Neural Weather Models: Advances & Insights

• Neural Weather Models are data-driven and hybrid forecasting systems that learn weather dynamics by mapping observations and reanalysis data to gridded predictions.
• They employ diverse architectures—including ConvLSTM, U-Net, transformers, and graph-based models—to capture multi-scale environmental processes and probabilistic outputs.
• Key applications include nowcasting, ensemble uncertainty quantification, and hazard prediction for extreme events like heat waves, tropical cyclones, and gusts.

Neural Weather Models (NeWMs) are data-driven or hybrid forecasting systems that learn weather-evolution operators, observation-to-forecast mappings, parameterization emulators, or downstream hazard decoders from reanalysis, observations, or numerical model output. In the literature represented here, the term covers regional and global state-prediction models, observation-driven short-range systems such as MetNet-2 and MetNet-3, hybrid replacements for parts of a GCM or regional workflow, and statistical post-processors that convert NeWM environmental fields into calibrated forecasts of hazards such as convective gusts, tropical-cyclone intensity, or extreme heat (Espeholt et al., 2021, Andrychowicz et al., 2023, Belochitski et al., 2021, Leclerc et al., 31 Mar 2025).

1. Scope and taxonomy

Category	Representative papers	Typical role
End-to-end field forecasters	MetNet-2 (Espeholt et al., 2021), MetNet-3 (Andrychowicz et al., 2023), ConvLSTM-attention model (Tekin et al., 2021)	Map recent atmospheric state or observations to future gridded fields
Probabilistic and UQ layers	Spread prediction with deep convnets (Grönquist et al., 2019), distribution-based ResNets (Clare et al., 2021), conformal error bars (Gopakumar et al., 2024)	Estimate spread, predictive distributions, or calibrated intervals
Hybrid physical-neural components	GFS-physics emulator (Belochitski et al., 2021), neural interpolation for regional boundaries (Jackaman et al., 17 May 2025), CNN post-processing of WRF (Sayeed et al., 2020)	Replace, augment, or bias-correct parts of NWP/GCM pipelines
Hazard-oriented post-processing	Extreme heat (Lopez-Gomez et al., 2022), Swiss wind gusts (Leclerc et al., 31 Mar 2025), TC intensity (Gomez et al., 25 Aug 2025)	Decode NeWM output into task-specific extremes
Interpretable or diagnostic weather ML	M-ENIAC (Brecht et al., 2023), WSINDy (Minor et al., 1 Jan 2025), gridpoint relaxation (Perkan et al., 13 Jun 2025), LRP predictability (Barnes et al., 2020), latent-physics hypothesis (Craig et al., 22 May 2026)	Diagnose dynamics, sensitivity, and physical structure

A persistent misconception is that NeWMs are synonymous with a single architecture or with complete replacement of numerical weather prediction. The cited literature does not support that view. Some systems are direct neural forecasters; others are observation-driven nowcasting and day-ahead models; others replace the atmospheric physics suite of a GCM, improve boundary transfer into a limited-area model, or post-process a coarse global NeWM into regional hazard probabilities (Belochitski et al., 2021, Jackaman et al., 17 May 2025, Leclerc et al., 31 Mar 2025). The category is therefore better understood functionally than architecturally.

Another common confusion concerns “standalone” versus “hybrid” use. MetNet-2 and MetNet-3 forecast directly from rich observation and analysis inputs (Espeholt et al., 2021, Andrychowicz et al., 2023), whereas the GFS emulator keeps the numerical dynamical core and replaces the atmospheric physics parameterizations with a shallow neural network (Belochitski et al., 2021). The Swiss gust and global TC-intensity studies are even more modular: Pangu-Weather or FourCastNet v2 provide the environmental forecast, and separate statistical or neural post-processors produce the final hazard distribution (Leclerc et al., 31 Mar 2025, Gomez et al., 25 Aug 2025).

