PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial intelligence pathways from weather to climate

Tom Beucler, J. David Neelin, Hui Su, Shivanshi Asthana, Chris Bretherton, Will Chapman, Costa Christopoulos, Spencer K. Clark, Aditya Grover, Ignacio Lopez-Gomez, Tapio Schneider, Adam Subel, Oliver Watt-Meyer

arXiv:2610.09770v1physics.ao-phcs.AIcs.LG

TL;DR

Climate projection asks how Earth-system statistics respond to altered conditions, including forcings that may be unobserved, rather than forecasting specific weather decades ahead. The paper reviews AI forecasting across timescales and identifies requirements for credible climate responses. It reports that AI can reduce time-to-solution by advancing fewer variables and timesteps, while emphasizing unresolved limits in testing against unseen climates.

  • Problem

    Climate projection must estimate responses to altered, sometimes unobserved forcings, and long rollouts alone do not ensure reliable forced responses or climate statistics.

    Source: §1 · p. 2, §3 · p. 7, §4.1 · p. 9

  • Method

    The paper reviews AI roles and models across weather and climate timescales, setting requirements for independently manipulable forcing agents and stress tests in out-of-distribution regimes.

    Source: §1 · p. 2, §4.1 · pp. 9–10

  • Results

    AI reduces time-to-solution relative to optimized GPU-ported models by computing fewer timesteps, prognostic variables, and levels.

    Source: §7.1 · p. 22

  • Takeaways & Limitations

    The paper points toward transparent combinations of learned and physical components, with models specializing by regime, process, and task.

    Source: §7.1 · p. 22

    The historical record contains one realization of the slow forced response with covarying forcing agents, and tests against held-out simulations establish fidelity to the reference model, not to Earth.

    Source: §4.1 · p. 10

Abstract

from arXiv · show

Deep learning has made rapid advances in weather forecasting: autoregressive models trained on atmospheric reanalyses now rival dynamical models across nowcasting, medium-range, and subseasonal-to-seasonal lead times, producing well-calibrated ensemble forecasts at reduced cost. We review these advances and consider their extension to climate horizons, where the challenge shifts from initial-condition skill to producing reliable statistical responses under altered forcings. AI-powered climate prediction systems must produce credible forced responses to drivers (e.g., greenhouse gases, land-use change) typically outside the observed record. We propose two minimum requirements for AI in climate modeling: (i) external forcing agents must enter explicitly enough to support interventions in which they vary independently; and (ii) robustness must be stress-tested in out-of-distribution regimes, including extremes and counterfactual trajectories. Using leading AI autoregressive emulators and hybrid physics-AI models, we identify development and coupling challenges. Comparing the reported throughput of these models with that of GPU-ported dynamical models highlights how AI can reduce time-to-solution by advancing only the target variables at the required resolution and using longer time steps, rather than integrating a full high-frequency, multivariate state. Diverse AI downscaling strategies can partially substitute for explicit fine-scale resolution, paving the way toward inexpensive local hazard assessment across prediction horizons.

1 Introduction

AI weather forecasting has advanced from nowcasting and medium-range prediction to subseasonal lead times, while remaining reliant on dynamical-model data and infrastructure. The paper examines the distinct challenge of using AI to estimate climate responses under altered, sometimes unobserved conditions.

1 Introduction

  • Data-driven models now rival dynamical models at subseasonal lead times of about one month on probabilistic scores, at a fraction of their computational cost [5].Earlier advances included nowcasting and medium-range forecasting.Source: §1 · p. 1
  • AI weather models remain dependent on dynamical modeling for reanalyses, initial conditions, and longer-timescale forecast inputs.Observation-driven efforts also rely on products that embed physical modeling.Source: §1 · pp. 1–2
  • Climate projection estimates how Earth-system statistics change under altered boundary conditions, including greenhouse gases, aerosols, and land use, rather than predicting specific weather decades ahead.Many such conditions are unobserved or counterfactual, making forced responses difficult to learn from observations alone.Source: §1 · p. 2
  • The paper traces AI’s distinct roles as a weather predictor, a climate-response model component, and a downscaling tool, asking how to represent rare events and unseen forcings credibly.It examines event-scale structure, changing predictability across timescales, and climate-modeling and evaluation choices, including design choices and failure modes.Source: §1 · p. 2

2 Data-driven forecasting with AI

The section surveys AI forecasting as a set of choices about forecast objects, prediction methods, and scale. It covers deterministic and probabilistic forecasting, data assimilation and end-to-end systems, regional refinement, and foundation models.

