North American Mesoscale Model

The North American Mesoscale Model (NAM) is a regional numerical weather prediction model run by the National Centers for Environmental Prediction (NCEP), part of the U.S. National Weather Service, for short-range operational forecasting over North America.^[1] NAM traces its lineage to the Eta model, which NCEP's predecessor, the National Meteorological Center, made operational in 1993; the system was renamed North American Mesoscale on 25 January 2005.^[2]^[3] Since 2011, NAM's dynamical core has been the Non-hydrostatic Multiscale Model on the B grid (NMMB); it previously ran on the Weather Research and Forecasting Non-hydrostatic Mesoscale Model (WRF-NMM) from 2006 to 2011.^[2]^[4] NCEP is scheduled to discontinue NAM on 6 October 2026, replacing it with the Rapid Refresh Forecast System (RRFS).^[5]

Operation

NAM runs four times daily, with forecasts initialized at 00:00, 06:00, 12:00, and 18:00 UTC. The 12 km parent domain, covering all of North America, produces forecasts out to 84 hours. Four fixed 3 km domains, nested one-way inside the parent domain and covering the contiguous United States (CONUS), Alaska, Hawaii, and Puerto Rico, run out to 60 hours. A relocatable 1.5 km fire-weather nest, placed within the CONUS or Alaska domain, runs out to 36 hours; its location is set by NCEP Service Centers or, during the summer wildfire season, the National Interagency Fire Center.^[1]^[6]

NAM uses a partially cycled, six-hour data-assimilation window with hourly analysis updates for the 12 km parent domain and the 3 km CONUS and Alaska nests, using the NCEP Gridpoint Statistical Interpolation (GSI) analysis system. The non-cycled Hawaii, Puerto Rico, and fire-weather nests are initialized from the parent domain's first guess.^[6] Since August 2014, the analysis has used a hybrid ensemble–variational method, incorporating background-error covariances from NCEP's global ensemble Kalman filter alongside the static variational background-error covariance.^[6]^[2] In practice, this lets the analysis weight observations using an estimate of the atmosphere's day-to-day uncertainty, drawn from that day's global ensemble, rather than relying solely on a fixed, climatological estimate of typical forecast error.

Principles

NAM's dynamical core, NMMB, is a non-hydrostatic model formulated on an Arakawa B grid using a hybrid sigma-pressure vertical coordinate, with 60 vertical layers and a model-top pressure of 2 millibars.^[6]^[7] NMMB was developed at NCEP as the successor to the regional-only WRF-NMM, extending the same modeling approach to allow global as well as regional configurations.^[4]

The 12 km parent domain uses the Betts–Miller–Janjic parameterized convection scheme; the higher-resolution nests run without parameterized convection, relying on the model's grid-scale (Ferrier–Aligo) microphysics to represent convective storms explicitly.^[1]^[6]^[8] This distinction reflects each domain's resolution: at 12 km grid spacing, individual thunderstorms are smaller than a single grid cell and must be approximated, while the 3 km nests are fine enough to simulate many storms directly. Turbulent mixing uses the Mellor–Yamada–Janjic level-2.5 boundary-layer scheme,^[9] radiative transfer uses the Rapid Radiative Transfer Model, and land-surface processes use the Noah land-surface model.^[10]^[6]

Variants

At its introduction in October 2011, the CONUS and Alaska nests ran at 4 km and 6 km horizontal grid spacing, respectively, and the fire-weather nest ran at 1.333 km over the CONUS or 1.5 km over Alaska.^[4] NCEP's 21 March 2017 upgrade reduced the CONUS nest to 3 km, the Alaska nest to 3 km, and standardized the fire-weather nest at 1.5 km; the Hawaii and Puerto Rico nests were unchanged at 3 km. The same upgrade replaced the previous twelve-hour assimilation cycle, which updated its analysis every three hours, with the six-hour cycle with hourly updates described above; this remains NAM's configuration as of 2026.^[11]^[2]

NCEP also distributes NAM output on a number of coarser fixed grids for aviation and other specialized users, including 20 km, 32 km, and 40 km domains over North America and downscaled guidance grids matching the resolution of the National Digital Forecast Database.^[12]

Usage

NAM guidance is distributed to National Weather Service forecast offices and to the public through NCEP's Open Data servers in GRIB2 format.^[12] Downscaled NAM fields, matched to the resolution of local National Digital Forecast Database grids, supply a portion of the input used by NWS offices to build their digital forecasts.^[4] The 3 km NAM nests are also used as time-lagged members of the High-Resolution Ensemble Forecast (HREF) system, alongside members of the High-Resolution Window and, more recently, the High-Resolution Rapid Refresh.^[13] NAM output also feeds NCEP's Real-Time Mesoscale Analysis (RTMA) and Unrestricted Mesoscale Analysis (URMA): the CONUS and Alaska analyses blend a downscaled forecast from the corresponding NAM nest with the High-Resolution Rapid Refresh, while the Hawaii and Puerto Rico analyses, which fall outside the HRRR domain, rely on a downscaled NAM nest forecast as their sole first guess.^[14] The Short-Range Ensemble Forecast (SREF) has likewise drawn part of its initial conditions for its NEMS-NMMB members from the NAM Data Assimilation System (NDAS) since 2015, with no further change documented before SREF's own retirement alongside NAM.^[2]^[15] A separate statistical post-processing system, NAM Model Output Statistics (NAM MOS), generates site-specific text forecasts from NAM output. All of these NAM-dependent products are scheduled to be discontinued alongside NAM itself.^[16]

