WORKS BEST
Monitored locations
Where the latest PM10 reading is available, recent history provides a powerful anchor for the next-hour forecast.
Measured, not imagined
Promising where it matters. Honest where it doesn’t.
We test on later dates the model never saw during training. That gives agencies a clearer picture of how it may behave in the real world.
5.9µg/m³ average errorJuly–December 2025 held-out test
72%elevated-hour detectionPM10 at or above 50 µg/m³
423K+forecast comparisonsPrediction vs. actual hourly reading
124California sitesUsed across the broader research pipeline
What “5.9” means
A practical everyday signal.
For each test, the system used information available at one hour to forecast the PM10 reading at the next hour. Across more than 423,000 comparisons, the prediction differed from the measured value by about 5.9 µg/m³ on average.
That average describes forecasts at monitored California locations with a current PM10 reading. It does not claim equal accuracy at every point between monitors.
Actual reading: 30 µg/m³
A typical-sized error: about 6 µg/m³
Illustration of the average error—not a guaranteed range for every forecast.
Transparent limits
Strongest for day-to-day use.
The model is designed as a supplemental information layer. Sudden wildfire smoke and windblown dust remain harder than ordinary conditions.
USE WITH CARE
Fast-moving spikes
The system catches many elevated hours but can still miss the size or timing of severe smoke and dust events.
NOT YET CLAIMED
Every point on a map
Accuracy between sparse monitors needs local field validation before district-wide grid forecasts can be presented confidently.
Evaluate AeroSentinel against your own data.
We can design a chronological, station-aware pilot for your district.