How it works

From weather history to a reason you can read.

Snowvy runs a layered snowpack model on each slope class of a place it has built, keeps a record of how each layer formed, and turns that record into short dated sentences.

Modelled, not an avalanche forecast. Snowvy explains the snowpack; it never says a slope is safe. Read your regional avalanche bulletin and make your own decisions.

1. Split the mountain into virtual slopes

Terrain comes from the Copernicus GLO-30 elevation model at about 30 m. Each cell is described by its elevation, slope, aspect, sky view, horizon and forest cover. Cells that share a 200 m elevation band, one of eight aspects (or flat), a steepness class (low, moderate, or steep at 30° and above) and forest cover form a virtual slope. The snowpack is simulated once per virtual slope, which is what makes hourly physics affordable over whole ranges.

2. Run the snowpack hour by hour

Weather for the slope model comes from Open-Meteo; the regional estimate uses ECMWF open data (IFS forecasts). It is adjusted to each slope's elevation, sun and shelter. Then every hour the model adds, compacts, warms, wets and refreezes layers:

Process	What the model tracks
New snow	Density from temperature and wind; one layer per storm
Wind drift	Which layers can be moved, and how wind-packed snow becomes denser
Heat and melt	Conduction through the pack, the surface energy balance, and ground heat at the base
Water	Meltwater and rain held in each layer, moving down, and refreezing
Crusts	Wet layers that refreeze into melt-freeze crusts, and stay crusts
Grain change	Rounding, faceting under strong temperature gradients, and depth hoar
Surface hoar	Growth on clear, calm, humid nights, destruction by wind, sun or melt, and burial

The formulations come from published, peer-reviewed snow science, including the Crocus model (Vionnet et al. 2012), dry metamorphism from Brun et al. (1992), crust formation from Jamieson (2006) and surface hoar from Horton et al. (2015).

3. Remember how every layer formed

Each layer keeps the date it formed and how (snowfall, wind-packed snowfall, surface hoar, refrozen melt or refrozen rain), plus a list of dated episodes such as melt, rain, refreeze, crust formed, faceting and buried, with the temperatures, gradients and water amounts that drove them. Buried layers are named by the day they were buried, the way forecasters talk about them.

4. Look for the known warning signs

Snowvy checks each layer against structural warning signs taken from avalanche research on simulated profiles (Schweizer & Jamieson 2007; Monti et al. 2014): large or persistent grains, sharp changes in grain size or hardness, and how deep the layer is. It reports the strongest of these concern types:

• Buried surface hoar
• Crust–facet combination
• Persistent weak layer
• Wind slab
• New snow (30 cm or more in 72 h)
• Wet snow
• Surface crust

5. Write the reason, with numbers you can check

Explanations are filled in from fixed templates using only the recorded episodes, like the phrase catalogue of the Swiss avalanche bulletin. Nothing is invented by a language model. A real example from the pilot:

About 80 cm down lies the layer buried on 5 Feb: surface hoar that grew on the snow surface during clear, calm nights around 1 Feb. It is now covered by about 80 cm of snow (snowfalls on 5 Feb, 11 Feb, 15 Feb), partly wind-deposited. 5 of 6 structural warning signs are present.

Steep, open north-facing slope at 2,200 m above Canal Roya (Pyrenees), 15 Feb 2026, run on historical weather.

What it does not do yet

• It does not compute stability measures such as critical crack length, and it does not give danger levels.
• Validation against measured snow depth and against forecasters' reported layers is under way. Until it is published, treat the model as an informed explanation, not a measurement.
• Detailed slope modelling is built place by place. Where it does not exist yet, Snowvy shows a regional estimate, labelled as one, or nothing. Each ski area page says whether its model is ready.

Key references

• Vionnet et al. 2012, Crocus. doi:10.5194/gmd-5-773-2012
• Brun et al. 1992. doi:10.3189/S0022143000009552
• Jamieson 2006. doi:10.1029/2005RG000176
• Horton et al. 2015. doi:10.5194/tc-9-1523-2015
• Schweizer & Jamieson 2007. doi:10.1016/j.coldregions.2006.08.011
• Herla et al. 2024. doi:10.5194/nhess-24-2727-2024

Join the waitlist