DeGaP: A Deep Gaussian Process Surrogate Model for Cleaning Data from Spatially Distributed Sensor Networks

The DeGaP model addresses the critical challenge of cleaning, gap-filling, and quantifying uncertainty in real-time magnetic field data. Developed as a Deep Gaussian Process (GP) surrogate model, DeGaP preserves high-frequency geophysical signals while removing noise and filling data gaps In Phase I, the DeGaP model successfully filled data gaps in high-frequency magnetic field measurements from the MagStar network, preserving essential geophysical features while quantifying uncertainty in the filled-in data. The model leveraged a novel spatiotemporal GP kernel that integrated a Spectral Mixture Kernel for capturing temporal complexities and a Radial Basis Function Kernel for spatial dependencies. Results showed that DeGap model accuracy in reconstructing high frequency components in magnetic field data under diverse geomagnetic conditions. Phase 2 will enhance and operationalize DeGaP by: 1. Advancing DeGaP by integrating it into real-time data processing workflows, delivering cleaned and gap-filled magnetic field data to NASA, NOAA, and other stakeholders. 2. Enhancements will include improved computational efficiency using GPU-accelerated techniques, expanded robustness to handle extended data gaps and noisy inputs, and an embedded anomaly detection module powered by GP uncertainty estimates. 3. The project will also develop user-friendly visualization tools for accessing cleaned data, uncertainty metrics enabling rapid decision-making for the subject matter expert in the loop to engage in ensuring data quality of the cleaned magnetic field data. The project will deliver a real-time, gap-filled magnetic field data stream with validated robustness under diverse geomagnetic conditions. DeGaP will directly support NASA’s heliophysics missions, model validation efforts, and space weather monitoring programs by ensuring reliable, high-quality data for scientific analysis and operational decision-making.