Optimizing Cloud Seeding with a Denoising Diffusion Model

Aerosol-cloud interactions (ACI) are a major source of uncertainty in estimating the total climate forcing of the Earth system. Natural and anthropogenic variations in the aerosol distribution, or aerosol perturbations, are a vital tool for constraining ACI. On the other hand, accurately estimating the specific effects of aerosol perturbations on cloud formation, precipitation efficiency, or radiative transfer, requires robust statistical analysis to bridge idealized numerical simulations and observations. In this talk, we introduce a novel machine learning (ML) approach to investigate how the spatial and temporal variability of aerosol perturbations affects cloud and precipitation properties. Specifically, we implement a conditional denoising diffusion model to sample different trajectories of aerosol perturbations in two different dynamic environments with super-droplet microphysics: a 1D kinematic driver (KiD) simulation and a prescribed 2D stratocumulus flow. Unlike model-based reinforcement learning (RL) and optimal control methods, denoising diffusion models are adept at handling high-dimensional stochastic environments while offering flexible conditionality and long-horizon scalability. Our findings demonstrate that cloud seeding can significantly influence observed precipitation patterns depending on the size and timing of perturbation as well as the characteristics of the background environment. Additionally, we discuss the assumptions required to extend the diffusion model planning algorithm to a large eddy simulation (LES). The methodology introduced here also has broader applications, including designing cloud seeding trajectories for uncrewed aerial vehicles (UAVs) in field campaigns such as the CLOUDLAB project [Henneberger et al., 2023], and identifying the optimal aerosol size distribution for marine cloud brightening experiments [Hoffmann and Feingold, 2021].