Improving Efficiency and Realism of Synthetic Flight Tracks Generated by Randomized Gaussian Processes

Abstract

Collision Risk Models (CRMs) estimate air collision probabilities to assess safety of new air traffic procedures, requiring

large amounts of historical flight data that are often unavailable. Randomized Gaussian Process (RGP) regression

generates many synthetic flight tracks from few historical tracks to address this issue. However, this algorithm introduces

two major limitations: synthetic track generation is computationally slow at a large scale, and unrealistic synthetic

altitude profiles reveal dramatic overshoots—landing flights ascending thousands of feet during descent—and excessive

smoothing of level-flight segments. This project utilizes 50 historical arrival tracks from landing flights on Runway 14 at

Zurich Airport. First, parallel computing was implemented using Python and the joblib library, employing 8 CPU cores to

generate flight tracks concurrently. Next, individual seed tracks are investigated showing how the nonparametric nature

of Gaussian Process regression causes altitudes to connect through a smooth curve between sparse design points.

Overshoots and loss of level segments can be explained by insufficient selection of design points near end altitudes and

level-to-descent transitions. Including these critical points preserves the overall level-descent shape of synthetic altitude

profiles, reducing dramatic overshoots while maintaining variation.