2. Representations and architectures

NeWM architectures in this corpus span recurrent convolutional models, U-Net families, graph and transformer systems, neural PDE solvers, and lightweight task-specific decoders. The ConvLSTM weather model of 2021 is a representative recurrent seq2seq design: a stacked ConvLSTM encoder–decoder with convolutional feature attention and a context matcher that aggregates encoder hidden states across time (Tekin et al., 2021). The attention mechanism weights variables at each grid cell and time step, while the context matcher sums encoder hidden states and aligns them with the decoder. This places it squarely in the pre-transformer tradition of spatio-temporal neural forecasting.

U-Net-like models appear repeatedly, but in distinct roles. A modified 3D U-Net is used to estimate ensemble spread from a small number of NWP trajectories, treating the atmosphere as a volumetric field over horizontal space and pressure levels (Grönquist et al., 2019). MetNet-3 uses a much larger hybrid architecture with learned topographical embeddings, a U-Net backbone, and modified MaxVit blocks, operating simultaneously on a $2496\,\mathrm{km} \times 2496\,\mathrm{km}$ local context at $4\,\mathrm{km}$ resolution and a $4992\,\mathrm{km} \times 4992\,\mathrm{km}$ large context at $8\,\mathrm{km}$ resolution (Andrychowicz et al., 2023). In localized radar nowcasting, a UNet with Axial Transformer augments multiscale convolution with axial attention over row and column dimensions, improving PSNR and SSIM on the processed HKO-7 subset relative to ConvLSTM, cGAN, and plain UNet baselines (Sonawane et al., 28 Apr 2025).

Global geometry is handled in several ways. The extreme-heat study trains convolutional NWMs on the cubed sphere at approximately $200~\mathrm{km}$ resolution and forecasts surface temperature anomalies globally from 1 to 28 days ahead (Lopez-Gomez et al., 2022). ConvCastNet uses spherical padding on a $3^\circ$ latitude–longitude grid to maintain compact contiguous receptive fields near the antimeridian and poles (Perkan et al., 13 Jun 2025). M-ENIAC treats spherical geometry through a coordinate embedding $$(\lambda,\varphi)\mapsto (\cos\varphi\cos\lambda,\cos\varphi\sin\lambda,\sin\varphi)$$ inside a physics-informed neural network that solves the barotropic vorticity equation directly on the sphere rather than learning a forecast operator from examples (Brecht et al., 2023).

The architectural spectrum also includes explicitly hybrid components. The GFS emulator replaces radiative transfer, cloud macro- and micro-physics, shallow and deep convection, boundary-layer processes, gravity wave drag, and land-model components with a single shallow neural network, while retaining the host GCM dynamics (Belochitski et al., 2021). The regional neural interpolation operator of 2025 uses residual 1D CNN blocks plus an explicit fine-grid flow map $A$ to transfer coarse global information to a fine limited-area model, framing nesting and spin-up as a learnable dynamics-aware interpolation problem rather than static geometric resampling (Jackaman et al., 17 May 2025).

At the opposite interpretive extreme, the “physics of AI weather models” study does not introduce a new forecaster but analyzes GraphCast and Aurora as encoder–processor–decoder systems whose residual processor layers may admit a latent-space dynamical interpretation. It reports strong CKA alignment among GraphCast variants and substantial alignment between GraphCast and Aurora, then proposes—explicitly as a hypothesis rather than a theorem—that these systems may implement a latent-space particle description and gradient flow toward a learned free-energy functional (Craig et al., 22 May 2026).

3. Data regimes and learning objectives

NeWMs in this literature are trained on heterogeneous data regimes. Reanalysis-centered systems use ERA5 directly or through WeatherBench, as in global extreme-heat forecasting, WeatherBench distributional forecasting, early ensemble studies on 500 hPa geopotential, and regional ConvLSTM forecasting (Lopez-Gomez et al., 2022, Clare et al., 2021, Scher et al., 2020, Tekin et al., 2021). Observation-driven short-range systems instead learn from dense and sparse sensors: MetNet-3 combines MRMS radar, OMO station observations, GOES imagery, the current HRRR assimilation state, geographic coordinates, learned topographical embeddings, elevation, current time, and requested lead time (Andrychowicz et al., 2023). Operational or quasi-operational regional studies use MEPS over the Nordic region (Gopakumar et al., 2024), WRF plus KMA ASOS stations over South Korea (Sayeed et al., 2020), SwissMetNet gust observations plus Pangu-Weather fields over Switzerland (Leclerc et al., 31 Mar 2025), or IBTrACS best-track intensity paired with Pangu-Weather and FourCastNet v2 outputs (Gomez et al., 25 Aug 2025).

