CorX Labs
New: version 0.5 Rain forecasts Every town Hourly & 14 days Hurricane tracks Apache-2.0 Research model

Pluvium

Open-source rain forecasts for every town and village (hour by hour for three days, day by day for 14) and hurricane tracks, built from the forecasts the world already makes. Made in Jamaica, for the Caribbean and any country between 60° S and 60° N.

Package: pluvium 0.5.5 · Python ≥ 3.10 · Engine: NatureV1 / iHELIX · Developer: Nathan · Origin: Jamaica


Pluvium is an AI forecasting system built by CorX Labs in Jamaica, in three parts. Pluvium’s network forecasts rain at every town hour by hour for three days: the chance each hour, whether it will be drizzle, showers, heavy rain or a thunderstorm, when it will start, how long it will last and how much will fall. Pluvium Blend combines six operational systems (ECMWF ENS, HRES and AIFS, and NOAA GEFS, GFS and GraphCast) into each day’s rain out to 14 days, and keeps learning from every day it can check. Pluvium Cyclone forecasts hurricane tracks from the same six systems.

Over a full year, April 2025 to March 2026, in seven places on four continents, Pluvium Blend had the best skill of every system, raw or post-processed, for rain of at least 1 mm on days 1 to 7 in all seven. It does not lead at 50 mm, where post-processed AIFS often wins at short range, and it under-forecast Hurricane Melissa’s rain over Jamaica. Pluvium Cyclone beats ECMWF HRES and NOAA GFS but is behind the National Hurricane Center’s official forecast. Every loss is published beside the wins.

A research model, not an official forecast. Pluvium is new and no independent party has verified it. Its flood levels are a rule that has not been checked against flood reports, and its alerts are marked as exercises. Do not use Pluvium for real decisions. Follow your national meteorological service, in Jamaica, the Meteorological Service of Jamaica and ODPEM. For hurricanes, the National Hurricane Center.

Products
Network, Blend, Cyclone
Rain, hourly
3 days, every town
Rain, daily
14 days, Pluvium Blend
Hurricane tracks
N. Atlantic & E. Pacific, 5 days
Engine
NatureV1 / iHELIX, 12 blocks
Mesh
6,075 points, ~11 km over land
Full-year test
7 places, Apr 2025–Mar 2026
Full size
Pluvium-500, 57.9M params
Covers
Any country, 60° S–60° N
Bulletins
English, Spanish, French, Patois
Package
pluvium 0.5.5
Licence
Apache-2.0

1. What it tells you

For each place, every hour, out to three days, in the words a person would use: “light drizzle”, “moderate showers”, “heavy rain”, “thunderstorms”, “torrential rain”.

Will it rain?

The chance of rain in each hour, and of any rain or any heavy rain in the next 6, 12, 24, 48 and 72 hours: computed exactly from the chain, not by pretending the hours are independent.

What kind of rain?

Drizzle, rain or showers, and how hard: light, moderate or heavy. A separate head gives the chance of thunder.

When does it start?

The hour rain will most likely start, in local time and with a window, and the hour it will be heaviest.

How long, and how much?

How long it will last, and the day’s total as a median and a bad case (the 90th percentile), with the chance of 10, 25, 50 and 100 mm and the wettest three hours.

Why is it raining?

The cause: a cold front or shear line, a tropical wave, a tropical cyclone, afternoon storms or trade-wind showers.

Lightning, and could it flood?

Lightning within 10 km in the last 15 minutes, from space. Flood risk with its reason, from the rain over the town’s river catchment, a rule, not yet verified.

What pluvium forecast writes

Files written by the pluvium forecast command and what each contains
FileWhat is in it
towns.csvOne row per place: the chance of rain now and in 6, 12, 24, 48 and 72 hours, heavy rain, thunder, the main weather, the likely start and its window (local time), the heaviest hour, 24-hour totals (median, 90th percentile, chance of 10, 25, 50 and 100 mm), the wettest 3 hours, rain over the river catchment, flood risk and its reason, lightning now, the cause, and a one-sentence summary
parishes.csvOne row per parish or province, with a headline
towns.geojsonThe same as points, for maps
hourly.npzThe hourly chance of rain, of each intensity and of thunder, and the expected rain, per place
forecast.json, report.txtMetadata and a readable digest

Bulletins come in English, Spanish or French, and in Jamaican Patois for Jamaica. Alerts are written in CAP 1.2, the common alerting format, and carry the status “Exercise” rather than “Actual”.

2. How it works

The global models are good at the large-scale weather but blur island rain. Their grids are 9–25 km or coarser, convection is parameterised, and the daily cycle of showers is often at the wrong hour: GraphCast, for one, puts tropical rain too close to local midday. Radar nowcasting needs radar, and the Caribbean has almost no open radar. Pluvium is built for what the region does have: the global models’ coarse forecasts, satellite rain, hourly airport reports and lightning seen from space.