2 Data-driven forecasting with AI

  • AI forecasting frames future states or derived observables as conditional prediction objects, whose choice shapes losses, sampling, rollouts, and evaluation.The section spans deterministic predictions, ensembles, local-observable distributions, and forecast sensitivities.Source: §2 · p. 2

2.1 Deterministic AI forecasting

  • Deterministic autoregressive models iterate learned state increments to predict future states, but central-tendency losses and stabilizing choices can smooth extremes and grid-scale structure.Models can nevertheless achieve stable, skillful global forecasts across architectures, with forecast stride and multi-step training affecting rollout behavior.Source: §2.1 · pp. 2–3

2.2 Probabilistic AI forecasting

  • Probabilistic systems sample plausible trajectories from a conditional distribution, with ensemble members sharing an initial state and differing through random inputs or generators.The resulting ensemble spread estimates forecast uncertainty.Source: §2.2 · p. 3
  • Direct CRPS-trained ensembles offer comparatively cheap sampling but may leave multivariate dependence uncalibrated, while iterative samplers can yield coherent samples at longer inference times.FourCastNet3 illustrates that fast direct inference can also retain spatially correlated, spectrally faithful structure; reanalysis-trained models may inherit a muted butterfly effect.Source: §2.2 · pp. 3–4

2.3 From improved data assimilation to end-to-end forecasting

  • AI can modify assimilation through learned priors or latent-space estimation, or learn directly from observations; differentiable forecasters also provide gradients for sensitivity studies.Reported gradients recover physically meaningful sensitivity structures for extreme events with computational savings relative to traditional adjoint workflows.Source: §2.3 · p. 4

2.4 Zooming in: Regional modeling and forecast refinement

  • Regional models evolve high-resolution dynamics, whereas forecast refinement conditions local or process-specific estimates on coarse forecasts; each strategy has distinct constraints.For tropical cyclones, post-processing recovered much of the maximum-wind and minimum-pressure forecasting skill from the coarse forecast environment, while observation-based downscaling produced finer wind fields.Source: §2.4 · pp. 4–5

2.5 Foundation models

  • Foundation models seek transferable representations through broad pretraining, but added data can hurt misaligned targets and does not guarantee tail behavior, physical consistency, or robustness.Process-specific heads or post-processing may still be needed for practical targets such as evapotranspiration or precipitation.Source: §2.5 · p. 5

3 Data-informed subseasonal-to-seasonal forecasting

S2S forecasting bridges weather and climate: as atmospheric initial-condition predictability fades, forecasts must represent probabilistic outcomes and the climate statistics their rollouts approach. The section reviews S2S benchmarks and two development routes: extending autoregressive weather AI and using S2S tests to improve hybrid climate models.

3 Data-informed subseasonal-to-seasonal forecasting

  • S2S forecasts are necessarily probabilistic as deterministic atmospheric predictability fades, and their long rollouts test whether models approach appropriate climate statistics.Slowly varying ocean and sea-ice states can retain predictive information, but that information also decays; climate statistics may be nonstationary under external forcing.Source: §3 · pp. 5, 7

3.1 Towards objective benchmarking at S2S scales

  • Weather-trained models can lose skill rapidly at S2S leads, often falling below climatology after about 10 days, while newer benchmarks test against stronger baselines.The 2021 WMO challenge compared probabilistic temperature and precipitation forecasts with climatology and a recalibrated ECMWF benchmark; the 2025 AI Weather Quest evaluates weekly probabilistic forecasts, where post-processing and hybrid methods have performed strongly.Source: §3.1 · p. 7

