On the Stochasticity of Aerosol-Cloud Interactions within a Data-driven Framework

Aerosol-cloud interactions (ACI) pose the largest uncertainty for climate projections. Among many challenges of understanding ACI, the question of whether ACI is deterministic or stochastic has not been explicitly formulated and asked. Here we attempt to answer this question by predicting cloud droplet number concentration Nc from aerosol number concentration Na and ambient conditions. We use aerosol properties, vertical velocity fluctuation w', and meteorological states (temperature T and water vapor mixing ratio q_v) from the ACTIVATE field observations (2020 to 2022) as predictor variables to estimate Nc. We show that the climatological Nc can be successfully predicted using a machine learning model despite the strongly nonlinear and multi-scale nature of ACI. However, the observation-trained machine learning model fails to predict Nc in individual cases while it successfully predicts Nc of randomly selected data points that cover a broad spatiotemporal scale, suggesting the stochastic nature of ACI at fine spatiotemporal scales.