How Pluvium makes a forecast: inputs, one Caribbean mesh, every town, one chain of rain states On the left, five inputs: the global models' forecasts (GFS, ECMWF IFS and AIFS), the rain falling now from the GOES satellite and MRMS radar, the last three hours of airport reports, physics computed rather than learned (mountain rain and sea breezes), and the place itself (terrain, coast, sun and season). They are read onto one Caribbean mesh of 6,075 points, about 11 kilometres apart over land and 80 at sea, with 43 frames, and processed by twelve iHELIX blocks, the engine NatureV1 runs on. Each block combines geodesic attention, whose heads each have a radius from 15 to 500 kilometres, index attention that finds places that look alike anywhere, and a recurrence through the forecast's frames. Every town then reads the mesh around it with its own terrain, coast and climate. One chain of four rain states per town and hour, dry, light, moderate and heavy, gives the chance of rain each hour, when it starts, the heaviest hour and how long it lasts and, with the other heads, how much falls, thunder, flood risk and why it rains. Inputs One Caribbean mesh Every town, every hour Answers Global forecasts GFS · ECMWF IFS · AIFS Rain falling now GOES satellite · MRMS radar Airport reports the last three hours Physics, computed mountain rain · sea breezes The place itself terrain · coast · sun · season A mesh of 6,075 points about 11 km apart over land, 80 km at sea · 43 frames Geodesic attention each head its own radius, 15–500 km Index attention finds places that look alike, anywhere Recurrence through time a gated delta rule along the frames × 12 iHELIX blocks the engine NatureV1 runs on Rain each hour When it starts Heaviest hour How long How much Thunder Flood risk Why it rains Each town reads the mesh its own terrain, coast and climate from a 35 km terrain patch One chain of rain states a 4 × 4 table of chances per hour Dry Light Mod. Heavy now → each state an hour later sampled for windows and totals Other heads amounts · thunder · kind · cause Physics is computed, not learned: Smith–Barstad mountain rain, sea-breeze convergence, fronts, and the rain now carried along the steering wind.
Information forward Encode → process → decode → chain Every answer from one chain
How Pluvium makes a forecast. The global models’ fields, the rain falling now, airport reports and computed physics are read onto one Caribbean mesh and processed by 12 iHELIX blocks, the engine NatureV1 runs on. Every town then reads the mesh around it, and one chain of rain states per town gives every hourly answer.

One chain, every answer

Most AI rain models write a separate number for every hour. Pluvium’s network writes, for every hour and town, a 4 × 4 table of chances: from dry, light, moderate or heavy rain now to each of them an hour later. One chain then gives every answer consistently, the chance of rain each hour, the hour rain most likely starts, the hour it turns heavy, when it stops, the chance of any rain in the next 6 to 72 hours, and totals and the wettest 1, 3 and 6 hours, by sampling paths of the chain.

Trained through the instruments’ eyes

Neither source of truth is perfect. A satellite smears a downpour over an 8 km box and misses light showers; an airport sees one spot. So training scores what each instrument reported through its own learned confusion matrix, a differentiable hidden Markov model , and a satellite miss is never taught to the model as “it was dry”.

Physics given, not learned

Mountain rain

Smith–Barstad linear theory, solved on 2 km tiles for every forecast hour with learned time scales: rain on the windward slopes, a shadow in the lee.

Sea-breeze storms

Where the sea breezes meet over each island in the afternoon, and when the sea-breeze front passes, from gravity-current theory.

Saharan air

An index for the Saharan air layer: the dry, warm mid-levels that choke summer showers.

Waves and fronts

Tropical-wave troughs at 700 hPa, cold fronts and shear lines, with their arrival times at every place.

Storm rain

Cyclones are too rare in any training period to learn from, so a storm-rain prior with shear asymmetry is an input.

The daily clock

Instability (CAPE and CIN), column humidity and the land and sea daily cycles, aimed at the hour of afternoon storms, which global AI models often get wrong.

For the first 12 hours the rain already falling is carried along the steering wind, the core idea of radar nowcasting, done with satellite rain because the region has so little open radar.

3. Pluvium Blend: every operational system in one forecast

Pluvium’s network reads one driving model at a time. Pluvium Blend reads all six operational systems at once (ECMWF ENS, NOAA GEFS, ECMWF HRES, ECMWF AIFS, NOAA GraphCast and NOAA GFS) and, where it runs, Pluvium’s own network. It writes one calibrated forecast of each day’s rain for every place, out to 14 days.

  1. Post-process every system on its own

    A small network turns one system’s forecast into the chance of at least 0.2, 0.5, 1, 2 … up to 150 mm. The chances fall with the threshold by construction, so they never contradict each other. One copy is trained per system, which gives that system’s own post-processed forecast. Another copy reads every system at once.

  2. Pool them with about 70 weights

    The pool averages the components’ log-odds with one weight per component, per band of thresholds and per lead. Its weights are fitted on alternate weeks the components never trained on, so no component is judged on truth it learned from. A pool this small cannot overfit, and it can always fall back to its best component. A missing system, GraphCast stops at day 10, simply drops out.

  3. Beat every rival after equal treatment

    Each system’s component is that system’s post-processed forecast, scored on its own. Pluvium Blend has to beat each of them in that form.

Why two layers: inside the training year, one network over all six systems was no better than the best single system: it learned combinations that did not carry over to a new season. The pool beat that network in about 20 of 24 comparisons. Training a blend for a new country takes a year of archived runs and satellite truth: about two hours of downloads and under an hour of CPU.

