Parameterizing Lagrangian cloud microphysics using machine learning

Abstract

Parameterizations in weather and climate models are essential for representing sub-grid scale processes that cannot be explicitly resolved due to computational constraints. However, they introduce significant uncertainties in the prediction of key atmospheric variables such as precipitation, cloud cover, and the estimation of the radiative balance. As kilometer-scale models eliminate the need for many parameterization schemes, micro-scale phenomena such as cloud microphysical processes emerge as important sources of model error. Cloud microphysical processes are represented via bulk schemes, which, while computationally efficient, rely on simplified assumptions that can lead to substantial errors in forecasts. This thesis presents the development of machine learning (ML) emulators aimed at improving the accuracy of these parameterizations by leveraging Lagrangian models, specifically the superdroplet method, which provides a more physically consistent representation of cloud microphysical processes. I introduce SuperdropNet, an MLbased emulator trained on superdroplet simulations, designed to predict the evolution of bulk moments while bypassing the assumptions of the bulk schemes. I introduce new training techniques and employ methods such as autoregressive training to develop a physicsinformed ML emulator. SuperdropNet demonstrates a stable performance across a wide variety of initial conditions and outperforms previously available ML emulators for droplet collisions. The emulator was also evaluated in an online scenario by coupling it with the ICOsahedral Nonhydrostatic (ICON) model. Coupling ML emulators to an atmospheric model is usually a labourious task that impedes the process of developing an effective emulator. I solved the technically challenging task of coupling a Python-based ML model to the FORTRAN codebase of ICON. I tested multiple coupling mechanisms for flexibility, ease of use, and speed, and found a C-based interface to be the most suitable. Encouragingly, when coupled to a warm bubble test scenario, SuperdropNet demonstrated long-term stability and predicted physically plausible quantities. This research underscores the potential of ML emulators to bridge the gap between the computational efficiency of bulk moment schemes and the physical accuracy of Lagrangian models. Such an approach could be extended to other ill-represented cloud microphysical processes in atmospheric modeling, such as sedimentation and cloudaerosol interactions, ultimately leading to more reliable weather and climate predictions.