Autonomous Assessment of Airspace Operations

NASArsquo;s 2018 SBIR solicitation topic A3.02 requests ldquo;Autonomous systems to produce any of the following system capabilities: Prognostics, data mining, and data discovery to identify opportunities for improvement in airspace operations.rdquo; Identifying opportunities for improvement is a critical ongoing need in the air traffic management domain, for which achieving high levels of performance is a daily concern. In Phase I, Mosaic ATM developed and delivered a machine learning model to predict terminal area (TRACON) transit times for flights arriving to Dallas/Fort Worth International Airport (DFW) based on a broad array of flight and weather input data. We also delivered a visualization capability that enables analysts to connect observed variations in TRACON transit times with their most important causes. For Phase II, we will build upon our Phase I outcomes by expanding both the sophistication of our methodologies and the performance domains we address. We will develop deep learning models of NAS performance metrics for two challenging domains: the capacity of New York area airports and the capacity of en-route airspace in the Eastern US. We will advance the state of the art in explainable deep learning for complex systems by following a rigorous process for extracting understandable basis vectors, by iteration over the internal variables of a hybrid neural network. By doing those two things, we will achieve a foundation for traffic manager decision support enhancements, to provide guidance on how to configure the Time-Based Flow Management (TBFM) system and the Traffic Flow Management System (TFMS) to maximize system-level performance. Finally, we will implement a user interface that is integrated with the NASA Sherlock ATM data warehouse, to enable NASA domain analysts to explore the models and the important features affecting performance, in a collaborative way.