Dissertation Defense: Savannah Ferretti

Title: Make it Rain—Understanding Thermodynamic Controls on South Asian Monsoon Rainfall

Abstract:  Seasonal rainfall from the South Asian monsoon strongly influences water resources, agricultural productivity, and economic stability across the Indian subcontinent, affecting the lives of more than two billion people. Accurately representing this rainfall in climate models, however, remains a longstanding challenge. Persistent biases in rainfall intensity, timing, and spatial organization are especially pronounced at regional scales, suggesting that the thermodynamic processes governing convective precipitation are still not fully understood. This dissertation investigates those processes using a hierarchy of physics-based and interpretable data-driven approaches that progressively move from evaluation of existing thermodynamic diagnostics toward discovery of new relationships directly from data.

In the first chapter, we examine how well existing thermodynamic diagnostics explain South Asian monsoon rainfall variability at subregional scales. A process-oriented diagnostic based on the buoyancy of an entraining plume is evaluated across five subregions of the South Asian monsoon domain. While a strong precipitation--buoyancy relationship is observed across all five subregions, and the diagnostic generally captures the direction of month-to-month rainfall changes, its ability to explain those changes consistently across subregions is much more limited. Clear physical explanations for the rainfall changes emerge in only two subregions, where variability in lower free-tropospheric moisture, rather than undilute buoyancy, appears to dominate the thermodynamic control on precipitation. In the remaining subregions, the diagnostic either struggles to accurately predict rainfall changes or does not provide a clear explanation for the observed variability. These results suggest that prescribed layer-averaged thermodynamic diagnostics can capture some important aspects of South Asian monsoon rainfall variability, but are often too restrictive to fully represent the thermodynamic structure relevant for precipitation across the broader domain.

The limitations of the physics-based diagnostic motivate the development of a more flexible approach for identifying thermodynamic controls on precipitation. To address this need, the second chapter introduces integration kernel learning, a machine-learning methodology for representing nonlocal structure in geophysical data. The method separates how nonlocal information is aggregated across horizontal space, height, and/or time from how local predictions are made, allowing physically interpretable weighting patterns to be learned directly from data. The framework is demonstrated using South Asian monsoon rainfall. As a first step, a series of neural networks is used to determine whether horizontal, vertical, or temporal structure contributes most strongly to precipitation prediction. This analysis shows that vertical thermodynamic structure contains substantially more predictive information than horizontal or temporal structure at the scales considered here, motivating the use of learned vertical kernels. These kernels identify which pressure levels are most important for different thermodynamic predictors, providing direct physical insight into how atmospheric structure regulates precipitation variability. The learned weighting patterns reveal distinct roles for boundary-layer and lower free-tropospheric humidity, while also producing compact thermodynamic features that provide a basis for equation discovery.

Building on the compact thermodynamic features produced by the learned kernels, the third chapter uses symbolic regression to discover analytic equations relating atmospheric structure to precipitation. A hierarchy of learned equations reveals how monsoon rainfall is regulated by moisture, instability, and surface coupling. At the simplest level, precipitation is strongly controlled by lower free-tropospheric moisture, with rainfall remaining suppressed until humidity exceeds a critical threshold and then increasing rapidly. A more complex equation reveals that this moisture control competes with an instability pathway associated with entrainment-related suppression, producing the strongest rainfall when moisture becomes the dominant thermodynamic control. Surface properties further modify rainfall through a correction term that reduces precipitation over land and cooler ocean regions, where thermodynamic structure alone would otherwise favor stronger precipitation. Together, these results demonstrate that regional monsoon rainfall variability can be represented through a compact set of interpretable equations that capture interacting thermodynamic controls across precipitation regimes.