AirFM-DDA: Air-Interface Foundation Model in the Delay-Doppler-Angle Domain for AI-Native 6G

Researchers introduce AirFM-DDA, a foundation model for 6G wireless networks that processes channel state information in the Delay-Doppler-Angle domain rather than traditional space-time-frequency representations. The model uses window-based attention instead of computationally expensive global attention, achieving superior generalization on channel prediction tasks while reducing computational costs by an order of magnitude.

AirFM-DDA represents a significant advancement in applying foundation model architecture to wireless communications infrastructure. The core innovation lies in reparameterizing channel state information into a domain that naturally separates multipath components along physically meaningful axes—delay, Doppler shift, and angle of arrival. This structural clarity enables more effective machine learning compared to existing approaches that operate on entangled space-time-frequency representations where distinct signal paths become superimposed and difficult to distinguish.

The broader context reflects the wireless industry's shift toward AI-native 6G design, where foundation models could replace traditional signal processing algorithms. Previous wireless foundation models relied on global attention mechanisms, which scale quadratically with input size and become prohibitively expensive for real-time network applications. AirFM-DDA's window-based attention mechanism exploits the natural clustering of multipath dependencies, reducing computational overhead while maintaining performance.

For network operators and equipment manufacturers, this development offers tangible benefits: faster model training, lower inference latency, and reduced power consumption in base stations. The model's robustness across extreme conditions—high mobility, large delay spreads, severe noise, and aliasing—suggests practical deployment potential without requiring scenario-specific retraining. This generalization capability particularly matters for 6G rollout, where infrastructure must handle diverse real-world conditions efficiently.

The competitive advantage accrues to organizations that can integrate these models into network-on-chip implementations. As 6G standards crystalize around AI-native principles, foundation models operating in physically meaningful domains may become essential infrastructure, similar to how transformers became foundational to modern software systems.

• →AirFM-DDA operates in Delay-Doppler-Angle domain to explicitly separate multipath components, improving channel representation learning
• →Window-based attention reduces computational costs by ~10x compared to global attention while maintaining performance
• →Model demonstrates strong zero-shot generalization across unseen scenarios and datasets without retraining
• →Maintains robustness under extreme conditions including high mobility, severe noise, and extreme aliasing
• →Represents shift toward AI-native 6G infrastructure where foundation models replace traditional signal processing