Accuracy

A 2014 Developmental Testbed Center study compared quantitative precipitation forecasts from NAM's full 12 km parent domain against the Global Forecast System, evaluating forecasts issued from 18 December 2008 to 15 December 2009 — a period predating NAM's 2011 switch from the WRF-NMM to the NMMB dynamical core. Using the Gilbert skill score, a traditional grid-to-grid metric, GFS consistently outperformed NAM. Using the fractions skill score, a spatial neighborhood-based method, the result reversed, particularly in winter: NAM's higher resolution produced a spatial distribution of precipitation that more closely matched observations than GFS's, even where the two verification approaches disagreed on which model was more skillful overall. The authors used the contrast to illustrate how traditional verification metrics can penalize higher-resolution forecasts that are spatially realistic but imprecisely placed.^[17]

A 2018 peer-reviewed evaluation of cool-season quantitative precipitation forecasts over mountainous terrain in the western United States compared the NAM 3 km CONUS nest against the High-Resolution Rapid Refresh, the Global Forecast System, the Short-Range Ensemble Forecast, and a convection-permitting NCAR ensemble, using SNOTEL station observations and PRISM precipitation analyses as verification. The study found that the higher-resolution, convection-permitting systems — including the NAM CONUS nest — were more deterministically skillful than coarser models, particularly over narrow interior mountain ranges where they better resolved topography and orographic precipitation.^[18]

A 2025 evaluation at NOAA's Hazardous Weather Testbed compared NAM's 3 km CONUS nest against the experimental Rapid Refresh Forecast System control member and three High-Resolution Window configurations (WRF-ARW, NSSL, and FV3 variants), using blind subjective ratings of severe-weather forecasts over a five-week experiment. NAM tied with the HRW FV3 configuration for the lowest subjective ratings of composite-reflectivity forecasts, though it rated ahead of HRW FV3 on environmental-field forecasts. The evaluators concluded that the combined subjective and objective results supported RRFS's implementation and the retirement of the legacy convection-allowing models it replaces, including NAM.^[19]

Development history

Origins: the Eta model (1993–2005)

The system that became NAM originated as the Eta model, named for its step-mountain ("eta") vertical coordinate, developed at the National Meteorological Center (NCEP's predecessor). An 80 km version became operational in 1993, and a higher-resolution, roughly 29 km "meso-Eta" configuration followed in 1994, described in a paper by Thomas Black of the National Meteorological Center.^[3]^[2] The Eta Forecast Model and its associated Eta Data Assimilation System (EDAS) were together referred to as Eta/EDAS.^[20] NCEP officially renamed the operational Eta forecast system North American Mesoscale (NAM) on 25 January 2005, with EDAS renamed the NAM Data Assimilation System (NDAS).^[2]

Transition to WRF-NMM (2006)

In June 2006, NCEP replaced the Eta dynamical core with the Weather Research and Forecasting Non-hydrostatic Mesoscale Model (WRF-NMM), while retaining the same 12 km domain. The change also replaced the Eta three-dimensional variational analysis with the unified Gridpoint Statistical Interpolation (GSI) analysis system.^[21]^[2] Through this WRF-NMM era, EMC's own verification found NAM's large-scale synoptic forecasts — measured by 500 hPa height and 250 hPa wind root-mean-square error over CONUS — persistently less accurate than the Global Forecast System. A package of model and data-assimilation upgrades implemented in March and December 2008 narrowed, but did not close, this gap. By early 2009, NAM's average 500 hPa height error had fallen by roughly 20 percent from 2006–07 levels, and its 250 hPa wind errors were close to parity with GFS.^[22]

NEMS-NMMB and the high-resolution nests (2011–2017)

On 18 October 2011, NCEP replaced WRF-NMM with the Non-hydrostatic Multiscale Model on the B grid (NMMB) running within the new NOAA Environmental Modeling System (NEMS) — the first operational implementation of NEMS at NCEP. The same upgrade introduced the CONUS, Alaska, Hawaii, and Puerto Rico nests — initially at 4 km, 6 km, 3 km, and 3 km grid spacing, respectively — and the relocatable fire-weather nest described above.^[4] A hybrid ensemble–variational data-assimilation scheme, drawing background-error information from NCEP's global ensemble, was implemented on 12 August 2014.^[2]^[6] On 21 March 2017, NCEP implemented what it characterized as NAM's last major science upgrade: it reduced the CONUS and Alaska nests to 3 km grid spacing, revised the model's microphysics and convection schemes to correct an excessive-precipitation bias exposed during Hurricane Joaquin, and replaced the twelve-hour assimilation cycle with the six-hour, hourly cycle described above. EMC subsequently redirected its regional modeling development toward a convection-allowing system built on the FV3 dynamical core, the eventual basis for RRFS.^[11]^[1]

Retirement and replacement by RRFS (2026)

On 12 May 2026, NCEP announced it would discontinue NAM, along with the Short-Range Ensemble Forecast, the High-Resolution Ensemble Forecast, the High-Resolution Window, and NAM MOS, replacing them with the Rapid Refresh Forecast System (RRFS) and its ensemble counterpart, REFS.^[16] The original effective date, 31 August 2026, was subsequently revised: an update issued 6 July 2026 moved the parallel data feed to on or about 11 August 2026 and operational implementation — coinciding with NAM's retirement — to 6 October 2026.^[5] The deterministic RRFS runs at 3 km grid spacing over a North America domain (2.5 km for Hawaii and Puerto Rico), with a 13 km North America output and a relocatable 1.5 km fire-weather configuration, consolidating the convective-scale guidance previously split among NAM, HiResW, and their nests into a single modeling system.^[5]