Learning Mechanisms from Observations

Gaining scientific knowledge primarily involves observing the world and interpreting those observations. In past centuries, scientists relied on their knowledge and modelling skills to derive insights. With the advent of modern computing and algorithmic advances, however, purely data-driven methods have become possible. However, these methods tend to neglect current scientific knowledge.

Our goal is to contribute to these Mechanism Learning (MEL) methods by enabling them to incorporate prior knowledge, e.g., by restricting the functional search space of possible mechanisms while learning to fit the given data.  In contrast to manual approaches, these methods are expected to suffer less from human bias in model selection, while still allowing optimisation to be based on known expert knowledge.

This project investigates and improves MEL methods, with a focus on dynamical systems widely used in groundwater research. Our goals include determining the functional search space, incorporating existing problem-specific knowledge into the learning process, framing the search as a (Bayesian) model selection problem, modelling and representing uncertainty within the learned mechanism, and finding a balance between data fit and simplicity to guide the learner towards efficient solutions.