Workshop: 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science
Authors: Azton Wells, Nesar Ramachandra, and William Hicks (Argonne National Laboratory); Gideon MacFarland (Argonne National Laboratory, Northwestern University); and Michael Buehlmann, Giulia Campitiello, Salman Habib, and Katrin Heitmann (Argonne National Laboratory)
Abstract: We introduce a multi-agent cosmology system designed to access and analyze the largest cosmological simulations in the world. The system is designed as an interactive, end-to-end analysis suite with the goal of accelerating and streamlining complex analysis of cosmological simulations. Instead of benchmarking, in this work we compare the workflow of a real scientific team, reflecting a typical academic researcher group, to one scientist operating the MACosmo system. Both the scientific team and the agentic team found meaningful new information: the agentic system team isolated a distinct population of cool-core clusters residing in dynamically disturbed halos that a dark-matter-only classification would misclassify, while the scientific team derived new observables that aid robust classification through observational data. This work suggests that carefully designed multi-agent systems, used as collaborators, can accelerate and broaden the analysis of cosmological simulations while reaching conclusions that supplement an expert team.
Back to 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science Archive Listing
Back to Full Workshop Archive Listing