Convective Self-Aggregation and the Madden-Julian Oscillation: Theory, Cloud-Resolving Simulations, and Machine Learning

This dissertation studies the organization and predictability of tropical convection through two phenomena: convective self-aggregation and the Madden-Julian Oscillation (MJO). We ask what drives tropical convection to self-organize without external forcing, and what serves as the source of predictability in such organized systems.

Convective self-aggregation is the spontaneous organization of random convection into large moist and dry regions in cloud-resolving simulations with uniform sea surface temperature (SST) and no rotation. To study its mechanisms, we develop a layerwise moist static energy (LMSE) variance framework that quantifies the contribution of each physical process at each altitude, complementing conventional frameworks that integrate MSE over the column and lose the vertical dimension. In a cloud-resolving simulation, the development of self-aggregation is associated with an increase in the LMSE variance, and both the diabatic and adiabatic production of this variance are dominated by the boundary layer (the lowest 2 km). This result agrees with previous available potential energy analyses showing that the boundary layer is key to self-aggregation.

We then apply the LMSE framework across climates. Previous research concluded that radiative feedbacks are essential to self-aggregation. In 2D cloud-resolving simulations at SSTs of 280--320 K, we find that the dominant mechanism changes with warming: radiative production dominates the LMSE variance production below 300 K, whereas above 300 K the adiabatic contribution switches from negative to positive and becomes dominant. When we disable radiative feedbacks by horizontally homogenizing the radiative heating rates, self-aggregation disappears at 280--290 K but still occurs at 310--320 K, consistent with the LMSE diagnosis and with 3D simulations. Self-aggregation is thus radiatively driven in cold climates and adiabatically driven in warm climates.

Finally, we study the MJO, a planetary-scale, intraseasonal envelope of tropical clouds and rainfall. MJO theories disagree on which spatial scales are essential, and we use deep learning to test this question. We develop a deep convolutional neural network (DCNN) that forecasts the MJO indices RMM and ROMI up to 25 and 33 days ahead, respectively, comparable to most operational subseasonal-to-seasonal models. Spectral analysis of the learned features shows that large-scale patterns dominate the forecast signals, and models using only large-scale input perform comparably to models using unfiltered input, supporting the large-scale view of the MJO. However, models using only small-scale input still produce skillful forecasts 12--22 days ahead. The DCNN achieves this by reconstructing the large-scale envelope of the small-scale activity, consistent with the multi-scale view of the MJO.