Invited talk | Weiwei Yang: From GridSFM to Asemic: Foundation Models for Feasible Grid Decisions

AI has a power grid problem. The power grid has a compute problem. Data centers are arriving faster than transmission lines can be built, so the fastest new capacity is whatever the existing grid can already carry. Finding it is the hard part. AC optimal power flow is nonconvex on its own; siting and switching add discrete choices on top. Each solve is slow, and the search is combinatorial.

Asemic builds physics foundation models to break this compute bottleneck. We start from GridSFM, the first foundation model for the power grid, which showed that one set of weights can learn AC optimal power flow across various grids. We extend it toward what grid decisions need: utility-scale networks and post-fault dynamics, not just steady state. This talk follows the path from curating public data into realistic grid topologies to building models that respect the physics by design. Then we put them to work on the Texas grid and ask three questions: how much new load can the existing wires carry, what does that do to everyone’s electricity bill, and what happens when something breaks?