AirLume models the atmospheric physics behind lightning formation to forecast strike risk along a flight route, before crews would otherwise see it coming.
Aircraft are struck by lightning one to two times a year on average. Current systems identify a strike only once it has happened, triggering a mandatory inspection regardless of damage.
Without route-specific risk data, crews avoid entire storm systems with detours of 100 or more miles, adding fuel burn and delay for risk that may be concentrated in a small stretch of the route.
63% of recorded strikes occur in conditions that flight crews did not associate with adverse weather, sometimes 25 miles from the nearest radar return. Visual and radar cues alone are not sufficient.
AirLume doesn't forecast weather, it forecasts lightning formation, using the same atmospheric data already available to flight planning, run through plasma physics models rather than pattern-matching on cloud cover.
Calculates the voltage at which atmospheric breakdown occurs from pressure and gap distance at a given altitude.
Models how free electrons multiply in an electric field to estimate charge buildup and discharge likelihood.
Computes field strength at cruise altitude from real-time atmospheric data, ahead of any visible discharge.
Runs risk assessment at each cruise level so a route can be re-planned by altitude, not only by heading.
| Component | Result | Target | Status |
|---|---|---|---|
| C physics engine | 88% on 181 historical strikes | 80%+ | Met |
| Real-time performance | 88.2 ms average per route | <100 ms | Met |
| Ada safety monitor | 100% alert classification, 20/20 | DO-178C dissimilar redundancy | Met |
| ML enhancement layer | 55% recall, 5-fold cross-validation | 80% | Below target |
The ML layer's shortfall is attributed to a 100-sample training set, too small for a Random Forest classifier to generalise. The physics engine and Ada safety monitor, which carry the core prediction and certification logic, both met their targets independently.
Pulls and validates atmospheric data from OpenWeatherMap on a 15–20 minute refresh cycle.
Runs Paschen's Law and Townsend Avalanche calculations across waypoints in under 100ms, roughly 60x faster than the equivalent in Python.
Independently re-derives risk using a separate empirical model, flags any disagreement beyond 15%, and enforces ARINC 653 timing budgets, the dissimilar redundancy DO-178C requires.
Orchestrates the pipeline, persists results, and renders the route as a colour-coded waypoint map.
The full technical report covers design iterations, the physics model, and complete test results. Happy to walk through any part of it.