SESAR 3 JU exploratory research, 2023–2026
SynthAIr developed and evaluated generative models that reproduce the statistical and operational properties of real flight data, so AI can be trained and tested when real data is scarce, sensitive or rare.
Pictured: each line is a synthetic landing at Oslo Gardermoen (SESAR Innovation Days 2024 poster).
The results in 30 seconds
95–97%
of real-data performance kept by turnaround and delay forecasts trained only on synthetic flights.
Compared with training on real data (full bar)
A share of what real data achieves, not accuracy: even real data only reaches R² ≤ 0.44 here. This generator also has the smallest privacy margin of the four.
1.7 million European flights
CEAS Aeronautical Journal, 2026
>6×
higher PR-AUC for predicting flight diversions after adding 1,000 synthetic ones. Only about 1 flight in 480 diverts.
PR-AUC, higher is better
Best of three tuned generators (0.28 to 0.44), on one month of New York flights.
127 diversions among about 61,000 flights
Aerospace Science and Technology, 2026
5%
of Dublin’s data, fewer than 1,000 landings, brought a model pretrained on Zürich close to one built from all of it.
Share of Dublin’s training data
One airport pair and a diffusion model. Rarer west-facing approaches stay underrepresented.
Zürich to Dublin
SESAR Innovation Days 2025
All three were measured against real flight data. This is early-stage research at TRL 1, not an operational system. The project met all 17 validation criteria; the seven on operational impact are qualitative and preliminary (D5.2).
See the evidenceThe challenge
Machine learning needs large, shareable and varied datasets. Operational aviation data is often none of these. SynthAIr explored whether generated data can fill the gap, and measured where it can and cannot.
European flight records from OAG, U.S. on-time data from BTS, ADS-B trajectories from OpenSky, and the EUROCONTROL R&D Archive.
Ten models for tabular records and trajectories, from statistical copulas to transformers and diffusion.
Fidelity, diversity, privacy and flyability, plus utility: train on synthetic data, test on real data.
Six ATM use cases, and general-purpose embeddings for analytics and transfer between airports.
See the data
Landing trajectories around Oslo Gardermoen. Drag the divider to compare the real training data with what a SynthAIr model generated from it.
Colours mark the five trajectory clusters. A visual match is only a first check: see the full evaluation in UC4. Synthetic Oslo landing trajectories are available as an open dataset on Zenodo.
Real
Synthetic
Use cases
From waiting at airport security to the next departure: what was tested in each use case, and what came out of it.
How long will an aircraft stay on the ground between flights?
Read the use caseCan generated trajectories stand in for real ones in ATM studies?
Read the use caseResults
Three findings from the final deliverables and papers, with their limits. All results are exploratory, at Technology Readiness Level 1.
REaLTabFormer kept the most predictive value on all three tasks, but offered the smallest privacy margin. Gaussian Copula did the reverse. The right choice depends on how sensitive the data is.
Distance to closest record is a useful early indication of privacy, not a formal guarantee.
Compare the models| Utility | Privacy | |||
|---|---|---|---|---|
| Turnaround | Departure delay | Arrival delay | Baseline protection | |
| REaLTabFormer | 0.97 | 0.96 | 0.95 | 0.23 |
| TabSyn | 0.93 | 0.76 | 0.64 | 0.30 |
| CTGAN | 0.74 | 0.68 | 0.57 | 0.34 |
| Gaussian Copula | 0.42 | 0.56 | 0.53 | 0.54 |
Only 127 of about 61,000 flights in the dataset were diverted. In the journal study, adding synthetic diversions from optimised generators lifted PR-AUC from 0.066 on real data alone to between 0.28 and 0.44.
Gaussian Copula consistently underperformed on utility in that study, and the data lacks weather and traffic features that are likely to drive diversions.
Read UC5A diffusion model pretrained on Zürich and fine-tuned on Dublin got close to a Dublin baseline with 5% of the local data, and reached baseline-level performance at around 20%.
Rarer approaches, such as west-facing arrivals, remain underrepresented after transfer.
See transfer learning| Dublin data | e-distance | DTW |
|---|---|---|
| Baseline | 0.662 ± 0.170 | 29.099 ± 0.106 |
| 0% | 1.228 ± 0.445 | 118.222 ± 0.352 |
| 5% | 0.693 ± 0.183 | 23.211 ± 0.084 |
| 20% | 0.681 ± 0.178 | 22.941 ± 0.081 |
| 50% | 0.679 ± 0.175 | 18.428 ± 0.067 |
| 100% | 0.668 ± 0.171 | 15.876 ± 0.060 |
Models
The latent representations these models learn also work as general-purpose embeddings for clustering, anomaly detection and core-set extraction.
Autoregressive generation that treats each flight record as a sequence of tokens
Diffusion in a VAE latent space for efficient mixed-type synthesis
Conditional training aimed at rare categories and imbalanced data
Stable to train, with a light compute footprint
Models dependencies separately from marginals; the strongest privacy margin of the four generators compared on European data
Time-frequency processing with transformer priors for global coherence
Temporal convolutions with a flexible mixture prior
Adversarial training that preserves temporal relationships
Continuous normalizing flows with deterministic, efficient sampling
Denoising diffusion adapted to flight data; the strongest transfer between airports
Open science
Code, data and reports are public, so others can reproduce the work and build on it.
The project
SynthAIr ran from 1 September 2023 to 28 February 2026 as a SESAR 3 Joint Undertaking exploratory research project. An external advisory board of airlines, air navigation service providers, airports and researchers shaped its technical direction through three workshops.
Meet the team Events and videos
This project has received funding from the SESAR 3 Joint Undertaking under grant agreement No 101114847 under European Union’s Horizon Europe research and innovation programme. Contact: Massimiliano Ruocco (SINTEF), massimiliano.ruocco@sintef.no .