Foundation Models for Nuclear Physics

Foundation Models are transforming numerous scientific fields and offering promising opportunities across various disciplines, including nuclear physics. While their application in this domain is still in its early stages, they have the potential to drive both methodological and computational advancements. In this presentation, I will explore ongoing studies that align with the broader vision of developing foundation models for nuclear physics. A key focus will be on their potential application at the Electron-Ion Collider, where they could integrate fast simulation and multi-task reconstruction within a unified framework. This approach would enable high-fidelity event-level detector response simulation, accurately reproducing low-level features from each subdetector while simultaneously enhancing multiple downstream reconstruction tasks. Ultimately, this could lead to improved performance, as well as more robust and integrated physics analyses. I will also discuss key challenges that currently limit the feasibility of such models. Finally, I will highlight concrete examples of pre-trained models supporting agentic human-in-the-loop workflows, assisting with data retrieval, summarization, and experimental guidance. These approaches, based on pre-trained models, are already beginning to enhance research workflows within the nuclear physics community.