GridFM

GridFM is a Microsoft Research initiative to build small foundation models for the electric power grid, applying the same recipe that drives modern language and weather models to the physics of AC power flow. The first release in the family, GridSFM, predicts complete AC Optimal Power Flow (AC-OPF) solutions in milliseconds — bus voltages, generator dispatch, branch flows, and a feasibility verdict — directly from a grid’s topology and operating conditions.

Traditional AC-OPF solvers are accurate but slow, taking minutes to hours on real-world grids with tens of thousands of components. As power systems grow more volatile under datacenter expansion, renewable variability, electrification, and extreme weather, operators need to evaluate thousands of scenarios in seconds, not hours. GridFM is designed to close that gap without sacrificing physical fidelity.

What’s new (May 2026):

• GridSFM is open source. The model architecture, training pipeline, and warm-start integration with PowerModels.jl are available on GitHub (opens in new tab); checkpoints are on Hugging Face (opens in new tab).
• Two model tiers. GridSFM-Open (~15M parameters, grids up to ~4,000 buses, MIT licensed) for research and prototyping, and GridSFM-Premier (~100M parameters, grids up to ~80,000 buses) for production-scale networks. Both share the same architecture.
• A companion open-data pipeline. We also released a five-stage pipeline that builds OPF-solvable transmission models for all 48 contiguous US states and six multi-state regions — including the full Western (5,076 buses) and Eastern (21,697 buses) Interconnections — using only public data (OpenStreetMap, EIA, US Census).
• 54 reproducible grid models are publicly released alongside the code, lowering the barrier to transmission-level research without proprietary or critical-infrastructure-restricted data.