Authors: Olivera Kotevska (Oak Ridge National Laboratory (ORNL)), Antigoni Georgiadou (Oak Ridge National Laboratory (ORNL)), Tushar Athawale (Oak Ridge National Laboratory (ORNL)), Arthur Lorenzon (Federal University of Rio Grande do Sul), Peter Coveney (University College London), Haotong Qin (Hong Kong Polytechnic University), Bronson Messer (Oak Ridge National Laboratory (ORNL))
Abstract: Scientific AI on leadership-class HPC sits at the intersection of three concerns the community usually addresses separately: energy efficiency, uncertainty quantification (UQ), and verifiable trust. Compression and quantization change model behavior in ways that matter for scientific validity. Energy optimizations can quietly erode privacy and safety properties. And the HPC, UQ, and AI-safety communities rarely meet in one room. This BoF session brings them together to identify shared challenges, gaps in current practice, and concrete next steps toward AI systems for science that are simultaneously efficient, uncertainty-aware, and verifiably trustworthy on the world’s largest computing systems.
Long Description: Goals. This BoF has three goals. First, to convene three communities that rarely intersect HPC energy-efficiency researchers, scientific UQ practitioners, and AI safety/security researchers around the shared problem of trustworthy scientific AI on leadership-class systems. Second, to surface concrete cases where the three axes interact and trade off against each other: where compression breaks uncertainty guarantees, where energy optimizations open security surfaces, or where verifiable trust constrains efficient deployment. Third, to identify what a coordinated research agenda across these three axes would look like, including shared benchmarks, common vocabularies, and follow-on community structures.
Topic. Energy efficiency for HPC AI is no longer a research curiosity it is a deployment constraint. Leadership-class facilities face hard power-budge ceilings; data-center electricity is increasingly the rate-limiting resource for scientific computing; post-exascale roadmaps explicitly target watts-per-FLOP as the central scaling axis. Scientific AI is therefore increasingly evaluated on energy efficiency alone joules per token, FLOPS per watt, model size after compression. But on leadership-class systems doing actual science (autonomous laboratories, materials discovery, fusion control), efficiency cannot be the only metric. Quantization and pruning shift model behavior in ways that affect scientific validity. Uncertainty quantification matters not only as a numerical property but as something scientists must see and reason about uncertainty visualization is the practitioner-facing side of UQ, and it determines whether a scientist using an HPC-deployed model can trust the result. Energy-saving choices can change privacy guarantees, attack surface, and decision-integrity properties. The three axes are in tension: efficiency optimizations may erode UQ and trust; UQ and trust may require methods that consume more energy. The community needs an integrated framework that treats them jointly - not three separate ones.
Relevance to the HPC audience. This BoF directly addresses three top-of-mind SC concerns. Energy: every leadership-class facility faces power-budget pressure, and AI workloads are increasingly the dominant draw. Uncertainty: scientific AI deployment requires methods to compute, communicate, and visualize when model outputs can be trusted. UQ itself is compute and energy intensive demanding HPC. Verifiable trust: as AI is integrated into autonomous instruments and cross-facility workflows, verifiability becomes a deployment requirement. The session is relevant to system architects, performance analysts, UQ practitioners, scientific visualization researchers, scientific application developers, AI/ML researchers on compression and deployment, and security researchers focused on cyberinfrastructure.
Expected outcomes. We will invite participants to contribute to a community position paper that identifies the most pressing research priorities and potential solutions at the intersection of the three HPC pillars: energy efficiency, UQ, and verifiable trust. Led by the BoF organizers, this paper will serve as a foundational step toward building a sustained research community that fosters collaboration across these traditionally separate areas. We envision this effort growing into a dedicated workshop and broader community initiative at future Supercomputing conferences. To support continued engagement, we will develop a BoF website that documents participants, relevant publications, software tools, and other resources related to the BoF theme. The website will serve as a community hub, facilitating knowledge sharing and collaboration beyond the conference itself.