An early-stage research effort in scientific understanding
AI2Science explores a simple idea: trained models in science should be studied not only for what they predict, but for what they may reveal about the structure of the world.
This is early-stage work that will evolve over the next few months. We are starting with one concrete project, VizFold, and using it to shape the broader philosophy behind AI2Science.
First project
VizFold is the first project built around the AI2Science philosophy. It focuses on protein folding as a testbed for asking whether high-performing models can also become sources of scientific insight.
The goal is not only better structure prediction, but understanding what folding models learn about sequence, geometry, interaction, and organization.
VizFold is where AI2Science becomes concrete: a first attempt to study learned biological representations as possible carriers of mechanistic structure.
Pointers
Open science
AI2Science is forming in public. As the work develops, we want the philosophy, code, experiments, and project artifacts to be inspectable, shareable, and reproducible.
Openness matters especially at this stage. The ideas are still taking shape, and we want that evolution to be visible.
Institutional home
AI2Science is supported by Georgia Tech's Center for AI in Sience and Engingeeing (ARTISAN), a research center committed to open, interpretable, and computationally ambitious science. ARTISAN provides the broader ecosystem for infrastructure, collaboration, and long-horizon research in AI-enabled scientific discovery.
This is an emerging direction with a clear philosophy, one initial project, and room to grow substantially over the coming months.
ARTISAN Center Website