From Prediction to Causation: Understanding theDrivers of Maximum Deep Convective Systems Area

Deep convective systems (DCSs) play a crucial role in the tropical hydrological cycle and radiative budget (Stephens et al., 2023; Roca et al., 2014). In particular, the largest and longest-lived of those cloud systems contribute to a high fraction of the extreme precipitation in the Tropics (Roca and Fiolleau, 2020). Therefore understanding what drives these types of systems is crucial. To that end, Abramian (2023) developed a new method to predict the maximum area of DCSs using the DYAMOND-Summer simulation with the cloud-resolving global model SAM, and the TOOCAN algorithm to track cloud systems. The method uses simple machine learning models, trained on information on the early stage of the systems and their surrounding environment, including dynamical and thermodynamical variables, morphological features of the systems and the characteristics of their neighbors.
We improve this method by incorporating an integrated gradients (IG) approach, which provides a more precise quantification of the importance of each input variable directly from the neural network model. Furthermore, we embedded the neural network outputs into a causal discovery framework by identifying the variables that explain the most variance, using the IG method. These key variables were then subjected to a causal discovery analysis, enabling the identification of causal drivers that influence the maximum extent of the systems at various stages of their lifecycle.
This approach improves both interpretability and includes causal inference to avoid non-causal relations. Preliminary results suggest that during the early stages (0.5 hours after the onset of the DCS), the strength of vertical velocity and upper tropospheric saturation explain most of the variance in the system’s maximum area. Interestingly, the presence of neighboring systems also plays a significant role, likely because a smaller number of neighbors allows more moisture and energy to be available for the DCS to grow. In contrast, during the later stages (around 3.5 hours after the DCS onset), when the area reaches its maximum, neighboring systems no longer contribute significantly to the variance. At this stage, thermodynamic factors, particularly
moisture and temperature, emerge as the primary drivers, with the 2-meter temperature playing a particularly important role, suggesting a potential role of cold pools in determining the maximum area of the system.