3.2 From weather to S2S: The FuXi-S2S approach

  • FuXi-S2S extends an autoregressive weather strategy to global daily forecasts out to six weeks and adds learned latent-space perturbations to generate ensembles.The perturbations are flow-dependent and sampled during inference, unlike conventional ensembles that perturb initial states; the method is associated with improved probabilistic skill relative to ECMWF S2S for total precipitation and outgoing longwave radiation.Source: §3.2 · p. 7
  • FuXi-S2S extends skillful MJO prediction from about 30 to 36 days while better maintaining MJO amplitude and phase and capturing realistic teleconnections.The authors connect these MJO improvements to improved global precipitation skill and describe a Pakistan floods case in which the model predicted extreme rainfall anomalies earlier than ECMWF S2S; saliency maps highlighted precursor patterns in the tropical central Pacific and western Indian Ocean.Source: §3.2 · p. 7

3.3 From climate to S2S: Lessons from CliMA-S2S

  • CliMA-S2S combines a coupled land–atmosphere model with postprocessing, using ERA5-based calibration to estimate uncertainty from a single deterministic trajectory.Its week-3 probabilistic forecasts capture large-scale geopotential-height and 2-m-temperature structures and show positive skill relative to climatology; climate development also uses process-specific observational diagnostics.Source: §3.3 · p. 8, Figure 2 · p. 8
  • S2S development exposed stratospheric, sea-ice-initialization, and gravity-wave biases, and many resulting choices were retained in later coupled CliMA configurations.Using only a constant freezing condition instead of a two-dimensional sea-ice state produced polar temperature and pressure biases persisting beyond the first forecast week; gravity-wave treatment affected surface pressure and upper-level momentum within a week.Source: §3.3 · pp. 8–9

4.1 What must an Earth system model represent?

Climate modeling requires reliable responses to unseen, time-dependent forcings, not merely sustained long rollouts. Credible approaches must represent independently variable forcing agents and account for processes and feedbacks that shape climate responses.

4.1 What must an Earth system model represent?

  • Climate projections must estimate spatially resolved fields and their responses to changing greenhouse gases, aerosols, and carbon-cycle feedbacks, from which ECS, TCR, and TCRE can be derived.The central challenge is reliable response to unseen, time-dependent forcings, including counterfactual scenarios.Source: §4.1 · p. 9
  • Relevant forcing agents must enter as explicit, independently manipulable inputs so their effects remain identifiable when varied separately.Greenhouse gases and aerosols primarily affect climate through radiative heating; in concentration-driven configurations their concentrations can be prescribed, whereas emissions-driven models must predict them through chemistry and biogeochemistry.Source: §4.1 · p. 9, Figure 3 · p. 10
  • Forcing pathways span different timescales: greenhouse gases persistently modify weather and variability, while aerosol–cloud and cloud-feedback processes can evolve on weather timescales.Land- and water-use changes act through slowly evolving boundary conditions and changes in surface properties and fluxes.Source: §4.1 · p. 9, Figure 3 · p. 10
  • Pure AI models learn component or coupled-system updates, whereas hybrid models retain physical equations for selected processes and learn unresolved or uncertain processes from data.Hybrid models commonly learn subgrid-scale parameterizations while prescribing processes such as resolved fluid dynamics and radiative transfer.Source: §4.1 · pp. 9–10
  • Both strategies face scarce observations of forced responses, and controlled OOD tests establish agreement with a reference model rather than with Earth.Most current AI emulators represent fewer processes than mature Earth system models; hybrid replacements offer a near-term route to CMIP7 participation, while coupled emulators need additional components.Source: §4.1 · p. 10

4.2 AI climate-model benchmarking and AIMIP Phase 1

AI climate models can reproduce major historical atmospheric features, but historical performance alone does not establish credibility under unseen boundary conditions or forcings. AIMIP provides a common evaluation framework spanning historical behavior and out-of-distribution responses.

