NSF CI Compass Virtual Workshop Report: AI Meets CI - Intelligent Infrastructure for Major & Midscale Facilities

The NSF CI Compass virtual workshop AI Meets CI: Intelligent Infrastructure for Major & Midscale Facilities (January 12–14, 2026) brought together NSF Major and Midscale Facility staff and a broader community to discuss how AI is currently used, what fails in practice, and what it would take to make AI-enabled capabilities supportable as part of facility cyberinfrastructure. The workshop included facility case studies, tool-focused discussions, and an interactive session focused on adoption and trust. This report synthesizes these discussions into cross-cutting themes, day-by-day highlights, and operating guidance intended to help facility leaders and program managers make well-grounded decisions.

The workshop covered a range of facility workflows, including user and staff support over institutional documentation (Rubin, NEON); real-time data processing pipelines with embedded ML components (Rubin active optics system (AOS) and alert processing, CMS triggers and reconstruction); metadata and provenance generation as part of data acquisition (MagLab, CHESS); QA/QC and anomaly detection (OOI, NEON); data discovery and semantic search across heterogeneous holdings (DesignSafe/NHERI); and code assistance for development and analysis tasks (IceCube, multiple facilities). The workshop also addressed cross-cutting concerns, including security for retrieval-augmented systems (Trusted CI), adoption and human factors (Day 3 interactive session), and the gap between working prototypes and supported services.

Three overarching messages emerged from these discussions.

First, facilities should treat AI adoption as a service-management problem, not a tool selection exercise. The primary challenge is not getting a prototype to work once. The challenge lies in operating a capability over time with clear boundaries, traceability, and a defensible review posture.

Second, the limiting factors are often “below the model.” Machine-actionable metadata, provenance, documentation stewardship, and governance boundaries determine whether AI-enabled workflows are safe and repeatable. Where these foundations are weak, AI tends to amplify confusion: it can produce plausible outputs that are difficult to verify, and it can encourage informal use that outpaces policy and security controls.

Third, human factors are not secondary to AI adoption. Adoption is constrained by incentives, perceived risk, and validation capacity. As tools become more capable and more autonomous (i.e., able to chain multiple steps, retrieve information, and take actions without human intervention at each step), staff and users need better ways to keep track of what the system did, what evidence it used, and what is in scope for automation.

Citation: E. Deelman, C. F. Vardeman II, P. Balaprakash, G. Broderick, D. Brower, D. Butcher, K. Chard, J. Duarte, J. Glatstein, L. Guy, T. Habermann, M. Hasan, M. Holl, D. Howard, B. Hurwitz, J. Ibarra, J. Jackson, K. Kee, K. Kumar, A. Mandal, R. Mishra, A. Murillo, J. Nabrzyski, D. Paine, D. S. Pandya, R. Ramachandran, R. Redmon, B. Riedel, C. Shimizu, E. Sokol, W. Sun, E. Taylor, N. Virdone, S. Volkova, and D. Wampler, NSF CI Compass Virtual Workshop Report: AI Meets CI – Intelligent Infrastructure for Major & Midscale Facilities.
Zenodo, Jul. 15, 2026. doi: 10.5281/zenodo.21383345.