GenCast: Diffusion-based ensemble forecasting for medium-range weather

• Weather uncertainty arises from partial observations, imperfect models, and nonlinear dynamics, making scenario ranges and their likelihoods essential for decision-making.Applications include tropical-cyclone evacuation, renewable-energy planning, and heatwave preparation.Source: §1 · p. 1
• NWP ensembles generate multiple possible forecasts, but state-of-the-art ENS remains prone to errors, slow to run, and time-consuming to engineer.ENS is ECMWF’s state-of-the-art NWP-based ensemble forecast within its broader Integrated Forecast System.Source: §1 · p. 1
• The described probabilistic approach iteratively refines noisy candidate states and autoregressively predicts successive weather states using independent noise samples.Each independently sampled noise trajectory yields a forecast trajectory, with refinement repeated before advancing through 30 time steps.Source: Figure 1 · p. 2, §1 · p. 2
• ML weather prediction improves accuracy and efficiency for non-probabilistic forecasts but typically targets mean trajectories, leaving uncertainty under-modeled and producing blurry states at longer lead times.Most methods minimize mean squared error, while prior ML-based ensembles have not rivaled operational ensemble forecasts such as ENS.Source: §1 · p. 3

Contribution

• GenCast produces more accurate global 15-day ensemble forecasts than ECMWF’s ENS at 0.25° resolution, with a single forecast taking about 8 minutes on a Cloud TPUv5.Ensembles can be generated in parallel.Source: §2 · p. 3

Weather-state representation

• The model represents six surface and six atmospheric variables across 13 pressure levels on a 0.25° grid, forecasting 30 successive states at 12-hour intervals.The forecast horizon is 15 days.Source: §2 · p. 3

Diffusion forecasting

• GenCast uses conditional diffusion to iteratively refine noise into future atmospheric states conditioned on the previous two states, repeating stochastic sampling to generate ensemble trajectories.Forecast states are sampled autoregressively.Source: §2 · p. 3

Architecture

• At each refinement stage, an encoder maps the latitude-longitude grid to a six-times-refined icosahedral mesh, a graph transformer processes k-hop neighborhoods, and a decoder returns the grid representation.The architecture consists of an encoder, processor, and decoder.Source: §2 · pp. 3–4

Training and initialization

• GenCast is trained on 40 years of ERA5 reanalysis with a single-step diffusion denoising objective, then rolled out autoregressively for 15-day forecasts initialized with ERA5 and EDA-derived perturbations.The perturbations represent uncertainty in the initial conditions.Source: §2 · p. 4

Sample realism

• GenCast samples at 1- and 15-day lead times are sharp and have spherical harmonic power spectra closely matching ERA5, unlike the blurry ensemble mean and GraphCast states.The ensemble mean is blurry because MSE-trained models approximate the ensemble mean and lose high-frequency power.Source: §3 · p. 4

Nonlinear derived quantities

• GenCast exhibits minimal wind-speed bias, whereas GraphCast exhibits large negative bias, because nonlinear quantities should be computed directly from samples rather than ensemble means.Wind speed is the L2-norm of the u and v wind components, so a nonlinear function of an average need not equal the average of that function.Source: §3 · p. 4

Tropical cyclone forecasts

• Seven days before Typhoon Hagibis made landfall, GenCast’s predicted cyclone trajectory spans many scenarios, while shorter lead times yield narrower uncertainty about landfall timing and location.The example concerns Typhoon Hagibis, which caused widespread destruction across eastern Japan and more than $17 billion in damages.Source: Figure 2n

ENS baseline

• ENS was regridded from 0.2° to 0.25° resolution, and both ENS and GenCast were evaluated using 50-member ensembles.Comparisons covered the surface variables and eight tropospheric pressure levels available for all 50 ENS members.Source: §4 · pp. 4, 6

Forecast visualization

• At long lead times, GenCast tropical-cyclone trajectories showed substantial spread, while shorter lead times produced a narrow range of trajectories.The visualization follows Typhoon Hagibis forecasts at 7-, 5-, 3-, and 1-day lead times.Source: §4 · p. 5, Figure 2 · p. 5

ML baseline

• GraphCast supplied the ML baseline, with ensemble spread generated from ERA5 initial-condition perturbations plus ad hoc Gaussian Process perturbations.Because GraphCast is deterministic, additional perturbations were needed to improve ensemble dispersion and skill.Source: §4 · p. 6

