OYA AI Lite

What OYA AI Lite actually computes

This is a working prototype of the modelling layer behind OYA AI, running entirely in your browser on real Jamaican terrain. Nothing here is a video or a rendering of a storm that already happened. Every number on the screen is integrated forward from the starting conditions you set, through the equations below, on a grid of the real island.

1. The ground

• Elevation is read from AWS Terrain Tiles at zoom 11 — about 73 m per pixel at this latitude — which carries SRTM and NED over land and ETOPO1 under the sea. Every 265 m model cell averages roughly thirteen of those samples, which is both finer and quieter than sampling a point. The island comes out at 11,005 km² against a surveyed 10,991 km², and the Blue Mountains reach 2,203 m against a surveyed 2,256 m: a sharp summit loses a little height when a 265 m cell averages across it.
• Bathymetry is the same source, carried on a separate coarser grid that reaches 90 km offshore so the model can see where the shelf ends. The shelf break is found by marching outward from each coastal cell until the water passes 180 m, which is what makes the south coast's broad shallow shelf behave differently from the north coast's narrow drop-off.
• Parishes come from geoBoundaries ADM1 at full resolution, derived from OpenStreetMap.
• Slope, relief and distance to the coast are computed from the elevation grid when the page loads.
• Drainage is built by filling depressions with a priority-flood pass, then routing D8 flow downhill and accumulating contributing area, so every cell on the island knows how much land drains through it and how high it sits above the nearest gully or river.
• Exposure: , modelled by spreading each parish's 2011 census total over the settlement points in that parish with a distance-decay kernel, weighted towards lower and flatter ground.

2. The storm

Intensity is bounded by the sea, not by a slider. The ceiling is the empirical maximum potential intensity of DeMaria and Kaplan (1994):

Wind shear is subtracted from that ceiling, and the storm relaxes towards whatever is left. Once the centre crosses the coast the Kaplan and DeMaria (1995) inland decay model takes over, and rough terrain accelerates the decay:

The wind field itself is the Holland (1980) gradient wind profile, with the Holland shape parameter B recovered from the peak wind and the pressure drop, and the radius of maximum wind from Willoughby, Darling and Rahn (2006):

On top of that: the storm's forward motion is added to the right of its track, which is why one side of a hurricane is always worse than the other; the flow turns inward by an inflow angle that steepens over land; the gradient wind is reduced to the surface and converted to a three-second gust; and the whole vortex drifts north-west from the beta effect even when you steer it elsewhere.

3. Terrain, which is where small islands are different

A global model at 25 km resolution sees Jamaica as a handful of cells and cannot tell the Blue Mountains from Portmore. At 265 m the difference decides everything:

• Wind is reduced by surface roughness over land, then increased again over exposed ridges and reduced in sheltered hollows, scaled by how far each point stands above its own surroundings.
• Rain follows an R-CLIPER-form radial profile, then is multiplied by the vertical velocity forced when that wind crosses the terrain. Windward slopes get up to three and a half times the open-ground rainfall; lee slopes get less than half. That is why the same storm drops 200 mm on the Liguanea plain and far more on the ridge above it.

4. Water

Storm surge is the sum of the inverse-barometer rise, wind setup across the shelf, and wave setup from the significant wave height the wind and fetch can build:

That head is then pushed inland cell by cell, losing height to friction faster through built-up ground than across open land, and it only reaches places the water can actually flow to. Fetch and depth in that wind-setup term are measured from the real sea floor at each stretch of coast, not assumed. The march out to sea stops at land as well as at the shelf break, because a ray that clips a cay would otherwise count kilometres of dry ground as nought metres of water and roughly double the surge. The setup term itself is the closed shallow-water solution rather than its linear limit, so it saturates instead of stacking ten metres of water on a shelf that cannot hold it. A Gilbert-scale storm builds about 1.8 m on the model's south coast, against the 6 to 9 feet in the storm report.

Rain flooding uses the SCS curve number method for how much of the rain runs off rather than soaking in, routes that runoff down the D8 network, converts accumulated volume into a peak discharge through a time of concentration that grows with catchment area, and turns discharge into a water level using the HAND method: the water surface at the nearest drainage line, minus how high your ground sits above it. Flat ground also ponds directly.

5. Slopes

Rain-triggered landslides use the infinite-slope factor of safety, with the saturation ratio driven by accumulated rainfall and upslope contributing area:

Below 1.0 the slope is expected to fail. In Jamaica this is not a footnote: more people have been killed by slope failure and flooding than by wind.

6. Damage

Wind damage uses a Weibull vulnerability curve per construction class, with the mix of classes shifting towards engineered building in denser areas. Flood damage uses a stage-damage curve for a single-storey dwelling. A slope failure under a building is treated as close to a total loss. The three are combined as independent causes, so a house cannot be destroyed twice.

7. Loss of life, and the warning that changes it

Drowning follows Jonkman, Vrijling and Vrouwenvelder (2008): the share of the exposed population that dies is lognormal in water depth,

applied to the fraction of a cell's residents who are actually in the water, and scaled up where the water arrives fast, which is what surge does and rainfall flooding does not. Wind deaths are taken from the population in dwellings the vulnerability curve has already destroyed; landslide deaths scale with how far the factor of safety has been pushed below 1, because a slope at 0.98 moves a few faces and a slope at 0.6 moves the hillside. No household is killed twice: the four causes are combined as independent risks.

Those last three rates are calibrated, not published. They are set so that a Gilbert-track run lands at 42 deaths against the 45 Jamaica actually recorded in September 1988, with gusts of 255 km/h, 1.8 m of surge against the 6 to 9 feet in the storm report, and US$5.0 bn of damage. Treat them as honest guesses with one known anchor, not as measurements.

Warning lead time then multiplies through all four. Evacuation effectiveness rises with the hours people are given and saturates, because no warning saves nobody and even a perfect one cannot move everyone:

Surge responds to it almost completely, wind least of all — people who stay are in the building when it fails. Move the slider at the top of the screen and watch the death toll move with it. On the model's Gilbert run, going from no warning to 24 hours takes the toll from 81 to 38. That gap is the entire reason OYA AI exists: the hazard is fixed, the warning is not.

Where this is known to be wrong. One anchor is not a validation set, and the model over-predicts the very worst cases. Run the Melissa scenario and it kills several hundred people, where the real October 2025 storm killed of the order of forty-five. Two things it does not represent account for most of that gap: Jamaica had roughly three days of warning, beyond the range this evacuation curve was fitted over, and forty years of building code sit between 1988 and 2025 that the vulnerability curves treat as a single construction mix. Read the extreme scenarios as what the physics does to the exposure, not as a forecast of a death toll.

What this prototype is not

It is not a forecast. It does not ingest live observations, satellite data or model guidance, and it will not tell you what the weather is going to do next week. The exposure layer is modelled from census totals and settlement points, not a building footprint register. Replacement costs, construction mix and the three calibrated fatality rates are stated assumptions, and they are in the table above so you can argue with them. One historical anchor is not a validation set.

What it is: the physics and the method, working end to end on real terrain, at the resolution a small island actually needs. The full OYA AI system extends this with live observations, assimilated model guidance, building footprints, road and utility networks, agent-based scenario testing, and validation against the historical record of events Jamaica has already lived through.

Back to OYA AI  ·  Talk to Maestro AI Labs