The training objective varies sharply with task. MetNet-2 and MetNet-3 formulate forecasting as categorical prediction over discretized output bins and optimize cross-entropy, with MetNet-3 also using auxiliary MSE on 617 HRRR state channels (Espeholt et al., 2021, Andrychowicz et al., 2023). WeatherBench distribution-based neural networks similarly bin $Z500$ and $T850$ into 100 classes and train with sparse categorical cross-entropy, then recover a mean forecast by taking the expectation of the predicted categorical distribution (Clare et al., 2021). ConvCastNet and the ConvLSTM-attention weather model use MSE-style deterministic losses over multi-step rollouts (Perkan et al., 13 Jun 2025, Tekin et al., 2021).

Several papers alter the objective to target tails or calibration rather than average error. The extreme-heat study compares mean squared error with exponential losses tailored to emphasize extremes and finds that transfer learning for a few epochs is enough to improve heat-wave skill with almost no skill reduction in the general temperature prediction task (Lopez-Gomez et al., 2022). The Swiss gust study parameterizes hourly gust distributions with a generalized extreme-value family and trains neural models by minimizing CRPS, while the VGAM baseline uses maximum likelihood (Leclerc et al., 31 Mar 2025). The TC-intensity study models intensification targets $4\,\mathrm{km}$0 and $4\,\mathrm{km}$1 and uses Gaussian distributional regression with closed-form CRPS (Gomez et al., 25 Aug 2025).

Some of the most distinctive workflows concern sparse supervision or symbolic discovery rather than conventional operator learning. MetNet-3’s densification masks station inputs with probability 25% during training but keeps them as targets, forcing the model to learn an implicit analogue of data assimilation from heterogeneous sensors (Andrychowicz et al., 2023). WSINDy, by contrast, does not train a neural forecast model at all; it constructs weak-form convolutional features and solves a sparse regression problem to identify PDE terms such as advection, divergence, Coriolis coupling, and pressure-gradient structure directly from simulated or assimilated atmospheric data (Minor et al., 1 Jan 2025).

4. Probabilistic forecasting and uncertainty

Uncertainty quantification is one of the clearest fault lines in NeWM research. Several studies treat it as a first-class objective rather than an add-on. A 2019 uncertainty paper predicts the spread of temperature forecasts from a small subset of numerical simulations using modified 3D U-Nets and related architectures, demonstrating at 850 hPa and 6-hour lead time an RMSE of 0.1060 for the baseline deep model versus 0.1346 for linear regression from initial spread (Grönquist et al., 2019). The target here is not the future weather state itself but an ensemble statistic, which situates the model as a neural surrogate for expensive ensemble uncertainty estimation rather than a full probabilistic forecaster.

A second line of work converts deterministic neural forecasters into ensemble systems. In the 500 hPa geopotential study, random initial perturbations, retraining, inference-time dropout, and singular-vector-based perturbations all improve the ensemble-mean forecast relative to the unperturbed neural network, with retraining giving the best RMSE and CRPS among the tested methods (Scher et al., 2020). That paper is explicit that the resulting neural ensemble is still systematically less skillful than state-of-the-art NWP, but it establishes a durable point: ensembling improves both uncertainty information and the mean forecast.

A third line emits forecast distributions directly. WeatherBench distribution-based neural networks predict a full 100-bin distribution at each grid point for $4\,\mathrm{km}$2 and $4\,\mathrm{km}$3 rather than a single scalar, and combine specialist subnetworks through stacking (Clare et al., 2021). MetNet-3 is similarly probabilistic, using 256 bins for surface variables and 512 bins for precipitation, which makes its outputs marginal predictive distributions rather than ensembles (Andrychowicz et al., 2023). The Swiss gust and TC-intensity studies go further by imposing parametric extreme-value or Gaussian predictive laws and training distribution parameters with proper scoring rules (Leclerc et al., 31 Mar 2025, Gomez et al., 25 Aug 2025).

