Rafael Ferreira da Silva is a Distinguished Research Scientist and Section Head for Data and AI Systems in the Computer Science and Mathematics division at Oak Ridge National Laboratory, where he leads research in AI, data and agentic workflows, and advanced computing systems. His work spans HPC/AI workflows, provenance capture, and autonomous science. He is the Founder and Executive Director of the Workflows Community Initiative and a co-Founder of the Autonomous Science Network and is a Senior Member of IEEE and ACM.
Presentation Title:
From Lab Bench to Supercomputer: Building Agentic Scientific Workflows, and Learning to Trust Them
Presentation Abstract:
We present two agentic systems built at Oak Ridge National Laboratory and use them to argue for what the field still needs. The first is an end-to-end agentic framework for high-throughput plant phenotyping in which a cloud-hosted Co-Scientist Agent and a Compute Agent on the Frontier exascale supercomputer run in separate security domains, bridged by a token-authenticated streaming channel, with end-to-end provenance captured at under 1% execution-time overhead. The second is AgentBuild, which treats agent construction as a workflow stage and compiles a Rietveld refinement agent over GSAS-II from a rubric, curriculum, and knowledge base authored by the domain scientist, so that a base-model swap becomes a re-tune instead of a rebuild. Both systems make the same point, i.e., that bringing agentic AI to leadership-class science is less a modeling problem than an infrastructure one. We then turn to a two-year community roadmap for trustworthy autonomous science, which finds that we can now generate candidate discoveries faster than we can verify them, and we close on the verification, validation, and provenance milestones that gap makes urgent.