A Microsoft Research project introduced a machine learning pipeline that predicts geomagnetically induced current risks for 66,935 U.S. substations. By combining solar wind data with local geological and infrastructure features, the system generates location-specific risk estimates 30 to 60 minutes before potential solar storm impacts occur on the ground.
Predicting geomagnetically induced currents
The system uses a three-stage pipeline to address the complexity of how solar activity interacts with Earth's magnetic field. A gradient-boosting model processes solar wind measurements from the L1 Lagrange point alongside geological conductivity and grid infrastructure data. This allows the model to estimate the rate of magnetic field change for individual assets, accounting for variables like resistive bedrock and transmission line orientation.
Researchers utilized a system of 50 AI agents to assist in exploring features and model configurations. The project relied on public data from NASA, the U.S. Geological Survey, and INTERMAGNET to train predictors for the Auroral Electrojet and Disturbance Storm Time indices. These indices help the model track large-scale geomagnetic storm strength and rare, high-intensity activity that specifically threatens physical infrastructure.
Performance and current limitations
Evaluation across the 2020-2026 period showed the model outperformed empirical solar-wind-based approaches. For major geomagnetic events, the system achieved a 76.5 percent detection rate, though false-alarm rates increased alongside storm severity. Testing indicated that performance is currently most effective at northern latitudes where geomagnetic signals are strongest.
The development remains a research project rather than a deployed utility tool. The author notes that the system requires further validation with operational utility data and real-world grid observations before it can be used for active grid management. Current results demonstrate the feasibility of using physics-grounded machine learning for specific infrastructure assessments.
Original source
This report summarises the source below. Analysis is labelled separately; product and research claims remain attributed to their source.
Read the original at Microsoft Research