Models

What We Do

We take the highest-resolution wind model available for Maui and apply AI corrections trained on 19 local weather stations. The result: forecasts that account for Maui’s complex terrain, not just the broad trade wind pattern.

What We Do

We use the highest-resolution wind model available for Maui — the UH WRF at 2 km — to produce hourly wind forecasts for 5 coastal regions. The model simulates atmospheric physics on a fine grid, capturing Maui’s trade wind patterns and terrain effects.

Why Corrections Matter

Even the best wind model for Maui (UH WRF at 2 km) systematically overpredicts wind by 4–6 knots at sheltered regions like South Maui, West Maui, and East Maui. A forecast that says “15 knots” when it’s actually 8 isn’t useful. Our AI models learn these terrain biases from a year of observations and correct for them automatically.

How Accurate

Tested against held-out observations from 19 stations across 4 corrected regions.

Wind speed accuracy by region

Accuracy is the typical (RMS) error on held-out test data. Individual forecasts may vary.

At sheltered sites, the physics model overpredicts by 4–6 knots on average. Our correction reduces this systematic bias to near zero while cutting overall forecast error by 50–70%. On the exposed North Shore, the physics model is already accurate (within 4 kt) so no correction is applied.

What We Display

• Wind maps – Base physics model output, useful for seeing island-wide patterns and trends at a glance
• Regional forecasts – AI-corrected wind speeds for the first 2 days, then base physics model output for days 3–5. When a correction is applied, you’ll see an annotation like “↓35%” — meaning the corrected speed is 35% lower than what the physics model predicted
• North Shore – Always shows physics model output — the model is already accurate on the exposed trade wind coast

What We Display

• Wind maps – Color-coded wind speeds across the island with animated playback through the day
• Regional forecasts – Hourly wind speed, direction, and gusts for each of Maui’s 5 coastal regions, out to 5 days

How It Works

Two layers:

• Physics model (WRF 2 km) – The UH/PacIOOS WRF atmospheric model simulates wind physics on a 2 km grid, producing hourly forecasts out to 5 days. We extract winds at each of Maui’s 5 coastal regions from nearshore over-water grid points.
• AI terrain correction – Per-station models trained against local observations learn how the physics model’s spatial patterns relate to its systematic errors at each location. The correction uses only model data — no dependence on live station observations — so accuracy is the same whether stations are reporting or not.

• Water points only – Grid points over land are excluded
• Nearshore focus – Only points within 3 km of the coastline
• Averaged readings – Each region samples 13–15 grid points

How It Works

• Physics model (WRF 2 km) – The UH/PacIOOS WRF atmospheric model simulates wind physics on a 2 km grid, producing hourly forecasts out to 5 days. We extract winds at each of Maui’s 5 coastal regions from nearshore over-water grid points.

• Water points only – Grid points over land are excluded
• Nearshore focus – Only points within 3 km of the coastline
• Averaged readings – Each region samples 13–15 grid points

The Base Model

Most weather apps rely on global models (GFS, ECMWF) where grid cells are larger than the terrain features that shape Maui’s wind — valleys, ridgelines, headlands. These models capture broad patterns (“strong trades all week”) but can’t distinguish a windy North Shore from a calm South Shore on the same afternoon. We use the 2 km WRF run by UH specifically for Maui and Oahu — the highest-resolution public wind model available for this area.