LPC Physics Forum

Thursday 5 Feb 2026, 06:00 → 14:00 US/Central

Sunrise (WH11NW) (FNAL/Zoom)

https://cern.zoom.us/j/96898909340
contact lpc-pf@fnal.gov for the passcode

1 Archi and AI Agents in Scientific Computing: Mechanics, Applications, and Open Challenges

AI has progressed from models that generate text to agents that retrieve information, execute code, and verify their outputs. Just as physicists adopted neural network architectures for particle identification by training them on our data, we can now apply modern AI agent frameworks to automate tasks in scientific computing. This talk explains the mechanics of these systems, including retrieval, tool use, and orchestration, and argues that their immediate value here lies in low-risk tasks such as documentation search, code generation, and debugging. I will discuss Archi, a retrieval agent developed at MIT in collaboration with CMS, to illustrate the engineering behind these platforms. Finally, we will outline open problems that demand domain expertise to enable reliable automation, including detecting stale documentation, linking information across fragmented sources (logs, wikis, code), and establishing rigorous verification for AI-generated results.

Speaker: Luca Lavezzo (MIT)

Archi_fermilab_feb_26.pdf

video1412951645.mp4