulate the butterfly effect, with error growth differing structurally from physics-based models, challenging claims of unlimited atmospheric predictability. |
| t20 | Testing the Limit of Atmospheric Predictability with a Machine Learning Weather Model | arXiv | 2025-04 | Recent study using ML backpropagation to probe the Lorenz predictability horizon, examining whether data-driven models can approach or extend the classical two-week deterministic forecast limit. |
| t21 | Generative AI to quantify uncertainty in weather forecasting | Google Research | 2024-03 | Google Research explanation of how generative diffusion models address the Lorenz chaos problem by sampling ensemble members from a learned probability distribution rather than propagating a single deterministic state. |
| t22 | Evaluation of five global AI models for predicting weather in Eastern Asia and Western Pacific | npj Climate and Atmospheric Science | 2024-09 | Independent homogeneous evaluation of Pangu-Weather, FourCastNet v2, GraphCast, FuXi, and FengWu on identical ERA5 initial conditions, including typhoon track verification, providing third-party skill comparisons across architectures. |
| t23 | NeuralGCM harnesses AI to better simulate long-range global precipitation | Google Research | 2026-01 | Google follow-up on NeuralGCM progress in precipitation simulation, the most difficult variable for global-scale models, illustrating how hybrid physics-ML approaches address out-of-distribution and tail-event forecasting challenges. |
| t24 | Machine learning for numerical weather prediction | ECMWF | 2025 | ECMWF internal review of ML history at the centre, covering the 2021 roadmap, AIFS ensemble development (AIFS ENS 1 implemented July 2025), and CRPS-optimised training versus MSE-trained deterministic models. |
| t25 | Deep Learning and Foundation Models for Weather Prediction: A Survey | arXiv | 2025-01 | Comprehensive taxonomy of transformer, GNN, physics-AI, and domain-specific weather models, situating all major architectures (FourCastNet, GraphCast, AIFS, NeuralGCM, Aurora, ClimaX) within a unified framework for comparative analysis. |
Academic & arXiv
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| a1 | Learning skillful medium-range global weather forecasting | Science | 2023-12 | GraphCast (Lam et al., DeepMind) outperforms ECMWF HRES on 90% of 1,380 test metrics using a multi-scale GNN trained on 221 ERA5 variables, marking the first definitive ML victory over operational NWP. |
| a2 | Accurate medium-range global weather forecasting with 3D neural networks | Nature | 2023-07 | Pangu-Weather (Bi et al., Huawei) introduces 3D Earth-Specific Transformers and a multi-timescale inference strategy that surpasses HRES on several variables, published in Nature with open weights. |
| a3 | GenCast: Diffusion-based ensemble forecasting for medium-range weather | arXiv / Nature (2024) | 2023-12 | Price et al. introduce a graph-based diffusion ensemble model that outperforms ECMWF ENS on 15-day probabilistic metrics, establishing generative ML as a credible route to ensemble forecasting. |
| a4 | Neural general circulation models for weather and climate | Nature | 2024-07 | Kochkov et al. (Google) fuse a differentiable dynamical core with learned ML physics, producing a hybrid model that matches GraphCast on medium-range skill and reproduces realistic climate statistics over decades. |
| a5 | A Foundation Model for the Earth System | arXiv / Nature (2025) | 2024-05 | Bodnar et al. (Microsoft) train Aurora on over one million hours of diverse geophysical data; it outperforms operational forecasts for air quality, ocean waves, tropical cyclone tracks, and high-resolution weather at much lower compute. |
| a6 | AIFS -- ECMWF's data-driven forecasting system | arXiv | 2024-06 | Lang et al. describe ECMWF's own GNN-transformer system, which went operational in February 2025, marking the first major centre to operationalise a purely data-driven global weather forecast. |
| a7 | WeatherBench 2: A benchmark for the next generation of data-driven global weather models | arXiv / JAMES (2024) | 2023-08 | Rasp et al. provide the community's standard independent evaluation framework for ML weather models, with ERA5 at 0.25-degree resolution as ground truth and a continuously updated live leaderboard. |
