Predicting Eastern Mediterranean lightning: evaluating microphysical and thermodynamic indices using a machine learning approach

Lightning serves as a fundamental indicator of convective intensity and an important mediator of atmospheric dynamics. Accurately modeling the potential for lightning occurrence is essential for understanding storm electrification and improving short-range forecasting. The Lightning Potential Index (LPI) is a physically based diagnostic parameter that quantifies the potential for charge generation within convective clouds. It combines model-resolved updraft velocity and precipitating ice content, thereby directly representing the mechanisms responsible for non-inductive charging (Yair et al., 2010). In contrast, thermodynamic indices such as the K-Index (KI) and Convective Available Potential Energy (CAPE) reflect the environmental instability and likelihood of convection, but lack an explicit representation of microphysical electrification processes (Peppler, 1988). Additionally, accumulated precipitation serves as a proxy for the integrated intensity of the storm systems. In this study, we evaluate the skill of this suite of atmospheric predictors - meaning LPI, KI, CAPE, and precipitation – all computed from WRF ensemble simulations, in reproducing observed lightning activity over the Eastern Mediterranean. Five case studies were selected, representing different synoptic conditions in winter. A comprehensive processing pipeline was developed to co-register model outputs and ground-based lightning detections from the ENTLN network onto a uniform 4 × 4 km grid and 3-hour temporal intervals. Spatially, all parameters were averaged per grid cell. Temporally, precipitation was summed, while other variables (LPI, KI, CAPE) were averaged over each period. All datasets were smoothed with a Gaussian kernel to reduce spatial noise and enable direct comparison across domains. Preliminary analyses indicate that thermodynamic indices and accumulated precipitation exhibit broad spatial footprints, significantly overestimating the areal extent of lightning activity. While LPI also displays a tendency towards broader coverage than observed, it demonstrates the highest degree of spatial localization among the examined parameters. To further quantify predictive skill, we employ a machine learning approach based on Random Forest algorithm. The spatial model matrices are decomposed into discrete single-cell vectors, utilizing the full suite of parameters. These features are used to classify the binary occurrence of lightning (presence/absence), independent of flash multiplicity, establishing a robust data-driven mapping between storm microphysics and lightning probability. ReferencesY, B. Lynn, C. Price, V. Kotroni, K. Lagouvardos, E. Morin, A. Mugnai, and M. d. C. Llasat (2010), Predicting the potential for lightning activity in Mediterranean storms based on the Weather Research and Forecasting (WRF) model dynamic and microphysical fields, J. Geophys. Res., 115, D04205, doi:10.1029/2008JD010868.Peppler, R. A. (1988). A review of static stability indices and related thermodynamic parameters. ISWS Miscellaneous Publication MP-104.‏‏