# RRFS (Rapid Refresh Forecast System) Data Assessment

## Overview

RRFS is NOAA's next-generation convection-allowing ensemble forecast system, intended to replace HRRR, RAP, NAM-NEST, and HREF. As of March 2026, RRFS is in operational evaluation — the ensemble post-processing products are publicly available on AWS S3, but the full deterministic model output with pressure-level fields is not yet publicly distributed.

## Data Availability

### AWS S3 Bucket: `noaa-rrfs-pds`

Two public prefixes with identical content:
- `rrfs_a/refs.YYYYMMDD/HH/enspost/` (primary)
- `rrfs_public/refs.YYYYMMDD/HH/enspost/` (mirror)

### Run Schedule
- 4x daily: 00Z, 06Z, 12Z, 18Z
- Data available ~2 hours after init time
- Recent data goes back about 10 days before being pruned

### Available Products (CONUS)

| Product | File Pattern | Contents | Usefulness for Propagation |
|---------|-------------|----------|---------------------------|
| `avrg` | `refs.tHHz.avrg.fNN.conus.grib2` | Ensemble mean | Only APCP (precipitation accumulation). **Not useful** — missing temp, dewpoint, wind, pressure, refractivity profile |
| `eas` | `refs.tHHz.eas.fNN.conus.grib2` | Exceedance probability | APCP/ASNOW exceedance probabilities. **Not useful** for propagation scoring |
| `ffri` | `refs.tHHz.ffri.fNN.conus.grib2` | Flash flood risk index | PPFFG exceedance. **Not useful** |
| `lpmm` | `refs.tHHz.lpmm.fNN.conus.grib2` | Ensemble mean (alternate) | APCP only. Same limitation as `avrg` |

### Forecast Hours
- f01 through f07 (1-7 hour forecasts)
- ~1.5MB per file for ensemble mean CONUS

### Grid
- 3km CONUS domain (same as HRRR)
- Lambert Conformal projection
- GRIB2 format with standard .idx files

## Key Finding: RRFS Public Data is NOT Sufficient for Propagation Scoring

The publicly available RRFS data on S3 contains **only ensemble post-processing products** — primarily precipitation statistics. It does NOT include:

- Surface temperature (TMP:2m)
- Dewpoint temperature (DPT:2m)
- Surface pressure (PRES:surface)
- Boundary layer height (HPBL)
- Precipitable water (PWAT)
- Wind speed/direction (UGRD/VGRD:10m)
- Cloud cover (TCDC)
- Pressure-level profiles (TMP/DPT/HGT at 1000-700mb) for refractivity gradient computation

These fields are essential for propagation scoring (all 9 factors). HRRR provides all of them.

## RRFS vs HRRR Comparison

| Feature | HRRR | RRFS (Public) |
|---------|------|---------------|
| Resolution | 3km | 3km |
| Update frequency | Hourly | 4x daily (00/06/12/18Z) |
| Forecast range | 18h (48h from 00/06/12/18Z) | 7h (from public data) |
| Surface fields | Full (TMP, DPT, PRES, wind, cloud, precip) | APCP only |
| Pressure levels | 8 standard levels (1000-700mb) | Not available publicly |
| Refractivity profile | Derivable from pressure-level T/Td/P | Not possible |
| S3 bucket | `noaa-hrrr-bdp-pds` | `noaa-rrfs-pds` |
| Byte-range requests | Yes (.idx files) | Yes (.idx files) |
| Historical depth | ~2 days on S3 | ~10 days on S3 |

## Recommendation

**Continue using HRRR as the primary data source.** RRFS public data is not yet suitable for propagation scoring due to missing atmospheric profile and surface fields.

Monitor for:
1. NOAA making full RRFS deterministic output public (likely when RRFS becomes the operational replacement for HRRR)
2. Additional fields being added to the ensemble post-processing products
3. NOMADS distribution of full RRFS output (currently returns 403)

When RRFS full output becomes available, key advantages would be:
- Ensemble spread information (confidence intervals on propagation scores)
- Potentially improved boundary layer and moisture forecasts from the FV3 dynamic core
- Longer forecast horizon for propagation prediction

## Meteorologist Feedback (March 2026)

From a meteorologist familiar with Gulf region propagation:

**Model choice validation:**
- HRRR is confirmed as the best short-term (0-48h) choice for low-level inversions and water vapor content. Higher vertical resolution in the boundary layer and tighter grid spacing give it the best handle on moisture — critical since humidity is our highest-weighted scoring factor (20%).
- **NAM should be avoided** — substantially over-estimates low-level water content, especially in the western Gulf region. Over-does fog and water vapor discontinuities. Would produce false-positive ducting predictions in exactly the areas where ducting matters most.
- **GFS is better for days 2-5** forecasts, with a more pessimistic (and more correct) stance on moisture. Worth considering if we add a multi-day propagation forecast view.

**RRFS transition notes:**
- HRRR will be operationally replaced by RRFS in the next few years
- The meteorologist expects RRFS files may use netCDF format (current experimental output on S3 is GRIB2 — the full operational output format may differ)
- The transition should be "fairly straightforward" — the scoring algorithm itself won't change, only the data ingestion layer
- RRFS is already available experimentally but (as documented above) the public output lacks the fields we need

**Implications for our architecture:**
- Keep HRRR as primary source — validated by domain expert
- Don't use NAM for anything moisture-related
- The `Propagation.Scorer` module is data-source-agnostic by design — when RRFS becomes available, only the fetch/extraction layer needs to change
- If we add forecast beyond 48h, consider GFS as a supplementary source for days 2-5

## References
- AWS S3: `s3://noaa-rrfs-pds/`
- NOAA RRFS: https://gsl.noaa.gov/focus-areas/unified_forecast_system/rrfs
- NOAA EMC RRFS: https://github.com/NOAA-EMC/rrfs-workflow