New in version 0.3

  • It learns every day. pluvium blend learn re-weighs the components and the pool after each day that can be checked, towards whoever has been right.
  • It sees every heavy day. The components train on every day with 25 mm or more observed or forecast, weighted so the chances stay unbiased.
  • Record storms. A forecast beyond anything in the training record is handed back, in part, to what the raw systems say, but only where that helped. Each place withholds its four wettest training days, forecasts them unseen, and keeps the hand-over only if it improved them. It stayed on in Jamaica, the Caribbean, the Maritime Continent, the Great Plains and the southern Andes, and switched itself off in the Alps and the Pacific Northwest.
  • Alerts in plain terms. Every place and day gets an alert tier for 50 mm (watch, warning, extreme), a flag when the forecast is beyond the training record, how rare the amount is there (“about once in 6 years”), and how many raw systems call for 50 mm.
  • Maps that look like weather. Scenarios arrange the calibrated amounts in space like the ECMWF ensemble’s members, so a long-range map is not a smooth blur.

4. Pluvium Cyclone: hurricane tracks

Pluvium Cyclone finds every tropical cyclone in the North Atlantic and the eastern North Pacific in all six systems (in every member of the ECMWF and NOAA ensembles, and in HRES, AIFS, GFS and GraphCast) and combines their tracks with weights learned through the season from the forecasts that have already verified. It needs no GPU and nothing to train.

For each storm it writes a bulletin and a data file: the track to five days, a circle that should hold the centre two times in three at each lead, the chance of the centre passing within 150 km of each listed place, and scenarios when the tracks split. Track only, not intensity.

pluvium cyclone --root ~/storms                        # every storm alive now: track, cone, passage chances
pluvium cyclone --root ~/storms --time 2025-10-25T22   # as it would have been issued then

Against the National Hurricane Center

Settings were chosen on three seasons and tested once on a fourth: the East Pacific in 2025, 167 forecasts of 18 storms. Mean track error on the forecasts all four made, with what a forecaster has in hand at advisory time. The best in each column is outlined.

Mean track error in nautical miles by lead for Pluvium Cyclone and the National Hurricane Center’s forecasts, East Pacific 2025 test season, early guidance; lower is better
Track error, nautical miles: lower is better12 h24 h36 h48 h60 h72 h96 h120 h
Pluvium CycloneCorX Labs, six public systems
NHC official forecastNational Hurricane Center18273440465273100
NHC consensus TVCNNational Hurricane Center19273543536492127
NHC corrected consensus HCCANational Hurricane Center19303443515878107
Forecasts compared130110927664533930
  • Against single models it wins: lower error than ECMWF HRES at all 8 leads (8 with 95% confidence) and than NOAA GFS at all 8 (5 with 95% confidence).
  • Against the Center’s consensus it is level.
  • Against the official forecast it lost on the test season, at every lead: 125 against 100 nautical miles at five days. It was ahead on the North Atlantic 2025 season, but its settings were chosen with that season in view, so that is not the result to trust.
  • Google DeepMind’s cyclone model is better than both.
  • The uncertainty holds up: the circle meant to hold the centre two times in three held it 70% of the time, and the passage chances scored a Brier skill of 0.47, against 0.42 for the raw cloud of ensemble members.
  • Hurricane Melissa: the chance of the centre passing within 150 km of Kingston passed 40% five days before it arrived and 90% three days before.

5. Results

Every number here was measured on real, archived forecasts: no reanalysis anywhere. Truth is CMORPH satellite rain, adjusted to rain gauges. Each rival is scored raw and after the same post-processing Pluvium gets, because beating a raw global model is easy and beating it after equal treatment is the real test. Losses are reported beside wins.

A full year, seven places, 14 days

Trained on April 2024 to March 2025 and scored on every day from April 2025 to March 2026, in six regions on four continents and in Jamaica at 8 km, against ECMWF ENS, HRES and AIFS and NOAA GEFS, GFS and GraphCast. Pick a range of days:

Rain of at least 1 mm, days 1–3, a full year in seven places

Brier skill of Pluvium Blend (dark) beside the best of every other system, raw or post-processed (light), April 2025 to March 2026.

Higher is better

Rain of at least 1 mm, days 4–7, a full year in seven places

Brier skill of Pluvium Blend (dark) beside the best of every other system, raw or post-processed (light), April 2025 to March 2026.

Higher is better

Rain of at least 1 mm, days 8–10, a full year in seven places

Brier skill of Pluvium Blend (dark) beside the best of every other system, raw or post-processed (light), April 2025 to March 2026.

Higher is better

Rain of at least 1 mm, days 11–14, a full year in seven places

Brier skill of Pluvium Blend (dark) beside the best of every other system, raw or post-processed (light), April 2025 to March 2026.

Higher is better
In how many of the seven places Pluvium Blend beat every other system, raw and post-processed, on each score and lead group, April 2025 to March 2026
In how many places it beat every other systemDays 1–3Days 4–7Days 8–10Days 11–14
Brier skill, 1 mm7 of 77 of 76 of 76 of 7
Brier skill, 10 mm7 of 77 of 77 of 73 of 7
Brier skill, 25 mm6 of 76 of 77 of 74 of 7
Brier skill, 50 mm3 of 73 of 73 of 75 of 7
Critical success index, 1 mm7 of 76 of 76 of 72 of 7
Critical success index, 10 mm6 of 76 of 77 of 73 of 7
ROC area, 1 mm7 of 77 of 76 of 74 of 7
CRPS6 of 77 of 77 of 73 of 7
RMSE4 of 76 of 76 of 74 of 7
Anomaly correlation6 of 75 of 73 of 72 of 7
  • Rain of 1 mm or more: the best of every system in all six regions on days 1 to 10, and in 5 of 6 on days 11 to 14. Against each system after the same post-processing it was better with 95% confidence in all 72 comparisons on days 1 to 7, and never worse.
  • Where it does not lead: 50 mm, where post-processed AIFS was ahead on days 1 to 7 in the Alps, the Caribbean and the Great Plains; anomaly correlation beyond a week in the mid-latitudes, where raw ensembles keep more; and winter on days 1 to 3, where it led in 2 of 6 regions (most within 0.004).
  • Jamaica, with Pluvium’s network as a seventh component: it led at 1 mm on days 1 to 7. Post-processed GraphCast led on days 8 to 10. On its own, Pluvium’s network scored +0.137 on days 1 to 3, behind GraphCast (+0.179) and AIFS (+0.172) after the same treatment; as a component it helps a little.

