betaFeedback

WorksPhysics and Astronomy

Towards operational data-driven regional weather forecasting: extreme-event skill and near-real-time verification with HRRRCast

Abstract

Improving convective-scale skill in data-driven regional weather models requires better representing rare high-impact events, training at native resolution, broadening variable output, and demonstrating performance in real-time settings. Here, we evaluate these challenges using HRRRCast, a data-driven convection-allowing emulator trained on the High-Resolution Rapid Refresh (HRRR) analysis data. HRRRCast combines native 3-km training, moderate sigmoid-weighted loss reweighting for reflectivity and precipitation, and broader system improvements including expansion of the training archive from 3 to 9 years (2016–2025), increased vertical resolution, expanded outputs, and architectural upgrades. The weighted loss improves high-threshold skill while preserving common-event behavior. Native-resolution training avoids aliasing from 6-km subsampling and substantially improves fractions skill score at convective scales, with gains that persist across pooling windows. HRRRCast has also been running in near-real-time mode for the past year. In seven months of independent observation-based verification, HRRRCast ensemble means improve several near-surface metrics relative to operational HRRR and maintain strong precipitation and reflectivity performance. Single-member reflectivity skill is strongest at 20–30 dBZ; at 40 dBZ, skill is near parity at 6-km pooling and better at pooling radii ≥\geq30 km. Precipitation skill remains competitive through 1 in/h and extends to about 3 in/h in the best recent configuration. A focused comparison with HREF further shows useful low-cost probabilistic guidance for 3-h precipitation. At storm scale, HRRRCast exhibits a displacement–structure trade-off, with generally better storm placement than storm morphology.

Significance Statement

This study examines how data-driven regional weather forecasting can be made more operationally useful. Using HRRRCast as a testbed, we show that improved treatment of rare high-impact events, native 3-km training, and a much longer training archive improve storm-related forecast skill, hourly precipitation performance, and several near-surface measures, while also providing useful low-cost ensemble guidance. The results show that the field can move beyond proof-of-concept demonstrations toward near-real-time practical complementarity with traditional weather models. They also identify an important remaining challenge: improving the fine-scale structure of predicted storms.

Introduction

Convection-allowing models (CAMs) have become essential tools for predicting high-impact weather phenomena such as severe thunderstorms, heavy precipitation, and flash flooding. Operating at km-scales, CAMs explicitly resolve many convective processes without relying on convective parameterization schemes, thereby producing more realistic forecasts of mesoscale weather features. The High-Resolution Rapid Refresh (HRRR) model, operational since 2014 and running at 3-km resolution over the contiguous United States (CONUS), exemplifies this capability and provides crucial guidance for operational forecasters [7].

Despite their success, physics-based CAMs like HRRR remain computationally expensive. A single 18-hour HRRR forecast requires substantial high-performance computing resources, limiting the frequency of model runs, the ensemble size, and the ability to explore alternative configurations for research. These computational constraints motivate the development of fast, data-driven emulators that can complement traditional numerical weather prediction (NWP) systems.

In Abdi et al. [ [1], we introduced HRRRCast, an ensemble data-driven emulator trained on HRRR analyses data to produce rapid convection-allowing forecasts over CONUS. HRRRCast demonstrated that using a simple architecture, one could generate ensemble forecasts with competitive skill at light-to-moderate reflectivity thresholds (20–30 dBZ), while significantly reducing computational cost relative to HRRR. The model was trained on sub-sampled 6-km data spanning three years and incorporated future GFS boundary conditions to improve long-lead skill. While this initial version represented a significant step toward operational machine learning-based regional forecasting, several limitations remained: (i) training on sub-sampled 6-km data introduced aliasing that corrupted fine-scale spectral structure in the training targets; (ii) performance at extreme precipitation and reflectivity thresholds lagged behind HRRR; and (iii) the limited training dataset (3 years) constrained the model’s ability to capture the full climatological range of CONUS weather regimes.

In this paper, we use HRRRCast to examine several ingredients needed for more operationally useful regional AI weather forecasting. The key contributions are:

  1. Native-resolution training: We demonstrate what is gained by training on the native 3-km HRRR grid rather than on sub-sampled 6-km fields.

  2. Extreme-event optimization: We demonstrate that a moderate sigmoid-weighted loss for reflectivity and hourly precipitation improves high-threshold skill while preserving common-event behavior.

  3. Improvements for operational performance: We expand the training archive to nearly a decade, broaden multivariate outputs and vertical levels, and introduce architectural upgrades.

  4. Operational/observational verification: We demonstrate that the resulting system is broadly comparable to operational HRRR in independent observation-based verification, including lead-time die-off diagnostics and threshold-dependent behavior.

  5. Storm-scale error characterization: We demonstrate and interpret a consistent displacement–structure contrast, including lead-time-dependent fragmentation and object-scale morphology biases.

HRRRCast has been running in near-real-time for nearly one year. The observation-based verification with the Model Analysis Tool Suite (MATS) [21] therefore focuses on the most recent seven months, when the native 3-km HRRRCast system was running. Over this period, HRRRCast single-member forecasts are broadly comparable to operational HRRR across multiple verification metrics, while its ensembles provide additional probabilistic guidance at minimal computational cost. Because rapid ensemble guidance is a key intended use case, we also include a focused comparison with HREF for probabilistic precipitation forecasting. These results show improved HRRRCast skill for relative humidity, dewpoint, and 10-m winds. Hourly precipitation skill is broadly competitive through 1.00 in and improves further in the best recent configuration. Reflectivity skill also improves at 20 and 30 dBZ; at 40 dBZ, behavior is scale dependent, with near parity at 6-km pooling and HRRRCast advantages at larger pooling radii. Fractions skill score (FSS) analyses further show that native 3-km training yields substantial gains at convection-permitting scales that persist across pooling windows. Together, these results position HRRRCast as a viable complement to physics-based CAMs for rapid ensemble generation, real-time decision support, and exploratory sensitivity studies.

Beyond aggregate skill gains, we examine residual storm-scale errors in AI regional weather models. This analysis reveals a consistent contrast between storm placement and storm morphology, with HRRRCast generally showing better displacement behavior than structural realism at longer leads. Using subhourly lead-time-conditioned inference as a diagnostic, we further show that temporally smooth evolution can coexist with persistent storm-structure bias. § 6 then evaluates these behaviors and discusses possible mechanisms.

The remainder of this paper is organized as follows. § 2 describes the updated methodology, including the training loss design and architectural enhancements. § 3 summarizes the expanded training dataset and preprocessing pipeline. § 4 presents the main near-real-time verification results, including observation-based diagnostics and a focused ensemble comparison against HREF; supplemental RMSE verification against HRRR analyses is deferred to Appendix § 20.§ 2. § 5 presents attribution and ablation studies, including native-resolution and loss-reweighting comparisons. § 6 focuses on storm-scale behavior and interpretation, including displacement, structure, and subhourly-cadence limitations. § 7 presents conclusions and § 8 outlines directions for future work.

Methodology

This section describes the key methodological improvements in HRRRCast relative to the original model presented in Abdi et al. [2026]. The updates include: (i) training on the native 3-km HRRR grid rather than sub-sampled 6-km data, (ii) an expanded training dataset spanning nearly a decade of HRRR and GFS analyses, (iii) a sigmoid-weighted balanced loss function to improve extreme-event skill, and (iv) architectural improvements required for 3-km training.

Native 3-km grid training

The original HRRRCast was trained on HRRR data sub-sampled to 6-km resolution to manage memory and computational constraints. While this approach enabled rapid prototyping, spatial coarsening introduces aliasing artifacts that corrupt fine-scale spectral structure in the training targets. By training directly on the native 3-km HRRR grid, the model learns physically consistent fine-scale structure characteristic of convection-allowing scales.

To accommodate the four-fold increase in grid points relative to 6-km training, we employ memory-efficient engineering strategies including mixed-precision training (float16 storage with selective float32 computation), gradient checkpointing, and optimized data loading pipelines. These techniques enable end-to-end 3-km model training on eight H100 GPUs without sacrificing model capacity or batch size.

Expanded training dataset and multivariate outputs

Relative to the original model, the training corpus is expanded from three years to nearly a decade, substantially broadening coverage of sampled weather regimes across seasons and event types. In parallel, the multivariate learning setup is expanded to include finer vertical structure and a larger prognostic variable set, with additional diagnostic fields derived for hazardous-weather applications.

Conceptually, this update shifts HRRRCast from a narrower proof-of-concept configuration toward a more complete CAM-emulator training target. Detailed data curation, variable definitions, vertical level design, and accumulated-precipitation target construction are described in § 3.

Sigmoid-weighted loss for extreme events

A fundamental limitation of standard mean squared error (MSE) training is the tendency to optimize the conditional mean of the target distribution, which systematically underweights rare, high-intensity events [19, 18]. In our training distribution, more than 90% of samples are low-threshold events, so uniform MSE primarily optimizes common cases and under-emphasizes extremes. For precipitation and reflectivity forecasting, this class imbalance mainly degrades high-threshold skill.

To address this, we implement a sigmoid-weighted balanced MSE loss applied selectively to composite reflectivity (REFC) and accumulated precipitation (APCP). The weighting function applies a smooth multiplicative factor that increases from approximately 1×1\times for common events to 3×3\times for rare, high-intensity events (defined as roughly 4σ4\sigma to 10σ10\sigma in normalized anomaly space). For REFC, σ≈5\sigma\approx5 dBZ in the training normalization, so the ramp begins in the lower convective-reflectivity range (reflectivity values around 20 dBZ, consistent with Figure 15); APCP uses the same 4σ4\sigma–10σ10\sigma criterion in its own transformed, variable-specific normalized space. This design provides controlled emphasis on extreme events while avoiding the unbounded weight growth that can destabilize training or produce unrealistic forecast amplitudes.

Candidate weighting functions and cumulative distribution comparisons

Figure 15. Left: candidate weighting functions explored for extreme-event training. A softplus-based formulation increased weights too aggressively and led to unrealistically large reflectivity and precipitation amplitudes. Right: REFC cumulative-distribution comparison at 18-h forecast lead for different weight caps. The 3×3 \times weighting produces a distribution that remains much closer to the reference behavior than the more aggressive 5×5 \times option, while still providing stronger extreme-event emphasis than 1×1 \times. This supports the final choice of a global sigmoid weighting function capped at 3×3 \times. For REFC, the ramp onset near 20 dBZ corresponds approximately to the 4σ4\sigma threshold used in § 2.3 (σ≈5\sigma\approx5 dBZ).

