Learning to Land Anywhere: Transferable Generative Models for Aircraft Trajectories
Olav Finne Præsteng Larsen 2 , Massimiliano Ruocco 1,2 , Michail Spitieris 1 , Abdulmajid Murad 1 , Martina Ragosta 1 · 2025
Read Learning to Land Anywhere: Transferable Generative Models for Aircraft Trajectories by Olav Finne Præsteng Larsen 2 , Massimiliano Ruocco 1,2 , Michail Spitieris 1 , Abdulmajid Murad 1 , Martina Ragosta 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-04155.txt.
Introduction
Olav Finne Praesteng Larsen, Massimiliano Ruocco, Michail Spitieris, Abdulmajid Murad, and Martina Ragosta test generative models of aircraft landing trajectories. The models are trained at a busy airport, and the question is whether they can transfer to airports with little data. Trajectory data underpin traffic management tools, yet secondary and regional airports often lack enough records for learning or large simulations.
Diffusion and flow-matching models pretrained on Zurich landings were fine-tuned on Dublin data in amounts from none to all. Diffusion models matched competitive performance with only five percent of local data and reached baseline level near twenty percent, beating scratch training throughout. Flow-based variants gained less evenly, marking diffusion transfer as the strongest route.
Catalog record
| CREATOR | Olav Finne Præsteng Larsen 2 , Massimiliano Ruocco 1,2 , Michail Spitieris 1 , Abdulmajid Murad 1 , Martina Ragosta 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-04155
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning to Land Anywhere: Transferable Generative Models for Aircraft Trajectories Olav Finne Præsteng Larsen 2 , Massimiliano Ruocco 1,2 , Michail Spitieris 1 , Abdulmajid Murad 1 , Martina Ragosta 1 Abstract Access to trajectory data is a key requirement for developing and validating Air Traffic Management (ATM) solutions, yet many secondary and regional airports face severe data scarcity. This limits the applicability of machine learning methods and the ability to perform large-scale simulations or “what-if” analyses. In this paper, we investigate whether generative models trained on data-rich airports can be efficiently adapted to data-scarce airports using transfer learning. We adapt state-of-the-art diffusion- and flow-matching–based architectures to the aviation domain and evaluate their transferability between Zürich (source) and Dublin (target) landing trajectory datasets. Models are pretrained on Zürich and fine-tuned on Dublin with varying amounts of local data, ranging from 0% to 100%. Results show that diffusion-based models achieve competitive performance with as little as 5% of the Dublin data and reach baseline-level performance around 20%, consistently outperforming models trained from scratch across metrics and visual inspections. Latent Flow Matching and Latent Diffusion models also benefit from pretraining, though with more variable gains, while Flow Matching models show weaker generalization. Despite challenges in capturing rare trajectory patterns, these findings demonstrate the potential of transfer learning to substantially reduce data requirements for trajectory generation in ATM, enabling realistic synthetic data generation even in environments with limited historical records. I Introduction The development of robust Air Traffic Management (ATM) systems increasingly relies on the availability of high-quality trajectory data. Such data underpins the design of machine learning algorithms, the validation of new operational procedures, and large-scale simulation studies. However, while major hub airports typically record large volumes of flight data, many secondary and regional airports suffer from severe data scarcity. This imbalance limits the ability to test and validate new ATM solutions across diverse operational contexts. Moreover, the exploration of rare events, edge cases, and “what-if” scenarios requires datasets far larger and more varied than what is available from real-world operations. Recent advances in generative modeling have shown promise in addressing this challenge by synthesizing realistic flight trajectories. Techniques such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models have demonstrated their ability to capture the distributional characteristics of real flight data and generate plausible synthetic samples. Nevertheless, these approaches typically assume access to a sufficient amount of local training data. In practice, this assumption often fails for airports with limited operations or incomplete historical records, restricting the applicability of generative methods where they are needed most. In this paper, we investigate whether transfer learning can mitigate these challenges. Specifically, we explore whether generative models trained on a data-rich airport can be effectively adapted to airports with scarce data. By pretraining on a large landing trajectory dataset from Zürich Airport and fine-tuning on limited data from Dublin Airport, we evaluate the transferability of state-of-the-art generative architectures, including Diffusion Models, Flow Matching, and Latent Flow Matching. The main contributions of this work are as follows: • We present, to the best of our knowledge, the first systematic study of transfer learning for generative trajectory models in the ATM domain. • We demonstrate that diffusion-based models can achieve competitive performance with as little as 5% of the target airport’s data and reach baseline-level performance around 20%, significantly reducing local data requirements. • We provide a comparative analysis of different generative approaches, highlighting the strengths of diffusion-based architectures and the limitations of flow-matching methods for transfer learning. • We discuss the implications of transfer learning for enabling realistic trajectory generation in data-scarce environments, thereby broadening the applicability of synthetic data in ATM research and development. By showing that generative models can transfer knowledge across airports, this work paves the way toward scalable synthetic trajectory generation, supporting the development of data-driven ATM systems even in operational contexts where historical records are limited. II State of the Art Research on synthetic flight landing trajectories lies at the intersection of deep generative modeling and aviation operations. Progress in deep generative models has been rapid, spanning Variational Autoencoders (VAE) [ 8 ] , Generative Adversarial Network (GAN) [ 4 ] , Diffusion Models (DM) [ 20 , 6 , 21 ] , and, most recently, Flow Matching (FM) [ 12 ] . Latent variants such as Latent Diffusion Models (LDM) [ 17 ] and latent flow matching (LFM) [ 2 ] combine strong sample quality with improved efficiency by shifting generation to a compressed space. While most architectural innovation has emerged in computer vision, a growing body of work adapts these ideas to spatio-temporal trajectories, where temporal dependencies, multi-modality, and physical constraints are central. In particular this work focus primarily on advances in trajectory generation in the ATM domain. This is currently a growing research area aiming to address challenges related to data accessibility, privacy, and the complexity of flight data [ 3 ] . Aircraft trajectory generation can be formulated as the challenge of generating a set of fully timestamped coordinates for an aircraft that reflect physical behavior and the statistical properties of the spatio-temporal bounds [ 15 ] . The existing literature on generative flight trajectory models divides them into two categories: model-driven and data-driven approaches [ 15 , 25 , 13 ] . Model-driven approaches rely on mathematical models of flight, ensuring that the trajectories follow the laws of aerodynamics and other operational constraints [ 15 ] . Data-driven methods use machine learning techniques to model and generate flight trajectories directly from data [ 15 ] . These methods, such as TimeGAN [ 25 ] , TCVAE and VampPrior [ 10 ] , have shown promise in capturing temporal dependencies and generating realistic landing trajectories and [ 13 ] that applied the proposed TimeVQVAE architecture [ 11 ] to generate complete flight trajectories and introduced new evaluation methods tailored to trajectory generation Interestingly, [ 10 ] generates trajectories by groundspeed and track. In contrast, [ 25 ] uses longitude and latitude, which is the most common approach in other domains, demonstrating that there are different approaches to modeling the data to be learned. While the application of data-driven generative models in aviation remains underdeveloped, significant advances have been made in other fields, particularly urban traffic, where state-of-the-art generative models such as VAEs [ 1 ] , GANs [ 27 ] , and DMs [ 28 , 29 ] have successfully learned complex trajectory data distributions. Recent advancements in DMs [ 28 , 9 , 24 ] have demonstrated significant potential for spatio-temporal data generation. Although aviation and urban traffic differ due to the three-dimensional and sparse nature of airspaces compared to the strict topological constraints of urban traffic [ 29 ] , the DMs’ ability to handle complex, high-dimensional distributions [ 21 , 6 ] suggests they can overcome the challenges posed by flight trajectories. These are challenges such as the need for physical realism, airspace restrictions, and various weather conditions [ 15 , 5 ] . While newer DMs have been proposed in urban traffic [ 29 , 24 ] , these are specifically tailored to the urban traffic domain by incorporating road constraint information into the models to improve performance. Therefore, DiffTraj