4.2 AI climate-model benchmarking and AIMIP Phase 1

  • Climate model selection must include long-rollout diagnostics and climate statistics because lower short-range forecast error does not guarantee smaller climate bias or a more accurate forced response.The paper notes that most autoregressive AI climate models are optimized with weather-prediction losses over lead times of a few days, whereas climate performance is defined using multi-year or multidecadal statistics.Source: §4.2 · p. 10
  • AIMIP Phase 1 evaluates atmospheric simulations for 1979–2024 under prescribed historical sea-surface temperatures across mean climate, trends, El Niño responses, variability, and OOD generalization.It establishes a common AMIP-style framework, output format, and training constraints for evaluating AI weather and climate models as atmospheric climate models.Source: §4.2 · p. 11
  • Several AI models reproduce major historical atmospheric features and responses to prescribed El Niño variability, but some underestimate historical warming and responses diverge in OOD experiments.The reported OOD experiments include spatially uniform sea-surface-temperature perturbations; the authors conclude that historical AMIP performance is necessary but insufficient for climate-model credibility.Source: §4.2 · p. 11
  • AIMIP and related benchmarks complement rather than replace the CMIP7 experimental hierarchy, and evaluation should include climate, variability, extremes, transient responses, and physical budgets.Depending on model scope, forced-response tests should assess the sign and magnitude of ECS, TCR, and TCRE.Source: §4.2 · p. 11

4.3 Hybrid physics-AI Earth system modeling

Hybrid physics–AI models retain physical structure while learning selected components, and CliMA extends this strategy across the Earth system. Its approach calibrates against climate statistics and quantifies uncertainty in learned parameters and projected statistics.

4.3.1 From hybridizing existing models to constraining AI models

  • Hybrid atmospheric models retain a dynamical core while adding data-driven components, including learned parameterizations and ocean closures.Examples include a differentiable Neural GCM trained against satellite observations and an ICON-A configuration whose learned parameterization improved target-process agreement after joint recalibration.Source: §4.3.1 · p. 11

4.3.2 The CliMA approach

  • CliMA combines conservation-respecting process-based models with data-driven components across atmosphere, oceans, and land, and trains against observed climate statistics.Its physical framework is intended to support out-of-distribution generalizability where training data are unavailable, while GPU-efficient components support ensembles for calibration and uncertainty quantification.Source: §4.3.2 · p. 11
  • CliMA uses derivative-free ensemble Kalman methods to calibrate model components against climate statistics, including for chaotic or nondifferentiable systems.Ensembles smooth noise-induced extrema in finite-time statistical losses, and indirect observations such as cloud cover and liquid water path can constrain processes that are not directly observable.Source: §4.3.2 · pp. 11–12
  • CliMA’s calibrate–emulate–sample framework reuses calibration ensembles to train a misfit emulator, then samples the parameter posterior with MCMC.This reduces model evaluations for posterior sampling by orders of magnitude relative to direct MCMC, enabling uncertainty propagation into climate statistics such as climate sensitivity.Source: §4.3.2 · p. 12
  • CliMA uses physical scaffolding and conservation constraints around learned small-scale processes to distinguish cloud responses to warming from direct greenhouse-gas effects.The approach addresses observational confounding between temperature and greenhouse-gas changes, a distinction needed for scenarios in which those drivers decouple; snow models likewise operate within water-mass and energy guardrails.Source: §4.3.2 · p. 12
  • CliMA’s hybrid design spans atmosphere, land, ocean, and coupling components, with trainable closures constrained by conservation laws to support generalization under altered forcing and counterfactual scenarios.The components include an atmospheric core, ClimaLand, Oceananigans, and ClimaCoupler.Source: §4.3.2 · p. 12

4.4 Autoregressive Earth system emulation

Autoregressive emulators can reproduce reference-model climate behavior efficiently, but credible responses to changing forcings and coupled ocean dynamics remain challenging. The section reviews atmospheric emulation and learned-ocean approaches, including their long-rollout tests and coupling requirements.

4.4.1 AI atmospheric model emulation: Lessons from ACE

  • 130× higher inference throughput and 700× lower energy use per simulated year are reported for ACE2 relative to its C96 SHiELD reference model.The estimates use runtime and rated hardware power.Source: §4.4.1 · p. 12
  • ACE2 reproduces ERA5’s time-mean climate, global trends, SST-forced ENSO variability, and tropical precipitation statistics including MJO-associated eastward propagation.When trained on ERA5, ACE2 also reproduces the global spatial distribution of tropical cyclones.Source: §4.4.1 · p. 13, Figure 4 · p. 13
  • ACE2-SOM reproduces an unseen intermediate 3×CO2 climate, including the extreme-precipitation tail and its warming sensitivity, more accurately than ∼400 km SHiELD.The comparison uses 50-year samples regridded to a common 4° Gaussian grid.Source: §4.4.1 · p. 13, Figure 5 · p. 14
  • AMIP-only and equilibrium-climate training did not ensure correct responses when SST and CO2 varied independently, while ACE2S-SHiELD+ improved separation of their responses on held-out forcing tests.ACE2S-SHiELD+ adds ramped-SST, random-CO2 simulations to its training mix.Source: §4.4.1 · pp. 14–15, Figure 6 · p. 14
  • Short-range training loss and long-range climate error can select different checkpoints, so climate emulators can be evaluated using long rollouts during training.The reported inference error uses eight five-year simulations after each epoch.Source: §4.4.1 · p. 15, Figure 7 · p. 15