Hurricane Melissa: the clearest failure

On 28 October 2025 Melissa crossed Jamaica as a category 5 hurricane. CMORPH measured 154 mm on average over the island’s cells, and up to 447 mm.

Hurricane Melissa: the chance of 50 mm over Jamaica, two days ahead

Averaged over Jamaica’s 450 cells, for 28 October 2025. It came true everywhere: every cell had 50 mm or more, 154 mm on average. Raw forecasts beside Pluvium Blend and Pluvium’s network.

Higher is better

Here a taller bar was the better forecast, and Pluvium came last. Calibration trained on a year without such a storm pulled a record forecast back towards normal. The daily table now reports how many raw systems call for 50 mm, so a storm they agree on is no longer hidden.

Four days ahead raw AIFS gave 100% and Pluvium Blend 42%. Melissa decided the year’s amount scores in Jamaica: in autumn the Blend lost CRPS and RMSE to post-processed GraphCast and raw ENS. In the other three seasons its CRPS was best or within 0.01 of the best. Making the pool itself keep a record forecast the systems agree on is the next step.

Version 0.3, aimed at the weak spots

The first full year found three weak spots: 50 mm, winter in mid-latitude climates, and structure at long range. Version 0.3 was aimed at them and scored on the same year by the same rules:

  • Heavy rain (50 mm): mean skill on days 1 to 3 went from +0.190 to +0.213, better in 7 of 7 places. Against each post-processed system at 50 mm, 79 of 161 comparisons are better with 95% confidence and none worse.
  • Rain or no rain (1 mm): on days 1 to 7, 81 of 84 comparisons with single post-processed systems are better with 95% confidence and none worse.
  • Days 11 to 14 got slightly worse: +0.0394 to +0.0390 at 1 mm. The online weights for that range learn from few forecasts, and cost a little.
  • Maps: on days 11 to 14 one scenario keeps 0.88 of the weather’s spatial variability, against 0.84 for the raw ECMWF ensemble.
  • A caution: the weak spots were found on this same year, so a second look at it is weaker evidence than a fresh year. One rule, how each place decides about record storms, was settled after five of the seven places had been scored.

Jamaica, July–August 2025: Pluvium’s own network against the six

The first test of Pluvium’s hour-by-hour network itself, at its small size (0.9M parameters, about two hours of training on a CPU) and driven only by GFS, against the six operational systems and two blends of them, on 450 land cells covering Jamaica. Skill scores are percentages, so a taller bar is a better forecast. Choose a day:

Will it rain? Rain of at least 1 mm, day 1

Brier skill score against the climatology of the same weeks, as a percentage. 0% is no better than climatology. Jamaica, July–August 2025.

Higher is better

Pluvium is the best single system on day 1; only the blend that includes Pluvium scores higher.

Will it rain? Rain of at least 1 mm, day 2

Brier skill score against the climatology of the same weeks, as a percentage. 0% is no better than climatology. Jamaica, July–August 2025.

Higher is better

Pluvium is the best single system on day 2; again only the blend that includes it scores higher.

Will it rain? Rain of at least 1 mm, day 3

Brier skill score against the climatology of the same weeks, as a percentage. 0% is no better than climatology. Jamaica, July–August 2025.

Higher is better

Day 3 is closer: AIFS, GFS and the blend of six edge ahead of Pluvium, and the differences are within the noise except against ENS and HRES.

Telling rainy days from dry ones: day 1, at least 1 mm

ROC area after calibration, as a percentage. 50% is a coin toss, 100% is perfect; bars start at 50%. Jamaica, July–August 2025.

Higher is better

A system with no skill would sit on the baseline. Pluvium ranks rainy and dry days better than any single operational system.

