Emulation of Bin Microphysical Processes with Machine Learning

The warm rain formation process is critical for understanding and simulating both weather and climate. It is recognized that cloud processes and feedbacks present large uncertainties in weather and climate simulations. For example, the simulated coverage and properties of clouds can highly regulate the radiative balance of the atmosphere, and precipitation formation is critical for the hydrologic cycle and precipitation extremes. The process depends on small scale interactions between cloud droplets and precipitation that must be crudely represented in bulk cloud schemes in large scale weather and climate models. This work examines whether machine learning emulation can replace traditional emulation (regression, curve fits) commonly used in these models to more directly trace critical cloud processes to detailed physical calculations. In this study, we aim at replacing the crude representation of coalescence and collision processes of cloud drops in a microphysical parameterization that generates warm (liquid) precipitation. Such processes are usually described in bulk and empirical form in most climate models, such as a fit to more detailed models. Here we replace the bulk scheme of such processes with two computationally expensive detailed treatments of the stochastic collection processes, and explore emulating them using machine learning techniques.

To build a training dataset for the machine learning emulator, a global General Circulation Model (GCM), the Community Atmosphere Model (CAM) simulation was run with two different detailed representations of the collision-coalescence process. One uses a collection kernel on bins from a bin microphysical scheme. The other method is a detailed super-droplet method tracking interactions between each drop size. A diverse set of machine learning models, including deep dense neural networks, and generative adversarial networks, are trained to estimate the resulting mass and bulk number concentration tendencies. A random hyperparameter search provides information to optimally tune each machine learning model, and then each scheme is validated with a temporally independent test set. Evaluations of each machine learning model examine their ability both to minimize approximation error and reproduce the correct distributions of each tendency. Machine learning interpretation techniques applied to each machine learning model and the bin microphysics schemes reveal how closely the machine learning models replicate the sensitivities of each scheme. Finally, we compare the CAM runs using the machine learning emulators with the original runs to determine whether the machine learning models can produce the correct climate within the model over the duration of the run.