DRIFT: Joint Channel Estimation and Prediction Towards Pilotless 6G Non-Terrestrial Networks

Researchers propose DRIFT, a lightweight AI framework for channel estimation and prediction in 6G non-terrestrial networks that reduces pilot overhead by up to 12% while requiring minimal computational resources suitable for satellite implementation. The approach uses data-driven processing after initial pilots, achieving significant spectral efficiency gains with fewer than 200k multiply-accumulate operations.

DRIFT addresses a fundamental engineering challenge in next-generation satellite communications: balancing accuracy with computational feasibility. As 6G networks expand to include non-terrestrial components like LEO satellites, traditional pilot-based channel estimation wastes valuable spectrum through repetitive overhead signaling. The research demonstrates that intelligent prediction models can maintain channel accuracy while reducing this waste, freeing spectrum for actual data transmission.

The motivation stems from LEO satellite constraints where power budgets directly limit onboard processing capability. Previous AI-based predictors often require millions of operations per inference cycle, making them impractical for satellites with stringent thermal and power limitations. DRIFT's architecture—combining convolutional and LSTM variants—achieves competitive performance at roughly 1% of typical neural network complexity, making orbital deployment feasible.

For telecommunications infrastructure developers and satellite operators, this innovation reduces operational costs by improving spectrum utilization without requiring complete hardware redesigns. The 12% spectral efficiency gain translates directly to increased data capacity per orbital asset, affecting both consumer broadband expansion and enterprise connectivity pricing models. The framework's robustness across different channel models and training-test mismatches suggests practical deployment readiness beyond controlled laboratory conditions.

Looking forward, the satellite communications sector will likely adopt similar lightweight ML approaches as 6G standards emerge. This research establishes that computational constraints need not prevent intelligent signal processing in space systems, potentially accelerating satellite operator adoption of AI-driven optimization across link budgets, power management, and resource allocation.

• →DRIFT achieves 12% spectral efficiency gains in LEO satellite networks by eliminating redundant pilot transmissions after initial channel characterization
• →The framework operates with fewer than 200k multiply-accumulate operations, enabling deployment on power-constrained satellite hardware
• →Lightweight LSTM and convolutional architectures maintain prediction accuracy while reducing computational overhead compared to standard deep learning models
• →The approach demonstrates robustness across different channel models and training-test mismatches, suggesting practical deployment viability
• →Joint channel estimation and prediction reduces pilot overhead specifically for 6G non-terrestrial networks, improving spectrum utilization efficiency