TPPC weekly seminar (in-person): Sven Krippendorf - Agentic AI for Theoretical Physics: Targeted Discovery and Completeness in the Type IIB Flux Landscape

Wednesday 25 Feb 2026, 12:30 → 14:00 Europe/London

Large Language Models are increasingly useful as agentic systems: they can plan, call tools, run and sanity-check code, critique drafts and calculations, and iteratively improve research outputs under human supervision. I’ll sketch a practical workflow for theoretical physics—literature digestion, project planning, symbolic/numerical pipelines, verification hooks, and writing—and how automation can speed up work while keeping trust and reproducibility in view.

As a concrete case study, I focus on the Type IIB flux landscape, which poses both an inverse problem (find vacua with desired properties) and a completeness problem (map out what exists in a controlled region of moduli space). On the inverse side, I present conditional generative models (e.g. CVAEs) that produce flux vacua with targeted outputs such as specified superpotential values and strict tadpole constraints, reaching regimes that are hard to access with MCMC-style searches. On the completeness side, I describe a tailored enumeration method applied to a two-modulus Calabi–Yau orientifold, revealing detailed vacuum distributions, hierarchical mass scales, and rare small-W_0 vacua.

Overall, the message is that agent-based automation + generative inverse design + systematic enumeration can make the landscape more navigable—and provides a template for AI-assisted discovery in theoretical physics.