A major caution, however, is that probabilistic output is not equivalent to calibrated uncertainty. The conformal-prediction study on Hi-LAM makes this explicit: for nominal 90% coverage, the raw Gaussian intervals of the probabilistic model achieved 74.62% empirical coverage, whereas conformalized deterministic and conformalized probabilistic versions reached 91.09% and 91.08%, respectively; the conformalized probabilistic model was also sharper, with average standardized interval width 0.96 versus 1.13 for residual conformalization of the deterministic model (Gopakumar et al., 2024). One of the central lessons of current NeWM work is therefore that uncertainty needs its own validation pipeline—coverage, sharpness, spread–error behavior, or proper scores—rather than being inferred from the mere presence of a variance head.

5. Extremes, hazards, and downstream use

Extreme-event forecasting has become a major test of NeWMs because coarse training data, deterministic losses, and spatial smoothing can erase the very signals that matter most operationally. The global extreme-heat study is explicit on this point: it trains convolutional NWMs to forecast surface temperature anomalies globally from 1 to 28 days ahead, compares MSE with extreme-aware exponential losses, and finds that losses tailored to emphasize extremes yield significant skill improvements in the heat-wave prediction task with almost no skill reduction in general temperature prediction (Lopez-Gomez et al., 2022). The same abstract notes that a symmetric exponential loss reduces smoothing with lead time and that the best NWM shows positive regressive skill relative to the ECMWF subseasonal-to-seasonal control forecast after two weeks.

Two 2025 hazard studies make a complementary argument: coarse global NeWMs can still be operationally valuable for local extremes if they are used as environmental backbones rather than direct hazard resolvers. In Switzerland, Pangu-Weather provides hourly forecasts on a $4\,\mathrm{km}$4 global grid, which are clipped over the country and post-processed into clusterwise gust distributions constrained by generalized extreme-value laws; all post-processing models achieve CRPSS above 0.35 across lead times, neural networks outperform VGAM by a mean relative improvement of 27%, and the CNN is best overall (Leclerc et al., 31 Mar 2025). The paper is explicit that thunderstorms are too localized to be directly resolved at native NeWM scale, so the successful strategy is to predict the larger-scale storm-supporting environment and decode that environment statistically into regional gust risk.

The tropical-cyclone intensity study reaches a similar conclusion globally. Because Pangu-Weather and FourCastNet v2 are trained on coarse ERA5-like fields, their native fields smooth wind maxima and minimum pressure in the inner core; the paper therefore trains post-processors on NeWM outputs to predict best-track-style $4\,\mathrm{km}$5 and $4\,\mathrm{km}$6 changes (Gomez et al., 25 Aug 2025). The best overall probabilistic performance comes from a masked Pangu-Weather ANN with test CRPS 0.3144 and test RMSE 0.4476, while convolutional models overfit and do not clearly outperform simple multilayer perceptrons. A notable result is that even linear models extract predictive information from NeWM outputs beyond what is available in initial conditions alone, which shifts the question from “can a NeWM resolve TC intensity directly?” to “does the NeWM forecast encode enough environmental information for a downstream decoder?”

These extreme-event studies also counter a second misconception: direct end-to-end hazard resolution is not the only meaningful benchmark for NeWMs. In practice, a fast global forecast of the environment can be operationally useful even when the final impact variable is produced by a separate statistical or neural module (Leclerc et al., 31 Mar 2025, Gomez et al., 25 Aug 2025).

