LF Energy Summit Europe 2026

We present GridSFM, an open-source dataset and foundation model that enables building and operating transmission-scale power grid models from open data to machine learning inference.

Starting from publicly available sources in the United States (OpenStreetMap, U.S. EIA, and U.S. Census), our pipeline constructs geographically grounded, AC-OPF-solvable U.S. power grid models at continental scale. Importantly, our modeling pipeline can be extended to other locations where similar open data is available.

On top of this data layer, GridSFM introduces a physics-informed foundation model trained across a diverse mix of grid topologies and operating scenarios to predict AC-OPF solutions in milliseconds, enabling fast feasibility screening and solver acceleration.

This session will walk through the full stack: from open data ingestion and model construction to ML training and deployment. We will share lessons learned building reproducible, large-scale grid datasets and integrating physics constraints into learning systems.