The weighting is applied only to REFC and APCP because these variables exhibit the most pronounced heavy-tailed distributions and are most critical for operational severe weather applications. All other forecast variables (temperature, winds, geopotential height, etc.) continue to use standard MSE. The rationale for this variable selection, comparisons with alternative weighting formulations, domain-size context, and the tuning experiments used to select the final schedule are provided in Appendix § 12.§ 6.

Architectural improvements

Several architectural enhancements complement the training and data improvements. We introduce dedicated preprocessing branches that process HRRR, GFS, and (for diffusion training) noised-HRRR inputs through separate residual stems before fusion, enabling more flexible input handling. These modifications, combined with memory-efficient engineering (gradient checkpointing, spatially grouped convolutions), enable stable training at native 3-km resolution. We also replace squeeze-and-excitation (SE) blocks with Convolutional Block Attention Modules (CBAM; Woo et al., [22]), which provide both channel-wise and spatial attention. Complete architectural specifications are provided in Appendix A.

Dataset

This section summarizes the data sources and preprocessing choices used for training and verification, with emphasis on aspects most relevant for interpreting § 4 results. Complete details on data sources, variable selection, preprocessing, normalization, version harmonization, and quality control procedures are provided in Appendix A.

Data sources

The primary training target is the HRRR model, which has been operational over CONUS since 2014 at 3-km horizontal resolution with hourly cycling [7]. Over the 2016–2025 training period, HRRR underwent several version updates that introduced changes to model physics, data assimilation procedures, and output variable specifications.

GFS data provide large-scale atmospheric context and future boundary conditions. We interpolate GFS data to the HRRR grid and include them as additional input channels, enabling the model to perform physically consistent downscaling implicitly from coarse global guidance to convection-allowing scales.

Training/testing splits and verification periods

The production 3-km training dataset is split chronologically into a training period from 2016-10-01T00 through 2025-06-05T23 and a testing period from 2025-06-06T00 through 2025-10-05T23. Additionally, the near-real-time MATS verification archive considered in this study spans 2025-10-05T23 through 2026-10-09T23, and the focused seven-month subset discussed in § 4.1 is 2026-03-20T00 through 2026-10-09T23.

Training variables and vertical resolution

The model predicts a comprehensive prognostic variable set consisting of 3D fields (temperature, geopotential height, u/v wind components, specific humidity, and vertical velocity) on 20 pressure levels spanning 1000–200 hPa, compared with 12 levels in the original model, to better resolve the vertical thermodynamic and kinematic structure relevant to convection; 2D surface and near-surface fields (2-m temperature, 2-m dewpoint, 10-m u/v winds, mean sea level pressure, and surface pressure); accumulated hourly precipitation; and composite reflectivity. Relative to the original model, we add 18 surface/integrated variables, including hourly accumulated precipitation (APCP), and treat them as prognostic variables in the same autoregressive framework (i.e., they are both network inputs and prediction targets at subsequent steps). In addition to these prognostic fields, we compute a broad set of surface and upper-air diagnostic variables relevant to hazardous-weather forecasting, bringing HRRRCast closer to the breadth of output variables available from HRRR.

Accumulated precipitation targets

Accumulated variables introduce several training challenges. In the HRRR analyses dataset, APCP (and even instantaneous precipitation rate) is zeroed out, so precipitation targets must be taken from HRRR forecasts rather than analyses. If reliable analyzed instantaneous precipitation-rate fields were available, they would provide a more direct supervision target. In future work, it may be preferable to use MRS hourly precipitation as the training target to reduce known HRRR precipitation biases and to improve consistency with other variables trained against HRRR analyses. We also found that the vertical velocity field (VVEL) in HRRR analyses contains spurious values; therefore, we use the 1-h forecast VVEL field instead of the analyzed field.

Another challenge is selecting the accumulation period for precipitation targets. HRRR provides hourly accumulated precipitation, whereas HRRRCast is trained to predict three lead times (1-h, 3-h, and 6-h). One alternative is to predict each corresponding accumulation directly. However, this effectively creates three strongly correlated precipitation targets with different sparsity/climatology. We therefore predict only the last-hour accumulation as the prognostic precipitation target. This choice is aligned with hourly autoregressive stepping, keeps precipitation as a single variable in training, and still allows 3-h and 6-h accumulations to be reconstructed consistently by summing hourly forecasts.

Preprocessing and normalization

All input and target fields are normalized using climatological statistics computed from the training dataset. For variables with highly skewed distributions (e.g., precipitation), we apply variance-stabilizing transformations prior to normalization to improve numerical conditioning and training stability. Native 3-km training requires careful attention to storage and I/O efficiency; we store preprocessed data in float16 format and employ mixed-precision training with selective float32 computation in loss evaluation and gradient-sensitive operations. These engineering considerations are detailed in Appendix § 12.

Results

Results in this section focus on near-real-time observation-based verification against MRMS through MATS, followed by a focused operational comparison against HREF.

Near-real-time verification against observations using MATS

A key distinction from retrospective verification is that the near-real-time archive spans three native 3-km HRRRCast configurations, with the strongest version operating only during the final two-month period. Aggregate statistics should therefore be interpreted with that model evolution in mind. Solid lines denote the full approximately seven-month archive; where shown, dashed lines denote the final two-month subset. The near-real-time ensemble used 9 members for most of the archive and 18 members after July 1, 2026. Member 0 uses unshifted GFS forcing, while the other members use phase-shifted GFS boundary-condition inputs; see Section A.7 for details.

Composite-reflectivity CSI in the lead-time die-off summaries depends on threshold (Figure 1a–Figure 1c). At all reflectivity thresholds, the HRRRCast ensemble mean (orange) remains consistently higher than operational HRRR across most lead times, with the expected decline in CSI as forecast range increases. For deterministic single-member comparisons, HRRRCast member 0 is higher than HRRR at 20 dBZ, approximately equal at 30 dBZ, and slightly lower at 40 dBZ. This threshold dependence is consistent with greater difficulty at the most intense convective cores. The fixed-lead threshold curves (Figure 1d–Figure 1f) further show that the single-member crossover threshold shifts lower with lead time. The ensemble mean, by contrast, retains a clear advantage over HRRR through 40 dBZ across all three lead times, with the three curves converging as skill approaches zero at 45 dBZ, suggesting that ensemble pooling extends meaningful CSI improvements into the intense-convection regime even as the deterministic advantage narrows. It is worth noting that the CSI diagnostics here use strict grid-to-grid matching rather than neighborhood-based verification and are therefore more sensitive to small displacement errors.

Composite-reflectivity CSI over the seven-month evaluation period

Figure 1. Composite-reflectivity CSI over the seven-month evaluation period. Top row: lead-time die-off at 20, 30, and 40 dBZ; bottom row: threshold-dependent CSI at 6-, 12-, and 18-h leads. The HRRRCast ensemble generally outperforms operational HRRR, but the two converge at 40–45 dBZ as CSI approaches zero; HRRRCast member 0 remains slightly lower at these highest thresholds.

Hourly precipitation CSI in the lead-time die-off summaries likewise reveals threshold-dependent behavior (Figure 2 and Figure 3). At 0.01 and 0.10 in, the HRRRCast ensemble (orange) is generally better than HRRR operational across most leads in both the full seven-month archive and the mature subset, and HRRRCast member 0 also becomes competitive after the first few forecast hours. At 0.25 and 0.50 in, the ensemble maintains a clear advantage over HRRR through much of the forecast in both archives, with the mature subset showing that the best native 3-km HRRRCast configuration strengthens that advantage further. At 1.00 in, the distinction between the two archives is especially important: in the full seven-month aggregate, the ensemble is near parity to modestly better than HRRR, whereas in the mature subset the HRRRCast ensemble is clearly better than HRRR across all plotted leads. HRRRCast member 0 remains below the ensemble mean at this threshold, but in the mature subset it is still competitive even at 1.00 in, especially at longer leads.

Six-panel line plots of CSI versus forecast lead time, with panels (a) CSI at 0.01 in, (b) CSI at 0.10 in, (c) CSI at 0.25 in, (d) CSI at 0.50 in, (e) CSI at 1.00 in, and (f) CSI at 2.00 in

Figure 2. Lead-time die-off of hourly-precipitation CSI across thresholds. Each panel shows CSI versus forecast lead for a different precipitation threshold. The final two-month subset (dotted lines) shows the strongest HRRRCast gains.

Six-panel line plots of threshold-dependent hourly-precipitation CSI, with the top row labeled Full seven-month archive and panels (a) 6h forecast, (b) 12h forecast, and (c) 18h forecast, and the bottom row labeled Final two-month subset with panels (d) 6h forecast, (e) 12h forecast, and (f) 18h forecast

Figure 3. Threshold-dependent hourly-precipitation CSI for 6-, 12-, and 18-h forecasts. The top row shows the full seven-month archive; the second row shows the mature recent subset. In the latest configuration, HRRRCast shows a clear improvement in precipitation skill up to about 3 in/h, and the performance advantage becomes more pronounced at 12- and 18-h lead times.

The additional fixed-lead threshold summaries (6-, 12-, and 18-h forecasts) reinforce this picture. Across these panels, the mature-subset curves show clearer HRRRCast improvement than the full seven-month aggregate, with the ensemble generally above HRRR across much of the threshold range and with the strongest separation at 12 and 18 h. The full-period curves are more mixed at some thresholds, especially at 6 h, but still indicate broadly competitive precipitation CSI and useful discrimination into the more extreme tail.

For near-surface thermodynamic and wind RMSE (Figure 4a–Figure 4d), HRRRCast shows robust skill gains for relative humidity, dewpoint, and 10-m wind speed across most lead times. Relative humidity RMSE is consistently lower for HRRRCast, and the two-month dashed subset indicates that the mature 3-km configuration further widens that advantage. Dewpoint RMSE shows a similar lead-time evolution, with HRRRCast maintaining lower error from the shortest leads through the day-1 range. Wind-speed RMSE improvements are likewise systematic, including the longer lead times where aggregate errors are largest. For 2-m temperature RMSE, the full-period aggregate is more mixed: HRRRCast is near parity to slightly better at several shorter and intermediate leads, while HRRR operational remains slightly lower at parts of the curve. Over the final month, however, HRRRCast has lower RMSE across all four near-surface variables, including 2-m temperature, suggesting that the mixed full-period temperature result is influenced by earlier near-real-time configurations rather than by the mature model alone.

Lead-time die-off verification for near-surface variables

Figure 4. Lead-time die-off verification for near-surface variables. Solid lines denote the full approximately seven-month evaluation window, and dashed lines denote the final two-month subset.

Vector wind bias diagnostics (Figure 4e–Figure 4f) indicate that HRRRCast substantially reduces the strong negative zonal-wind bias present in HRRR operational across most lead times in both the full-period and final-month summaries. For meridional wind, both systems stay closer to zero than in the zonal component, but HRRRCast generally remains less negatively biased through most of the forecast.

