SALAMA: A Machine Learning approach for lightning forecasting through post-processing of simulation data

Thunderstorms constitute a major hazard to society and economy. Especially in light of the expected increase of extreme weather events due to climate change, reliable thunderstorm warnings become ever more important. However, as lightning is not directly computed in numerical weather prediction (NWP) simulations, the appearance of thunderstorms in forecast output remains elusive. In this work, we introduce SALAMA, a tool to identify signatures of lightning activity in NWP simulations using a feedforward artificial neural network (ANN). It infers in a reliably calibrated manner the probability of lightning occurrence at some point in space and time, given only a set of local input parameters that are extracted from NWP simulations and related to thunderstorm development. We train the neural network with ensemble forecasts from ICON-D2-EPS during the summer period of 2021. The skill of SALAMA is measured through established scores from meteorology and machine learning. We study in detail how the forecast skill depends on the lead time of the forecast as well as the spatial scale of the forecast objects and put particular emphasis on a careful estimation of model uncertainty. Even with a relatively simple ANN architecture and local input parameters, we find a forecast skill superior to traditional approaches in the literature. SALAMA is ready for operational use.