Evaluation protocol

• Each model was evaluated against its corresponding analysis: ENS against HRES-fc0 and ERA5-trained ML models against ERA5.This follows the protocol described in Lam et al. (2023, Appendix Section 5.2).Source: §4 · p. 6
• ENS retained up to 2 hours of additional assimilation-window look-ahead over ML models for most variables during the 2019 test period.ML forecasts were initialized at 06 and 18 UTC to limit look-ahead to 3 hours, whereas ENS had 3–5 hours; SST was excluded from this difference.Source: §4 · p. 6
• Verification against a single deterministic analysis can favor under-dispersed ensembles at short lead times by ignoring initial-condition uncertainty.Near zero lead time, identical initial conditions equal to the deterministic analysis would optimize such scores.Source: §4 · p. 6

CRPS comparison

• GenCast’s CRPS improvements over ENS are largest at lead times up to around 3–5 days, for surface variables, and for temperature and specific humidity at higher pressure levels, where gains reach 10–30%.The CRPS score measures how well forecast marginal distributions represent ground truth and is minimized in expectation when they reflect true predictive uncertainty.Source: §5 · pp. 6–7
• GenCast significantly outperforms ENS on 97.4% of 1320 variable, lead-time, and level combinations, with 99.8% of targets better at lead times beyond 36 hours.Figure 3’s scorecard evaluates all variables and eight pressure levels; precipitation is excluded from the main results because of concerns about ERA5 precipitation-data quality.Source: §5 · p. 6, Figure 3 · p. 7
• GenCast has better CRPS than GraphCast-Perturbed on 99.6% of targets, while GraphCast-Perturbed beats ENS on only 27% of targets.GraphCast-Perturbed’s advantages over ENS are concentrated at shorter lead times.Source: §5 · p. 7

Ensemble-mean RMSE

• GenCast’s ensemble-mean RMSE is as good or better than ENS’s on 96% of targets and significantly better on 82%.Ensemble-mean RMSE evaluates how the ensemble mean matches ground truth but does not account for forecast uncertainty.Source: §5 · p. 7

6. Ensemble calibration

• GenCast is generally as well calibrated as or better calibrated than ENS according to spread–skill ratios and rank histograms.Its spread–skill scores are usually close to 1, the ideal value for a perfectly calibrated ensemble.Source: §6 · p. 8
• GenCast’s spread–skill scores are usually close to 1, indicating ensemble spread commensurate with forecast errors on average.Values above 1 indicate over-dispersion, while values below 1 indicate under-dispersion and over-confidence.Source: §6 · pp. 7–8
• GenCast tends to have flat rank histograms, indicating that observations are broadly indistinguishable from ensemble members.A flat histogram is the ideal rank-histogram pattern for an ensemble forecast.Source: §6 · p. 8
• GraphCast-Perturbed is consistently overconfident, with spread–skill scores below 1 and ∪-shaped rank histograms.These patterns indicate under-dispersion, with truth values ranking mostly near the ensemble tails.Source: §6 · p. 8

7. Extreme weather events

• Relative Economic Value measures forecast gains over climatology, normalized by the gain from a perfect forecast, across cost/loss ratios.The optimal preparation decision occurs when the event probability exceeds the cost/loss ratio.Source: §7 · p. 8

7.2. Tropical cyclones

• 11%, 38%, and 106% higher relative economic value than ENS is achieved by GenCast for tropical-cyclone preparedness at 1-, 3-, and 5-day lead times.At a 0.01 cost/loss ratio, GenCast consistently yields greater REV across other ratios and lead times up to 7 days.Source: Figure 4 · p. 9, §7.2 · p. 10

7.1. Local surface extremes

• 1.9× better relative economic value than ENS is obtained by GenCast for predicting 99.99th-percentile high temperature at a 0.05 cost/loss ratio.GenCast also maintains an advantage across other cost/loss values, 5- and 7-day lead times, event percentiles, and several extreme-weather variables.Source: §7.1 · p. 8, Figure 4 · p. 9
• GenCast significantly improves on ENS for extreme temperature, wind-speed, and mean-sea-level-pressure exceedance events across tested thresholds, except for some low-pressure improvements.The comparison uses Brier skill scores for 99.99th, 99.9th, and 99th percentiles and all lead times.Source: §7.1 · p. 9