Pluvium against six operational forecast systems and two blends on Jamaica: Brier skill scores and ROC areas as percentages (higher is better) and CRPS in millimetres (lower is better), with the best value in each row marked
ScorePluviumCorX LabsENSECMWFGEFSNOAAHRESECMWFAIFSECMWFGraphCastDeepMind · NOAAGFSNOAABlend of sixAll sixBlend + PluviumSix + PluviumNever rainReference
Rain at all, at least 1 mmBrier skill, higher is better
Day 1July–August 20255.6%•6.9%3.0%•8.2%•7.3%•8.7%9.0%•13.6%−20.1%
Day 2July–August 20252.9%•9.1%1.1%•6.9%•6.4%•9.1%9.5%13.4%−20.9%
Day 3July–August 20251.4%•4.3%−0.1%•6.3%4.9%5.8%6.1%8.1%−21.0%
Wetter days, at least 10 mmBrier skill, higher is better
Day 1July–August 20254.7%6.8%6.3%7.7%7.4%8.8%5.9%4.7%6.3%
Day 2July–August 20257.4%6.8%7.2%9.0%7.0%3.5%3.8%3.6%•6.1%
Day 3July–August 20258.8%9.1%8.0%9.3%8.2%6.1%5.4%3.8%7.1%
Telling rainy days from dry ones: day 1ROC area, higher is better; 50% is chance
At least 1 mmJuly–August 202566.9%70.2%62.4%68.8%70.1%72.5%72.9%76.7%—
At least 10 mmJuly–August 202574.2%66.7%58.5%57.0%69.3%53.6%72.6%76.6%—
How much falls: daily totalsCRPS in mm, lower is better; raw forecasts
Day 1July–August 20251.451.801.941.761.832.42———
Day 2July–August 20251.671.582.411.921.762.52———
Day 3July–August 20251.681.602.361.941.732.56———
The test Pluvium lost, at least 1 mmBrier skill, higher is better
Day 1May–June 20255.9%7.1%3.9%7.0%8.2%8.2%8.7%4.6%−21.2%
Day 2May–June 20252.2%6.7%−2.3%6.4%8.3%7.4%9.4%3.8%−20.9%
Day 3May–June 2025−1.4%4.1%−1.1%5.1%6.6%4.0%6.8%7.8%−20.4%

★ Best in the row, among every system and blend (the never-rain reference is not counted). A dashed outline marks the best single forecast system when a blend is ahead of it. • Pluvium better than this system with 95% confidence (moving-block bootstrap over days). Brier skill scores and ROC areas are shown as percentages. CRPS compares raw forecasts, and only Pluvium was trained on the CMORPH truth, so part of its lead there is calibration. The blends have no single amount forecast, so no CRPS. Source: Pluvium’s operational benchmark; tested model pluvium-small (0.9M parameters), driven by GFS.

July and August 2025: protocol frozen before the data were downloaded

The scorecard above in words: 62 daily 00 UTC runs, every system calibrated with the same logistic regression on its 30 previous runs:

  • Rain at all (at least 1 mm): the highest skill of any single system on days 1 and 2. With 95% confidence it beats ECMWF ENS and HRES on all three days, AIFS and GraphCast on days 1 and 2, and the blend of all six on day 1. Against GEFS and GFS it is ahead on days 1 and 2 in 86–97% of bootstrap resamples, but the 95% interval still includes no difference. On day 3 it ties everything except ENS and HRES.
  • 10 mm or more: a tie with the best system on every day.
  • 25 and 50 mm: no system, ECMWF’s included, does meaningfully better than a forecast that never says rain. Rain this heavy at one 8 km cell on one day is too rare (0.5% and 0.1% of cell-days) for any of them.
  • Daily totals: Pluvium’s CRPS is 40–50% smaller than ECMWF ENS’s and GEFS’s on every day: 0.87 mm against 1.45 and 1.80 on day 1. This compares raw outputs, and only Pluvium was trained on CMORPH, so part of the gap is calibration.
  • It adds information: adding Pluvium to the blend of all six raises skill at 1 mm from +0.090 to +0.136 on day 1 and from +0.095 to +0.134 on day 2.

May and June 2025, the test Pluvium lost

The first test, also with its protocol fixed in advance. After calibration Pluvium was worse than the best operational system on 9 of 12 threshold-day comparisons (95% confidence) and tied on the other 3. Its raw forecasts still discriminated best, ROC area 0.76 at 1 mm and 0.84 at 10 mm on day 1, and had the best CRPS, 0.94 mm against 1.82 for ECMWF ENS.

Why it lost. Every system is calibrated on the 30 runs before each test day, so early-May forecasts were calibrated on April. April 2025 was unusually wet on Jamaica (2.6 mm per cell and day, against May’s 0.8), and Pluvium had never trained on an April. The calibration learned that any rain Pluvium forecast meant a heavy event, and on the first days of May turned modest forecasts into 96–98% chances of 25 mm, on days when at most 6% of cells saw that much. The protocol then moved from isotonic to logistic calibration, frozen in a commit before July and August were downloaded. Both protocols, for both periods:

Brier skill score for at least 1 mm on Jamaica, days 1, 2 and 3, for both test periods under both calibration protocols
Test periodCalibrationPluvium, days 1 / 2 / 3Best operational system, days 1 / 2 / 3
May–June 2025isotonic−0.032 / −0.034 / +0.021+0.082 / +0.083 / +0.066
May–June 2025logistic+0.051 / +0.026 / −0.048+0.097 / +0.107 / +0.072
July–August 2025isotonic+0.126 / +0.118 / +0.014+0.083 / +0.084 / +0.067
July–August 2025logistic (frozen)+0.127 / +0.118 / +0.052+0.087 / +0.091 / +0.063

The July–August calibration windows contain no April. That is luck of the calendar, not a fix, which is why training on every month is first on the list below.

The test Pluvium lost: at least 1 mm, day 1, May–June 2025

Brier skill score as a percentage, isotonic calibration (the protocol fixed for this test).

Higher is better

After a wet April it had never trained on, calibration turned Pluvium’s forecasts into false alarms, and it fell below climatology.

Between versions

Each version of pluvium-small was trained from scratch on the same Jamaican data with the same 32-minute budget on a 4-core CPU, then scored on held-out autumn 2021 with emulated forecast errors. Skill is against climatology unless stated.