| a8 | WeatherBench: A Benchmark Data Set for Data-Driven Weather Forecasting | Journal of Advances in Modeling Earth Systems | 2020-11 | Foundational benchmark paper by Rasp et al. that established ERA5-based evaluation of data-driven global weather models and catalysed rapid progress by providing common baselines. |
| a9 | FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators | arXiv | 2022-02 | Pathak et al. (NVIDIA/Berkeley Lab) introduce Adaptive Fourier Neural Operators applied to 0.25-degree global weather prediction, the first model to achieve HRES-competitive skill at that resolution. |
| a10 | Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect? | Geophysical Research Letters | 2023-10 | Selz and Craig demonstrate that current deterministic ML models fail to reproduce rapid upscale error growth from small perturbations, meaning they sidestep rather than respect Lorenz's intrinsic predictability limit. |
| a11 | Atmospheric Predictability Beyond 30 Days with Machine Learning | arXiv | 2025-04 | Tests whether ML models can push deterministic predictability beyond Lorenz's two-week estimate, revisiting the theoretical limit in the context of modern data-driven approaches. |
| a12 | Numerical models outperform AI weather forecasts of record-breaking extremes | arXiv | 2025-08 | Zhang, Fischer et al. show empirically that ECMWF HRES consistently outperforms GraphCast, Pangu-Weather, and FuXi on record-breaking heat, cold, and wind events, quantifying the tail-event failure mode of current ML models. |
| a13 | Robustness of AI-based weather forecasts in a changing climate | arXiv | 2024-09 | Tests AIFS, GraphCast, and Pangu-Weather on pre-industrial, present-day, and +2.9 K future climates, finding skillful short-range forecasts but systematic cold biases in warmer states that expose out-of-distribution fragility. |
| a14 | ClimaX: A foundation model for weather and climate | arXiv / ICML (2023) | 2023-01 | Nguyen et al. present the first weather-climate foundation model using transformer pretraining on CMIP6 datasets, demonstrating generalisation to tasks unseen during pretraining including climate projections. |
| a15 | Evaluation of five global AI models for predicting weather in Eastern Asia and Western Pacific | npj Climate and Atmospheric Science | 2024-09 | Independent regional evaluation of Pangu-Weather, FourCastNet v2, GraphCast, FuXi, and FengWu against ERA5 for typhoon track and intensity, providing geographically focused accuracy assessment beyond global averages. |
| a16 | Comparing and Contrasting Deep Learning Weather Prediction Backbones on Navier-Stokes and Atmospheric Dynamics | arXiv | 2024-07 | Systematic controlled comparison of GNN, transformer, and Fourier-operator architectures on both synthetic Navier-Stokes and real ERA5 data, isolating the effect of architecture from training choices. |
| a17 | An update to ECMWF's machine-learned weather forecast model AIFS | arXiv | 2025-09 | Describes AIFS 1.1.0, the operational version deployed August 2025, documenting incremental skill improvements, new variables, and the correction of a precipitation forecast issue in the initial release. |
| a18 | Neural general circulation models for modeling precipitation | Science Advances | 2026-01 | Yuval, Kochkov et al. extend the NeuralGCM framework by training directly on satellite-based precipitation observations, demonstrating improved simulation of extremes and the diurnal cycle over existing GCMs. |
| a19 | Data-driven ensemble forecasting with the AIFS | ECMWF Newsletter | 2024-10 | Describes ECMWF's two parallel approaches to ML ensemble forecasting (diffusion-based and CRPS-trained AIFS), both using the same encoder-GNN-transformer architecture as the deterministic system. |
| a20 | FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale | arXiv | 2025-07 | Latest generation of the NVIDIA FourCastNet series applies geometric deep learning to probabilistic global forecasting at scale, extending the Fourier-operator lineage into ensemble territory. |
