Process-Based Climate Model Development Harnessing Machine Learning: III. The Representation of Cumulus Geometry and Their 3D Radiative Effects

Process-scale development, evaluation, and calibration of physically based parameterizations of clouds and radiation are powerful levers for improving weather and climate models. In a series of papers, we propose a strategy for process-based calibration of climate models that uses machine learning techniques. It relies on systematic comparisons of single-column versions of climate models with explicit simulations of boundary-layer dynamics and clouds (Large-Eddy Simulations [LES]). This paper focuses on the calibration of cloud geometry parameters (vertical overlap, horizontal heterogeneity, and cloud size) that appear in the parameterization of radiation. The solar component of a radiative transfer (RT) scheme that includes a parameterization for 3D radiative effects of clouds (SPARTACUS) is run in offline single-column mode on an ensemble of input cloud profiles synthesized from LES outputs. The space of cloud geometry parameter values is efficiently explored by sampling a large number of parameter sets (configurations) from which radiative metrics are computed using fast surrogate models that emulate the SPARTACUS solver. The sampled configurations are evaluated by comparing these radiative metrics to reference values provided by a 3D RT Monte Carlo model. The best calibrated configurations yield better predictions of TOA and surface fluxes than the one that uses parameter values computed from the 3D cloud fields: The root-mean-square errors averaged over cumulus cloud fields and solar angles are reduced from similar to 10 Wm(-2) with LES-derived parameters to similar to 5 Wm(-2) with adjusted parameters. However, the calibration of cloud geometry fails to reduce the errors on absorption, which remain around 2-4 Wm(-2).

Automated summary

The process-scale development, evaluation, and calibration of physically based clouds and radiation parameterizations using machine learning techniques can improve weather and climate models. Calibrating cloud geometry parameters can lead to better predictions of TOA and surface fluxes, but does not reduce errors on absorption.