7.2. Tropical cyclones

• Tropical-cyclone probabilities are derived from the proportion of GenCast or ENS ensemble members whose tracked cyclone centres occupy a location at a given time.TempestExtremes is applied to ensemble forecasts and corresponding ERA5 or HRES-fc0 tracks.Source: §7.2 · pp. 9–10

8.1. Spatially pooled evaluation

• ∼25% improvement over ENS occurs for GenCast’s short-lead max-pooled 2 m temperature CRPS, with performance approaching ENS from approximately 8 days.The improvement is consistent across spatial scales, whereas GraphCast-Perturbed performs substantially worse than ENS and degrades as pooling size increases.Source: §8.1 · p. 10
• 98.1% of average-pooled and 97.6% of max-pooled CRPS targets favor GenCast over ENS, versus 23.9% and 7.0% for GraphCast-Perturbed.Across 5,400 targets spanning variables, levels, lead times, and spatial scales, GenCast’s relative performance increases at larger scales, indicating stronger spatial-structure prediction.Source: §8.1 · p. 10

8.2. Regional wind power forecasting

• ∼20% improvement over ENS CRPS persists through 2 days for regional wind power, with 10–20% improvement from 2–4 days and statistically significant gains to 10 days.The simplified experiment aggregates idealised wind-power estimates across wind farms within 120 km, 240 km, and 480 km pooling regions; GenCast improves substantially more than GraphCast-Perturbed.Source: §8.2 · p. 11

Overall findings

• GenCast outperforms ECMWF’s ENS in accuracy and decision-making value across variables, vertical levels, lead times, and application settings.It generates stable, skillful 15-day global forecasts despite training explicitly on 12-hour steps, with ensembles whose uncertainty is generally well calibrated.Source: §9 · p. 11

Operational considerations

• Operational deployment may require higher resolution, training on operational ENS inputs, and efficiency techniques such as distillation for faster sampling.GenCast currently operates at 0.25° and is computationally more expensive than an equivalent deterministic MLWP model because diffusion sampling requires multiple function evaluations per timestep.Source: §9 · p. 11

Broader contribution

• GenCast still depends on NWP-based data assimilation for its initial conditions, while the work demonstrates generative AI can model complex weather distributions reliably for consequential decisions.The authors characterize these results as opening a new front in weather forecasting with greater accuracy, efficiency, and accessibility.Source: §9 · p. 11

A. Task definition and general approach

• GenCast combines an MLWP forecast model with traditional NWP-based state inference using ERA5 and ERA5 EDA initial conditions.The model uses ensemble analysis to approximate the current atmospheric-state distribution before forecasting.Source: A. Task definition and general approach · p. 19
• The model represents the atmospheric state as an 84 × 720 × 1440 array containing 6 surface variables and 6 atmospheric variables at 13 pressure levels on a 0.25° grid.Because this is only a partial state representation, GenCast conditions on both the current and previous states, X_0 and X_-1.Source: A. Task definition and general approach · p. 19
• GenCast generates 15-day trajectories at 12-hour steps by autoregressively sampling each state conditioned on the previous two states.Repeated sampled trajectories support Monte Carlo estimates of probabilities and expectations in the high-dimensional forecasting problem.Source: A. Task definition and general approach · p. 19

B.1. ERA5 training and evaluation data for GenCast

• The ERA5 archive represents global weather from 1959 to the present at hourly resolution for hundreds of static, surface, and atmospheric variables.It includes deterministic ERA5 reanalysis and ERA5 EDA, which has 9 ensemble members and 1 control member at lower native resolution.Source: §B.1 · p. 20
• GenCast uses selected ERA5 variables on a 0.25° equiangular grid at 13 WeatherBench pressure levels, subsampling inputs to 6-hour resolution and extracting 12-hour sequences.The pressure levels range from 50 to 1000 hPa, and training sequences use either 00/12 UTC or 06/18 UTC times.Source: §B.1 · p. 20
• The 12-hour training step avoids the bimodal distribution caused by ERA5’s 12-hour observation-assimilation windows and targets the next assimilation window.Accumulated variables such as precipitation are aggregated over the preceding 12 hours rather than subsampled.Source: §B.1 · p. 20
• The dataset covers 1979–2019, with development training on 1979–2017, validation on 2018, and frozen-model testing on 2019 after retraining through 2018.The GraphCast-Perturbed baseline used data from 1979 to 2018 at 0.25° resolution under a consistent split protocol.Source: §B.1 · p. 20

B.2. Data preprocessing.

