Speaker
Description
Large synchrotron datasets present a growing challenge: data volumes increasingly outpace human capacity for systematic exploration and discovery. We present an agentic AI framework that couples LangGraph-based multi-agent orchestration with Tiled/bluesky for exploring and searching data. Specialized agents autonomously browse Tiled node hierarchies, retrieve datasets and metadata, apply analysis routines, and synthesize findings across large collections of experimental runs — enabling natural-language-driven data exploration without requiring users to write bespoke analysis scripts for each query.
Beyond dataset navigation for light source, we deploy deep research agents to build and enrich scientific knowledge graphs. These agents traverse literature, experimental metadata, and prior analysis results to construct structured representations of materials knowledge, linking experimental observables to processing conditions and published findings. The combined system enables queries that span both raw data and accumulated scientific context.
We demonstrate the approach on SAXS/GISAXS datasets at the Advanced Light Source, where agents identify structural trends, flag anomalous scans, and surface candidate datasets contextualized against the broader materials literature.
This work raises practical questions around trust, reproducibility, and human oversight that the facility software community will need to address collectively.