| a21 | WeatherBench 2: A Benchmark for the Next Generation of Data-Driven Global Weather Models (JAMES published version) | Journal of Advances in Modeling Earth Systems | 2024-06 | Published final version of the WeatherBench 2 framework, providing metrics, baselines, and methodology used by all major ML weather model evaluations as of 2024-2026. |
| a22 | Butterfly Effects and Finite Predictability in AI-Based Weather Prediction | ESS Open Archive | 2025-07 | Revisits Selz and Craig's butterfly-effect findings using a taxonomy of three types of butterfly effect, sharpening the understanding of which predictability limits current AI models do and do not respect. |
| a23 | A Practical Probabilistic Benchmark for AI Weather Models | arXiv | 2024-01 | Brenowitz et al. propose CRPS-based probabilistic evaluation specifically designed for ML weather models, filling a gap in WeatherBench 2 which focused on deterministic metrics. |
| a24 | SAFE: A Novel Approach to AI Weather Evaluation through Stratified Assessments of Forecasts over Earth | arXiv | 2025-10 | Proposes geographically stratified evaluation of AI weather models to surface regional performance differences masked by global-average RMSE, exposing heterogeneous skill across latitudes and terrain. |
| a25 | Do machine learning climate models work in changing climate dynamics? | arXiv | 2025-09 | Systematic OOD evaluation of state-of-the-art ML climate models under distribution-shifted scenarios, documenting where models fail to generalise and informing credible near-term directions for climate-scale ML. |
VC & Analyst Reports
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| v1 | Global AI-Based Weather Modelling Market Size 2025–2034 | Custom Market Insights | 2026-04 | Provides one of several competing market-size estimates for AI weather modelling, with sector breakdown and driver analysis useful for triangulating the true commercial scale of the opportunity. |
| v2 | AI in Weather Prediction Market Report 2030: Industry Insights | Knowledge Sourcing Intelligence | 2025-09 | Sizes the AI weather prediction market at USD 608 million in 2025 growing to USD 891 million by 2030 at 7.92 percent CAGR, offering a conservative anchor against more bullish projections. |
| v3 | AI-Based Weather Modelling Market Size Report by 2033 | Transpire Insight | 2026-02 | Projects the market reaching USD 7.2 billion by 2033 at a 26.4 percent CAGR, the most bullish mainstream estimate, and identifies hybrid physics-AI models as the highest-growth model segment. |
| v4 | AI-Based Climate Modelling Market Size, Forecasts 2025–2034 | GM Insights | 2024-12 | Sets the 2024 baseline at USD 266 million with a 23.1 percent CAGR, notes weather forecasting held over 45 percent of segment revenue, and names Google, AccuWeather, and Microsoft as holding over 31 percent combined market share. |
| v5 | [AI-Based Climate Modelling Market | Industry Report, 2033](https://www.grandviewresearch.com/industry-analysis/ai-based-climate-modelling-market-report) | Grand View Research | 2025-01 |
| v6 | [AI-Based Climate Modelling Market Size | CAGR of 24%](https://market.us/report/ai-based-climate-modelling-market/) | Market.us | 2025-01 |
| v7 | Climate resilience technology: An inflection point for new investment | McKinsey & Company | 2025-09 | McKinsey's primary strategic framing of the climate resilience investment opportunity, estimating a USD 1 trillion private-capital opportunity by 2030 and citing 27 US billion-dollar weather disasters in 2024 as a demand driver. |
| v8 | State of Climate Tech 2024 | PwC | 2024-12 | PwC's fifth annual climate tech investment survey, tracking 12,000+ startups and 52,000 deals, finding AI-centred climate ventures raised USD 6 billion in Q1-Q3 2024, doubling their share of total climate tech funding from 7.5 percent to 14.6 percent. |
| v9 | Climate tech's future may be AI-powered | PwC | 2025-01 | C-suite companion piece to the 2024 State of Climate Tech report, explicitly positioning AI-driven weather modelling and climate adaptation as the growth pocket within an otherwise contracting market. |
