GridSFM: A Foundation Model for Solving AC Optimal Power FlowNew

GridSFM presents a pretrained foundation model combining physics-informed fine-tuning with a 15-million-parameter graph neural network architecture for solving AC Optimal Power Flow across diverse grid topologies. The model was pretrained on 54 different power grid configurations ranging from 500 to 4,000 buses, and demonstrates strong zero-shot generalization achieving 2.45% generation-cost error on a held-out 10,000-bus case without performance degradation as system scale increases. This work targets the power systems domain, applying machine learning to optimize grid operations and power flow calculations.