Unsupervised learning of microphysical process rates using generative machine learning models

Cloud microphysics schemes employed in atmospheric and climate models have traditionally used moments of the droplet size distribution to parameterize process rates such as collision-coalescence. However, it is not clear that prognostic moments form an optimal basis set for representing microphysical processes across the range of conditions experienced in the atmosphere. In addition, microphysical parameterizations employing prognostic moments require developing closure schemes, which are known to suffer from both parametric and structural uncertainty in their representations of inherently higher dimensional cloud processes. These uncertainties limit model fidelity and lead to forecasting errors in models.

Recently, machine learning methods applied to numerous physical systems have demonstrated skill in discovering governing equations directly from high-dimensional observational data. Here we investigate how generative machine learning methods such as variational auto-encoders can be used to learn optimal predictors for microphysical process rates in an unsupervised manner, by simultaneously learning lower dimensional representations of droplet size distributions and predicting their dynamic evolution. We discuss how these findings could be used to either guide moment-based scheme development, or replace them with alternative formulations.