Ben Riegler

In order to make Machine Learning methods useful for scientific discovery, they should be tailored to the problem setting of scientific research. This usually includes an underlying physical system, scarce and expensive data, availability of domain knowledge and the need for uncertainty quantification. This PhD project aims to develop adaptive modeling techniques leveraging meta-learning algorithms to accelerate scientific research. Meta-learning enables learning at a higher level, extracting patterns from related tasks to improve performance on new, unseen tasks. This allows such models to learn key properties of an underlying physical system, making Meta-Learning attractive in scientific settings. By incorporating techniques from Bayesian deep learning, we can infuse uncertainty estimation and online learning capabilities into Meta-Learning frameworks. Uncertainty quantification is vital where mistakes are costly and can aid scientists in their decision-making process. This interaction of scientists with Machine-Learning models is highly desirable and should be facilitated by model design. Additionally, a Bayesian approach allows for the incorporation of domain specific knowledge in form of prior information. This project will work on methods to translate prior knowledge into model properties, model architectures as well as simulated training data. In particular, methods refining abundant physics-informed simulator data with expensive real data will be explored. Settings to be investigated include but are not limited to complex physical systems such as weather and climate, PDEs as well as engineering and cosmology applications.