Inferring the controlling factors of ice aggregation from targeted cloud seeding experiments

Overview

Ice aggregation in supercooled stratiform clouds constitutes a critical microphysical process governing precipitation formation and cloud development. Although ice aggregation is recognized as essential to cloud evolution, quantification of aggregation rates in situ has remained technically limited due to challenges in tracking individual ice crystals. This study employs targeted glaciogenic seeding experiments combined with deep learning-based detection to measure aggregation rates directly in natural clouds and identify the dominant environmental and microphysical factors controlling aggregation efficiency.

Methods and approach

In situ measurements were conducted within persistent supercooled stratiform clouds using the CLOUDLAB glaciogenic seeding framework. Ice crystals were nucleated upwind of measurement locations, with downwind sampling after known advection intervals enabling estimation of crystal age. A deep learning detection algorithm (IceDetectNet) quantified individual monomers comprising ice aggregates to determine initial ice crystal number concentration (ICNCt0). The analysis examined multiple potential controlling variables: ICNCt0, temperature, ice crystal size, aspect ratio, and turbulence characteristics. Three independent methodological approaches were employed to identify dominant factors: causal inference analysis, physical equation derivation, and machine learning model evaluation. A suite of 11 machine learning models and a physically based formulation were compared for aggregation rate prediction.

Results

Initial ice crystal number concentration (ICNCt0) emerged as the dominant factor controlling aggregation rates across all three analytical approaches. However, the relationship between aggregation rate and ICNCt0 exhibited subquadratic dependence with a mean exponent of approximately 0.92 (95% confidence interval: 0.88–0.97), diverging from theoretical predictions of quadratic dependence. This discrepancy may indicate involvement of smaller ice crystals in aggregation processes, though this remains unconfirmed. Among machine learning models, CatBoost demonstrated superior statistical performance metrics. The physically based formulation, while not achieving optimal statistical fit, demonstrated greater robustness across sensitivity analyses, suggesting superior generalization potential.

Implications

The identification of ICNCt0 as the primary aggregation control factor provides empirical validation for concentration-dependent aggregation mechanisms in natural clouds. The subquadratic exponent deviation from theoretical expectations indicates incomplete understanding of the aggregation process and suggests that existing parameterizations may require refinement to account for the contribution of smaller ice crystal populations.