• EDA perturbations are bilinearly upsampled from 0.5° to 0.25°, while SST NaNs over land are replaced with the global minimum SST; one ENS initialization was omitted because of persistent download errors.ERA5-based preprocessing also uses ERA5 SST where interpolated EDA SST is unavailable at some locations.Source: §B.2 · pp. 20–21

B.3. Data Availability

• The study uses ERA5 and ERA5 EDA, ECMWF ENS and HRES, and external power-generation datasets obtained from their stated public archives.The listed sources include the Copernicus Climate Data Store, ECMWF’s TIGGE archive, the Global Power Plant Database, and NREL turbine models.Source: §B.3 · pp. 21–22

C. Initial conditions; C.1. Deterministic analysis; C.2. Ensemble of data assimilations (EDA) analysis; C.3. Deterministic analysis plus EDA perturbations; C.4. Additional ad hoc Gaussian process perturbations

• GenCast ensembles are initialized by combining deterministic ERA5 analyses with EDA-member-minus-mean perturbations, the method that performed best for the diffusion model beyond the shortest lead times.Deterministic analysis alone provides no ensemble dispersion, while EDA analysis is better but remains under-dispersed and lower-resolution; additional Gaussian-process perturbations did not improve diffusion-model skill.Source: §C.3 · p. 22, §C.1 · p. 22, §C.2 · p. 22

C.4. Additional ad hoc Gaussian process perturbations

• Gaussian-process initialization perturbations use a 1200 km spherical correlation length and variable-specific scales, but adding them to diffusion-model initial conditions generally reduces skill except at the smallest scale factors.The perturbations are sampled independently for selected variables, use a scale factor of 0.085, and were selected by CRPS sweeps over scale factors and correlation lengths.Source: §C.4 · pp. 22–23

D. Diffusion model specification

• The key model innovation is a learned neural diffusion model for p(X_t|X_t−1, X_t−2), which samples normalized residuals rather than directly sampling X_t.The model follows Karras et al. (2022) in several diffusion choices but changes the noise distribution, training noise-level distribution, and loss weightings.Source: D. Diffusion model specification · pp. 23–24

D.1. Sampling process

• Forecast samples are produced by integrating the probability-flow ODE with stochastic churn and noise inflation, using 20 second-order DPMSolver++2S steps and 39 denoiser evaluations per generated timestep.Each solver step is conditioned on the two previous states, and the final step requires only one denoiser evaluation.Source: §D.1 · p. 24

D.2. Noise distribution on the sphere

• Noise is sampled as isotropic Gaussian noise in the spherical-harmonic domain, yielding a flat resolved power spectrum and improved spherical isotropy over independent grid-cell noise.The noise is projected onto the discrete latitude-longitude grid by an inverse spherical harmonic transform.Source: §D.2 · p. 24

D.3. Denoiser architecture; D.4. Training the denoiser

• The denoiser uses a GraphCast-related encoder-decoder but replaces its processor with a 16-block graph transformer on a 41,162-node icosahedral mesh, while training predicts the next 12-hour state with area- and variable-weighted denoising loss.Noise levels are encoded through Fourier features and conditional layer normalization; training samples noise levels matched to the generation schedule rather than using Karras et al.’s log-normal distribution, and does not feed the model its own previous predictions.Source: §D.3 · p. 25, §D.4 · pp. 25–26

D.4.1. Resolution training schedule

• GenCast trains first at 1° for 2 million steps, then fine-tunes at 0.25° for 64,000 steps, preserving message-signal scale during resolution upsampling.The first stage uses a 5-refined mesh, the second a 6-refined mesh, and encoder message sums are divided by 16 during fine-tuning.Source: §D.4.1 · p. 26

D.5. Sampler hyperparameters

• Sampling uses DPMSolver++2S with stochastic churn and noise inflation under the Karras et al. (2022) noise-level schedule.The solver requires 2N−1 function evaluations for N noise levels, with settings specified in Table D2.Source: §D.5 · p. 27, Table D2 · p. 27

E.1. ENS initialisation and evaluation times

• ENS is evaluated at the same 06/18 UTC validity times as GenCast, with interpolated lead-time CRPS that overestimates ENS performance, especially at short leads.GenCast uses 06/18 UTC initialisations, whereas ENS uses 00/12 UTC because matching archived ENS forecasts are unavailable.Source: §E.1 · p. 28