4.4.2 AI ocean climate modeling

  • Coupled climate projections require ocean circulation, deep-ocean heat uptake, and carbon storage; explicit component coupling enables separate development, but changing-climate generalization remains untested.A near-term objective is coupled emulation across CMIP DECK experiments and standard emissions scenarios, using limited reference-model data.Source: §4.4.2 · pp. 15–16
  • Samudra learns a 5-day autoregressive prognostic map from OM4 rather than integrating its equations at 15-minute steps, and remains stable in multicentury control-like runs.Its finer-resolution versions recover progressively finer western boundary currents and mesoscale eddies, while deep-ocean variability remains challenging.Source: §4.4.2 · p. 16
  • SamudrACE produces centuries-long simulations at about 1° resolution, reproduces CM4 mean climate and seasonal sea ice, and generates coupled variability absent with prescribed boundaries.Its ENSO amplitude is realistic, but it favors approximately three-year oscillations over CM4’s broader low-frequency spectrum.Source: §4.4.2 · p. 16

5 What makes AI-powered climate models faster?

The section asks when AI reduces climate-simulation time and explains that the reported throughput advantage largely reflects choices about what is computed, at what resolution, and with what update interval.

5 What makes AI-powered climate models faster?

  • The throughput comparison is deliberately simple and restricts dynamical models to GPU-ported configurations because hardware, precision, state, resolution, and task differences impede direct comparisons.The descriptive fit is not a scaling law and mixes mature GPU ports with research AI inference implementations, excluding training and training-data generation costs.Source: §5 · pp. 16, 18
  • SamudrACE reproduces approximate Niño3.4 variability amplitude, ENSO precipitation response, and seasonal sea-ice climatology, but favors approximately three-year ENSO oscillations over CM4.The caption describes five 200-year rollouts and a 40-year hold-out run.Source: Figure 8 · p. 17
  • For full-state models, normalized throughput is expected to increase approximately with grid spacing squared times update interval, while AI can choose longer intervals without resolving every stability substep.Source: §5 · p. 17
  • Accounting for update interval reduces mean absolute model-level offset from 0.81 to 0.24 in log-space, raises weighted R2 from 0.84 to 0.96, and changes the fitted exponent from 3.14 to 1.01.The fitted exponent is with respect to ∆x in panel (a) and (∆x)2∆t in panel (b).Source: §5 · p. 17, Figure 9 · p. 18
  • AI’s reported time-to-solution gains arise largely from computing less: predicting fewer task-relevant variables, using compressed vertical representations or longer time steps, and not explicitly prognosing fine-scale structure.The comparison cannot isolate intrinsic algorithmic efficiency.Source: §5 · p. 18
  • For local hazards, a coarse scenario-dependent model can supply context while downscaling reconstructs the high-resolution variables needed for hazard and impacts modeling.This approach avoids prognosing the full high-frequency, multivariate climate state at high resolution.Source: §5 · p. 18

6 AI-powered downscaling strategies

AI downscaling spans generative refinement, statistical correction, and emulation, with evaluation and robustness tests needed to assess whether these approaches remain credible beyond observed conditions.