Together, these MATS results show that HRRRCast maintains clear reflectivity CSI gains at all thresholds, provides broadly competitive to improved precipitation CSI, and improves the background near-surface state (humidity, dewpoint, temperature, wind speed), especially in the mature final-month configuration.

In Appendix § 20, we provide a supplemental HRRR-analysis-based RMSE verification for variables not yet supported by the observational MATS pipeline (Appendix § 20.§ 2; Figure 17). We also show day-to-day variability and daily model cycles over a two-week window (Figure 21–Figure 20), along with a representative forecast example illustrating the newly added fields (Appendix § 20.§ 5; Figure 22).

Selected RMSE verification fields against HRRR analyses: (a) 80-m zonal wind (m s$^{-1}$); (b) 80-m merid. wind (m s$^{-1}$); (c) CAPE (J kg$^{-1}$); (d) CIN (J kg$^{-1}$); (e) 10-m zonal wind (m s$^{-1}$); (f) 10-m merid. wind (m s$^{-1}$); (g) 2-m temperature (K); (h) 2-m dewpoint (K); (i) Mean sea-level pressure (Pa); (j) Hourly precip. (mm); (k) Composite refl. (dBZ); (l) Visibility (m); (m) Total cloud cover (%); (n) Mid cloud cover (%); (o) Low cloud cover (%); (p) High cloud cover (%)

Figure 17. Selected RMSE verification fields against HRRR analyses.

Daily model cycle: composite-reflectivity CSI versus MRMS at multiple thresholds

Figure 20. Daily model cycle: composite-reflectivity CSI vs MRMS (multiple thresholds).

Daily model cycle: RMSE for selected surface variables vs METAR

Figure 21. Daily model cycle: RMSE for selected surface variables vs METAR.

Example HRRRCast forecast evolution for the nor’easter storm

Figure 22. Example HRRRCast forecast evolution for the nor’easter storm on 2026-09-26 at 1-, 3-, and 6-h forecast leads, showing hourly precipitation, reflectivity, CAPE, relative humidity, mean sea-level pressure, 500-hPa geopotential height, and total cloud cover.

Focused ensemble comparison against HREF

Here, we examine whether HRRRCast’s probabilistic products are competitive with an established ensemble forecast system. To address that question, we perform a focused MATS verification of HRRRCast ensemble performance relative to the operational High-Resolution Ensemble Forecast system (HREF) over a two-and-a-half-month period (2026-07-27 through 2026-10-09). HREF is a multi-model, convection-allowing ensemble that combines different operational forecast systems with varying initial conditions, model physics, and perturbation strategies, and it also includes several time-lagged members. Because the full HREF is still computationally expensive, it is produced on a 6-h cycle and provides probabilistic guidance for precipitation and severe-convection hazards. This comparison is narrower in scope than the HRRR-based verification and is limited to 3-h accumulated precipitation over a two-and-a-half-month period.

Lead-time verification for 3-h accumulated precipitation (top row of Figure 5) shows that HRRRCast has lower RMSE than HREF at every shared lead, with mean RMSE values of 1.627 and 1.787, respectively. The bias comparison is more nuanced: HREF maintains a persistent wet bias, whereas HRRRCast maintains a persistent dry bias of comparable magnitude. The spread diagnostic reveals a clear calibration contrast. HREF spread exceeds its RMSE at all leads, indicating an over-dispersive ensemble, whereas HRRRCast spread remains far smaller than its RMSE, indicating under-dispersion. A complementary CRPS comparison (bottom panel of Figure 5) also favors HRRRCast, consistent with its lower mean error and narrower ensemble distribution, although that advantage does not imply uniformly better threshold-specific calibration.

Focused HREF comparison panels: (a) RMSE; (b) Bias; (c) Spread; (d) Brier Score at 0.635 mm; (e) Brier Score at 7.62 mm; (f) Brier Score at 25.4 mm; (g) Reliability at 0.635 mm; (h) Reliability at 7.62 mm; (i) Reliability at 25.4 mm; (j) Time series of RMSE/CRPS over the full verification period

Figure 5. Focused HREF comparison for 3-h accumulated precipitation. 1st row: lead-time dependence of mean RMSE, mean bias, and ensemble spread for HREF operational and the HRRRCast ensemble. 2nd row: Brier Score lead-time curves for event thresholds of 0.635, 7.62, and 25.4 mm. 3rd row: corresponding reliability diagrams. Final two panels: RMSE and CRPS time series over the full verification period. Shaded bands indicate variability. HRRRCast has lower RMSE across the shared lead times; the spread panel shows HREF is more dispersive, while calibration differences depend strongly on event threshold.

Threshold-dependent probabilistic skill is more mixed. At 0.635 mm, HRRRCast has slightly lower Brier score at the shortest leads, but HREF is slightly better from 12 to 18 h. At 7.62 mm, HRRRCast is modestly better at nearly all shared leads, with only minimal separation at 15 h. At 25.4 mm, HRRRCast has the lower Brier score at every shared lead.

Reliability behavior (bottom row of Figure 5) explains this threshold dependence. At the lightest threshold (0.635 mm), HREF remains close to the one-to-one line, whereas HRRRCast transitions from mild underconfidence at the 25% bin to clear overconfidence at the 50% and 75% bins. At 7.62 mm, both systems are overconfident, but HREF stays closer to perfect reliability across the occupied bins. At 25.4 mm, the ranking reverses: HREF calibration degrades sharply, with its highest-probability bin dropping to an observed relative frequency of only about 0.16, whereas HRRRCast remains monotonic and reaches about 0.39. Overall, HREF is better calibrated for light-to-moderate 3-h precipitation, whereas HRRRCast is more reliable for the heaviest events.

We also compare ensemble calibration with a separately trained model optimized using a CRPS objective in Appendix § 20.§ 3 (Figure 19). The CRPS model is better calibrated on most variables than HRRRCast, whose diffusion-based ensemble remains somewhat under-dispersive. However, calibration of APCP and REFC remains challenging even for the CRPS-based model, highlighting the difficulty of rare-event probabilistic forecasting.

Ten-member spread-error ratio plots for diffusion and CRPS-optimized ensembles

Figure 19. Ten-member spread-error ratio as a function of forecast lead time for the diffusion and CRPS ensembles. The ideal value is 1.0. The diffusion model remains under-dispersive across many fields and lead times, while the CRPS model is substantially closer to calibrated spread and therefore better matches the spread-error criterion.

Ablation Studies

Having established near-real-time deterministic and ensemble performance, we next isolate two contributors to these gains: native-resolution training and moderate loss reweighting. These attribution analyses also clarify the scale dependence of HRRRCast skill, including stronger 40-dBZ performance when neighborhood-based FSS metrics are used.

The attribution analyses in the next two subsections use single-member forecasts. For clarity, we use the following model labels: HRRR 3 km denotes native operational HRRR output; HRRR 6 km denotes the same HRRR forecast field subsampled to 6 km for verification; HRRRCast 3 km denotes the native-grid HRRRCast model; and HRRRCast 6 km denotes the separately trained 6-km HRRRCast model, everything else being the same.

Attribution of skill gains: native-resolution training

Native-grid training provides consistent gains over the 6-km baseline and competitive reflectivity skill against operational HRRR, with the largest benefits at finer scales.

When the HRRR analysis is sub-sampled from its native 3-km grid to a coarser resolution for training, the absence of a proper anti-aliasing filter folds sub-grid energy back across the spectrum, corrupting fine-scale spectral structure in the training targets. A model trained on this data learns these aliasing-contaminated spectral characteristics rather than the physically consistent fine-scale structure of the native grid. Training HRRRCast directly on the native 3-km HRRR grid avoids these subsampling-induced aliasing artifacts. Consistent spectral evidence for this aliasing behavior and improved small-scale variance retention is shown in Figure 6 and Appendix § 20 (Figure 16).

Power spectral density comparison for composite reflectivity

(a) Composite reflectivity

Power spectral density comparison for convective inhibition

(b) Convective inhibition

Power spectral density comparison for surface visibility

(c) Surface visibility

Figure 6. Representative PSD comparisons illustrating the native-grid advantage. At the highest resolved wavenumbers, the 6-km curves commonly turn upward relative to the native 3-km HRRR analysis, consistent with subsampling-induced aliasing. Native-grid HRRRCast retains more small-scale variance and remains closer to the 3-km analysis across these variables.

Power spectral density comparisons for selected variables for 1h lead time

Figure 16. Power spectral density comparisons for selected variables for 1h lead time

Verification against observations using the fractions skill score (FSS) confirms substantial improvement at fine spatial scales. At a 6-km pooling window, native-grid HRRRCast improves markedly over the 6-km baseline across all reflectivity thresholds examined. For single-member forecasts, FSS is generally better than operational HRRR at 20 and 30 dBZ, while 40-dBZ skill is more nuanced, with near parity at intermediate leads and slight degradation at longer leads.

Figure 7 illustrates representative lead-time FSS behavior from the 3-km versus 6-km comparison set. In the top row, HRRRCast at native 3 km (solid red) remains above the 6-km HRRRCast baseline (dashed red) across lead times at both 20 and 30 dBZ, with the clearest separation at 30 dBZ; it also generally outperforms HRRR at both evaluation resolutions (black curves). At 40 dBZ (top-right panel), native-grid training still improves on the 6-km baseline, but the comparison against HRRR-3km narrows to near parity at intermediate leads and slightly favors HRRR at longer leads.

Fractions skill score lead-time curves comparing native-grid HRRRCast, 6-km HRRRCast, and HRRR

Figure 7. Fractions skill score (FSS) lead-time curves comparing native-grid HRRRCast (3 km), 6-km HRRRCast, and HRRR (each evaluated on both 3-km and 6-km grids). Native 3-km training improves HRRRCast skill at convection-permitting scales and retains benefits across reflectivity and 1-h precipitation thresholds. Top row: reflectivity thresholds (20/30/40 dBZ). Bottom row: 1-h precipitation thresholds (0.6/2.7/11.5 mm h−1^{-1}).

The corresponding 1-h precipitation curves in the bottom row show a similar overall pattern: native-grid training generally improves HRRRCast relative to the 6-km baseline. The clearest and most consistent gains appear at light-to-moderate thresholds (0.6 and 2.7 mm h−1^{-1}), although relative to HRRR-3km these curves are often near equal to slightly lower at mid-to-late leads. At the heaviest threshold (11.5 mm h−1^{-1}; bottom-right panel), deterministic HRRRCast-3km degrades substantially after the first few lead times.

