Begin: Can begin immediately

First Supervisor: Prof. Dr. Heisenberg

Second Supervisor: Natasha Randall MSc.

Level: Master's thesis

Problem:

Flood forecasting involves inputting initial conditions (like rainfall and elevation values) into a model which outputs a flood prediction (such as flood risk or water depth). Traditionally, flood forecasts are produced by numerical models, which precisely simulate the physical processes of a flood event. Recently however, data-driven approaches have begun to be applied to flood prediction, using deep learning models like CNNs, RNNs and GANs to generate flood forecasts. Deep learning models require large amounts of data to train on, but geoscientific datasets are often very sparse and/or low quality.

Solution:

One solution to this problem is to combine theory- and physics-guided approaches with data-driven models, using knowledge of physical processes to essentially “fill in the gap” left by the lack of data. Additionally, deep learning models often make physically implausible predictions, so by integrating physical processes into a model, they can better learn the real underlying casual relationships.

Objective:

The objective of this thesis is to research and implement theory- and physics-guided approaches to flood prediction with deep learning models. Some examples of physics-guided methodologies include embedding physical laws in model architectures, using blocks that represent different stages in the flood process, regularising the loss function based on physical constraints like the conservation of water, or using graph neural networks (GNNs) to represent the dynamics of water flow.

Tasks:

• Identify a suitable dataset(s) and pre-process the data.
• Research state-of-the-art physics-guided approaches to deep learning.
• Develop a deep learning model for flood prediction that implements theory- and physics-guided methodologies.
• Evaluate and analyse the model's performance.
• (Optionally) Explore the explainability of the model and its capabilities to learn causal relationships.

This thesis is suitable for students who are interested in using AI to support environmental sustainability. They should have experience in coding and working with deep learning models, and be willing to learn some basics of hydrology. This thesis will be carried out in English. If you are interested, please get in contact via email.