Abstract. Ice aggregation in clouds plays a crucial role in cloud development and precipitation formation. Despite the significance of ice aggregation, direct in situ quantification of aggregation rates in natural clouds has been challenging due to the difficulty of tracking ice crystals. Here, we present in situ measurements of ice aggregation rates in persistent supercooled stratiform clouds. Using novel glaciogenic seeding experiments (CLOUDLAB), ice crystals are nucleated upwind and subsequently measured downwind after a known advection time in cloud, allowing us to estimate their age. A deep-learning-based detection algorithm (IceDetectNet) counts the individual monomers of aggregates to derive the initial ice crystal number concentration (ICNCt0). We considered several factors that may influence ice aggregation, including ICNCt0, temperature, ice crystal size, aspect ratio, and turbulence. Among these, ICNCt0 was found to be the dominant factor controlling aggregation rates by three independent approaches: causal inference, a physical equation, and machine learning models. We report, however, a subquadratic dependence of the aggregation rate on ICNCt0 (mean exponent ∼ 0.92; 95 % CI: 0.88–0.97), in contrast to theoretical expectations (quadratic dependence). One possible explanation is that aggregation may also involve smaller ice crystals, but this remains hypothetical. To predict aggregation rates, we evaluated 11 machine learning models and a physically based formulation. CatBoost achieved the best statistical performance, while the physical model proved more robust in sensitivity tests. These findings provide new insights into the microphysical and environmental controls of ice aggregation and establish a robust methodological foundation for studying aggregation processes in natural clouds.
Zhang et al. (Wed,) studied this question.