To examine scale dependence directly, Figure 8 compares pooling radii from 6 to 90 km at fixed thresholds. For 40-dBZ reflectivity, single-member HRRRCast is near parity with HRRR at 6-km pooling, but clearly outperforms HRRR at

Pooling-radius dependence of native-grid benefits

Figure 8. Pooling-radius dependence of native-grid benefits. Top row: 40 dBZ reflectivity for 6-, 30-, and 90-km pooling windows. Bottom row: 1-h precipitation at 2.7 mm h−1\mathrm{mm}\,\mathrm{h}^{-1} for 6-, 30-, and 90-km pooling windows. At larger pooling windows, HRRR skill is relatively insensitive to evaluation grid spacing, while HRRRCast continues to show a native-resolution advantage.

larger pooling radii (30 and 90 km). A similar, though weaker, pattern appears for moderate 1-h precipitation (2.7 mm h−1^{-1}).

This behavior is consistent with the verification setup. HRRR’s FSS is nearly insensitive to evaluation grid spacing for pooling radii ≥30\ge30 km because the HRRR 6-km curve is simply a subsampled version of the same HRRR 3-km forecast field, whereas HRRRCast 6 km is a separately trained model. Even so, HRRRCast retains a persistent native 3-km advantage across pooling scales. This supports a real benefit from native-resolution training, although the 3-km versus 6-km HRRRCast gap still reflects both resolution and model-configuration differences.

Attribution of extreme-threshold gains: loss reweighting

This subsection shows that moderate, targeted loss reweighting improves high-threshold skill without destabilizing performance on common events.

As described in § 2.3, the sigmoid-weighted balanced MSE loss applies a capped sigmoid ramp that increases the event weight from approximately 1×1\times to 3×3\times for anomalies in the 4σ4\sigma–10σ10\sigma range, restricted to REFC and APCP. The weighting scheme is intentionally moderate—substantially below the weights of up to 800×800\times used in some nowcasting systems [see e.g., Shi et al., 2017, Jin et al., 2024, Cao et al., 2025]. The candidate formulations explored during loss design are summarized in Figure 15.

The 6-km tuning experiments in Figure 9 show that introducing sigmoid weighting consistently improves FSS over an unweighted loss, especially after the first few lead times where rare-event penalties become most important. This behavior appears in both reflectivity and 1-h precipitation verification. For the reflectivity thresholds, the 3×3\times configuration is generally the best-performing weighted variant, especially for the higher thresholds. For the precipitation thresholds, the 3×3\times and 5×5\times variants are very similar, indicating that increasing beyond moderate weighting provides little additional benefit. These tests also reinforced the need for moderation: stronger weighting can improve threshold-based skill while simultaneously making the forecast distribution less realistic if the loss growth is too aggressive. Based on this behavior, we retained the capped global sigmoid strategy for the production 3-km training to improve extremes while limiting degradation on frequent, lower-intensity events. Mechanistically, the sigmoid weighting improves high-threshold skill by redistributing optimization weight across the reflectivity spectrum: because values below 20 dBZ dominate the training distribution, uniform weighting tends to prioritize low-reflectivity skill and under-emphasize higher-reflectivity events. Increasing the loss weight above 20 dBZ rebalances this objective,

Representative 6-km HRRRCast tuning experiment for extreme-event weighting in the training loss

Figure 9. Representative 6-km HRRRCast tuning experiment for extreme-event weighting in the training loss. Top row: reflectivity thresholds (20/30/40 dBZ). Bottom row: 1-h precipitation thresholds (0.6/2.7/11.5 mm h−1\mathrm{mm}\,\mathrm{h}^{-1}). Weighted-loss variants improve FSS relative to the unweighted model and move performance closer to HRRR; for reflectivity, the 3×3 \times variant is generally best, while for precipitation the 3×3 \times and 5×5 \times curves are generally very similar, supporting the use of capped sigmoid weighting for extreme-event forecasting.

improving higher-reflectivity forecasts while preserving useful low-reflectivity behavior. The REFC 18-h forecast CDF comparison (Figure 15, right panel) is consistent with this mechanism and supports selecting the moderate 3×3 \times configuration over the more aggressive 5×5 \times option.

Storm-scale behavior

Storm displacement

Object-based and field-based displacement metrics consistently indicate improved storm placement by HRRRCast relative to HRRR.

We focus here on an object-based diagnostic that is directly interpretable for storm-placement behavior. Figure 10 shows the 40 dBZ results using two complementary object-based diagnostics: center-of-mass (COM) displacement and matched-pair centroid distance. Together, these metrics quantify bulk centroid placement and object-to-object placement after matching. Across lead times, both metrics are generally lower for HRRRCast than HRRR, indicating improved storm-location fidelity. The phase-correlation displacement metric (third panel), which is not object-based and instead estimates bulk translational offset from the full reflectivity field, shows the same ranking: HRRRCast generally has smaller displacement than HRRR. Similar results have also been reported in Smith et al. [ [20].

Storm-placement diagnostics at 40 dBZ

Figure 10. Storm-placement diagnostics at 40 dBZ (left to right): center-of-mass (COM) displacement, matched-pair centroid distance, and phase-correlation displacement. Lower values indicate better placement fidelity. The first two panels are object-based, whereas phase-correlation displacement is a field-based translational metric. Across panels, HRRRCast (red) is generally better than HRRR (black) across lead times.

This ranking is consistent with recent real-time Hazardous Weather Testbed (HWT) evaluations [6]. In the representative case shown in Figure 11, HRRRCast aligns the convective line more closely with MRMS than HRRR in deterministic reflectivity, and the ≥40\ge40 dBZ probability swath is likewise better co-located than REFS (the ensemble Rapid Refresh Forecast System). This qualitative case is consistent with the displacement metrics above and highlights the operational value of probabilistic severe-convection guidance.

Representative HWT real-time example showing deterministic composite reflectivity and probability guidance

Figure 11. Representative HWT real-time example from the 2026 Spring Forecasting Experiment [6] valid 2026-05-19 06:00 UTC, corresponding to a 30-h forecast initialized from the previous day’s 00Z cycle, comparing HRRR, HRRRCast, REFS, and MRMS observations. Top row shows deterministic composite reflectivity. Bottom row shows probability guidance for reflectivity ≥40\ge40 dBZ. In this case, HRRRCast places the main convective swath closer to the observed MRMS line than REFS, consistent with improved storm-placement behavior.

Storm structure

Despite placement gains, structure diagnostics reveal persistent morphology deficits, particularly at higher reflectivity thresholds and longer lead times.

To complement the storm-displacement diagnostics above, we evaluate storm structure with object-based metrics designed to isolate morphology (overlap, fragmentation, and object-size realism). Figure 12 summarizes representative thresholds and lead-time behavior using two minimum object area thresholds: 144 km² (4×44 \times4 grid cells) and 576 km² (8×88 \times8 grid cells). These thresholds bracket the effective resolution of HRRR, which is approximately 7Δx≈217\Delta x \approx21 km [8], corresponding to a feature area of roughly 441 km². Therefore, the 576 km² threshold is expected to filter the population to mesoscale convective system (MCS)-scale systems that HRRR can resolve. To avoid altering morphology-sensitive metrics, objects are defined directly from threshold exceedance regions using 8-connectivity, without pre-smoothing or convolution. Duda and Turner [2023] likewise minimized smoothing by using convolution radius of just 1.

Objective storm-structure metrics at 40 dBZ for two object-size filters

Figure 12. Objective storm-structure metrics at 40 dBZ shown for two object-size filters. Columns are (left to right): matched-object intersection over union (IoU; higher is better), fragmentation bias (fcst/ref; 1 is ideal), and forecast 90th-percentile object area. The top row uses a 144 km2^2 minimum-area threshold (small-object filter; 4×44 \times4 grid-cell minimum), and the bottom row uses 576 km2^2 (MCS-scale filter; 8×88 \times8 grid-cell minimum). Across both thresholds, HRRRCast 3 km remains over-fragmented while HRRR remains under-fragmented. The object-area contrast is scale dependent: at 144 km2^2, HRRRCast 3 km tends to underestimate object size; at 576 km2^2, HRRR substantially overestimates object size, while HRRRCast 6 km is closest to MRMS.

Matched-object Intersection over Union (IoU) at 40 dBZ (first column) lies mostly in the 0.15–0.20 range across both rows. HRRRCast 3 km is generally lower than HRRR, while HRRRCast 6 km is consistently higher than HRRRCast 3 km and is often closer to HRRR. The gap is real but modest, indicating that once objects are matched, overlap quality is broadly similar.

Fragmentation bias at 40 dBZ (middle column) tells a more definitive story. We define fragmentation bias as the ratio of forecast to observed object number density, that is, object count per unit covered area: FBfrag=(Nf/Af)/(No/Ao)FB_{\mathrm{frag}} = (N_f/A_f)/(N_o/A_o). Equivalently, since the mean object area is Aˉ=A/N\bar{A} = A/N, this can be written as FBfrag=Aˉo/AˉfFB_{\mathrm{frag}} = \bar{A}_o/\bar{A}_f. Thus, values above 1 indicate that forecast objects are too small on average (over-fragmented), values below 1 indicate that forecast objects are too large or too merged (under-fragmented), and a value of 1 indicates the correct mean object size. HRRR consistently under-fragments relative to MRMS, with bias generally around 0.65–0.75, meaning it produces fewer but larger objects. HRRRCast 3 km, by contrast, is over-fragmented at both thresholds: the effect is stronger in the 144 km2^2 row (often about 1.3–2.0 after the first few leads) and milder in the 576 km2^2 row (roughly 1.0–1.4). HRRRCast 6 km is intermediate, remaining near neutral to mildly over-fragmented (about 1.0–1.2).

Comparing rows in Figure 12 clarifies how these structure biases depend on object scale. The largest row-dependent change appears in the object-area column: at 144 km2^2, HRRRCast 3 km yields smaller 90th-percentile (P90) object areas (consistent with many small fragments), whereas at 576 km2^2 HRRR yields substantially larger P90 areas than MRMS (∼4500\sim4500 km2^2 vs. ∼2500\sim2500 km2^2), consistent with merged, oversized MCS-scale objects. HRRRCast 6 km is closest to MRMS in the 576 km2^2 panel. Together, the row-wise comparison reinforces a complementary error structure: HRRRCast 3 km is too fragmented for smaller objects, while HRRR is too merged at MCS scales.

The representative HWT case in Figure 11 shows the same mixed behavior: good line placement but more fragmented high-intensity reflectivity structure than MRMS. This pattern is consistent with an important limitation of the current setup—predominantly 1-h supervision, which leaves the model to infer substantial sub-hour storm evolution and may degrade fine-scale convective organization.

