Rain on your route, the next three hours

Where it falls short

1. Small chances run low

It under-forecasts at low chances. When StormScope gave a chance between 0 and 20% (on average 9%), heavy rain came 28% of the time, over 86,504 stretches. At the top it is well calibrated: chances of 80 to 100% (on average 91%) saw heavy rain 91% of the time.

2. Whole trips: it misses most wet trips at the 50% line

As a yes or no for a whole trip, with the line at 50%, StormScope misses about 4 in 5 of the trips that met heavy rain: it flags 20% of them, with 4% false alarms. Radar persistence and HRRR flag more wet trips but raise far more false alarms: persistence 55% (40% false alarms), HRRR 62% (57% false alarms). GFS catches fewer than StormScope, 15% (28% false alarms).

44,928 trips (every route, leaving every 30 minutes up to two and a half hours after each forecast); radar measured heavy rain on at least one stretch of 9,874. Wet trips caught is the hit rate; false alarms is the share of flagged trips that stayed dry.

3. Where the edge comes from

Against radar persistence and HRRR, most of the edge comes from light echoes. Against persistence, 66% of the gain is on stretches where radar measured only a light echo (20 to 35 dBZ) on arrival, where StormScope gave a smaller chance of heavy rain than the yes or no forecasts did; against HRRR, 57%. On stretches that did get heavy rain it did slightly worse than persistence (those stretches give back 5% of the gain). Against GFS it is the other way round: 72% of the gain is on stretches that did get heavy rain.

Worked out after the result from the scored stretches; not part of the pre-registered rule.

One route, leave now or 20 minutes later

San Jose to San Francisco by US-101, forecast made at 05:00 PST on 1 Jan 2026 (13:00 UTC)

Leaving at 05:00 PST: highest chance 100%, and radar measured heavy rain on 23 of 34 stretches. Leaving 20 minutes later: highest chance 63%, and radar measured heavy rain on 6 of 34.

Example trips are chosen by a rule committed on 27 Oct 2026, before any forecast existed: for each route, the trip leaving at the forecast's start whose highest stretch chance is highest, the earliest on a tie. So they show the wettest-looking forecasts, not typical ones.

One stretch is one 3 km model cell the route enters, in driving order; the minutes are the routing engine's, with no traffic, so they do not change with the time of day.

The scorecard

624 forecasts, every half hour on the 13 days from January to March 2026 when heavy rain reached the Bay Area, each scored on 12 routes out of San Jose against what radar measured. A Brier skill score above zero means StormScope's chance was closer to what happened than the comparison forecast; 1 is perfect.

In the rule: five numbers, all must be above zero

Against radar persistence (MRMS), beyond 60 minutes

Against GFS, every lead

Reported only, not in the rule

Against radar persistence, 0 to 60 minutes

Against NOAA HRRR (reported only, not in the rule)

In brackets: 95% intervals from 1,000 resamples of the qualifying days. All five lower bounds are above zero as well; the intervals are reported, not part of the rule. The dashed rule is zero: no better than the comparison.

The table scrolls sideways for the four Brier scores: StormScope, persistence, HRRR and GFS.

Brier scores (lower is better; 0 is perfect) on the same stretches for all four forecasts. Heavy rain is MEASURED by radar; the four forecasts are FORECAST.

Reliability: when it says 30%, does it rain 30% of the time?

Across all leads, the up to 20% band averages 9% and sees heavy rain 28% of the time: read a small chance from StormScope as a real one. Bands with few stretches (the top bands at 2 to 3 hours) say little.

NVIDIA's own published numbers (theirs, not ours)

NVIDIA's own figures, from the model card, not ours: 929 forecasts run over the whole of the US, March to August 2025, scored over NVIDIA's own evaluation regions with different measures (storm-object and neighbourhood CSI, and a fractions skill score, over 33 km windows). They cannot be compared number for number with our Brier skill scores above. We ran the Flash variant; NVIDIA lists both.