Pluvium versions 1, 2 and 3 compared on the same data and budget, with emulated forecast errors
ScoreVersion 1Version 2Version 3
Hourly rain chance, skill+13.1%+16.0%+16.4%
Yes/no at 30%: critical success index0.1570.2000.214
Any rain in 24 h, skill+48.4%+49.3%+50.5%
Rain start within 2 h of the truth0.4540.4640.463
24 h totals, CRPS skill+15.2%+23.8%+44.6%
Heavy hours, skill+0.6%−3.2%−4.5%
Thunder at airports, skilln/a+2.6%+0.6%

Version 3, the default, nearly doubles the skill of 24-hour totals and has the best hourly skill , with 392 training steps against version 2’s 592, because its deeper engine costs more per step. Heavy-rain hours and thunder remain its weak spots.

What is still missing

  1. Keep a record storm the systems agree on. Melissa showed calibration shrinking a forecast that every raw system had right.
  2. 50 mm and winter. Version 0.3 narrowed both; post-processed AIFS still often wins heavy rain at short range.
  3. Train the real network. The network results come from pluvium-small, 0.9M parameters trained for about two hours on a CPU. Pluvium-500 is 60 times larger and designed for a day on one GPU. The first three GPU runs did not finish. The first stopped after 138 minutes. The second stopped right after it measured one forecast at 38.6 GB on the 96 GB card. The third trained on that card for the first time, then stopped when a data file was missing. Versions 0.5.4 and 0.5.5 fix what stopped them: a batch that cannot fit is not tried, a stopped run starts again and keeps the last 400 lines of its log, and every data file is checked before training starts. The first weights are now on Hugging Face: the stage A checkpoint with the lowest loss on held-out forecast times so far. Stage A is one step of the program. Training on real GFS runs (stage B), calibration and the scores on a held-out year come after it, so this page has no Pluvium-500 result yet.
  4. A fresh year. Version 0.3 was scored on the year its weak spots were found on.

Four short technical papers and a list of the ideas that were built, tested and dropped are published with the code. They are research results from one person’s project and are not peer reviewed.

6. Use it in your country

Nothing in Pluvium is tied to Jamaica. The engine works on the sphere, and every source it learns from is global: terrain, borders, towns, satellite rain, airports and the driving models. Any country between 60° S and 60° N can be set up in minutes: here, Trinidad and Tobago:

pluvium setup --country TT --root ~/pluvium-TT     # area, terrain, provinces and every town
pluvium blend train --root ~/pluvium-TT            # learn Pluvium Blend from a year of archived runs (CPU, ~2 h)
pluvium blend forecast --root ~/pluvium-TT         # 14 days of daily rain chances for every town
pluvium blend learn --root ~/pluvium-TT            # every day: update the weights from the days now verified
pluvium program --root ~/pluvium-TT --country TT   # train Pluvium's own network for the country (GPU)

pluvium blend forecast writes one row per place and day: the chance of 1, 10, 25 and 50 mm; the median, 90th percentile and expected rain; how many raw systems call for 50 mm; an alert tier; whether the forecast is beyond the training record; and how rare the amount is there.

  • --country takes ISO codes, one or several (DO,HT for Hispaniola). --box south,north,west,east sets the extent by hand; split large countries into regions.
  • Bulletins come in each country’s languages: English, Spanish or French, and Jamaican Patois for Jamaica.
  • Live “rain now” comes from GOES-East or GOES-West wherever they see, across the Americas and the Pacific. Elsewhere Pluvium forecasts without it, as a third of its training forecasts did.
  • Set up and tested from live data in Jamaica, Trinidad and Tobago, Sri Lanka (in the eastern hemisphere) and Ghana (which straddles longitude 0). An area that crosses the date line is refused with a clear message.

7. Install and run

pip install pluvium
pluvium now --country JM

pluvium now does whatever is still missing. It sets the country up in seconds, downloads a trained Pluvium Blend if one has been published for the country, or trains one if not (86 minutes for Jamaica on a four-core machine, nearly all of it downloads, leaving 41 MB on disk), then reads the newest run of each of the six systems and prints the forecast for the largest towns. After that a forecast takes about a minute. No GPU, and nobody fetches the systems’ files by hand.

pluvium now --country JM --place "Port Antonio"   # the places whose name contains this
pluvium now --country JM --parishes               # every parish: its typical place and the range across it
pluvium now --country JM --parish "St Ann"        # the largest towns of one parish
pluvium now --country JM --when                   # the first three days in six-hour periods, each with its chance
pluvium now --country JM --learn                  # learn first from the days whose rain has been published since
pluvium now --country DO,HT --storms              # two countries, and every tropical cyclone alive now
pluvium blend publish --root ~/pluvium-JM --repo you/your-repo   # share the blend you trained

In Python:

from pluvium.quick import forecast, brief
got = forecast("JM")
print(brief(got["result"], got["names"], got["population"]))

When in the day it rains

New in version 0.5. For each of the first three days, every place gets the six-hour period in which rain is most likely, in its own clock (in Jamaica: 7 pm to 1 am, 1 am to 7 am, 7 am to 1 pm and 1 pm to 7 pm). The chance for a period is the blend’s chance of a wet day, spread over the day by the systems’ members mixed half and half with when it usually rains at that place.