E.2. Verification metrics

• Verification uses finite-ensemble CRPS, ensemble-mean RMSE, spread/skill, rank histograms, Brier skill score, and relative economic value.CRPS and RMSE favor smaller values; spread/skill near 1 indicates ideal dispersion, while Brier skill score and relative economic value range from climatological zero to perfect one.Source: §E.2.1 · p. 28, §E.2.2 · p. 29, §E.2.3 · p. 29

E.2.3. Spread/skill ratio; E.2.4. Rank histogram

• Spread/skill and rank histograms diagnose ensemble dispersion, with ideal forecasts having spread/skill approximately 1 and flat rank histograms under exchangeability.U-shaped rank histograms indicate under-dispersion, whereas ∩-shaped histograms indicate over-dispersion when bias is small.Source: §E.2.3 · p. 29, §E.2.4 · p. 29

E.2.5. Brier skill score; E.2.6. Relative Economic Value

• Brier skill score evaluates binary-event probabilities against climatology, while relative economic value measures cost-loss decision savings across thresholds and user ratios.Relative economic value is maximized over probability thresholds and is bounded between 0 for climatology and 1 for a perfect forecast.Source: §E.2.5 · p. 30, §E.2.6 · p. 30

E.3. Statistical tests

• GenCast–ENS differences are tested with paired stationary block bootstraps using 10,000 resamples and bias-corrected confidence intervals, rejecting equality when intervals exclude zero.The procedure aligns differing initialisation times, selects blocks automatically, and linearly imputes one missing ENS initialisation.Source: §E.3 · pp. 32–33

E.4. Local surface extremes evaluation

• Local surface extremes are defined by high and low climatological percentiles from seven years of 6-hourly data, spanning approximately 25 days to 7 years and 8 months of return periods.Thresholds are evaluated separately at each latitude and longitude using ERA5 for GenCast and HRES-fc0 for ENS.Source: §E.4 · p. 33

E.5.1. Cyclone evaluation

• The probabilistic cyclone evaluation independently tracks forecasts, extracts strike probabilities, and uses REV to penalize cyclone existence and position errors.TempestExtremes is applied consistently to ENS, GenCast, and ground-truth datasets; ENS is downsampled to 12-hourly resolution for comparison.Source: §E.5.1 · pp. 33–34
• ERA5 contains 1.51 cyclones per timestep versus 1.85 for HRES-fc0, but REV normalizes base-rate differences and GenCast still beats ENS beyond 1 day using HRES-fc0 truth.HRES-fc0 therefore has a 23% higher cyclone base rate than ERA5.Source: §E.5.1 · p. 34, Figure E1 · p. 34

E.5.2. Cyclone tracker

• TempestExtremes detects cyclone candidates from pressure minima and upper-level warm cores, then stitches physically plausible trajectories using movement, duration, intensity, elevation, and latitude criteria.For 12-hourly data, the tracker uses five time slices and a 12° stitch range; ten days of ground truth are prepended, and results are reported through day 9.Source: §E.5.2 · p. 35

E.6. Spatially pooled evaluation

• Spatially pooled evaluation uses spherical icosahedral regions spanning 120–3828 km, with area-weighted average- and max-pooling CRPS to assess multiple spatial scales.Average-pooling emphasizes larger structures, whereas max-pooling emphasizes positive tails such as extreme wind or temperature.Source: §E.6 · p. 36

E.7. Regional wind power forecasting evaluation

• Regional wind-power evaluation interpolates 10 m wind forecasts at 5344 wind farms, converts them through an idealized 2 MW turbine curve, and scores aggregated power with CRPS.Evaluations use 120, 240, and 480 km groups; the database captures approximately 40% of global wind-farm capacity as of 2020.Source: §E.7 · p. 39

F.2. Ensemble calibration

• GenCast’s spread-skill and rank-histogram calibration is generally as good as or better than ENS’s and substantially better than GraphCast-Perturbed’s.GenCast’s spread-skill scores approach 1 from 2–3 days onward, while GraphCast-Perturbed is substantially under-dispersed.Source: §F.2 · pp. 41–42

F.3. Ensemble-Mean RMSE

• GenCast is as good as or better than ENS in 96% of ensemble-mean RMSE cases and significantly better in 82%.Additional results show ENS consistently outperforming GraphCast-Perturbed at long lead times, while GenCast generally outperforms GraphCast-Perturbed.Source: §F.3 · p. 44, Figure F5 · p. 45, §F.6 · p. 50