Source: NVIDIA, model card file mrms_forecast_skill.md, Hugging Face nvidia/stormscope-goes-mrms at revision dcbe14a. Higher is better; 1 is perfect.

Method

Model

NVIDIA Earth-2 StormScope-GOES-MRMS, variant 3km_10min_flash, Hugging Face nvidia/stormscope-goes-mrms at dcbe14a506e849b1afec9962dbf774f07d7aece8 (NVIDIA Open Model License)

Code

NVIDIA Earth2Studio at commit c00d096ae190f71fc1be3bf24e2124f1d227a9ef (Apache-2.0), the coupled satellite and radar rollout from its own nowcasting example

Software

torch 2.13.0 built for CUDA 12.6 and NATTEN 0.21.7, from Earth2Studio's own lock file; rented NVIDIA H100 GPUs. A rerun of the same seeds gave byte-identical forecasts.

Forecasts

624: every :00 and :30 UTC on the 13 qualifying days, 16 members and 18 steps of 10 minutes each, over a Bay Area crop (36.8 to 38.2 N, 123.0 to 121.2 W). None skipped.

Qualifying days

13 days in January to March 2026 with at least 50 cells of heavy rain in the Bay Area on the radar: 01-01, 01-03, 01-04, 01-05, 01-06, 02-10, 02-11, 02-12, 02-16, 02-17, 02-18, 02-19, 03-31.

Heavy rain

Radar (MRMS composite reflectivity, MEASURED) at or above 35 dBZ anywhere in the 3 by 3 cells (9 km) around a stretch, at the minute the car gets there. StormScope's chance is the share of its 16 members that show the same.

Comparisons

Radar persistence: the radar picture at the forecast's start, held still. NOAA HRRR: its 15-minute reflectivity, from the latest run at least an hour old. NOAA GFS: its rain rate at or above 5.62 mm an hour, from the latest run at least four hours old. All three are yes or no.

Pass rule

Brier skill score above zero against persistence at 60 to 120 and 120 to 180 minutes, and against GFS at all three leads: five numbers, all above zero. HRRR is reported, not part of the rule.

Pre-registration

Written and committed on 27 Oct 2026, before any forecast was made; SHA-256 8b9f6dc740b34bd557ebd5aa9cef8fde73bf7211fda7f98fcc45ed9b9e5cb4bf.

Addenda

A1 (27 Oct): the scorer and data downloads committed before any GPU ran. A2 (27 Oct): a dry run showed the scorer reading radar's "no echo" value as missing; the text already said no echo, and the code was fixed to match, before any forecast existed. A3 (28 Oct): after a one-hour benchmark, all 624 candidate forecasts were run.

Example trips

The selection rule was committed on 27 Oct 2026, before any forecast existed.

Added afterwards

The light-echo breakdown was worked out after the result and is not part of the rule.

Limits

• Radar reflectivity is not rain on the road. 35 dBZ is about 5.6 mm an hour by the Marshall-Palmer relation; that conversion is an assumption.
• StormScope reads GOES-East, which sees California at a slant, and radar coverage is thinner behind the Santa Cruz Mountains. Neither is measured here.
• Thirteen rain days in one winter around one city.
• Route times come from MapMap routing on OpenStreetMap with no traffic.

Data

• Radar: NOAA MRMS, regridded and thresholded by MapMap
• Satellite (model input): NOAA GOES-19 ABI and GLM
• Comparison forecast: NOAA HRRR
• Comparison forecast: NOAA GFS
• Routes: MapMap routing on © OpenStreetMap contributors

NOAA has not reviewed this page and does not endorse it.

MapMap is a member of the NVIDIA Inception program. This is our own test of NVIDIA Earth-2 StormScope, a model available to developers; NVIDIA has not reviewed or approved it. NVIDIA and NVIDIA Earth-2 are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.

Forecast labels: NVIDIA Earth-2 StormScope, NOAA HRRR and NOAA GFS are FORECAST; radar is MEASURED; the route strip is MODELLED from the forecast.