Checked on 61 past runs for Jamaica: on wet days the mix placed the rain 7.8% better than the usual pattern alone, while the members alone placed it 4.9% worse, so a blend without the usual pattern shows no periods. It rarely changes which period is the most likely one (66.7% right against 66.5% for the usual pattern), and the chance printed for a period is only as good as the daily chance.

The AI models it reads

pluvium upstream keeps unchanged copies of the published weights of the AI models whose forecasts Pluvium reads: DeepMind’s GraphCast, NOAA’s GraphCastGFS and AIGFS, and ECMWF’s AIFS Single. Each copy carries its publisher’s licence and a checksum for every file. Pluvium does not run these models; it reads the forecasts NOAA and ECMWF make with them.

A first checkpoint of Pluvium-500, the full-size model, is on Hugging Face. Open the checkpoint (CorxLabs/Pluvium-500, folder checkpoint/stage_a/best). Stage A is the training on reanalysis. Every 30 minutes the trainer scores the network on 32 forecast times from the held-out year, and this folder keeps the weights with the lowest loss so far. They are not the finished model. They are not calibrated, and the results on this page still come from pluvium-small.

To train your own, pluvium program runs the whole pipeline on one GPU (geography and towns, ERA5, CMORPH and airport reports, climatology, two training stages, calibration and scoring on a held-out year, publishing to the Hugging Face Hub, and the live forecast) and resumes where it stopped.

GPU_HOURS=24 HF_TOKEN=... HF_REPO=you/Pluvium-500 pluvium program

Live data

The live data feeds Pluvium reads and what each is used for
FeedUsed for
GFS (NOAA), ECMWF IFS and AIFS open dataThe driving fields: one forecast per model, mixed, so their disagreement widens the windows
GEFS, 31 members (optional)More members in the mix, so the spread of start times and totals reflects the atmosphere’s own uncertainty
GOES-East rain rateThe rain falling now, and the field carried along the wind; infrared when the rain-rate product is unavailable
MRMS Caribbean radar (NOAA)Rain now where radar reaches: Puerto Rico and the Virgin Islands
METAR airport reportsThe last three hours of weather, and the truth that online learning checks forecasts against
GOES-East Geostationary Lightning MapperLightning within 10 km of each town in the last 15 minutes

Learning from its mistakes

pluvium learn reads what the airports reported after each forecast and updates a corrector for every parish, lead time and part of the day. The corrector is pulled back towards “no correction” at every step, so it moves only on repeated evidence and relaxes when a pattern ends. A parish without an airport borrows its country’s correction.

Licence and data. The code is Apache-2.0. The data has its own terms: ERA5 (Copernicus Climate Change Service); CMORPH CDR and ISD (NOAA NCEI); GFS and GEFS (NOAA NCEP); NOAA’s real-time GraphCast; ECMWF open data, IFS, ENS and AIFS (CC-BY-4.0); GOES-East imagery and lightning (NOAA); METAR (NOAA Aviation Weather Center); GeoNames and geoBoundaries (CC-BY-4.0); and Terrarium elevation tiles (Mapzen and its sources).

8. When something goes wrong

A forecast system fails in operation in ways no benchmark shows. Pluvium checks for the failures that matter and says what it did about each in the forecast’s notes.

Failures Pluvium checks for in operation and what it does about each
FailureWhat Pluvium does
A driving model changes units (Celsius for Kelvin, hPa for Pa)Caught from physical ranges, converted, and noted
Missing fields, gaps or spikes, a frozen feedPhysical defaults, interpolation in time, or a flag in the notes
The satellite rain product is pausedFalls back to the infrared estimate and names the product used
No airport reports at allCarries on: it was trained with whole networks of reports dropped at random
A situation unlike training, a record hurricane, a new driving-model versionPulled towards climatology in proportion, with the inputs responsible named
An operational run is late or missingThe previous run at the matching lead; a system missing altogether drops out of Pluvium Blend’s pool
A storm bigger than anything in trainingHanded back, in part, to the raw systems where that helped; flagged in the daily table, with how many systems call for 50 mm
The network cannot runA statistical fallback forecast, labelled as such
Chances that contradict each otherRepaired before publishing, and the repairs counted
A half-written forecastA new forecast becomes the latest only when it is complete

9. Limits

Do not use Pluvium for real decisions. It is a research model, not an official forecast, and no independent party has verified it. For hurricanes, the National Hurricane Center is the authority.

  • One test year, looked at twice. Version 0.3 was scored on the year whose weak spots it was built for, and one of its rules was settled after most places had been scored.
  • Heavy rain is still the weak spot. Post-processed AIFS often wins 50 mm at short range, and a single season holds few heavy-rain days.
  • Record storms. Calibration pulled Hurricane Melissa’s rain back towards normal. The daily table now shows when the raw systems agree on 50 mm; read it.
  • The truth is imperfect. Satellite rain at 8 km and airport reports. Neither sees every shower, the satellite under-reports mountain rain, and it sees winter snow poorly.
  • A small network. Every network result comes from pluvium-small. Pluvium-500 has a stage A checkpoint on Hugging Face and no scores on this page.
  • Hurricane tracks only. Pluvium Cyclone forecasts where a storm goes, not how strong it gets, and it trails the National Hurricane Center’s official forecast.
  • Flood risk is a rule. Its thresholds have not been verified against flood reports, and its alerts carry the status “Exercise”.
  • Long range is hard for everyone. Skill at 1 mm on days 11 to 14 is small for every system, and the lead changes hands often.
  • Not polar. The satellite truth stops at 60° of latitude.