| v10 | WindBorne Raises $15 Million to Scale Its Balloon Constellation and Bring AI Weather Modeling to the Fight Against Climate Change | Business Wire / WindBorne Systems | 2024-06 | Announces WindBorne's USD 15 million Series A led by Khosla Ventures, the most significant dedicated VC round in AI weather data infrastructure in 2024, and describes WeatherMesh's benchmark claim against GraphCast. |
| v11 | This AI weather startup is out-forecasting government agencies | TechCrunch | 2026-06 | June 2026 profile of WindBorne's sixth-generation WeatherMesh model, reporting claims of exceeding ECMWF IFS accuracy on surface temperature and hourly rather than six-hourly update frequency, with 3 km resolution in the continental US. |
| v12 | Weather startup WindBorne Systems raised $15M led by Khosla | Axios Pro: Climate Deals | 2024-05 | Axios Pro's deal-focused coverage provides investor-level framing of WindBorne's strategic positioning at the intersection of low-cost hardware, ubiquitous communication networks, and AI forecasting demand. |
| v13 | Brightband sees a bright (and open source) future for AI-powered weather forecasting | TechCrunch | 2024-09 | Covers Brightband's emergence from stealth and USD 10 million Series A, framing the startup's bet on raw observational data and open-source methodology as a differentiated approach to commercialising AI weather forecasting. |
| v14 | A new startup is targeting a gap in AI weather forecasting | Latitude Media | 2025-03 | Energy-sector specialist coverage of Brightband's strategy of using raw (unprocessed) observational data to extend reliable forecasts beyond seven to ten days, targeting renewable energy operators as primary customers. |
| v15 | Why we invested: Jua - Redefining Weather Intelligence for the Energy Transition | Ananda Impact Ventures | 2025-06 | Investor thesis memo explaining the EUR 10 million Series A co-investment in Swiss weather-AI startup Jua, articulating the energy-transition demand case for better probabilistic weather forecasting for wind and solar operators. |
| v16 | Jua raises $16M to build a foundational AI model for the natural world, starting with the weather | TechCrunch | 2024-02 | Documents Jua's USD 16 million 2024 raise to build a physics foundation model claiming 20x the parameter count of GraphCast, placing this within the broader foundation-model investment wave and articulating risks around reliability and consistency. |
| v17 | Climate resilience technology: Capturing value in a $1T market | McKinsey & Company | 2025-09 | Quantifies the climate resilience tech opportunity at USD 1 trillion for private capital by 2030 and documents specific VC and PE deal activity in weather-event simulation, including General Atlantic's investment in Technosylva. |
| v18 | Gartner Hype Cycle Identifies Top AI Innovations in 2025 | Gartner | 2025-08 | Confirms that Gartner's 2025 AI Hype Cycle does not specifically name weather-AI models, instead prioritising AI agents and AI-ready data, signalling the technology remains below the enterprise-software radar for formal placement. |
| v19 | Climate centre stage in Gartner's Hype Cycle report | Sustainability Magazine / Gartner | 2024-11 | Covers Gartner's 2024 Hype Cycle for Environmental Sustainability, noting AI-driven sustainability initiatives as a portfolio element without placing AI weather-specific models, useful for mapping the gap in analyst coverage. |
| v20 | AI and Climate Tech Startups Dominated 2025's Largest Funding Rounds | Entrepreneur Loop | 2025-12 | Aggregates CB Insights and Crunchbase data showing AI startups received over USD 160 billion in the first three quarters of 2025 and AI captured 31 percent of all VC investment, providing the macro context in which weather-AI sits as a micro-niche. |
| v21 | AI Weather Forecasting 2026: Models, Accuracy & Results | ArticlesEdge | 2026-05 | Practitioner synthesis noting NOAA committed USD 50 million over five years to AI weather research and that running ECMWF HRES requires hundreds of millions in supercomputing versus a few dollars per hour on cloud GPU for ML equivalents. |
| v22 | Green and intelligent: the role of AI in the climate transition | npj Climate Action (Nature Portfolio) | 2025-06 | Peer-reviewed review contextualising AI weather and climate modelling within the broader climate-action investment landscape, noting IceNet outperforming ECMWF SEAS5 on sea-ice forecasting and DeepMind's 20 percent wind-energy value improvement. |