6 AI-powered downscaling strategies

  • Downscaling routes include dynamical regional simulation, empirical-statistical mapping, and emulation of fine-scale regional simulations; statistical refinement can also follow partial dynamical downscaling.Observationally trained statistical methods face out-of-distribution limits, while emulation depends on scarce fine-scale simulation targets.Source: §6 · pp. 18–19, Figure 10 · p. 19
  • Because coarse fields do not uniquely determine fine-scale structure, climate super-resolution increasingly models a conditional distribution of high-resolution fields rather than a single deterministic reconstruction.Diffusion and flow-based methods can represent multiple plausible outcomes while recovering fine-scale structure.Source: §6 · pp. 19–20
  • Generative downscaling evaluation should combine probabilistic, multivariate, extreme-value, physical-structure, temporal-coherence, spectral, and transfer diagnostics selected for the application.No single metric is sufficient, and nominal grid spacing alone may not establish that variance exists at the targeted scales.Source: §6 · p. 20
  • Paired super-resolution assumes sample-wise correspondence, whereas unpaired correction matches the model-output distribution to observations conditional on coarse information and auxiliary predictors.This distributional criterion is stricter than matching climatological marginals but more realistic than requiring sample-wise alignment.Source: §6 · p. 20
  • Unpaired AI bias correction remains less established, and its advantage over statistical methods is unclear; added value requires testing joint fidelity, forced trends, compute and data needs, and strong baselines.Statistical correction can better reproduce local marginals, AI can better recover spatial structure, and combining them can outperform either alone.Source: §6 · pp. 20–21
  • Pseudo-reality experiments test robustness by treating high-resolution model output as pseudo-observations and withholding scenarios or extremes from downscaler training.Assess the full bias-correction and super-resolution pipeline with complementary diagnostics, particularly for hazards dependent on coherent spatiotemporal evolution.Source: §6 · p. 21
  • Downscalers trained only on historical conditions can underestimate future climate-change signals, whereas including warmer climates in training improves their representation.Benchmarks should test complete pipelines under withheld climates and extremes, including forced-trend preservation and compute-to-skill tradeoffs.Source: §6 · p. 21

7 Conclusion

Extending AI weather forecasting to climate horizons requires credible responses to altered forcings, unseen extremes, and counterfactual scenarios; no single system resolves Earth system prediction across timescales. The conclusion outlines requirements, complementary roles for AI, near-term data opportunities, and a diverse ecosystem of hybrid models.

7.1 Summary

  • No single AI system resolves Earth system prediction across timescales, which requires more than extending weather-model rollouts.The challenge includes credible responses to altered forcings, unseen extremes, and counterfactual scenarios.Source: §7.1 · p. 21
  • Climate AI requires forcing agents to be explicit enough for independent manipulation and robustness evaluation beyond training, including extremes, transient adjustments, and climate shifts.The optimal mix of data-driven and physics-based components depends on the application.Source: §7.1 · p. 21
  • AI can enhance learned model physics, accelerate climate-model emulation, downscale projections for hazards, and fuse heterogeneous climate data streams.Generative downscaling can translate coarse 50–200 km climate information into kilometer-scale fields; fused microwave, laser, and lidar measurements can better constrain liquid and ice water path.Source: §7.1 · p. 21
  • AI reduces time-to-solution by computing fewer variables, levels, and timesteps, while synergistic learned and physical components can improve climate information.The conclusion points toward transparent combinations of learned and physically based components, while noting that optimizing their training may be more complex.Source: §7.1 · p. 22

7.2.1 ERA6 and CMIP7 as near-term drivers of progress

  • ERA6 and CMIP7 efforts may provide stronger training signals and richer scenarios, attribution experiments, and emissions-driven simulations for learning and testing forced responses.The climate efforts include scenarios involving SST patterns, weather extremes, carbon–water coupling, and possible tipping behavior.Source: §7.2.1 · p. 22

7.2.2 Towards an ecosystem of transparent, hybrid prediction systems

  • Earth system AI may develop as an ecosystem of specialized, mixable models; structural diversity supports ensembles, data fusion, hazard-specific prediction, and alternative model designs.The ICON-A-MLe framework mixes conventional and AI convection parameterizations based on learned confidence, and its AI parameterization was trained on E3SM data.Source: §7.2.2 · p. 22
  • Model diversity increases evaluation and maintenance burdens, creating tension between proliferation and consolidation around flexible, transparent, trusted models.Credible components should ideally report their training data, optimization objective, sample-wise skill and uncertainty, and training and inference costs.Source: §7.2.2 · p. 22

Related papers

Loading 2610.09770v1…