Researchers fitted growth models to 290 levitated ice crystals, using up to 500 seconds of mass measurements each.[1]
Adding mass dependence cut total squared errors in predicted mass ratios by about 60% against the Nelson–Baker model. Each ratio measures mass relative to its starting value. Adjusting surface attachment alone fitted less well.[1]
The improvement points to attachment and shape changing as crystals grow, inferred through mass measurements.[1]
Symbolic regression, an equation-search method, turned the learned relationship into a formula. Separate AIDA chamber measurements helped select its form, which improved low-temperature predictions. The neural model overpredicted growth beyond 1,000 seconds.[1]
Sources & further reading
- Discovering how ice crystals grow using neural ODEs and symbolic regression Science Advances · 25 Sep 2026 · Accessed 20 Oct 2026