Discussion: Storm structure in AI weather models

In this section, we discuss three candidate mechanisms behind storm-structure limitations in current AI regional weather models: the coarse autoregressive time step, the choice of training target, and the use of deterministic versus generative training objectives.

Base autoregressive cadence. Operational convection-allowing NWP models (e.g., HRRR) integrate prognostic equations with internal time steps on the order of seconds, enabling frequent dynamical adjustment of storm organization. Most current AI regional models, by contrast, evolve the atmosphere at ≥1\ge1-h intervals and must therefore infer the full sub-hour storm evolution from a single autoregressive step rather than resolving it. Within this fixed cadence, subhourly states can be generated either by direct lead-time-conditioned queries or by short-step autoregressive rollout: the latter can improve short-time continuity but may accumulate stepwise error with depth. High-frequency AI stepping is possible, but stable long-range short-step composition remains technically difficult. StormScope [16] uses 10-min stepping for nowcasting but reverts to 1-h stepping for longer leads, whereas STRATA [Hu et al., 2026] demonstrated stable 10-min rollouts up to 24 h only with large model capacity (∼330\sim330M parameters), autoregressive fine-tuning, and multiple stabilization measures, including local spatial processing, de-aliasing, overlapping-tile blending, and physical/stability constraints. On the other hand, physics-based CAMs can maintain much shorter internal time steps over longer integrations, which likely contributes to stronger long-lead storm morphology, though they also accumulate displacement error (§ 6.1).

We run an experiment with an experimental HRRRCast subhourly-inference configuration to show that subhourly inference is not equivalent to subhourly supervision. Figure 13 shows a representative 3-h case where HRRRCast is queried diagnostically at fractional lead times rather than run as a separately trained subhourly forecast system; these queries yield smooth frame-to-frame evolution, yet still exhibit storm over-connectivity and reduced fine-scale convective realism. In this example, HRRRCast shows stronger band continuity than both MRMS and HRRR, but with broader, more diffuse areal coverage and a missed southern separation that contributes to local over-connection. MRMS retains stronger cell-scale intermittency, whereas HRRR appears more segmented with sharper compact cores.

Representative subhourly composite-reflectivity evolution

Figure 13. Representative subhourly composite-reflectivity evolution for a 3-h forecast at 2024-05-01T00 comparing diagnostically queried HRRRCast fields with MRMS and HRRR at 15-min intervals. HRRRCast is temporally smooth but shows broader, more connected reflectivity than MRMS, whereas HRRR appears sharper and more segmented.

More broadly, there are at least three routes to subhourly AI forecasts: explicit subhourly state evolution, fractional lead queries from a lead-time-conditioned model as in HRRRCast, and temporal downscaling of a coarser forecast (e.g., Ingstad et al. [2026]). These approaches differ mainly in how fine-time structure is generated and constrained. The first and third rely on explicit subhourly supervision, whereas HRRRCast can generate temporally smooth intermediate states and some fine-scale variability without directly constraining those states with subhourly observations or analyses. For the supervised approaches, forecast or observational training data (e.g., Ingstad et al. [2026], Hu et al. [2026], Pathak et al. [2026]) are likely needed, rather than analyses or reanalyses alone, to construct temporally consistent fine-time forecast sequences. A fourth, fundamentally different route is physics-AI hybrid evolution, as in WeatherGFT [Xu et al., 2024], in which a fixed numerical PDE operator advances the atmospheric state at fine internal time steps and a parallel

neural branch corrects unresolved processes and accumulated error, enabling subhourly states without direct subhourly supervision.

Training target. When an AI model is trained against NWP model output—whether analyses or short-range forecasts—it may learn to reproduce spatial smoothing, effective-resolution limits, and structural artifacts the NWP system itself introduces. Training directly against radar observations removes this layer of NWP-induced distortion. Consistent with this mechanism, StormScope results show improved morphology under MRMS-supervised training relative to NWP-supervised targets [16]. HRRRCast trains on HRRR analyses, which are internally consistent but still inherit NWP effective-resolution and spectral characteristics; radar-supervised or blended observation/analysis targets are therefore a promising path for improving fine-scale storm realism.

Deterministic versus generative objective. Even with fine spatiotemporal resolution, deterministic models trained under pointwise losses tend toward conditional means, producing spatially blurred output. This smoothing occurs because pointwise losses reward averages over plausible realizations rather than any single sharp realization. Generative training objectives can mitigate this and recover sharper textures (e.g., diffusion or token-based generative approaches), but objective choice alone is not sufficient [4, 10, 2]. In practice, any benefit from a generative objective remains coupled to cadence and supervision target: 1-h stepping limits sub-hour evolution fidelity, and HRRR-analysis supervision constrains the distribution the model learns.

These three factors expose a fundamental tension in current AI weather modeling: models operating at coarse autoregressive cadences (≥1\ge1 h), such as HRRRCast, can achieve competitive storm placement out to 48 h but structurally under-resolve fine-scale convective organization, while short-step AI models that better resolve morphology are confined to the nowcasting regime by rapid error accumulation (≲6\lesssim6 h; e.g., Pathak et al. 2026). No present AI strategy simultaneously achieves both goals at extended convective leads. Absent a fundamental advance that simultaneously enables stable long-range autoregression with small timesteps and faithful representation of fine-scale convective dynamics, degradation of storm structure at extended leads (≥12\ge12 h) remains a fundamental limitation of current AI weather models rather than merely an architecture-specific deficiency.

Additionally, given the observed displacement–structure contrast between AIWP and NWP models, it is important to motivate a balanced evaluation framework in which storm-structure inspection is paired with storm-displacement metrics and other object-based and scale-aware diagnostics (e.g., FSS across thresholds/scales, neighborhood methods, and probabilistic reliability), so that model differences are treated fairly and operationally relevant skill is not underestimated.

Conclusions

This study shows that regional AI weather forecasting can become more operationally useful through targeted treatment of extreme events, native-resolution training, expanded output coverage, and sustained near-real-time verification. The principal findings are as follows.

Native 3-km training substantially improves convective-scale skill. Training directly on the native 3-km HRRR grid, rather than sub-sampled 6-km data, avoids the aliasing effects incurred in the earlier 6-km configuration. Fractions skill score verification shows consistent gains over the 6-km baseline across reflectivity and precipitation thresholds, with 20- and 30-dBZ reflectivity generally better than HRRR and 40-dBZ behavior that is near parity at 6-km pooling but better at larger pooling windows.

Sigmoid-weighted loss improves extreme-event forecasting. Applying a capped sigmoid weighting schedule selectively to composite reflectivity and hourly precipitation improves high-threshold FSS relative to an unweighted baseline. Across tuning experiments, the 3×3 \times cap provides the best trade-off between rare-event gains and preserving common-event behavior.

Expanded observational verification confirms broad skill improvements. Seven months of near-real-time observational verification using MATS demonstrate that HRRRCast ensemble means outperform operational HRRR in RMSE for relative humidity, 2-m dewpoint, 10-m wind speed. HRRRCast also outperforms HRRR in CSI at multiple reflectivity thresholds. For hourly precipitation CSI, HRRRCast is competitive through 1.00 in and shows especially strong improvement in the final two-subset of the near-real-time archive.

Storm-scale evaluation reveals a displacement–structure trade-off. Across § 6, displacement diagnostics generally favor HRRRCast, while structure diagnostics show persistent fragmentation and over-connectivity biases, consistent with limits imposed by hourly supervision cadence and training-target characteristics; this distinction is illustrated in Figure 13, where temporally smooth fractional lead-time inference can still coexist with storm-structure limitations.

Overall, HRRRCast shows clear skill gains from native-resolution training and targeted loss design, while fine-scale storm morphology remains the main outstanding limitation.

Future Directions

Several avenues for further improvement are identified by the current verification results and methodological choices.

Sub-hourly training cadence. An important limitation in this context is training supervision rather than model inference capability. HRRRCast can generate forecasts at sub-hourly intervals if needed because lead time is provided explicitly as a model input. However, training is currently dominated by hourly targets, which likely imposes a ceiling on storm-scale morphology fidelity relative to physics-based models that integrate at sub-minute time steps. As discussed in the storm-structure section, training with 10-minute or finer targets could provide a direct test of whether finer temporal supervision improves storm-structure realism.

Broader variable coverage in observational verification. MATS-based observational verification is currently limited to a subset of HRRRCast outputs. Extending MATS verification to cloud cover, visibility, CAPE, CIN, and 80-m wind variables would enable independent benchmarking of the full expanded output set beyond the HRRR-analysis-relative RMSE diagnostics presented here.

Ensemble calibration. Future work will focus on improving ensemble spread and threshold-dependent probabilistic calibration, especially for precipitation and other rare-event variables.

Acknowledgment

The authors gratefully acknowledge support from the Software Environment for Novel Architectures (SENA) effort at GSL under grant NA24OARX405C0022. Additional support for the first, third through sixth, and ninth authors was provided by NOAA cooperative agreement NA22OAR4320151 through the Cooperative Institute for Earth System Research and Data Science (CIESRDS). Additional support for the third and tenth authors was provided by NOAA cooperative agreement NA24OARX432C0007-T1-01. This research used resources of the NOAA Research and Development High Performance Computing System (RDHPCS; http://rdhpcs.noaa.gov). We would like to thank Liaofan Lin for reviewing the manuscript, David Dowell and Jeffrey Duda for helpful discussions.

Disclaimer

The statements, findings, conclusions, and recommendations are those of the author(s) and do not necessarily reflect the views of NOAA or the U.S. Department of Commerce.

Availability Statement

HRRRCast has been running in near-real-time for nearly one year. Forecasts are initialized with HRRR analyses, disseminated publicly via the NOAA-GSL AWS bucket, archived on NOAA HPSS, and visualized through DESI. Code, data access, visualization, and verification resources are available at:

  • https://github.com/NOAA-GSL/HRRRCast-live/tree/master

  • https://noaa-gsl-experimental-pds.s3.amazonaws.com/index.html#HRRRCast

  • https://sites.gsl.noaa.gov/desi

  • https://gsl.noaa.gov/mats

Methods and Data Details

Variable Set

The variable selection follows a utility-first expansion strategy similar to recent operational AI-forecast model updates [15]; we aim to cover as much of the operational forecast space as possible while preserving training consistency. Practically, this means retaining a physically complete atmospheric core, then adding surface and integrated fields that are valuable for hazardous-weather interpretation and downstream applications, but only when they are consistently available and quality-controlled across the multi-year archive.

Table 1 summarizes the full HRRRCast variable configuration used for training. The dynamic meteorological fields are treated as prognostic variables (ingested and predicted autoregressively), while static fields (terrain height and land mask) are included as fixed conditioning inputs.

