GridSFM: A Foundation Model for AC Optimal Power Flow

Grid Small Foundation Model (GridSFM) is a foundation model for power systems trained on 200 grids and over half a million scenarios. It predicts AC-OPF solutions in milliseconds: given a grid topology and loading conditions, it produces bus voltages, generator dispatch, branch power flows, and a feasibility classification without running a solver, and when higher confidence is needed its predictions serve as warm starts that accelerate conventional solvers. GridSFM is released in two tiers, GridSFM-Open (∼15M parameters, for grids up to a few thousand buses, research and prototyping) and GridSFM-Premier (∼100M parameters, for production scale grids up to tens of thousands of buses), both sharing the same architecture and trained on a broad open-data corpus of transmission topologies and operating scenarios spanning feasible and infeasible regimes via a multi-axis perturbation pipeline. Code is available at https://github.com/microsoft/GridSFM, and released checkpoints are hosted in the Hugging Face collection at https://huggingface.co/collections/microsoft/gridsfm.