10. Made in Jamaica, for island rain

Pluvium is built in Jamaica, for islands where a shower can soak one side of a mountain and leave the other dry, and where a global model’s grid square can be wider than a parish. It starts from Jamaica, its bulletins even come in Patois, covers every town and village in the Caribbean, and can be set up for any country that wants it.

It is the second weather model from CorX Labs. NatureV1 forecasts the whole atmosphere and hurricanes; Pluvium uses the same iHELIX engine at the scale of a town, for the rain people plan their day around. Hurricane Melissa, which struck Jamaica in October 2025, is the storm the project has learned most from. The engine grew out of HELIX, the lab’s linear-context architecture.

For warnings in Jamaica, follow the Meteorological Service of Jamaica and the Office of Disaster Preparedness and Emergency Management (ODPEM). Elsewhere, follow your national meteorological service. Pluvium is not a substitute for any of them.

11. Questions

What is Pluvium?

Pluvium is an open-source AI forecasting system built by CorX Labs in Jamaica and released under Apache-2.0, in three parts. Pluvium’s network forecasts rain at every town hour by hour for three days: the chance of rain, whether it will be drizzle, showers, heavy rain or a thunderstorm, when it will start, how long it will last and how much will fall. Pluvium Blend gives each day’s rain out to 14 days for any country between 60° S and 60° N. Pluvium Cyclone forecasts hurricane tracks. It is a research model, not an official forecast.

Is Pluvium an official forecast?

No. It is a research model that no independent party has verified, and its flood levels are a rule that has not been checked against flood reports. Do not use it for real decisions. Follow your national meteorological service; in Jamaica, the Meteorological Service of Jamaica and ODPEM. For hurricanes, follow the National Hurricane Center.

How accurate is Pluvium?

Over a full year, April 2025 to March 2026, in seven places (Jamaica, the Caribbean, the Maritime Continent, the Alps, the US Pacific Northwest, the US Great Plains and the southern Andes), Pluvium Blend had the highest Brier skill of every system, raw or post-processed, for rain of at least 1 mm on days 1 to 7 in all seven places. Against each of ECMWF ENS, HRES and AIFS and NOAA GEFS, GFS and GraphCast after the same post-processing, it was better with 95% confidence in all 72 comparisons on days 1 to 7 in the six regions, and never worse. It does not lead at 50 mm, where post-processed AIFS often wins at short range, and it under-forecast Hurricane Melissa’s rain over Jamaica: a 66% chance of 50 mm two days ahead, when every raw system said 83 to 100%.

Can Pluvium forecast hurricanes?

Pluvium Cyclone forecasts where tropical cyclones in the North Atlantic and the eastern North Pacific will go: a track to five days, a circle that should hold the centre two times in three, and the chance of the centre passing within 150 km of each place. Track only, not intensity. On its test season it beat ECMWF HRES and NOAA GFS at every lead and was level with the National Hurricane Center’s consensus, but it was behind the Center’s official forecast at every lead (125 against 100 nautical miles at five days), and Google DeepMind’s cyclone model was better still. For a live storm, follow the National Hurricane Center.

Can Pluvium forecast rain in my country?

It can be set up for any country between 60° S and 60° N, because nothing in it is tied to Jamaica. pluvium setup --country with an ISO code builds the area, terrain, provinces and every town, and Pluvium Blend can be trained for it from a year of archived runs: about two hours of downloads and under an hour of CPU. It has been set up and tested from live data in Jamaica, Trinidad and Tobago, Sri Lanka and Ghana, with bulletins in English, Spanish, French and Jamaican Patois. Its skill in a new country is unknown until it is scored there.

What is Pluvium Blend?

Pluvium Blend combines six operational systems (ECMWF ENS, HRES and AIFS, and NOAA GEFS, GFS and GraphCast), plus Pluvium’s own network where it runs, into one calibrated forecast of each day’s rain out to 14 days. Each system is first post-processed on its own; a pool weighs them, and its weights move after every verified day towards whoever has been right. A forecast beyond anything in the training record is handed back, in part, to what the raw systems say, in the places where that helped. It needs no GPU.

How is Pluvium related to NatureV1?

Pluvium’s network runs on the same engine as NatureV1: iHELIX, whose attention heads have physical radii in kilometres on the sphere. NatureV1 forecasts the whole atmosphere and hurricanes on a 1.5° global grid. Pluvium uses the engine on a Caribbean mesh of 6,075 points, about 11 km apart over land, with physics such as mountain rain and sea-breeze convergence built in, to forecast the rain at each town hour by hour.

Is Pluvium free and open source?

Yes. Version 0.5.5 is released under the Apache-2.0 licence, which allows commercial use, and installs with pip install pluvium. Pluvium Blend is trained for your own country on a CPU, and Pluvium Cyclone needs nothing trained. The full-size network, Pluvium-500, has a first checkpoint on Hugging Face, from stage A of its training. It is not calibrated, and this page has no scores from it yet. The data it learns from has its own terms, credited on this page.

Who built Pluvium?

Pluvium was built by Nathan Lyttle, the founder of CorX Labs and a solo developer from Jamaica, who also built NatureV1 and its iHELIX engine.

New

Open code, published results, including what it got wrong.

Pluvium 0.5.5 is on PyPI under Apache-2.0, and every result on this page is reported in full, the full year, the hurricane tracks, Hurricane Melissa and both of the network’s first tests.