Master Thesis (w/m/d): Thunderstorm Forecasting Using Diffusion Models on Satellite Data

Thunderstorms are associated with several extreme-weather events (e.g. lightning, heavy rain, hail). As a consequence of climate change, the severity of these events is expected to increase in the future. Spatiotemporally highly resolved short-term predictions of thunderstorms are an essential tool to limit significant social and economic damage.

Short-term forecasting has traditionally been a data-driven field, so naturally, machine learning (ML) models have emerged as promising alternatives to traditional extrapolation-based methods. Current models are based on convolutional neural networks (CNNs), which produce deterministic forecasts. However, weather systems are inherently chaotic, implying the need for probabilistic models to quantify uncertainty. On the other hand, diffusion models (DMs) have become state of the art in generative tasks, particularly in image generation. DMs are designed to produce diverse samples, making them well-suited for tasks where uncertainty needs to be explicitly modeled.

The primary objective of this master thesis is to improve upon current CNN-based thunderstorm forecasting by training DMs on satellite data. Specifically, this project aims to investigate the efficacy of DMs in short-term thunderstorm forecasting and compare their performance to CNNs in terms of predictive accuracy and uncertainty quantification.