15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

AI-Driven Auto-Formalization in Lean4 of the Physics of the Electron-Ion Collider

Not scheduled
15m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for Data Analysis and Theory AI/ML for Data Analysis and Theory

Speaker

Dr Wouter Deconinck (University of Manitoba)

Description

Extracting physics at the Electron-Ion Collider (EIC) means connecting measured observables to theory through chains of definitions, approximations, and fitting or unfolding procedures that are rarely stated precisely enough to be checked. We present EpsilonEridani, an open project that formalizes physics in Lean4 on top of Mathlib, Physlib and TauCeti, with the constraint that all code is written and reviewed by AI agents (currently Claude and Codex). Taking the approach of TauCeti, humans set direction through roadmaps and review rubrics held in separate repositories; agents register their intentions to avoid duplicated effort, open pull requests, and are reviewed by AI reviewers against rubrics covering scope, correctness, reuse, and design. We focus on formalizing the theoretical concepts underlying the full EIC physics scope presented in the EIC Yellow Report. The aim is a coherent library in which theory can be expanded by building new results on proven definitions and theorems rather than restating them. We then outline how to connect this theory to extraction techniques, by stating in Lean4 the assumptions under which kinematic reconstruction, unfolding, and fits recover structure functions and parton distributions from measured quantities. Supporting infrastructure links these statements to the software that generates and analyzes events: the HepMC3 event record, random-number engines for event generation, and bridges between mathematical real numbers and floating-point representations. We describe the agent workflow, how the Lean4 type checker acts as a hard verifier of agent output, and qualitative lessons on failure modes. We discuss how the human role shifts from writing proofs to auditing that formal statements match the intended physics, and how machine-checked theory can guide the extension of EIC theory and the validation of extraction methods.

Author

Dr Wouter Deconinck (University of Manitoba)

Presentation materials

There are no materials yet.