Physical mechanisms of deep convective system maximum area in a hierarchy of datasets

Deep convective systems (DCSs) play a crucial role in the tropical hydrological cycle and radiative budget [1,2]. In particular, the largest and longest-lived of those cloud systems contribute to a high fraction of the extreme precipitation in the Tropics [3] Therefore understanding what drives these types of systems is crucial.
To that end, Abramian et al. 2025 [4]  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 [5]. 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 investigated whether this method would work in observations too, using DCS tracks identified by TOOCAN in satellite data [6], combined with ERA5 data for the physical variables. For both observations and DYAMOND we used both a Lasso linear regression and two different deep neural networks.
Furthermore we aimed at understanding which physical variables constrain the most the maximum area of the systems, and to that end we used an explainable AI method, the integrated gradients ([7]) to assess which physical variables contributed the most to the model prediction.
Firstly, we managed to achieve good predictivity scores for both the non-linear models and both the datasets and we obtained quite robust results in terms of feature importance, with the pre-storm environmental CAPE and deep shear playing a pivotal positive role to achieve a large maximum area, while the presence of neighboring systems was one of the main negative contributors.
Finally, we tested the ML results by looking at composites of the most important variables in the observational dataset: for example pre-storm CAPE composite showed significantly higher than average values for the largest systems. 

[1] Nesbitt, S. W., R. Cifelli, and S. A. Rutledge, 2006: Storm Morphology and Rainfall Characteristics of TRMM Precipitation Features. 
[2] Bony, S., Semie, A., Kramer, R. J., Soden, B., Tompkins, A. M., & Emanuel, K. A. (2020). Observed modulation of the tropical radiation budget by deep convective organization and lower-tropospheric stability. 
[3] Remy Roca and Thomas Fiolleau. Extreme precipitation in the tropics is closely associated with
long-lived convective systems.
[4] S. Abramian, C. Muller, C. Risi, et al. How key features of early development shape deep
convective systems.
[5] Thomas Fiolleau and Remy Roca. An algorithm for the detection and tracking of tropical
mesoscale convective systems using infrared images from geostationary satellite. 
[6] T. Fiolleau and R. Roca. A database of deep convective systems derived from the intercalibrated
meteorological geostationary satellite fleet and the toocan algorithm (2012–2020). 
[7] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks.