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

Spatiotemporal data from space and ground-based instruments are used in research and as a validation source for physical models throughout NASA fundamental space science and sensors on the ground and in space. While there are a wide range of measurement types and uses, all data suffers from non-physical errors that must be identified, and removed, before its use in science and operations. It is important that these data gaps be filled with scientifically-derived proxies automatically, and with as little latency as possible, for timely hazard detection. Our proposed innovation is to develop a Deep Gaussian Process surrogate model (DeGaP) that is used to fill in gaps due to either bad or missing data, while retaining the high precision and fidelity of the original measurements along with uncertainty quantification (UQ) of the filled-in measurements. Deep Gaussian Processes are a powerful technique to model complex system behavior while enabling uncertainty to be fundamentally incorporated by virtue of the stochastic structure of the model. In Phase I, we will a) develop a prototype of the DeGaP model using a ground-based magnetic field sensor network, MagStar. b) quantify the accuracy and uncertainty of the model predictions in terms of spatial and temporal proximity to other magnetometers in the network c) host a web service for continuous dissemination of the DeGaP model output wherein gaps and transient disturbances in the magnetic field data are replaced and uncertainty quantification provided for the filled-in data. Potential applications for the Phase 1 development include the Space Weather Modeling Framework (SWMF) running at CCMC, World Magnetic Model at NOAA and space-based magnetometers. Potential private sector market for the product is the U.S. bulk power industry, entities that support US power utilities in regulatory response, and entities that need machine-learning (ML) ready data sets from sensor networks for ML models in production.