ATD-2 to TFDM — NASA's Machine-Learning Airport Scheduler Saves Over a Million Gallons of Jet Fuel

Aircraft that push back from the gate too early end up idling in long taxi queues, burning fuel and emitting carbon while going nowhere; airports have historically relied on static, physics-based estimates rather than predicting exactly when a runway slot will actually open.

Replace legacy physics-based taxi-time modelling with machine-learning models - including neural networks - trained on real surface-surveillance data, to predict taxi-out and taxi-in times and better time aircraft pushback from the gate.

NASA's Airspace Technology Demonstration-2 (ATD-2) project built an Integrated Arrival, Departure and Surface (IADS) system whose Surface Trajectory-Based Operations component was field-tested at Charlotte Douglas International Airport (KCLT), using data also drawn from Dallas/Fort Worth (KDFW), from September 2017 to September 2021.

The system saved more than 1 million gallons of jet fuel, an estimated $1.4 million in flight-crew costs, 933 hours of passenger delay (valued at roughly $4.5 million), and around 8 tons of CO2 per day at Charlotte Douglas. NASA formally transferred the technology to the FAA in September 2021, where it was folded into the Terminal Flight Data Manager (TFDM) programme for a phased rollout beginning with Charlotte and Phoenix and eventually reaching 27 hub airports, including Atlanta, Chicago, New York, Washington, San Francisco and Los Angeles.

The published machine-learning taxi-time models still carry meaningful uncertainty - roughly 10-15% for total and airport-movement-area taxi times, and about 20% for ramp taxi time - and national rollout to all 27 airports was projected to take five to ten years, so the measured benefits above remain concentrated at the original test sites rather than proven network-wide.