Improving Prognostic Moist Turbulence Parameterization with Machine Learning and Software Design

Abstract

The primary result of this work is that concepts from software design and machine learning may be used to improve moist turbulence parameterization in weather and climate models. We have seen relatively slow improvement of moist turbulence parameterization in past decades, and explore a radically different approach to parameterization involving machine learning. The core of the approach is to rely on a trusted source of training data, such as high-resolution models or reanalysis, to be used to train a machine learning algorithm to perform the closures normally defined by conventional parameterization. The Python packages \texttt{sympl} (System for Modelling Planets) and \texttt{climt} (Climate Modeling and Diagnostics Toolkit) are introduced. These packages are an attempt to rethink climate modelling frameworks from the ground up. The result defines expressive data structures that enforce software design best practices. It allows scientists to easily and reliably combine model components to represent the climate system at a desired level of complexity and enables users to fully understand what the model is doing. Random forest and polynomial regression are used as an alternate closure assumption in a higher-order turbulence closure scheme trained for use over the summertime Northeast Pacific stratocumulus to trade cumulus transition region. While the machine learning closures better match high-resolution model data over withheld validation samples compared to a state-of-the-art higher-order turbulence closure scheme, the resulting model is unstable when used prognostically. Within a first-order closure framework, an artificial neural network is trained to reproduce thermodynamic tendencies and boundary layer properties from ERA5 HIRES reanalysis data over the summertime Northeast Pacific stratocumulus to trade cumulus transition region. The network is trained prognostically using 7-day forecasts rather than using diagnosed instantaneous tendencies alone. The resulting model, Machine Assisted Reanalysis Boundary Layer Emulation (MARBLE), skillfully reproduces the boundary layer structure and cloud properties of the reanalysis data in 7-day single-column prognostic simulations over withheld testing periods. Radiative heating profiles are well-simulated, and the mean climatology and variability of the stratocumulus to cumulus transition are accurately reproduced. MARBLE more closely tracks the reanalysis than does a comparable configuration of the underlying forecast model. Similar results are obtained over the Southern Great Plains.