VariablesHRRR levels (hPa)GFS levels (hPa)Loss weight
Atmospheric variables:
Temperature (TMP)200, 300, 350, 400, 450, 500, 550, 600, 650, 700, 750, 800, 825, 850, 875, 900, 925, 950, 975, 1000250, 500, 850, 10001.0 per variable distributed among levels
Geopotential height (HGT)
U component of wind (UGRD)
V component of wind (VGRD)
Vertical velocity (VVEL)
Specific humidity (SPFH)
Surface/Integrated variables
Composite reflectivity (REFC)IntegratedIntegrated1.0
2-meter temperature (T2M)2m AGL2m AGL1.0
2-meter dewpoint temperature (D2M)2m AGL2m AGL0.7
10-meter U wind component (UGRD10M)10m AGL10m AGL0.3
10-meter V wind component (VGRD10M)10m AGL10m AGL0.3
80-meter U wind component (UGRD80M)80m AGL80m AGL0.1
80-meter V wind component (VGRD80M)80m AGL80m AGL0.1
Total cloud cover (TCDC)Entire atmo.Entire atmo.0.3
High cloud cover (HCDC)High cloudHigh cloud0.1
Mid cloud cover (MCDC)Mid cloudMid cloud0.1
Low cloud cover (LCDC)Low cloudLow cloud0.1
Visibility (VIS)SurfaceSurface0.3
Accumulated precipitation (APCP)Surface (1-h acc.)Surface (6-h acc.)1.0
Cloud ceiling height (HGTCC)N/AN/A0.3
Convective available potential energy (CAPE)SurfaceSurface0.3
Convective inhibition (CIN)SurfaceSurface0.2
Mean sea level pressure (MSLET or PRMSL)Sea levelSea level0.1
Surface pressure (PRES)SurfaceSurface0.02
Static variables
Orography height (OROG)SurfaceN/AN/A
Land mask (LAND)SurfaceN/AN/A

Table 1. HRRRCast input variables and pressure levels.

The atmospheric core consists of six 3D variables (TMP, HGT, UGRD, VGRD, VVEL, SPFH) on 20 HRRR pressure levels from 1000 to 200 hPa, with denser sampling in the lower troposphere (1000–800 hPa) to better resolve boundary-layer and convective structure. GFS conditioning for these 3D variables is provided on four standard levels (1000, 850, 500, and 250 hPa), supplying large-scale synoptic context without requiring full vertical matching often used in nested AI models such as [20].

The surface/integrated block includes near-surface thermodynamics and winds, cloud-cover fractions, visibility, precipitation, reflectivity, and instability diagnostics. The loss-weight column reflects practical forecast priorities: highest weights are assigned to core-impact variables such as composite reflectivity, hourly precipitation, and 2-m temperature; intermediate weights are used for dewpoint, cloud/visibility and CAPE-like fields; and lower weights are assigned to noisier or lower-priority channels (e.g., 80-m winds and pressure fields) to stabilize multi-target training. For the 3D atmospheric variables, the listed 1.0 weight is distributed across levels with larger weight near the surface [1].

Unlike recent work that applies variable-specific bounding layers (e.g., ReLU/HardTanh) to enforce output ranges [15], HRRRCast does not hard-constrain the network head. Instead, we clip in normalized space using variable-wise min/max limits estimated from the training dataset. This design choice is tied to the model formulation: the network predicts residual tendencies that are added to the current state xtx_t through an output skip connection, and several targets are represented in transformed space (log-based transforms and z-normalization). In this setting, a simple bounded activation at the network output constrains only the residual increment, while the final prediction after the additive update (xt+Δxx_t + \Delta x) can still move outside the intended range. That makes robust, physically consistent range enforcement difficult to impose uniformly across variables and can suppress needed large corrections during autoregressive rollout.

Beyond the prognostic fields, HRRRCast computes several classes of post-processed diagnostic variables from the native model output. These include near-surface thermodynamic diagnostics such as 2-m relative humidity, 2-m potential temperature; column and precipitation diagnostics such as precipitable water, conditional rain or freezing-rain masks, and warm- or cold-layer depth; wind diagnostics such as gust estimates and column-maximum wind; severe-convection diagnostics such as low-level shear, storm-relative helicity; and freezing-level diagnostics. This post-processing step broadens the usable forecast variable set without requiring the network to predict every downstream hazard-oriented field directly.

Dataset Curation and Harmonization

The dataset is expanded to nearly a decade of HRRR analyses and GFS forecasts, spanning October 2016 to October 2025. This period covers three HRRR versions (v2, v3, and v4) and three GFS configurations (pre-FV3 spectral-element GFS, v15, and v16). Incorporating multiple model generations introduces nontrivial distributional shifts arising from changes in physical parameterizations, data assimilation systems, and – most notably for GFS – the transition to a fundamentally different dynamical core.

A separate challenge arises from temporal inconsistency in GFS outputs. Versions prior to v16 provide forecasts at 3-hour intervals in the first 120 hours. This reflects a design choice tied to downstream applications (e.g., boundary condition updates for limited-area models such as RAP), rather than a deficiency in model fidelity. However, our training setup requires temporally aligned HRRR-GFS pairs at hourly resolution to construct a consistent learning target for the HRRR emulator. To achieve this temporal alignment, we apply interpolation to the 3-hourly GFS data.

Missing variables present an additional complication. HRRR v1 is excluded due to the absence of several key predictors required by the model. While older GFS versions also lack certain variables, this is less critical because HRRR serves as the primary source of high-resolution target fields; missing predictors in GFS can be partially mitigated through learned mappings, whereas missing targets in HRRR would directly degrade supervision. Specifically, pre-v16 GFS data lacks the following variables: REFC, TCDC, LCDC, MCDC, HCDC, VIS, and HGTCC.

We address these three issues—distributional shifts, missing variables, and coarse temporal resolution—using a unified strategy: explicit conditioning on model version via temporal metadata. Specifically, the network is provided with two binary indicator masks denoting whether a sample originates from HRRR v3 or v4, allowing it to implicitly infer v2. Another mask is used to indicate older GFS versions that use a spectral dycore instead of FV3. This approach enables the model to learn version-dependent corrections without requiring separate training pipelines. These auxiliary conditioning channels along with others for encoding temporal phase, and rollout-specific context are shown in Table 2.

CategoryChannelsPurpose
Seasonal/diurnal phasesin⁡(doy)\sin(\mathrm{doy}), cos⁡(doy)\cos(\mathrm{doy}), sin⁡(hour)\sin(\mathrm{hour}), cos⁡(hour)\cos(\mathrm{hour})Encodes the annual and daily cycle.
HRRR version masksv3 mask, v4 maskIdentifies the HRRR version.
GFS version maskv15 maskDistinguishes older GFS configurations.
Forecast controlslead time, diffusion stepForecast conditioning.

Table 2. Auxiliary conditioning channels.

While these are the primary issues encountered in constructing the expanded dataset, additional challenges are expected if native model-level (rather than pressure-level) data are used, because of inconsistencies in vertical coordinates and evolving model configurations.

Accumulated Precipitation and Transforms

HRRRCast predicts only the last-hour accumulated precipitation at each lead time. For example, for an 18-h forecast, the model predicts the 17–18 h accumulation at lead 18 h. Longer accumulations are reconstructed by summing hourly forecasts.

In the near-real-time semi-operational setup, APCP is initialized by the previous cycle’s 1-h forecast accumulation (F01) since it is trained as a prognostic variable; however, it may be more convenient to design the network to treat accumulated variables diagnostically. Additionally, we apply logarithmic transforms to precipitation and other highly skewed variables including visibility, CAPE, CIN, and cloud-ceiling height, followed by z-normalization. The implemented transforms are

log⁡(1+max⁡(x,0))for nonnegative variables,\log(1+\max(x,0)) \quad\text{for nonnegative variables,}

and

sign⁡(x)log⁡(1+∣x∣)for CIN.\operatorname{sign}(x)\log(1+\lvert x\rvert) \quad\text{for CIN.}

Hard clipping based on min/max limits computed from the training dataset is applied during rollout to reduce unphysical drift.

Training at Native 3-km Resolution

Training at the native HRRR resolution (1059×17991059 \times1799) increases data volume by roughly 4×4 \times compared to the prior 6-km setup. To manage this, normalized data are stored as float16 for I/O efficiency, while mixed-precision training is used with selective float32 compute in sensitive layers.

Training uses TensorFlow distributed data parallelism, with the effective global batch size scaling as the per-node batch size times the number of MPI workers. For the native-grid production run, we trained on four nodes with two NVIDIA H100 GPUs per node (eight H100 GPUs total). Completing training over the full 9-year dataset took approximately 1.5 months of wall-clock time.

For the native 3-km training, we use the AdamW optimizer with the learning-rate schedule that uses a 1-epoch warmup followed by cosine annealing over a 100-epoch run (see left panel of Figure 14). The learning rate ramps to a peak value of 2×10−32 \times10^{-3} at the end of warmup and then decays smoothly to nearly zero, reaching approximately 10−610^{-6} by the end of training. In addition, version-dependent sample weights are applied during training to gradually de-emphasize older HRRR versions while preserving their contribution to the expanded historical dataset. Figure 14 summarizes both the learning-rate schedule and the epoch-dependent HRRR version-weight schedule used in this study.

Learning-rate schedule and HRRR version weights

Figure 14. Left: learning-rate schedule used for native 3-km training (1-epoch warmup to a peak of 2×10−32 \times10^{-3} followed by cosine annealing over 100 epochs). Right: epoch-dependent HRRR version weights used during training. Samples from the newest HRRR version (v4) retain the largest weight, while older versions (v3 and especially v2) are progressively down-weighted. The weights are normalized so that all HRRR versions continue to contribute while emphasizing the most recent model climatology; the schedule keeps the total version weight fixed each epoch (v2+v3+v4 = 3) so the loss scale remains comparable over training, while using smooth decay to avoid over-dominating updates with v4 alone.

Architecture Details

Relative to the original HRRRCast model [1], the core network architecture was updated in several ways to improve skill at native 3-km resolution while preserving numerical stability during training.

Model configuration and memory optimization

As in the original HRRRCast model Abdi et al. [ [1], the model follows a U-Net-style encoder–processor–decoder architecture with skip connections, built entirely from 3×33 \times3 residual blocks.

