How our wind model works — and where it doesn’t.

1. How much sun the slope is getting. From the angle and direction of the ground under you — read off USGS elevation data — against where the sun actually is. Cloud cover damps it. If a ridge is in the way, the slope gets nothing, and we check sixteen directions because the sun moves.
2. How warm that has made it. Not instantly — the ground has to heat before it heats the air. We track that as a running balance, which is why the model produces a lag rather than flipping at sunrise.
3. Which way that pushes the air, and how hard. Warm air climbs the slope, cold air drains down it. Strength scales with how steep the ground is and how strong the heating or cooling has been.
4. Add the regional wind. As vectors, not numbers — then we say which of the two is in charge.

When the flow reverses

53 observed reversals at 17 stations across Colorado, Wyoming and Montana, on clear days; the stations report hourly, which alone contributes ±30 minutes. Still more than an hour of honest uncertainty — which is why the app shows a window rather than a time.

Full disclosure: a constant offset from sunrise, fitted to the same data, ties our model on clear days (64 vs 65 min). The physics earns its keep where that shortcut has no mechanism — cloudy mornings, when weak heating pushes the flip later — and that case is not yet validated, because measuring it needs cloudy-day data we are still collecting.

A correction we shipped in August 2026

Earlier versions predicted the instant the slope’s thermal balance crossed zero and presented that as “the flip.” But a zero crossing is zero metres per second — invisible on your face. What you feel, and what stations measure, is the new flow actually establishing, and that runs about 2.5 hours behind the crossing. The app was calling the flip early by roughly that much, with full confidence. It now predicts the felt flip, and the error bars above are measured against exactly that event.

How strong the slope flow is

860 clear-day station-hours at 34 stations, comparing observations against the national model to isolate the slope component; the coefficient is fitted on the 343 daytime hours where the direction model is validated. A sunny 20° slope works out to roughly 3 mph of thermal — a modest, real push, not a gale.

An earlier version published a much larger win here (1.10 vs 2.46 m/s). That fit had a statistical flaw — it compared magnitudes, which cannot cancel noise — and overstated the flow’s strength about threefold. The numbers above are from the corrected fit, and the arrows in the app shrank to match.

Which way it blows — and the part we got wrong

Night-time drainage direction is no better than a coin flip, and that is the opposite of what we expected. Cold air follows the drainage — steered by the shape of the valley — rather than the face you happen to be standing on. Steeper stations scored worse, which is what that explanation predicts.

So the app no longer states a drainage direction as fact. It tells you the slope is draining and roughly how hard, and says plainly that the compass direction is unreliable. The scent cone widens to match. We would rather lose the feature than have you trust it on a stalk.

• Narrow canyons. Flow follows the canyon, not the slope. This is the same effect that breaks drainage direction above.
• Downwind of a sharp ridge in strong wind. The surface wind can reverse entirely in the lee. Completely unmodelled.
• Basins where cold air pools. Drainage collects and stops. We will keep predicting movement that has settled.
• Timbered slopes. Wind under a canopy decouples from everything above it. This is a common bedding situation and it is our weakest case.
• A front moving through. The regional wind changes faster than the model assumes; give it a few hours.
• Sunlit snow. Changes the heat balance; we will overestimate.
• Sunrise and sunset times assume a flat horizon. In a steep drainage, real first light can be an hour later.