SALAMA 1D: Identification of thunderstorm occurrence from convection-permitting forecasts of vertical profiles using deep learning

Thunderstorms pose significant risks to society and the economy due to hazards such as heavy precipitation, hail, and strong winds, necessitating accurate forecasting to mitigate their impacts. Convection-permitting numerical weather prediction (NWP) models can explicitly resolve convective processes, but predicting thunderstorms from their output remains challenging since there is no obvious variable that directly indicates thunderstorm occurrence. Many approaches rely on combining multiple single-level variables, such as convective available potential energy (CAPE), which are derived from state variables like temperature, pressure, and specific humidity, and act as surrogates for thunderstorms. In this study, we present a deep neural network model that bypasses surrogate variables and instead directly processes the vertical profiles of state variables provided by convection-permitting forecasts. Our model, SALAMA 1D, analyzes ten different NWP output fields, such as wind velocity, temperature, and ice particle mixing ratios, across the vertical dimension, to produce the corresponding probability of thunderstorm occurrence. The model’s architecture is motivated by physics-based considerations and symmetry principles, combining sparse and dense layers to produce well-calibrated, pointwise probabilities of thunderstorm occurrence, while remaining lightweight. We trained our model on two summers of forecast data from ICON-D2-EPS, a convection-permitting ensemble weather model operationally run by the German Meteorological Service (DWD), using the lightning detection network LINET as the ground truth for thunderstorm occurrences. Our results demonstrate that, up to lead times of (at least) 11 hours, SALAMA 1D outperforms a comparable machine learning model that relies solely on derived variables. Additionally, a sensitivity analysis using saliency maps indicates that the patterns learnt by our model are to a considerable extent physically interpretable. This work advances NWP-based thunderstorm forecasting by demonstrating the potential of deep learning to extract valuable predictive information from high-dimensional NWP data while preserving model interpretability.