Physically Regularized Machine Learning Emulators of Aerosol Activation

Abstract:

The activation of aerosol into cloud droplets is an important step in the formation of clouds. Directly resolving this process in climate models is challenging due to the computational complexity required to resolve the necessary chemical and physical interactions. To address this, climate models use a variety of parameterizations which are more computationally efficient at the expense of reduced process representation and accuracy. Here, we explore how machine learning emulators can be used to bridge the gap in computational cost and parameterization accuracy and improve model simulations.

Specifically, we evaluate a set of emulators of a detailed parcel model that nearly explicitly resolves aerosol activation using physically regularized machine learning regression techniques. We directly incorporate physical information from existing fast parameterizations (Twomey, 1959, & Abdul Razzak and Ghan 2000) into the development of boosted regression tree-based and neural network-based model architectures. We find that the machine learning models can faithfully reproduce the parcel model simulated activation at lower computational cost and higher accuracy than many existing parameterizations. Furthermore, the more prior physical information is used in model development, the greater improvement in accuracy. This work enables the implementation of improved hybrid physical-machine learning models of aerosol activation into next generation climate models and demonstrates the value of physical constraints in machine learning model development.