Speaker
Description
Bridging the gap between a visiting user's X-ray Photon Correlation Spectroscopy (XPCS) experiment concept and its technical feasibility at Argonne National Laboratory's Advanced Photon Source (APS) often requires extensive consultation with beamline scientists. An integrated AI platform at APS Beamline 8-ID is introduced, consisting of multiple components. An agentic AI system autonomously explores publication records across global research center databases, selects publications by evaluating paper content for relevance to XPCS, and submits them to a human review queue before ingestion. Approved publications are embedded using SciBERT and stored in a Qdrant vector database, powering a retrieval-augmented generation (RAG) system that enables users to evaluate experiment feasibility through a conversational interface grounded in curated XPCS literature rather than general LLM knowledge. A document ranking system allows beamline scientists to assign priority weights to ingested publications, ensuring responses reflect the most relevant sources for the beamline. Users can refine hypotheses and inquire about experiment feasibility across multiple turns, such as whether 8-ID can maintain sub-angstrom resolution during high-temperature metallic glass aging over extended timescales, and receive scientifically grounded responses with source attribution. By providing users with on-demand access to beamline-specific expertise, the platform enables better experiment preparation, deepens users' understanding of experiment feasibility, and reduces burden on beamline scientists.
