Combining Analytical, Numerical, and AI Models to Improve Cloud and Convection Representation in Climate Simulations

Abstract

Accurate representation of small-scale processes, such as convection and cloud formation, remains one of the greatest challenges in climate modeling, even in kilometer-scale storm-resolving simulations. These processes are essential for determining large-scale atmospheric behavior, but computational constraints prevent their full representation in global climate models. For instance, deep convection may not exhibit convergent behavior with increasing resolution, and significant uncertainties persist in ice microphysics parameterization. This thesis explores these challenges in two parts: the first (Chapters 2 and 3) examines how small-scale convection and cloud processes influence large-scale atmospheric states in idealized radiative-convective equilibrium simulations, while the second (Chapter 4) investigates machine learning (ML) as a tool for efficiently emulating these processes in climate models. Chapter 2 addresses a fundamental question: what controls the vertical thermal structure of an equilibrium atmosphere? To answer this, I developed a refined zero-buoyancy plume model that analytically solves equilibrium atmospheric profiles given boundary conditions. The model highlights how plume-environment mixing influences vertical temperature profiles, upper-tropospheric convective mass flux, and cloud fraction. These findings align with convection-permitting simulations, which reveal that higher horizontal resolution---acting as a proxy for enhanced plume-environment mixing---leads to increased cloud fraction and mass flux in the upper troposphere. Chapter 3 explores the impact of microphysics scheme uncertainties on equilibrium atmospheric states, particularly focusing on deep convective overshoots into the tropical tropopause layer (TTL). We find that different microphysics schemes produce distinct heat balance regimes in the TTL. Two schemes lead to a “hard-landing” scenario, where frequent, strong convective overshoots induce significant cooling (~0.2 K day−1), while a third scheme results in a “soft-landing” scenario, with weaker overshoots and minimal cooling (~0.03 K day−1). This difference arises from variations in upper-tropospheric stratification driven by atmospheric cloud radiative effects (ACRE). The scheme producing the soft-landing scenario generates stronger ACRE, leading to a ~3K warmer, more stable upper-tropospheric layer that buffers convective updrafts. Chapter 4 demonstrates how ML can emulate these small-scale processes efficiently by learning directly from high-resolution simulations. Using data from superparameterized climate simulations, we train ML models to replace the embedded cloud-resolving models. While previous studies show that such hybrid ML-physics simulations can reproduce key climate statistics, they often suffer from online instability, particularly in setups with real geography and explicit cloud condensate coupling. By integrating an expressive U-Net architecture with cloud microphysics constraints, we achieve stable and skillful multi-year hybrid climate simulations with realistic cloud climatology and explicit cloud condensate coupling.