| v23 | [Mapping climate hazards: Advancing adaptation | McKinsey Global Institute](https://www.mckinsey.com/mgi/our-research/advancing-adaptation-how-evolving-hazards-could-shape-the-agenda) | McKinsey Global Institute | 2025-12 |
| v24 | AI-Powered Weather Forecasting: The New Frontier of Climate Intelligence | University of Chicago Sustainability Dialogue | 2025-11 | Articulates three strategic imperatives for institutions engaging with AI weather forecasting, positioning accurate forecasting as a national security and TCFD compliance issue, reflecting how enterprise buyers are framing procurement decisions. |
| v25 | State of Climate Tech 2025 | Net Zero Insights / State of Climate Tech | 2025-12 | 2025 annual survey concluding that AI has become foundational infrastructure rather than a supporting tool in climate tech, with 'adaptation' and 'resilience' now dominating investor discourse over 'environmental impact'. |
Blogs & Independent Thinkers
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| b1 | [AI Weather Hub | Karolina Stanisławska | Substack](https://aiweatherhub.substack.com/) | AI Weather Hub (Substack) |
| b2 | GenCast – AI meets ensemble forecasting | AI Weather Hub (Substack) | 2025-01 | Explains GenCast's conditional diffusion ensemble architecture and its relationship to the butterfly effect, cross-referencing the Nature paper with accessible independent analysis. |
| b3 | Why AI Weather? – AI Weather Hub | AI Weather Hub (Substack) | 2024-10 | Sets out the independent case for why ML weather forecasting is a distinct paradigm shift from statistical post-processing, contextualising it against the ChatGPT moment in AI. |
| b4 | AI and the future of weather forecasting – Phil Siarri | Philaverse (Substack) | 2025-07 | Named-author survey of operational deployment status at ECMWF, Met Office, NOAA, and Asian agencies, with model-by-model architecture summaries and citation of primary literature. |
| b5 | A Breakthrough Year for AI in Weather Forecasting: Insights and Opportunities | CLAI Ventures (Substack) | 2025-01 | Investor-practitioner analysis covering NeuralGCM and GenCast with specific benchmark figures, and noting the resolution gap between GenCast at 0.25° and ENS at 0.1°. |
| b6 | Exploring GraphCast – Open-Meteo | Open-Meteo (Substack) | 2024-04 | Practitioner API developer commentary on the ERA5-training versus GFS-initialisation mismatch in production GraphCast, and on NOAA's retraining effort to resolve it. |
| b7 | AI for weather forecasting – GeoAI Unpacked #2 | GeoAI Unpacked (Substack) | 2024-10 | Independent analysis highlighting the blurriness problem in AIFS versus IFS outputs, legacy Fortran infrastructure at national weather services, and the tension between accuracy metrics and visual realism. |
| b8 | So, which weather forecast is the best? – ActuallyWeather | ActuallyWeather (Substack) | 2025-12 | Independent real-world verification of ECMWF, GFS, HRRR, and GraphCast using location-specific skill scores, providing evidence outside vendor-reported benchmarks. |
| b9 | AI Weather Forecasting 2026: Models, Accuracy & Results | ArticlEdge | 2026-05 | Comprehensive independent synthesis citing primary benchmarks and noting the distribution-shift risk as AI models trained on 1979–2017 ERA5 are deployed into a systematically warmer 2026 atmosphere. |
| b10 | AIFS: a new ECMWF forecasting system | ECMWF Newsletter | 2024-01 | Primary institutional source documenting ECMWF's architectural choice of GNNs for AIFS, its ERA5 training regime, and its explicit comparison to GraphCast and Pangu-Weather. |
| b11 | Accuracy versus activity – ECMWF AIFS Blog | ECMWF | 2024-12 | Operational scorecard from ECMWF comparing AIFS, GraphCast, Pangu-Weather, and Aurora on RMSE and forecast activity metrics, including Aurora's instability beyond day 7. |
| b12 | Anemoi: a new framework for weather forecasting based on machine learning | ECMWF | 2024-10 | Documents ECMWF's open-source Anemoi framework enabling national meteorological services to build regional ML models on the same architecture as AIFS. |
