Microsoft Research GridSFM: foundation model solves AC optimal power flow 100× faster than DC approximation

Microsoft Research presented GridSFM on May 13, 2026 — a small foundation model that solves one of the most costly classical operational problems in the electricity sector. The model approximates AC optimal power flow (AC-OPF) in milliseconds on grids up to 80,000 nodes and identifies potential savings of $20 billion annually in congestion costs.

What does GridSFM specifically do?

AC-OPF is an optimization problem that determines the most economical way to produce and distribute electrical energy while respecting physical constraints (Kirchhoff’s laws, voltage limits) and operational constraints (generator capacities, ramping). GridSFM approximates the optimum in milliseconds for grids of 500 to 80,000 nodes — making it usable for real-time dispatch, while classical solvers require minutes.

How much faster is the model and how accurate is it?

Speed: 100× faster than DC approximation and 1,000× faster than full AC solvers at inference. The price for speed is a slightly higher cost gap — median 2.23%, mean 3.41% versus traditional solver solutions. Feasibility detection achieves 94.5% on feasible scenarios and 96.1% on infeasible ones. The warm-start scenario delivers a 1.66× geometric mean speedup over cold-start solving.

What economic value does it identify?

Microsoft Research estimates savings of up to $20 billion annually in congestion costs in the US electric grid, plus addressing 3.4 TWh of renewable curtailment (lost renewable generation due to grid constraints). The primary application is rapid screening of thousands of contingency scenarios that were previously limited by computational cost — transmission planning, reliability analysis and real-time dispatch optimization.

How was the model trained and what versions exist?

GridSFM was trained on 150+ base grid topologies and approximately 500,000 scenarios covering variable load profiles, outages and operational constraints. The broad training distribution enables generalization across different grids without training a separate model per topology. Two release tiers are available: GridSFM-Open for research grids up to 4,000 nodes (public) and GridSFM-Premier for production systems up to 80,000 nodes (by request).

The approach positions foundation models as a tool for scientific and engineering domains — a parallel with the Microsoft MatterSim model for materials science, which was experimentally confirmed through the synthesis of TaP on May 12.