The full stage configuration is

resolution=[3 km,6 km,12 km,24 km,12 km,6 km,3 km],blocks=[2,3,4,14,4,3,2],filters=[288,256,256,256,256,256,288].\begin{aligned} \texttt{resolution} &= [3\ \text{km}, 6\ \text{km}, 12\ \text{km}, 24\ \text{km}, 12\ \text{km}, 6\ \text{km}, 3\ \text{km}],\\ \texttt{blocks} &= [2, 3, 4, 14, 4, 3, 2],\\ \texttt{filters} &= [288, 256, 256, 256, 256, 256, 288]. \end{aligned}

The resulting model has approximately 50 million trainable parameters.

To make native-grid training tractable, we use gradient checkpointing (recomputation) in selected preprocessing, encoder, decoder, and output-head stages. We also use spatially grouped convolutions (2×22 \times2 tiling with overlap) in highest-resolution stages. Overlap-and-crop stitching preserves equivalence to full-domain same-padding convolutions while reducing memory pressure.

Because training uses normalized float16 data and mixed precision, numerically sensitive operations are explicitly computed in float32 (e.g., normalization, key dense/attention operations, and selected residual/output operations) to avoid instability.

Input preprocessing branches

We introduced dedicated input preprocessing branches before the main encoder–decoder path so that HRRR, GFS, and (for diffusion training) noised-HRRR channel groups are processed by separate residual stems and then fused. This design allows each input type to undergo specialized feature extraction before entering the shared encoder, improving the model’s ability to integrate multi-source information.

Attention mechanism upgrade: CBAM

We replaced the squeeze-and-excitation (SE) blocks used in the original model with Convolutional Block Attention Modules (CBAM) [22]. While SE blocks provide channel-wise attention that adaptively recalibrates feature responses by modeling inter-channel relationships, they do not explicitly capture spatial attention patterns. CBAM extends this capability by sequentially applying both channel and spatial attention mechanisms, enabling the network to focus on both what features are important (channel attention) and where they are located (spatial attention).

The CBAM module consists of two sequential sub-modules: a channel attention module that uses both average-pooled and max-pooled features to generate channel-wise weights, followed by a spatial attention module that uses channel-pooled features to produce a 2D spatial attention map. This sequential refinement is applied at multiple scales throughout the encoder-decoder architecture, allowing the network to adaptively modulate both feature channels and spatial locations at each resolution level.

Sigmoid-Weighted Loss: Design Details

Prior work has explored similar strategies in the nowcasting context, with some published schedules applying weights as large as 800× for the most extreme events [19, 13, 18, 5]. Our schedule is intentionally moderate to preserve skill on common events while improving performance at high thresholds.

The same weighting approach can in principle be applied to other forecast variables with skewed or heavy-tailed behavior, but weighting is restricted here to REFC and APCP because these variables drive the largest practical deficits in operational utility. To validate training stability and tune the weighting schedule, experiments were first conducted on the 6-km model before their adoption in the 3-km training.

This design choice should be interpreted in the context of domain size and sampling. Over a CONUS-scale training domain, most grid points and times correspond to weak or non-convective conditions, so extreme reflectivity and heavy precipitation events form a small fraction of the optimization signal. In that regime, explicit reweighting helps counteract the dominance of non-storm examples and preserves sensitivity to rare, operationally critical extremes.

By contrast, models trained on smaller, storm-focused domains (for example, WoFS-like windows centered on active convection) can have a much higher effective event density. When the training distribution is already concentrated on convective states, specialized extreme-event weighting may be less necessary to recover strong high-reflectivity skill. This interpretation is consistent with WoFSCast results [9], which are trained on storm-focused windows rather than broad CONUS domains and therefore do not face the same class-imbalance regime (a large background “sea of zeros”) that motivates stronger imbalance correction in our setup. Our weighting strategy can be viewed as a partial solution to extreme-weather prediction difficulties, and newer benchmarks such as Extreme Weather Bench [14] help the community measure and improve progress on this challenge.

To implement this loss design, we evaluated several candidate weighting functions in early experiments. A softplus-based weighting function was initially preferred, but its rapidly increasing tail encouraged unrealistically large reflectivity and precipitation values in the generated forecasts. We therefore replaced it with a capped sigmoid family that still emphasizes the targeted extreme-event range while preventing uncontrolled growth in the loss assigned to the most intense samples. Figure 15 summarizes the candidate weighting functions explored during this tuning process; we selected a global sigmoid formulation with a maximum weight of 3×3 \times, which provided the best balance between stronger extreme-event emphasis and stable forecast amplitudes. The REFC CDF comparison at 18 h (right panel of Figure 15) further reinforced this decision: the 3×3 \times configuration stayed substantially closer to the baseline distribution than 5×5 \times, while still shifting the tail more than 1×1 \times.

Near-Real-Time Configuration

The near-real-time HRRRCast system runs hourly on the NOAA Research and Development High Performance Computing System (RDHPCS). It produced a 9-member ensemble forecast for most of the near-real-time archive, but since July 1, 2026, it has produced an 18-member ensemble forecast. Extended 48-h forecasts are initialized at 00, 06, 12, and 18 UTC, while 18-h forecasts are generated at all other initialization times. Model inference is distributed across three 94-GB NVIDIA H100 GPUs. Total wall-clock compute time is approximately 48 min for the 18-h cycles and 1 h 56 min for the extended 48-h cycles. Final model output is delivered in standard GRIB2 format, and the full list of core output variables is provided in Table 1.

Ensemble members start from the same initial state but differ through the diffusion noise used to initialize the forecast [1]. To add diversity and avoid ensemble collapse, the GFS boundary-condition inputs are phase-shifted for each member; this shift is zero for the first 12 forecast hours and then increases by 1h in 12-hour blocks. The diffusion noises are also correlated across forecast hours, following common video diffusion practice, to avoid jumpiness during rollout.

Additional Verification and Sensitivity Results

Power spectral density comparisons

Figure 16 compares isotropic power spectral density (PSD) across selected near-surface and storm-related variables. Across variables, all forecast systems show the expected reduction in variance toward smaller scales relative to the analyses, with consistently shorter small-scale tails in the 6-km curves than in their 3-km counterparts. At the highest resolved wavenumbers, the 6-km spectra also commonly turn upward relative to the 3-km HRRR-analysis reference, a pattern consistent with aliasing introduced by subsampling. This signal appears not only in 6-km HRRRCast and HRRR forecasts, but also in the 6-km HRRR-analysis curves, indicating that subsampling alone can imprint high-wavenumber distortion.

These PSD diagnostics reinforce the rationale for native 3-km training: avoiding 6-km subsampling better preserves physically meaningful small-scale variance and limits aliasing artifacts before they are learned by the model. Consistent with this, HRRRCast generally stays closer than HRRR to the 3-km HRRR-analysis spectra at intermediate wavelengths, especially for reflectivity and near-surface winds, indicating improved mesoscale structure representation rather than only large-scale amplitude correction.

Supplemental verification against HRRR analyses

The near-real-time verification in § 4.1 provides robust observational benchmarking for variables with established MATS infrastructure (surface thermodynamic and wind fields, reflectivity, and precipitation). However, MATS verification has not yet been implemented for all variables included in HRRRCast, particularly the expanded set of diagnostic fields and cloud variables introduced to support hazardous-weather applications. To assess these newly added variables, we provide supplemental verification against HRRR analyses using RMSE as the primary metric.

Figure 17 shows representative RMSE curves for selected 2D diagnostics spanning low-level winds and convective metrics (80-m wind components, CAPE, and CIN), precipitation and visibility, and cloud-cover fields at multiple levels. The HRRRCast single member is competitive with HRRR across most fields, despite being a single deterministic trajectory from a diffusion model, which is known to sacrifice RMSE performance in favor of better probabilistic skill [17]. On the other hand, the HRRRCast ensemble is the best performer across nearly all fields; this archive uses 9 members for most cycles, expanding to 18 members after July 1, 2026. However, it is expected that a single member deterministic model trained with RMSE objective should outperform the HRRRCast ensemble in RMSE, as was the case in the previous version [1].

Restricting the RMSE comparison to the synoptic-cycle subset for which HREF was available (00, 06, 12, and 18Z; Figure 18) yields the same broad conclusion. The HRRRCast ensemble remains the lowest-RMSE system for nearly all variables shown, including dewpoint, precipitation, CAPE, CIN, and the mid-, low-, and high-cloud-cover fields, while HREF generally improves upon operational HRRR and often also upon the single-member HRRRCast forecast. The main exception is 2-m temperature, for which operational HRRR has the lowest RMSE over this matched subset, with the HRRRCast ensemble remaining competitive and HREF lying in between the ensemble and the single-member AI forecast. Overall, the matched HREF subset indicates that the ensemble benefit in HRRRCast is not simply an artifact of the broader sample: even when evaluated only at cycles where an operational convection-allowing ensemble exists, HRRRCast retains a consistent RMSE advantage across most of the expanded diagnostic variable set.

Selected RMSE verification fields for 2-m temperature, 2-m dewpoint, hourly precipitation, CAPE, CIN, mid cloud cover, low cloud cover, and high cloud cover

Figure 18. Selected RMSE verification fields for the subset of cycles when HREF was available (00, 06, 12, and 18Z), comparing HRRRCast, operational HRRR, and HREF on a matched sample.

Calibration comparison against a CRPS-trained baseline

To isolate the calibration trade-off, we compare the HRRRCast diffusion ensemble with a separate model trained with a Continuous Ranked Probability Score (CRPS) objective and without diffusion. The CRPS model follows the approach in Alet et al. [ [3]: it uses a single global noise vector of size 32, generates two ensemble members during training, and produces an almost-fair CRPS estimate. The spread-error ratio (SER), defined as the ratio of ensemble spread to RMSE, is near 1 for well-calibrated uncertainty and substantially below 1 for under-dispersion. Figure 19 shows a consistent pattern: the CRPS-trained ensemble remains closer to the ideal value of 1 for most variables and across most lead times, whereas the diffusion ensemble is systematically under-dispersive. This behavior is expected given the optimization target: CRPS-based training explicitly rewards calibrated ensemble distributions, so the model learns to produce more realistic spread relative to forecast error.

This result does not imply that the CRPS ensemble is uniformly preferable for our application because ensemble calibration is not the sole objective. The diffusion model remains attractive because its denoising formulation preserves sharper, more coherent storm structures and mesoscale gradients, which are especially important for reflectivity, precipitation, and hazardous-weather fields. Interestingly, even the CRPS-trained ensemble remains under-dispersive for the most important variables, namely RFC and APCP, indicating that rare, heavy-tailed convective regimes remain difficult to calibrate even with a CRPS objective. The CRPS-trained ablation does improve ensemble calibration overall, but incorporating CRPS-like objectives into a diffusion model is non-trivial because diffusion training operates on denoised samples rather than full forecast trajectories.

Timeseries diagnostics

Here, we plot the daily model cycle for an approximately two-week period (2026-08-11 to 2026-08-24), with forecasts initialized at 12Z each day and verified through the subsequent 24-h forecast window.