| b13 | GraphCast: AI model for faster and more accurate global weather forecasting | Google DeepMind Blog | 2023-11 | Primary vendor announcement for GraphCast containing the headline claim of 90% superiority over HRES across 1,380 targets, and the Hurricane Lee nine-day landfall prediction case study. |
| b14 | GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy | Google DeepMind Blog | 2024-12 | Primary vendor source for GenCast's diffusion-based ensemble design and the 97.2% superiority over ECMWF ENS claim, providing the vendor-side data that independent commentary cross-references. |
| b15 | Probabilistic weather forecasting with machine learning (GenCast) | Nature | 2024-12 | Peer-reviewed primary source for GenCast, confirming that prior ML ensemble methods failed on blurring and that GenCast is the first to outperform ENS at 0.25° resolution. |
| b16 | On Some Limitations of Current Machine Learning Weather Prediction Models | Geophysical Research Letters | 2024-06 | Massimo Bonavita's widely cited analysis documenting that ML forecast energy spectra differ from NWP and reanalysis, producing blurriness that standard RMSE metrics do not penalise. |
| b17 | Improving AI Weather Prediction Models Using Global Mass and Energy Conservation Schemes | Journal of Advances in Modeling Earth Systems | 2025-11 | Demonstrates that adding conservation-law constraints to FuXi reduces forecast error and corrects the drizzle bias, directly addressing the physical-consistency critique. |
| b18 | An update to ECMWF's machine-learned weather forecast model AIFS | arXiv / ECMWF | 2025-09 | Documents the August 2025 AIFS v1.1.0 update incorporating output bounding layers to prevent physically implausible outputs such as negative precipitation. |
| b19 | Evaluation of five global AI models for predicting weather in Eastern Asia and Western Pacific | npj Climate and Atmospheric Science | 2024-09 | Independent homogeneous comparison of five ML models under identical ERA5 initial conditions, finding FengWu leads for typhoon track prediction and that a multi-model ensemble rivals the best single model. |
| b20 | A fast physics-based perturbation generator of machine learning weather model for efficient ensemble forecasts of tropical cyclone track | npj Climate and Atmospheric Science | 2025-03 | Finds that FuXi attenuates small-perturbation growth compared to IFS, suggesting AI weather models may understate the butterfly effect rather than overcome it. |
| b21 | An Observations-focused assessment of Global AI Weather Prediction Models During the South Asian Monsoon | arXiv | 2025-09 | Evaluates seven AI models against 458 weather stations and satellite data, finding AIFS leads on most metrics but all models show substantially higher errors against ground observations than against reanalysis. |
| b22 | Validating Deep Learning Weather Forecast Models on Recent High-Impact Extreme Events | Artificial Intelligence for the Earth Systems (AMS) | 2025-01 | Case studies on the 2021 Pacific Northwest heatwave and 2021 winter storm find that ML models match HRES locally but underperform aggregated, and lack variables needed for humid heatwave health-risk assessment. |
| b23 | Physics-based models outperform AI weather forecasts of record-breaking extremes | Science Advances | 2026-04 | The most direct counter-evidence to headline ML accuracy claims, showing physics-based NWP retains an advantage for truly unprecedented extreme events outside training distribution. |
| b24 | AI-Driven Weather Forecasts to Accelerate Climate Change Attribution of Heatwaves | Earth's Future (AGU) | 2025-08 | Demonstrates a new application of AI weather models for near-real-time attribution of heatwaves to anthropogenic climate change, showing NeuralGCM's hybrid physics advantage for SST-dependent events. |
| b25 | Weather forecasting in a changing climate: the rise of AI and Machine learning? | ScienceDirect (journal article) | 2026-05 | Most recent practitioner review, documenting AIFS superiority for tropical cyclone track prediction and the open Anemoi framework, with a frank assessment of remaining resolution and coupling gaps. |