6. Diagnostics, interpretability, and open problems

Interpretability in NeWMs now spans at least four distinct agendas: forecast-error diagnostics, state-dependent predictability, symbolic equation discovery, and hypotheses about latent physical laws. The gridpoint-relaxation paper provides the most direct diagnostic framework. Using ConvCastNet, it relaxes selected forecast subdomains toward ERA5 during rollout and measures downstream skill gains, finding that an 8-day midlatitude forecast improves substantially when the stratosphere and boundary layer are relaxed, whereas relaxation of the tropical atmosphere has negligible effect on midlatitude and polar regions (Perkan et al., 13 Jun 2025). The same study couples these interventions to gradient-based saliency maps—gradients of absolute forecast error with respect to the initial state—showing overlap between regions of high error sensitivity and regions where relaxation most improves skill. This effectively imports adjoint-style diagnosis into a fully differentiable NeWM setting.

A second interpretive line focuses on regime dependence and attribution. The Seattle SST study is not a full weather model, but it establishes a methodological template that is directly relevant to NeWMs: forecast confidence can identify “forecasts of opportunity,” and layer-wise relevance propagation can highlight the input regions supporting those confident predictions (Barnes et al., 2020). In that study, overall testing accuracy is 62%, but the top 10% most confident predictions reach 82%, and the dominant relevance maximum lies in the ENSO region. The broader implication is that NeWM verification should not be restricted to unconditional mean scores; regime-conditioned skill and sample-specific relevance may be equally informative.

A third line seeks explicit symbolic structure. M-ENIAC uses physics-informed neural networks to solve the barotropic vorticity equation on the sphere and reproduces the historical 1950 forecast setup, beating the recreated ENIAC solver in RMS error for all four cases while remaining far from operational practicality (Brecht et al., 2023). WSINDy goes further in the direction of equation discovery, recovering compact interpretable models such as $4\,\mathrm{km}$7 for barotropic turbulence and shallow-water equations with recovered $4\,\mathrm{km}$8 and gravity coefficients very close to their generating values (Minor et al., 1 Jan 2025). This strand of work does not compete directly with foundation-style forecast systems, but it addresses a different scientific objective: exposing the governing balances rather than merely emulating them.

The most ambitious interpretive proposal is the 2026 “physics of AI weather models” paper. It reports that AI models’ day-5 forecast errors correlate more strongly with one another than with NWP systems, that CKA between GraphCast variants exceeds 0.94 while GraphCast–Aurora CKA exceeds 0.73 in the reported cases, and that processor layers in GraphCast and Aurora move from larger to smaller spatial scales with depth (Craig et al., 22 May 2026). On that basis, it hypothesizes that deterministic AI weather models may implement a latent-space particle description and gradient flow toward a learned free-energy functional. The paper is careful to frame this as a hypothesis, not a derivation of the primitive equations or a proof of latent thermodynamics.

Open problems are correspondingly varied. Calibration remains a live issue even for explicitly probabilistic models (Gopakumar et al., 2024). Extremes remain sensitive to coarse training resolution, smoothing, and objective choice (Lopez-Gomez et al., 2022, Gomez et al., 25 Aug 2025). Hybrid pipelines remain operationally important, which means NeWMs cannot be evaluated solely by end-to-end deterministic skill (Belochitski et al., 2021, Leclerc et al., 31 Mar 2025). Reproducibility is uneven: several studies note incomplete reporting of optimizer settings, batch size, epochs, or regularization, especially in hazard post-processing and nowcasting contexts (Leclerc et al., 31 Mar 2025, Sonawane et al., 28 Apr 2025, Sayeed et al., 2020). Finally, interpretability itself remains contested. Attention maps, relevance propagation, gradient saliency, and latent-alignment analyses can all identify meaningful structure, but none by itself resolves causality or guarantees that a model’s internal dynamics are physically faithful (Barnes et al., 2020, Perkan et al., 13 Jun 2025, Craig et al., 22 May 2026).

Taken together, these studies portray NeWMs not as a single model family but as a rapidly diversifying research program. Its center of gravity includes direct neural weather prediction, observation-driven forecasting, hybrid numerical–neural components, calibrated uncertainty layers, and post-processors for high-impact hazards. Its most mature strengths are low-latency inference, flexible objective design, and modular integration into forecasting pipelines; its central unresolved questions concern calibration, extremes, physical fidelity, and how best to interpret the latent dynamics that make these systems skillful.