Native AI Integration in 6G PHY

AI gets embedded directly into transceivers - neural receivers, learned modulation, AI channel coding built into the hardware itself.

Replace traditional blocks?No. Hybrid approach works best - use AI for complex stuff like interference handling and channel estimation, keep proven methods for basic processing like FFT.

Optimal balance Depends on use case. Critical applications lean conventional with AI backup. Dynamic scenarios go AI-first.

Reliability across conditions

• Meta-learning for quick adaptation
• Fallback to conventional when AI confidence drops
• Distributed learning across network deployments
• Robust training with diverse channel scenarios

The trick is modular designs that switch between AI and conventional processing based on real-time conditions and reliability needs.

Dear Murtadha Shukur ,

Establishing an AI-native air interface for 6G networks means replacing blocks in the signal processing chain on the physical layer with trained machine learning models.

https://www.rohde-schwarz.com/us/solutions/wireless-communications-testing/wireless-standards/6g/ai-and-ml-for-6g-networks/ai-and-ml-for-6g-networks_257029.html#:~:text=Establishing%20an%20AI%2Dnative%20air,with%20trained%20machine%20learning%20models.

Regards,

Shafagat

How can AI be natively integrated into the physical layer in 6G? One novel approach is to treat AI models themselves as structured signal processors, not just learning agents. In my recent preprint [1], I introduced Spectral Complexity Theory, where logical problems like SAT are embedded in frequency space using Fourier and Walsh transforms. This revealed latent geometric separability in their structure — allowing for predictive hardness analysis before solving. This idea can extend to AI integration in 6G by modeling neural network architectures spectrally, and matching their complexity to the spectral characteristics of the wireless channel. Instead of blindly deploying a static AI block, we analyze the channel’s deformation spectrum in real time, and select or morph the AI model based on spectral compatibility. This allows AI to dynamically adapt to the communication environment — as a signal-matching agent. Rather than replacing traditional signal processing entirely, hybrid spectral strategies could be deployed: Classical DSP handles stable, low-curvature signal conditions. AI models engage when spectral curvature or entropy exceeds thresholds — signaling higher complexity environments. This spectral alignment paradigm provides both robustness and interpretability, potentially allowing us to unify AI and physical-layer processing under a complexity geometry framework.

Reference

[1] Elsherif, Osama. (2025). Spectral Geometry of Boolean Satisfiability: A Fourier-Walsh Perspective on Complexity Landscapes. Preprint on ResearchGate