BEACONS and BEACONS-FM: Lighting a Path to Modular, Composable, Formally Verified Fusion Foundation Models
Ammar Hakim
Princeton Plasma Physics Lab
I will present our research program in constructing modular, composable, formally verified foundation models, with application to fusion. This program has three hierarchical parts, building on top of each other: (1) formally verified numerical solvers, (2) neural solvers that maintain bounded errors in extrapolatory regimes, and (3) foundation models that use a mixture-of-experts approach to construct super-networks to solve complex physics problems. Our key insights are derived from measure theory and category theory, allowing us to use continuity properties to decompose a function in a way that minimizes errors, even in regimes outside the training set. I will show how we can use these theoretical building blocks to construct deep networks, and in turn use these deep networks to build foundation models. As an example, I will show applications to hydrodynamics, culminating in full (2D and 3D) profile predictions for NSTX-U experimental shots. This work was performed in collaboration with Jonathan Gorard (Princeton University).