Reduced-order modeling for linearized representations of microphysical process rates
Submitter
Lamb, Kara Diane — Columbia University
Area of research
General Circulation and Single Column Models/Parameterizations
Journal Reference
Science
Cloud processes present a major challenge to our ability to model future climate. Here we explore how we can use new, data-driven approaches to determine optimal representations for cloud processes in climate models.
Impact
In this study, we explore a new, data-driven approach to developing simplified (bulk) representations of cloud microphysics for climate models, where a model “learns” how to represent the state of a cloud using a small number of variables. We use this approach to discover the minimum number of variables needed to accurately represent microphysical processes, with a focus here on collision-coalescence.
Summary
Clouds are made from many ice particles and liquid droplets, with a variety of sizes, shapes, and chemical compositions. Traditional approaches for representing small-scale (microphysical) cloud processes start by defining some set of variables, which is expected to contain key information about cloud properties. These variables are “moments” corresponding to important statistics of the population of cloud particles, such as total mass of cloud water, or total number of particles. One can use physical arguments to develop a set of equations that describe how these bulk properties evolve in time. Traditional moment-based models require much less computational power than more sophisticated microphysical models, but have limited accuracy. In this study, we explore an alternative approach, where a model “learns” how to represent the state of a cloud using a small number of variables. By investigating the geometry of this compressed representation, we find that we can represent the merging of cloud droplets (coalescence) using only three variables, so this approach may lead to more accurate models without requiring excessive computational resources. We can use this approach to design new models with favorable features, such as representing droplet coalescence using linear functions.
Reduced-order modeling applied to a higher-fidelity (bin or superdroplet) microphysical model can be used to learn a lower-dimensional, latent representation of the droplet size distribution. Process rates such as collision-coalescence can be learned as linear operators in this latent space representation, in order to predict the microphysical evolution of the cloud at the next time step. Image courtesy of KD Lamb.