Scientific discovery increasingly depends on connecting models, data, documents, and domain expertise into coherent workflows. This talk explores how agentic AI systems can be enhanced through knowledge graphs and operational experience to create workflows that are not only automated, but continuously improving. Using examples from nuclear materials research and the AERIS framework, we demonstrate how knowledge can be accumulated, curated, and reused across tasks, transforming isolated tools into collaborative scientific ecosystems. We discuss the challenges of knowledge curation, provenance, and governance, and present a vision for agentic systems that learn from both structured knowledge and prior operational outcomes while maintaining transparency and human oversight.

Talk part of the TPC26 BoF: “Agentic AI: Transforming HPC Research”
The session explores how LLM-driven agents can support code translation, modernization, performance modeling, tuning, and autonomous HPC workflow orchestration.

Moderator: Mohammad Alaul Haque Monil
Speakers:

• Geoffrey Fox, University of Virginia
• Ana Gainaru, Oak Ridge National Laboratory
• Murat Keçeli, Argonne National Laboratory
• Meifeng Lin, Brookhaven National Laboratory
• Keita Teranishi, Oak Ridge National Laboratory