Identification of thunderstorm occurrence in convection-permitting ensemble forecasts using deep neural networks

Vahid Yousefnia, Kianusch (2025): Identification of thunderstorm occurrence in convection-permitting ensemble forecasts using deep neural networks. Dissertation, LMU München: Fakultät für Physik

Abstract

Thunderstorms have potentially hazardous impacts on society and the economy due to accompanying phenomena, such as lightning, strong winds, and intense precipitation, creating a demand for accurate and timely thunderstorm forecasts. Thunderstorm forecasts several hours in advance are based on simulations of the future atmosphere via numerical weather prediction (NWP). However, as none of the NWP state variables, such as temperature, pressure, or specific humidity, directly indicates thunderstorm occurrence, surrogate variables like convective available potential energy or synthetic radar reflectivity are used as proxies instead. Surrogate variables of thunderstorm occurrence are typically derived from NWP state variables through the consideration of physical principles and empirical knowledge. In this thesis, however, we present a machine learning (ML) model based on deep learning which bypasses the use of such surrogate variables; instead, the model directly processes the vertical variation of the NWP state variables with height to infer the corresponding probability of thunderstorm occurrence. In addition, this thesis makes use of a convection-permitting ensemble NWP model, i.e., an NWP model which (1) allows for resolving atmospheric convection without parameterizations, and (2) generates multiple possible forecasts consistent with forecast uncertainty. While these two properties have individually shown promise for improving thunderstorm forecasts, their combined potential for this task has so far been less explored. Specifically, we train our model on forecasts of ICON-D2-EPS, a limited-area model for Central Europe run operationally by the German Meteorological Service (DWD), with observations from the lightning detection network LINET serving as the ground truth. With regard to model architecture, we employ considerations based on physics and symmetries to keep model size and inference times computationally efficient. For instance, a sparse layer encourages interactions at similar height levels, whereas a shuffling mechanism forces the model to learn pressure coordinates instead of non-physical patterns tied to the vertical NWP grid. Evaluating our model for lead times up to 11 hours, we find that it outperforms a baseline model relying on traditional thunderstorm surrogate variables, which shows the capability of deep learning methods to discover—on their own—skillful representations of thunderstorm occurrence in NWP data. A linear sensitivity analysis (saliency map) suggests that these patterns found in the data are to a considerable extent physically interpretable: our model has learned the climatological propagation direction of thunderstorms in the study region and relies on fine-grained structures, such as ice-particle content near the tropopause and cloud cover, as well as mesoscale structures related to atmospheric instability and moisture. As additional results, we quantitatively explain skill gains resulting from our use of ensemble data. Finally, we demonstrate how neural network models like ours help keeping thunderstorm occurrence predictable for longer lead times compared to models which do not rely on ML. This thesis primarily contributes to improving the skill of thunderstorm forecasts by combining high-resolution NWP and ensemble systems with deep learning. On the other hand, many concepts and methods derived here apply to general binary classification problems, especially when high class imbalance is involved. More generally, our results exemplify the usefulness of incorporating physical considerations and symmetry principles into ML architectures to achieve lightweight models.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: Deep convection, Numerical weather prediction, Machine learning, Interpretability, Binary classification
• Themengebiete: 500 Naturwissenschaften und Mathematik ; 500 Naturwissenschaften und Mathematik > 530 Physik
• Fakultäten: Fakultät für Physik
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 21. Juli 2025
• 1. Bericht­erstatter:in: Craig, George C.