Incorporating Machine Learning into Mission Operations

Abstract:

Mass budgets, data storage, and processor power have been factors limiting the scope of what can be done with data on-board spacecraft. Today, more powerful computers, miniaturized memory storage devices, and machine learning models allow more to be done, but their unvalidated combined use in-flight still presents a technical risk for major NASA missions. The Magnetospheric Multiscale (MMS) mission and its memory management system provide a testing ground for incorporating machine learning models into mission operations. We present continued work with the Ground Loop System (GLS) -- a hierarchy of empirical and machine learning models to aid in event detection and science discovery. As a first application, a Long-Short Term Memory (LSTM) neural network model that detects magnetopause crossings has been implemented into MMS's operational data stream. Within the first five months of operation, the model predicted 76% intervals identified as magnetopause crossings. We present the results of an improved LSTM model, as well as the second application within the GLS -- a physics-informed Bayesian model for region identification. The Bayesian model is capable of distinguishing between the solar wind, bow shock, magnetosheath, magnetopause, and magnetosphere. With it, region-specific event classification models can be trained. We further assess the two models' ability to detect magnetopause crossings and discuss implications for on-board memory management and mission design.