EAGER: Machine Learning (ML)-based Geomagnetic Field Reconstruction Framework

Magnetic storms dramatically alter Earth’s magnetosphere while transferring energy from the solar wind into the near-Earth environment. Understanding such processes requires a quantitative description of the geomagnetic field and electric currents. However, this is difficult because the magnetosphere is vast and only sparsely sampled by spacecraft. Traditional empirical magnetic field models provide useful estimates, but they are too rigid in their spatial structure and temporal evolution to capture the rapid reconfiguration of the magnetosphere during magnetic storms. A new approach leverages machine learning to identify measurements from past storms that are similar to a given event. This enables reconstruction of the magnetic field during individual events with much greater fidelity, but at the cost of increased complexity and computational resources. This project will make these advanced methods broadly accessible through open-source software, precomputed reconstructions, and user-friendly tools, supporting space science research and advancing space weather prediction. The overarching goal of this project is a new generation of empirical geomagnetic field algorithms that reconstruct the global structure of the magnetosphere during storms and substorms and enable their widespread use by the scientific community. Recent advances in machine learning and the growth of magnetospheric datasets have enabled data mining-based reconstructions. Traditional empirical models (collectively known as Tsyganenko models) are widely used for their simplicity but are limited by their statistical nature and rigid architectures, particularly for dynamic space weather phenomena such as storms and substorms and for event-specific studies. In contrast, data mining-based approaches provide significantly improved spatial and temporal fidelity but remain difficult to use due to their computational cost. This project will address this gap by making the first combined storm-substorm empirical model of the magnetospheric magnetic field widely available to the scientific community. The model and data mining algorithms will be released as open-source software in modern programming languages. Precomputed reconstructions will be generated at five-minute cadence for the period from 1995–2024, and a Run-on-Request tool will allow users to produce customized event reconstructions. The project will also establish the framework and computational infrastructure required for rapid development, maintenance, and proliferation of these empirical models and ML-based algorithms for the foreseeable future. These efforts will enable new data-assimilative and gray-box modeling approaches and provide derived data products to support next-generation magnetospheric research, including advances in space weather forecasting beyond global activity indices toward spatially resolved parameters. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2524858 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Grant Stephens | Institution: Johns Hopkins University, BALTIMORE, MD | Award Amount: $300,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2524858 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2524858.html