GridSFM: Microsoft's small foundation model speeds up power grid optimization by 100×

Microsoft has unveiled GridSFM, a small foundation model designed to solve AC optimal power flow (AC-OPF) problems—a notoriously complex optimization challenge that underpins modern grid operations. The model can approximate solutions in milliseconds, versus hours for traditional solvers, potentially unlocking significant economic and operational gains.

The Grid Optimization Challenge#

Power systems face mounting pressures: surging electricity demand, renewable energy integration, electrification of transport, and extreme weather. All of these require operators to answer a critical question: what operating points keep the grid running safely and cheaply? That question is answered by AC-OPF, a non-convex optimization problem that computes the cheapest generator dispatch while respecting voltage limits, thermal constraints, and power flow physics.

The stakes are enormous. AC-OPF decisions directly govern outcomes at the scale of $20 billion per year in congestion costs and multi-terawatt-hour renewable curtailment (lost renewable energy due to congestion). But solving AC-OPF at scale is computationally brutal: utility-scale grids can take hours to solve, forcing operators to choose between evaluating a few carefully selected scenarios accurately or running thousands of scenarios through approximations that ignore critical physics.

How GridSFM Works#

GridSFM sidesteps this bottleneck by using a block-structured discrete neural operator—representing the grid as a directed graph with buses as vertices and transmission lines as edges. The model was trained across 150+ different grid topologies and roughly half a million scenarios, spanning varying load profiles, multi-element outages, and different generator costs. This forces the model to generalize rather than memorize.

The approach yields two tiers:

• GridSFM-Open: research-scale grids up to 4,000 buses
• GridSFM-Premier: production-scale systems up to 80,000 buses

Performance and Accuracy#

Across 54 test grids, GridSFM-Open achieves a median cost gap of 2.23% versus solver ground truth (83% of scenarios within 5% accuracy). When used as a warm-start seed for traditional solvers, GridSFM achieves 1.66× speedup over cold starts and 1.59× faster than the industry-standard DC-OPF warm-start.