Advancing Climate Risk Modeling of Severe Convective Storms Through Deep Learning

Severe convective storms (SCS), including hail, tornadoes, straight-line winds, lightning, and heavy precipitation, represent a significant and evolving source of climate risk. SCS perils pose significant challenges for sectors such as insurance and finance, where accurate risk quantification is essential for underwriting, portfolio management, and resilience planning. Assessing the risk of these perils requires robust frameworks capable of capturing non-linear dynamics, spatial heterogeneity, and compounding effects. However, current modeling approaches often exhibit limited skills when restricted to narrow hazard scopes (e.g., hail-only) or coarse annual scales, limiting their ability to resolve seasonal and intra-seasonal variability. This research introduces a risk assessment framework that leverages deep learning architectures, specifically, a U-Net model augmented with attention mechanisms, to predict the frequency and severity of SCS perils. The model is trained on high-dimensional interpretable meteorological predictors calculated in-house from reanalysis and climate model data, and georeferenced hazard observations from diverse sources. Attention layers within the U-Net architecture enhance feature localization and interpretability, addressing challenges in modeling rare and spatially complex events critical for risk assessment. The framework produces peril-specific daily probabilities and climatological maps, allowing for modeling cross-peril correlation as well as multi-day outbreaks. By integrating physical understanding with data-driven modeling, this approach offers a scalable and interpretable solution for climate risk assessment to support applications such as underwriting, accumulation management, and risk mitigation.