What controls ice crystal aggregation in clouds

Researchers measured ice aggregation directly inside persistent supercooled stratiform clouds by creating ice crystals upwind and sampling them downwind after a known travel time. A deep-learning detector counted monomers in aggregates to establish initial concentrations, and several analytical approaches found that initial crystal concentration was the dominant control on aggregation rates. The observed relationship between aggregation rate and initial concentration was subquadratic, with a mean exponent near 0.92 (95% CI: 0.88–0.97), differing from the quadratic dependence expected from theory. Machine learning and a physical formulation were compared for prediction, with CatBoost performing best statistically and the physical model showing greater robustness in sensitivity tests.

What the study examined

This work focused on how ice crystals coming together inside clouds are controlled by environmental and microphysical factors. The team ran targeted seeding experiments in persistent supercooled stratiform clouds, making ice form upwind and measuring the same crystals downwind after a known advection time, which provided estimates of their age.

A neural network-based tool, IceDetectNet, was used to count the individual building blocks of crystal clusters so that an initial ice crystal number concentration could be derived. The study evaluated several candidate controls, including initial concentration, temperature, particle size and shape, and turbulence.

Key findings

Three independent analysis approaches — causal inference, a physical equation, and machine learning — all pointed to initial crystal concentration as the dominant factor controlling how quickly crystals stick together. The measured relationship between aggregation rate and initial concentration was weaker than many theoretical expectations: the average exponent was about 0.92 with a 95% confidence interval of 0.88–0.97, indicating a subquadratic dependence.

• Deep-learning detection enabled estimation of starting concentrations by counting monomer particles inside clusters.
• Among 11 machine learning models tested for predicting aggregation rates, CatBoost had the best statistical performance.
• The physically based model proved more robust when sensitivity was examined, suggesting complementary strengths between data-driven and physically based approaches.

Why it matters

Understanding the controls on how crystals combine matters because that process influences cloud development and how precipitation forms. By combining targeted seeding, an automated detection tool, multiple analysis methods, and both data-driven and physical prediction tools, the study offers a methodological framework for measuring and predicting aggregation behavior in natural clouds.

The finding that the dependence on starting concentration is subquadratic points to more nuance in how collisions and sticking play out in real clouds than some simple theories predict, and highlights the value of direct in situ measurements paired with modern analysis methods.