Machine learning and string theory

This project falls into the area of string theory and more specifically string model building. Recently, advanced methods of computer science, particularly machine learning techniques, have been introduced into string theory, as a way of understanding and organising the large data sets string theory produces. The aim of this project is to use techniques of machine learning to advance the understanding of string theory and its associated low energy theories.

Supervised machine learning will be used to explore large, pre-existing and new, data sets within string theory. This includes data sets which consist of Calabi-Yau manifolds and their cohomological data as well as vector bundle data and cohomology. A basic question is whether this data can be learned with existing machine learning methods. New methods and structures of neural networks might be required to handle data of this kind. Another aim is to extract new mathematical structures from machine-learning string data, in the way that has recently been demonstrated is feasible for line bundle cohomology data. These new mathematical structures are likely to facilitate bottom-up model building approaches which will be developed.

Reinforcement learning is another area of machine learning with promising applications in particle theory and string theory. A main aim of the project is to understand how reinforcement learning can be used for high energy theory model building and how large landscapes of physics models can be explored in this way. As a proof of concept, reinforcement learning will first be applied to well-established areas of high-energy theory model building, including building of models for fermion masses and inflationary model building. Based on this experience, reinforcement learning will be used in string model building. At first this will be carried out for landscapes of bundles or fluxes on a fixed manifold. Later, landscapes including classes of manifolds will be tackled.

The results from string theory reinforcement learning and the bottom-up methods identified through supervised learning will eventually be combined for a more targeted approach to model building, which aims at identifying, more readily than current possible, physically promising parts of the string landscape.