The Facility as Collaborator: Enabling AI-Driven Discovery at National-Lab Scale

Tom Uram of the Argonne Leadership Computing Facility describes how ALCF is positioning the facility as a collaborator for AI-driven discovery. Beyond the Aurora exascale system, ALCF deploys dedicated inference hardware including Sophia (NVIDIA A100), a SambaNova system, and forthcoming B200 and GH200 systems, backing an inference service in production since mid-2024 that serves over 35 open-weight LLMs, embedding models, and domain-specific science models through a standard API that integrates with agent harnesses. Remote job submission APIs and Globus Compute let agents orchestrate work across systems. Examples span time-sensitive analysis for the Advanced Photon Source with AI-based tomography segmentation returned to beamlines during experiments, between-shot analysis for a fusion facility using ALCF and NERSC, the OpenCosmo interface for in-place analysis of petabyte-scale cosmology data, the StructBio Reasoner drug-discovery workflow for intrinsically disordered proteins that found binding targets outperforming human predictions using the Academy agent framework, and ChemGraph, where planner, executor, and analyst agents manage an elastic pool of simulation jobs via Parsl. Uram also outlines the DOE Genesis Mission context, including the American Science Cloud effort to share models and infrastructure across DOE labs and the ModCon consortium for foundation model development. Audience discussion covers vendor lock-in with commercial LLMs, reproducibility of stochastic models, and how dynamic agentic workflows strain batch schedulers designed for deterministic jobs.