Our research asks two complementary questions: what can physics teach us about learning systems, and what can modern machine learning reveal about fundamental physics? We combine mathematical and theoretical analysis with computational experiments and automated search.

Physics of Learning

We treat learning systems as physical and mathematical objects. We investigate how data selection, optimisers and architectures shape learning dynamics, internal representations and scaling behaviour. Our approach combines ideas from statistical and theoretical physics with controlled numerical experiments: the aim is not only to observe scaling laws, but to understand the mechanisms behind them and turn that understanding into better learning systems.

Current work includes collective descriptions of neural-network training and the mechanisms behind neural scaling.

Selected work:

Machine learning for fundamental physics

We develop machine-learning and agentic systems for problems in quantum field theory, particle physics and string theory. This is a natural continuation of our earlier work on discovering symmetries, dualities and integrable structure. Detecting Symmetries with Neural Networks established the broad idea of learning mathematical structure from data; Integrability Ex Machina showed how an automated search could recover precise theoretical structures. Today’s agentic systems extend this direction by exploring executable representations, algorithms and physical models at scale.

With Yi Gu, we treat scattering-amplitude calculations as executable programs and use self-evolving program search to find more efficient and revealing representations. The search moves between ideas from amplitude theory—recursion, symmetry, basis reduction and shared computation—and combines them into new algorithms. Read the paper or explore the code and search trajectories.

String vacua at scale

JAXVacua is infrastructure for constructing and exploring large ensembles of string vacua efficiently. Developed over several years with Andreas Schachner and collaborators, it uses automatic differentiation, compilation and parallelisation in JAX to make previously inaccessible regions of the Type IIB flux landscape computationally tractable. Read the foundational paper.

With Zhimei Liu, we use conditional generative models to solve the inverse problem: rather than sampling vacua first and filtering afterwards, we generate flux configurations targeted at desired physical properties. Read the paper.

Numerical Calabi–Yau geometry

Ricci-flat Calabi–Yau metrics solve nonlinear geometric differential equations but are rarely known explicitly. Our numerical work uses machine learning to approximate these metrics across families of geometries, including their dependence on complex-structure moduli. This makes geometric quantities needed for string compactifications accessible to computation.

Verified and reusable physics

We are exploring formal verification through PhysLib, the community Lean library for physics. Formalisation does not by itself establish that a physical model is meaningful, but it makes definitions, assumptions and logical steps explicit and machine-checkable. In recent work with Joseph Tooby-Smith, we developed reusable PhysLib infrastructure for a certified classification problem in SU(5) model building. Read the preprint.

Further applications in cosmology and phenomenology

Alongside these two core research directions, we apply machine-learning and data-intensive methods to selected problems in cosmology and particle-physics phenomenology. With Nina Elmer, we study how errors in neural-network outputs can be quantified and when those outputs can be trusted, particularly in particle-physics applications.

Continuing collaborations include cosmological inference with Kai Lehman and eROSITA cluster science with Silas Zelmer. Earlier work in this strand has also addressed gravitational-wave signals and searches for new particles such as axion-like particles.

Selected work:

For talks and discussions across these areas, see the DAMTP Data Intensive Science Seminar and the former Physics Meets ML initiative.

Interested in joining us or seeking fellowship hosting? See Opportunities.