GridSFM: A new, small foundation model for the electric grid

Details

AC optimal power flow (AC-OPF) determines the cheapest generator dispatch that keeps a power grid stable — it underpins real-time dispatch, market clearing, and contingency analysis. The problem: AC-OPF is a non-convex optimization that can take hours to solve at utility scale, forcing grid operators to trade off solution quality for speed or rely on simplified approximations. GridSFM replaces the iterative solver with a small foundation model that predicts AC-OPF solutions in milliseconds. Beyond estimating generator dispatch and costs, GridSFM produces full AC system states — giving operators direct visibility into line congestion, voltage conditions, and system health. Microsoft frames the economic stakes: up to $20B/year in US congestion losses and 3.4 TWh of annual renewable curtailment are tied to how well these operating points are found. The model is released alongside an open US transmission-topology dataset that Microsoft published earlier. The goal is to give the research community a foundation to build grid simulators and planning tools without recreating data or models from scratch. For AI researchers and energy sector engineers, this is a credible application of foundation model techniques to critical infrastructure with measurable economic impact.