Causal Drivers of Continental Mesoscale Convective System Biases in Kilometer-Scale Simulations

Mesoscale convective systems (MCSs) play a central role in regulating the global energy and water cycles through their extensive cloud coverage and the associated redistribution of latent heat. Global convection-permitting models at kilometer scale have made substantial progress in representing several MCS characteristics, including bulk precipitation statistics and anvil extent. However, persistent deficiencies remain in simulating MCS populations in key tropical hotspots. Using four years of global ICON simulations at 2.5 km horizontal resolution, we identify systematic regional biases in MCS occurrence, with overestimated MCS initiation over the Amazon and Congo forest regions. In addition, simulated MCSs generally have less spatial extent than those identified in satellite-based observations.

In this talk, we investigate the physical drivers underlying these biases using causal machine learning approaches to identify environmental factors that control MCS initiation, size, and intensity. Preliminary observational analyses indicate that, over the Amazon basin, mid-level wind shear and column-integrated water vapor exert strong controls on MCS size and total precipitation. We compare these observed causal relationships with those inferred from the ICON simulations to assess whether the same controlling factors operate in the model. Discrepancies in the identified drivers provide insight into the mechanisms responsible for model biases, their impacts on simulated MCS structure and rainfall characteristics, and potential pathways on how to improve the modeling system.