SyntAIr - IMPROVED ATM AUTOMATION AND SIMULATION THROUGH AI-BASED UNIVERSAL MODELS FOR SYNTHETIC DATA GENERATION

Información del proyecto

SynthAIR

Identificador del acuerdo de subvención: 101114847

DOI

10.3030/101114847

Proyecto cerrado

Fecha de la firma de la CE 15 Junio 2023

Fecha de inicio 1 Septiembre 2023

Fecha de finalización 28 Febrero 2026

Financiado con arreglo a

• Climate, Energy and Mobility

Coste total

€ 1 215 003,75

Aportación de la UE

€ 993 550,00

Coordinado por

SINTEF AS
Norway

Periodic Reporting for period 2 - SynthAIR (SyntAIr - IMPROVED ATM AUTOMATION AND SIMULATION THROUGH AI-BASED UNIVERSAL MODELS FOR SYNTHETIC DATA GENERATION)

Período documentado: 2024-09-01 hasta 2026-02-28

Resumen del contexto y de los objetivos generales del proyecto

Air traffic management (ATM) is undergoing a deep digital transformation. SESAR has set ambitious targets for capacity, safety and environmental performance, and AI is widely seen as a key enabling technology. However, AI adoption in aviation faces a structural bottleneck: the lack of large, high-quality and representative operational datasets.

In ATM, data scarcity is structural. Data accumulates slowly, rare but safety-critical events such as go-arounds and diversions occur only a few times, and privacy, safety and commercial constraints limit sharing across airlines, airports, air navigation service providers and ground handlers. This makes many AI methods difficult to develop, validate and share.

SynthAIr investigated whether synthetic data can act as a practical substitute for restricted operational records. Synthetic data statistically mimics real operations without exposing the original data. If generated with sufficient fidelity, it can support AI training, privacy-preserving collaboration, rare-event augmentation and fast-time ATM simulation.

The project brought together SINTEF Digital, Delft University of Technology, Deepblue and EUROCONTROL. Over three years, the consortium evaluated nine generative AI models across six use cases, including delay forecasting, turnaround time prediction, traffic generation, diversion prediction, passenger flow prediction and schedule completion.

The objective was to provide the ATM community with an evidence-based assessment of where synthetic data is promising, where it remains immature, and what is needed before operational uptake.

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

SynthAIr worked on tabular flight records and four-dimensional flight trajectories. A major achievement was the development of two evaluation frameworks. For tabular data, F·U·P measures Fidelity, Utility and Privacy: statistical realism, downstream AI performance and protection against exposure of original records. For trajectories, F·D·R adds Diversity, measuring whether generated trajectories cover the full range of operational patterns.

For tabular data, four models were evaluated on 1.74 million European flight records: Gaussian Copula, CTGAN, TabSyn and REaLTabFormer. REaLTabFormer performed best, reaching 97% of real-data predictive utility for turnaround time forecasting, a Kolmogorov–Smirnov complement score of 0.991 and a detection indistinguishability score of 0.847. Models trained on synthetic data identified the same operational predictors as real-data models, with feature-importance alignment of 0.99. Privacy evaluation showed overfitting protection above 0.99. A complementary US experiment showed that TVAE-generated synthetic data could outperform real data in a predictive task.

For trajectories, SynthAIr compared TCVAE and TimeVQVAE. TimeVQVAE combines Short-Time Fourier Transform decomposition, vector-quantised codebooks and a bidirectional masked transformer prior. It achieved a Fréchet Inception Distance 274 times lower than TCVAE and an Autocorrelation Difference 51 times better, showing more accurate reproduction of temporal flight dynamics.

The project also conducted a cross-airport transfer learning study. A model pretrained on Zurich Airport data was fine-tuned with Dublin Airport data. Fine-tuning with only 5% of the target data, about one week of landing operations, outperformed training from scratch on the full Dublin dataset. This was replicated across four architectures and reduced data requirements by up to 80%.

Finally, trained generative encoders were reused as analytics tools. Their latent representations identified five operational communities, seven Dublin approach patterns and ten representative Heathrow arrivals. The project produced seven peer-reviewed publications, with open code and synthetic datasets.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

SynthAIr advanced the state of the art in four directions. First, it provided the first systematic evaluation of deep generative models for aviation tabular data across multiple architectures, use cases and a unified framework. REaLTabFormer and TabSyn reached 95–97% of real-data predictive utility across turnaround time, departure delay and arrival delay forecasting. The 0.99 feature-importance alignment shows that synthetic data preserved the operational drivers used by predictive models.

Second, SynthAIr moved trajectory generation beyond simple geometric similarity. TimeVQVAE showed that generative models can reproduce temporal dynamics of real flight paths, which is essential for ATM simulation fidelity.

Third, the transfer learning study provided a pathway for data-poor airports. Smaller and regional airports often lack enough historical data to train robust AI models independently. SynthAIr showed that a pretrained model from a data-rich airport can be adapted using only a small amount of local data. This suggests a scalable European model in which large hubs support pretraining and smaller airports benefit through fine-tuning.

Fourth, SynthAIr showed that generative models are also analytical instruments. Their embeddings can support anomaly detection, route characterisation, approach clustering and representative trajectory extraction without manual labelling.

The potential impacts are significant but conditional. Synthetic data can support training dataset augmentation, privacy-preserving collaboration and rare-event-rich simulation, with longer-term relevance for safer operations, more efficient airspace design and lower barriers to AI adoption. It can also reduce the imbalance between data-rich organisations and smaller aviation actors without direct pooling of sensitive records.

Operational uptake requires further work. Current models learn statistical regularities but do not fully enforce aircraft performance envelopes, weather effects or domain constraints. Rare-event modelling remains difficult because the most safety-relevant events are least represented in historical data. Transfer learning must be tested across more airports and operational environments. Stronger privacy validation, including membership inference testing and differential privacy, is also needed.

SynthAIr therefore does not deliver an operational product. It delivers a credible research pathway toward trustworthy, privacy-aware and scalable synthetic data use in European ATM.