In composite-reflectivity CSI (Figure 20), the HRRRCast ensemble mean remains above HRRR operational at 20, 30, and 40 dBZ in the aggregate and through many event cycles, while preserving similar event timing. Across thresholds, the HRRRCast ensemble-mean advantage is expressed mainly in amplitude: event peaks and post-peak shoulders are generally higher than HRRR, especially at 30 dBZ. At 40 dBZ, deterministic HRRRCast member 0 stays close to HRRR on many cycles, while the ensemble mean remains systematically higher, indicating that most of the high-threshold gain is concentrated in the ensemble signal rather than in a single deterministic trajectory.

Figures 21–20 show that HRRRCast broadly preserves the day-to-day and diurnal phase of HRRR operational while changing error amplitude in a variable- and threshold-dependent way. In the surface RMSE series (Figure 21), HRRRCast is consistently lower for 2-m relative humidity, 2-m dewpoint, and 10-m wind speed, and is also modestly lower for 2-m temperature in the aggregate. The largest separations occur during recurring daily high-error periods, where HRRRCast damps several of the strongest HRRR peaks.

Illustrative forecast example

To illustrate the forecast behavior of the newly added HRRRCast fields, we show the forecast evolution for a nor’easter storm case initialized on 2026-09-26. This example highlights the evolving precipitation pattern, instability structure, cyclone evolution, and the associated 500-hPa flow pattern that steers the system.

This example is intended to show how the newly added HRRRCast fields evolve during a realistic severe-weather episode, with the most prominent feature being the coexistence of two distinct storm regimes: a land-falling cyclone along the northeastern coast and a separate convective regime over the central Great Plains. The coastal system is most evident in the mean-sea-level-pressure minimum and the associated corridor of heavy precipitation and reflectivity that becomes increasingly focused along the coastline from the 1-h to 6-h forecast. In contrast, the Great Plains signal is expressed more as a broad region of elevated CAPE and cloud cover, with more diffuse precipitation and a less compact pressure anomaly, consistent with a strongly convective, warm-sector structure rather than a mature coastal cyclogenesis center. The 500-hPa geopotential-height field reinforces this interpretation by showing a trough-and-ridge pattern that supports both the East Coast cyclone pressure fall and the interior instability burst.

At the 1-h lead, the fields already suggest two separate dynamical centers: a coastal low-pressure system and a more inland convective plume over the Plains. By 3 h, the northeastern cyclone becomes more organized, as indicated by the strengthened pressure gradient, broader precipitation shield, and moistening of the near-surface environment, while the Plains region remains characterized by elevated instability and cloud growth. By 6 h, the land-falling northeastern storm is the dominant structure in the precipitation and pressure fields, while the Great Plains feature remains as a secondary maximum in CAPE and cloud cover, illustrating how multiple storm-scale mechanisms can coexist within the same large-scale flow regime.

References

  1. [1]D. Abdi, I. Jankov, P. Madden, V. Vargas, T. A. Smith, S. Frolov, M. Flora, and C. Potvin. Hrrrcast: A data-driven emulator for regional weather forecasting at convection-allowing scales. Artificial Intelligence for the Earth Systems, 5(2):500261, 2026. doi: 10.1175/AIES-D-25-0061.1. URL https://journals.ametsoc.org/view/journals/aies/5/2/AIES-D-25-0061.1.xml.DOI
  2. [2]S. Agrawal, M. A. Hassen, E. A. Brempong, B. Babenko, F. Zyda, O. Graham, D. Li, S. Merchant, S. H. Potes, T. Russell, D. Cheresnick, A. P. Kakkirala, S. Rasp, A. Hassidim, Y. Matias, N. Kalchbrenner, P. Gupta, J. Hickey, and A. Bell. An operational deep learning system for satellite-based high-resolution global nowcasting, 2025. URL https://arxiv.org/abs/2510.13050.
  3. [3]F. Alet, I. Price, A. El-Kadi, D. Masters, S. Markou, T. R. Andersson, J. Stott, R. Lam, M. Willson, A. Sanchez-Gonzalez, and P. Battaglia. Skillful joint probabilistic weather forecasting from marginals, 2025. URL https://arxiv.org/abs/2506.10772.
  4. [4]M. Andrychowicz, L. Espeholt, D. Li, S. Merchant, A. Merose, F. Zyda, S. Agrawal, and N. Kalchbrenner. Deep learning for day forecasts from sparse observations, 2023. URL https://arxiv.org/abs/2306.06079.
  5. [5]Y. Cao, S. Li, J. Feng, L. Chen, H. Li, and Y. Zhang. Enhancing machine learning models for nowcasting and short-term forecasting of precipitation with a novel probability-matching loss function. Geophysical Research Letters, 52, 2025. doi: 10.1029/2025GL119442.
  6. [6]A. Clark, I. Jirak, T. Galarneau, T. Supinie, K. Knopmeier, D. Harrison, J. Vancil, M. Silcott, D. Jahn, C. Karstens, A. Wade, C. Nixon, K. Halbert, G. Talkington, K. Hoegwind, C. Potvin, N. Yussouf, S. Ernst, J. Picca, M. Baldwin, J. Duarte, J. Martin, D. Imy, B. Matilla, and P. Burke. The 2026 noaa hazardous weather testbed spring forecasting experiment: Preliminary findings and results, 2026. URL https://doi.org/10.25923/gy19-2j41.
  7. [7]D. C. Dowell, C. R. Alexander, E. P. James, S. S. Weygandt, S. G. Benjamin, G. S. Manikin, B. T. Blake, J. M. Brown, J. B. Olson, M. Hu, T. G. Sminrnova, T. Ladwig, J. S. Kenyon, R. Ahmadov, D. D. Turner, J. D. Duda, and T. I. Alcott. The high-resolution rapid refresh (hrrr): An hourly updating convection-allowing forecast model. Part i: Motivation and system description. Weather and Forecasting, 37(8):1371 – 1395, 2022. doi: 10.1175/WAF-D-21-0151.1. URL https://journals.ametsoc.org/view/journals/wefo/37/8/WAF-D-21-0151.1.xml.
  8. [8]J. D. Duda and D. D. Turner. Using object-based verification to assess improvements in forecasts of convective storms between operational hrrr versions 3 and 4. Weather and Forecasting, 38(10):1971 – 1994, 2023. doi: 10.1175/WAF-D-22-0181.1. URL https://journals.ametsoc.org/view/journals/wefo/38/10/WAF-D-22-0181.1.xml.
  9. [9]M. L. Flora and C. Potvin. Wofcast: A machine learning model for predicting thunderstorms at watch-to-warning scales. Geophysical Research Letters, 52(10):e2024GL112383, 2025. doi: https://doi.org/10.1029/2024GL112383. URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2024GL112383. e2024GL112383 2024GL112383.
  10. [10]G. Franch, E. Tomasi, R. Wanjari, V. Poli, C. Cardinali, P. P. Alberoni, and M. Cristoforetti. Gptcast: a weather language model for precipitation nowcasting. Geoscientific Model Development, 18(16):5351–5371, Aug. 2025. ISSN 1991-9603. doi: 10.5194/gmd-18-5351-2025. URL https://dx.doi.org/10.5194/gmd-18-5351-2025.
  11. [13]Z. Jin, X. Xu, M. Bilal, S. Wu, and H. Lin. UResIhaM: Radar reflectivity inversion for smart agriculture with spatial federated learning over geostationary satellite observations. Computational Intelligence, 40(3):e12684, 2024. doi: 10.1111/coin.12684.
  12. [14]T. Mandelbaum, A. McGovern, D. Rothenberg, and U. Ali Syeda. brightbandtech/ExtremeWeatherBench: v1.0.0, 2026. URL https://github.com/brightbandtech/ExtremeWeatherBench. Software; community-driven benchmark suite for high-impact weather evaluation.DOI
  13. [15]G. Moldovan, E. Pinnington, A. P. Nemesio, S. Lang, Z. B. Bouallègue, J. Dramsch, M. Alexe, M. S. Cruz, S. Hahner, H. Cook, H. Theissen, M. Clare, C. O’Brien, J. Polster, L. Magnusson, G. Mertes, F. Pinault, B. Raoult, P. de Rosnay, R. Forbes, and M. Chantry. An update to ecmwf’s machine-learned weather forecast model aifs, 2025. URL https://arxiv.org/abs/2509.18994.
  14. [16]J. Pathak, M. S. Abbas, P. Harrington, Z. Hu, N. Brenowitz, S. Ravuri, A. Carpentieri, J. Leinonen, C. Adams, O. Henningh, N. Geneva, D. Durran, and M. Pritchard. Learning accurate storm-scale evolution from observations, 2026. URL https://arxiv.org/abs/2601.17268.
  15. [17]I. Price, A. Sanchez-Gonzalez, F. Alet, T. R. Andersson, A. El-Kadi, D. Masters, T. Ewalds, J. Stott, S. Mohamed, P. Battaglia, R. Lam, and M. Willson. Gencast: Diffusion-based ensemble forecasting for medium-range weather, 2024.arxiv.org/abs/2312.15796
  16. [18]S. Ravuri, K. Lenc, M. Wilson, D. Kangin, R. Lam, P. Mirowski, M. Fitzsimons, A. Atencia, B. Centeno, A. Nazaret, J. Stanway, N. Tomkins, et al. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597: 672–677, 2021. doi: 10.1038/s41586-021-03854-z.
  17. [19]X. Shi, Z. Gao, L. Lausen, H. Wang, D.-Y. Yeung, W.-k. Wong, and W.-c. Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. In Advances in Neural Information Processing Systems, volume 30, 2017. URL https://arxiv.org/abs/1706.03458.
  18. [20]T. A. Smith, M. Pope, S. Frolov, B. Basarab, D. Abdi, P. Madden, and I. Jankov. Bridging short- and medium-range weather forecasting with machine learning, 2026. URL https://arxiv.org/abs/2608.26822.
  19. [21]D. D. Turner, J. Hamilton, W. Moninger, M. Smith, B. Strong, R. Pierce, V. Hagerty, K. Holub, and S. G. Benjamin. A verification approach used in developing the rapid refresh and other numerical weather prediction models. Journal of Operational Meteorology, 8(3), 2020. doi: 10.15191/nwajom.2020.0803. URL https://doi.org/10.15191/nwajom.2020.0803.
  20. [22]S. Woo, J. Park, J. Lee, and I. S. Kweon. CBAM: convolutional block attention module. CoRR, abs/1807.06521, 2018. URL http://arxiv.org/abs/1807.06521.

Paper details

vlnc:W_6qkswnfx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents