Draft:Rapid Refresh Forecast System

Not to be confused with Rapid Refresh (weather prediction).

The Rapid Refresh Forecast System (RRFS) is a convection-allowing numerical weather prediction system being developed by the National Centers for Environmental Prediction (NCEP), part of the U.S. National Weather Service, for short-range operational forecasting over North America. RRFS and its companion ensemble system, the RRFS Ensemble Forecast System (REFS), are scheduled to become operational on 6 October 2026, replacing four existing NCEP systems: the North American Mesoscale Model (NAM), the High-Resolution Ensemble Forecast (HREF), the Short-Range Ensemble Forecast (SREF), and, for most domains, the High-Resolution Window (HiResW). Together, RRFS and REFS consolidate NCEP's convective-scale forecasting suite into a single hourly updated modeling system. Development of RRFS and REFS is led jointly by NWS's Office of Modeling and Development and NOAA's Office of Oceanic and Atmospheric Research, specifically its Global Systems Laboratory. A wider network of other NWS entities, academic institutions, federal agencies, and international partners has also contributed. RRFS runs on the FV3 dynamical core and, like NAM and other NCEP regional systems, uses the Gridpoint Statistical Interpolation (GSI) analysis system in a hybrid ensemble-variational configuration with hourly cycling.

Operation

The deterministic RRFS is scheduled to run hourly over a North America domain at 3 km horizontal grid spacing, producing forecasts to 84 hours for the four synoptic cycles (00:00, 06:00, 12:00, and 18:00 UTC) and to 18 hours for the other, non-synoptic hourly cycles. A companion five-member RRFS ensemble, distinguished from REFS below, is scheduled to run over the same North America domain for the four synoptic cycles, producing forecasts to 60 hours; its members vary in initial conditions, lateral boundary conditions, and model physics to sample a range of forecast outcomes.

Principles

RRFS is built on the FV3 dynamical core, distinguishing it from the WRF-ARW core used by the High-Resolution Rapid Refresh (HRRR), a separate NCEP convection-allowing model that is not being retired alongside NAM, SREF, HREF, and HiResW. Data assimilation uses the GSI hybrid ensemble-variational method with hourly cycling, the same general approach used for NAM's own analysis. A 2024 peer-reviewed study also tested a 2022-vintage experimental RRFS build with a separate radar-reflectivity ensemble-variational data-assimilation system, introduced 6 July 2022; whether this capability carried into RRFS's operational configuration has not been confirmed.

RRFS's physics parameterizations, as published by NOAA's Global Systems Laboratory, include the MYNN-EDMF scheme for planetary boundary-layer turbulence and shallow convection, the MYNN scheme for the surface layer, the Thompson-Eidhammer scheme for microphysics and climatological aerosols, the scale-aware Simplified Arakawa–Schubert (saSAS) scheme for deep convection, RRTMG for long- and short-wave radiation, and the RUC land-surface model. A built-in smoke-and-dust component, RRFS-Smoke Dust, uses RAVE fire-emissions data with the FENGSHA windblown-dust scheme; separate coupled models handle large lakes (FVCOM) and small lakes (CLM Lake).

Variants

Deterministic RRFS output is provided at 3 km grid spacing for the contiguous United States (CONUS) and Alaska, at 2.5 km grid spacing for Hawaii and Puerto Rico, and on a coarser 13 km North America output grid. A separate, relocatable fire-weather configuration runs at 1.5 km grid spacing over a 5-by-5-degree rotated latitude–longitude region.

REFS is a separate ensemble product-generation system, distinct from the five-member RRFS ensemble described above. It combines forecast output from the current and six-hour-old cycles of both the deterministic and ensemble RRFS. For the Alaska and CONUS domains, it also adds two further members drawn from the current and six-hour-old cycles of HRRR. NOAA's Global Systems Laboratory, a co-developer of RRFS, describes this same structure in general terms as a control member (the deterministic RRFS) plus six other members including HRRR, supplemented by multiple time-lagged members. REFS replaces HREF, which it resembles in product type. It differs in producing forecasts to 60 hours, versus HREF's 48 hours. It also generates products for all four domains — CONUS, Alaska, Hawaii, and Puerto Rico — at all four synoptic cycles, whereas HREF has produced non-CONUS domain products only twice daily.

Between the deterministic and ensemble RRFS and REFS, the new systems are scheduled to fully retire NAM and HREF, and to retire HiResW's CONUS, Alaska, Hawaii, and Puerto Rico domains — the same domains covered by deterministic RRFS — while HiResW's Guam domain, outside RRFS's coverage area, continues running unchanged; SREF is scheduled to retire alongside these systems as well.

Usage

RRFS and REFS output is scheduled for distribution through NOAA's NOMADS servers and the NOAAPORT satellite broadcast system, in GRIB2 format, the same channels used for NAM and the other systems being retired. A real-time parallel data feed is scheduled to become available on or about 11 August 2026, ahead of the systems' scheduled operational implementation on 6 October 2026. Once operational, RRFS and REFS are intended to serve as the primary guidance for the forecasting applications previously supported by NAM, the High-Resolution Ensemble Forecast, the Short-Range Ensemble Forecast, and the High-Resolution Window, including use by National Weather Service forecast offices and public dissemination of model output.

Accuracy

A 2024 peer-reviewed study compared a 2022-vintage experimental RRFS build against the operational High-Resolution Rapid Refresh (HRRR) on 32 widespread convective events over the eastern United States. RRFS overpredicted total storm area by 40–60 percent relative to observations. At a 35 dBZ reflectivity threshold, individual storm sizes were up to 90 percent larger than observed — a larger bias than HRRR showed on the same cases. RRFS did outperform HRRR at capturing the timing of convection initiation, particularly for cellular and cluster-type storms.

Three NOAA Testbed evaluations conducted in 2025 compared RRFS and REFS more directly against the systems they were scheduled to replace. At the 2025 Spring Forecasting Experiment, a separate evaluation compared the RRFS control member directly against the NAM Nest and three High-Resolution Window (HiResW) configurations — the deterministic systems RRFS was proposed to retire. In blind subjective ratings over the five-week experiment, RRFS scored higher on average than all four, with the NAM Nest and one HiResW configuration often rated lowest of the group; RRFS was consistently one of the better-performing models throughout. Objective verification against radar observations agreed: RRFS had a slightly higher critical success index and the lowest false-alarm ratio of the five systems compared, though participants also noted its simulated storms often appeared more intense than observed even where the objective bias statistics were neutral.

A separate comparison at the same experiment measured RRFS not against the systems it replaces, but against the existing High-Resolution Rapid Refresh (HRRR) — NOAA's established hourly-updated benchmark, which RRFS is not scheduled to replace. At the 2024 Spring Forecasting Experiment, hosted by NOAA's Hazardous Weather Testbed, HRRR had outperformed nearly every experimental model tested, including RRFS. By 2025, blind subjective evaluations of five deterministic convection-allowing models showed the opposite pattern. A GSL configuration using the Model Prediction Across Scales (MPAS) dynamical core received the highest mean rating, 6.3 out of 10. The operational HRRR followed at 6.1, then RRFS at 5.9 — up from RRFS's own 2024 rating of 5.4. A separate NSSL MPAS configuration and a NASA FV3 configuration tied for fourth at 5.8 each. None of these differences were statistically significant. Objective verification against radar observations produced a different result: the top-rated MPAS configuration had the worst objective skill score of the five. NOAA's Global Systems Laboratory characterized the subjective result as the first time in ten years of HRRR operations that a new system had outperformed it at this testbed, and said it gave forecasters confidence RRFS could replace NAM. In a parallel comparison, REFS rated slightly higher on average than the operational HREF for Day 1 storm-attribute forecasts, with REFS's ensemble radar data assimilation playing a key role in improving forecasts of deep convection through the first six hours.

At the Weather Prediction Center's 2025 Flash Flood and Intense Rainfall experiment, RRFS underperformed HRRR on warm-season precipitation but outperformed the other models tested, an improvement over its performance the previous year. REFS slightly outperformed HREF on several objective measures of deep convection, though its skill depended heavily on which convective parameterization scheme it used.

At the 15th Winter Weather Experiment, hosted by the Weather Prediction Center's Hydrometeorology Testbed over the 2024–25 winter season, evaluation of RRFS and REFS was severely limited by extended downtime of the RRFS ensemble that winter. For the few snowfall cases available, from November 2024, the RRFS control member performed subjectively similarly to HRRR. Freezing-rain cases were too rare to evaluate, and the sample was too small to draw any conclusions about REFS's ensemble performance at all. This is in contrast to NOAA's public summary, which described the sample as merely constraining rather than precluding conclusions about REFS.

Taken together, the SFE and FFaIR evaluations led NOAA to describe forecasters as having confidence that RRFS could replace NAM operationally; the WWE, limited by the ensemble's extended downtime, reached no comparable conclusion.

Development history

Origins within the Unified Forecast System (2018–2022)

RRFS originated as a convection-allowing regional application within NOAA's Unified Forecast System (UFS), the agency's broader effort to consolidate its numerical weather prediction models around shared modeling infrastructure. Development began around 2018, when NOAA's Global Systems Laboratory (GSL) and NCEP's Environmental Modeling Center (EMC) jointly established the core model infrastructure: an FV3 limited-area configuration, its Extended Schmidt Gnomonic (ESG) grid projection, and initial data-assimilation and lateral-boundary-condition options. The underlying goal was to replace NCEP's separate legacy regional models with a single, higher-resolution system built on infrastructure shared across NOAA's broader modeling programs, rather than developing another standalone model in isolation.

From 2018 onward, prototype RRFS configurations were evaluated each year at the Spring Forecasting Experiment (SFE), hosted by NOAA's Hazardous Weather Testbed. Participation grew quickly: the number of FV3-based configurations submitted to the SFE rose from 10 in 2020 to 24 in 2021 and 59 in 2022, as more research groups contributed prototype ensemble members. By 2023, RRFS's 3 km North America grid held roughly 925 million three-dimensional forecast points across 65 vertical levels — about 7.6 times the horizontal gridpoints, and 9.7 times the three-dimensional gridpoints, of HRRR's smaller CONUS domain. RRFS's physics suite also evolved during this period: a mid-2023 status update listed the Grell–Freitas scheme for convection, since replaced by the scale-aware Simplified Arakawa–Schubert (saSAS) scheme used in the operational configuration described above. By the time RRFS's design was finalized, it looked substantially different from the 2018 prototype it began as, reshaped by years of testbed feedback and physics revisions.

Schedule delays and the path to implementation (2022–2026)

RRFS's implementation date slipped repeatedly. A mid-2022 overview described operational implementation as planned for "late 2024." By mid-2023, NOAA's internal target had narrowed to September 2024 through March 2025, with a code freeze planned for late 2023 and a final science evaluation to follow through 2024 — though NOAA itself labeled this window a "best case," not a firm commitment. The same mid-2023 assessment also flagged persistent, unresolved warm-season biases in RRFS's precipitation and reflectivity forecasts, along with broader differences in simulated storm structure compared to HRRR — issues consistent with the iterative, testbed-driven refinement process described above. Operational challenges continued into the original target window itself: the RRFS ensemble suffered substantial downtime over the 2024–25 winter, the same season RRFS had once been expected to launch. NCEP's first public retirement notice for NAM, issued 12 May 2026, set the replacement date at 31 August 2026. An update issued 6 July 2026 revised this again, to 6 October 2026, alongside a parallel data feed beginning on or about 11 August 2026. This final decision followed the extensive 2025 NOAA Testbed evaluations detailed in the Accuracy section above, which compared RRFS and REFS against the legacy systems they were scheduled to replace and, per NOAA's own account, cleared RRFS's path to operations only after years of physics and data-assimilation refinement. Altogether, RRFS is set to reach operational status roughly two years behind its original schedule — a gap that reflects both unresolved technical issues along the way and NOAA's stated preference for extended, testbed-driven validation before committing a new system to operations.

Future development: RRFS-MPAS

NOAA has described the initial RRFS release — retroactively distinguished as RRFSv1 once work began on a successor — as an interim step toward a more capable design. Development is already under way on a second version, informally called RRFS-MPAS or "RRFS Version 2," which would replace the FV3 dynamical core with the Model Prediction Across Scales (MPAS) core. As of 2023, NOAA cited FV3's numerical design as poorly suited to convection-allowing scales over the long term, particularly its tendency toward spurious convection during frequent data assimilation. The same 2023 assessment noted that MPAS's own data-assimilation and physics infrastructure remained immature, and that legacy systems needed to retire before a planned moratorium on further changes to NOAA's WCOSS2 supercomputing system — making RRFSv1 a necessary intermediate step rather than a system NOAA could bypass in favor of MPAS directly. A 2023 presentation estimated RRFSv2's implementation for fiscal year 2027; NOAA's Global Systems Laboratory later projected a broader 2027–2028 timeframe. MPAS-based prototypes outperformed both RRFSv1 and the operational HRRR in several 2025 Testbed evaluations, reinforcing this direction. NOAA's own planning treats RRFSv1 as a way station rather than a destination, with a technically different successor already scheduled for the late 2020s.

See also

• North American Mesoscale Model – the system RRFS is scheduled to replace in October 2026
• High-Resolution Rapid Refresh – NCEP's existing hourly-updated convection-allowing model, not being replaced by RRFS but contributing members to REFS
• Rapid Refresh (weather prediction) – an unrelated, similarly named NCEP model
• Global Forecast System – supplies RRFS's lateral boundary conditions
• Numerical weather prediction

External links

• NOAA Global Systems Laboratory: Rapid Refresh Forecast System
• NCEP: RRFS Products

Category:Weather prediction Category:National Weather Service numerical models