F.8. Tropical cyclones

• GenCast significantly outperforms ENS across broad cost-loss ratios for tropical cyclones at 1 day and smaller cost-loss ratios through 7 days.The comparison uses each model’s own ground-truth cyclone tracks.Source: §F.8 · p. 55
• Despite two distribution-shift penalties, GenCast significantly (p< 0.05) outperforms ENS at small cost-ratios and lead times of 3–5 days.The result is robust to differences in cyclone base rates between ERA5 and HRES-fc0.Source: Figure F19 · p. 55, §F.8 · p. 55

F.9.1. Wind speed as a derived variable

• Deriving skillful wind-speed distributions from 𝑢 and 𝑣 tests whether GenCast models their cross-variable dependence structure.GraphCast’s expectation-based training can underestimate nonlinear derived quantities such as wind speed when uncertainty in wind direction shrinks the expected wind vector.Source: §F.9.1 · p. 56
• GenCast exhibits minimal bias for derived 10 m wind speed and maintains its CRPS advantage over ENS, unlike GraphCast-Perturbed at that variable.GraphCast and GraphCast-Perturbed show significant negative wind-speed bias at longer lead times despite minimal bias for 10u and 10v.Source: §F.9.1 · p. 56

F.9.2. Pooled evaluation

• 98.1% and 97.6% of 5400 pooled targets favored GenCast over ENS for average-pooled and max-pooled CRPS, versus 23.9% and 7.0% for GraphCast-Perturbed.The pooled evaluation aggregates wind components into wind speed and includes tp12hr across variables, lead times, and spatial scales; surface-only results also favor GenCast as pooling size increases.Source: §F.9.2 · p. 56

F.10. Regional wind power forecasting statistical significance

• GenCast’s regional wind-power CRPS improvement over ENS is statistically significant through 7-day lead times for all pool sizes and through 10 days with some exceptions.Figure F27 marks significant differences using p< 0.05.Source: §F.10 · p. 63, Figure F27 · p. 64

F.11. Ablation of EDA perturbations in GenCast initial conditions

• EDA-perturbed initial conditions have little effect on CRPS and RMSE, while deterministic initialization improves early spread-skill ratios but generally under-disperses after the first steps.The ensemble-initialized GenCast generally has better spread skill from approximately 2–3 days onward; GenCast without EDA perturbations still outperforms ENS on 96% of ensemble-mean RMSE targets and 97% of CRPS targets.Source: §F.11 · p. 64, Figure F28 · p. 65, Figure F29 · p. 66

F.12.1. Initialisations used for GenCast evaluation

• GenCast evaluations use 06/18 UTC initialization because 00/12 UTC forecasts benefit systematically from the ERA5 look-ahead and therefore have an advantage in CRPS and ensemble-mean RMSE.The 06/18 UTC choice is presented as the more conservative evaluation approach.Source: §F.12.1 · p. 67, Figure F31 · p. 68

F.12.2. Lead time interpolation in regional wind farm evaluation

• Lead-time interpolation can overestimate 06/18 UTC ENS performance, with 06/18 UTC ENS 6% worse than interpolated 00/12 UTC ENS at 12 hours, potentially advantaging ENS.This effect helps explain non-monotonic GenCast relative performance at the shortest lead times in Figure 4d.Source: §F.12.2 · p. 67, Figure F32 · p. 69

G.1. Tropical cyclone tracks

• Cyclone visualizations cover extreme 2019 cases including Idai, Hagibis, Halong, and Dorian, but are illustrative rather than representative comparisons of GenCast and ENS.Different initialization times can materially alter cyclone forecasts, as illustrated by ENS’s improved Dorian northward-curve prediction when initialized 12 hours later.Source: §G.1 · p. 70

G.2. Visualising forecasts of additional variables during Typhoon Hagibis; G.3. Global forecasts

• Additional visualizations show Typhoon Hagibis forecasts for 10 representative variables, while global visualizations show the same variable set across lead times using median-CRPS 2019 initialization times and the first ensemble member.The variables include humidity, geopotential, temperature, wind, 2-meter temperature, and mean sea-level pressure.Source: §G.2 · p. 75, §G.3 · p. 85

H. Author Contributions

• Author contributions span conceptualization, project management, data curation, model development, code, experiments and evaluation, and paper writing and editing across the listed authors.The contribution statement assigns multiple authors to each work category.Source: H. Author Contributions · p. 96