Generative Models for Aviation Data
The SynthAIr project has developed a comprehensive suite of generative models specifically designed for Air Traffic Management (ATM) applications. Our models address two fundamental data types in aviation: tabular flight records and time-series trajectory data.
Tabular Data Models
Five complementary approaches for generating synthetic flight records:
- REaLTabFormer: Transformer-based autoregressive generation achieving 94-97% utility retention
- TabSyn: Diffusion models in latent space for efficient mixed-type synthesis
- CTGAN: Conditional adversarial training optimized for rare events and imbalanced data
- TVAE: Variational autoencoders providing stable training with minimal computational requirements
- Gaussian Copula: Statistical approach offering strongest privacy protection
Time Series Models
Specialized architectures for aircraft trajectory generation:
- TimeVQVAE: Time-frequency domain processing with transformer priors for global coherence
- TCVAE with VampPrior: Temporal convolutional networks with flexible prior distributions
- TimeGAN: Adversarial training preserving temporal relationships and sequential patterns
- Flow Matching: Continuous normalizing flows for deterministic, efficient trajectory sampling
- Diffusion Models: Denoising diffusion in latent space for high-fidelity spatiotemporal generation