Network Architecture: Open RAN, AI-Native, NTN

Despite the compelling vision, O-RAN faces several open challenges:

• Fronthaul latency and jitter: The split 7.2x fronthaul must deliver IQ samples with sub-millisecond latency and tight jitter bounds. Ethernet-based transport (eCPRI) requires Precision Time Protocol (PTP) synchronisation to $< 65$ ns — difficult over multi-hop switched networks.

• Performance gap: Current O-RAN deployments show 5 -- 15% throughput degradation compared to integrated (single-vendor) solutions, primarily due to suboptimal cross-layer coordination across open interfaces.

• RIC real-time constraints: The near-RT RIC must ingest telemetry, run inference, and push control actions within 10 ms. This requires co-locating ML inference engines with the O-DU, raising questions about edge compute placement and cost.

• Security: Open interfaces expand the attack surface. The fronthaul carries raw IQ data, and compromise of the RIC could allow adversarial control of scheduling and beamforming.

In 5G, machine learning is applied as an overlay: the protocol stack is designed with model-based algorithms, and ML is used to tune parameters or replace specific blocks (e.g., CSI compression with autoencoders). The 6G vision goes further — an AI-native air interface where ML is integral to the standard from the outset.

Key research directions include:

1. Neural receivers: End-to-end learned detection that jointly handles channel estimation, equalisation, demapping, and decoding, potentially bypassing the conventional block-by-block processing chain. Early results show 1 -- 3 dB gains in highly non-linear channels (e.g., with low-resolution ADCs or strong phase noise).

2. Learned CSI feedback: At FR3/sub-THz frequencies with massive arrays, the CSI dimensionality explodes. Encoder-decoder architectures (e.g., CsiNet) can compress $N_t \times N_{\mathrm{sub}}$ channel matrices to a few hundred bits while preserving 90 -- 95% of the achievable rate.

3. Foundation models for wireless: Large pretrained models (analogous to LLMs) that capture wireless channel statistics across environments, frequencies, and mobility patterns, then fine-tuned for specific deployment scenarios.

4. On-device training: Federated learning at the network edge, where UEs collaboratively train shared models without exchanging raw data — connecting to the over-the-air computation framework of Section 33.5.

Embedding AI into the air interface standard raises fundamental questions that the research community is actively debating:

• Standardisation of ML models: Should 3GPP standardise the model architecture (e.g., "the encoder shall be a 4-layer CNN with dimensions ..."), or only the interface (input/output dimensions, quantisation, latency constraints)?

• Reproducibility and interoperability: If the Tx uses one vendor's neural encoder and the Rx uses another vendor's decoder, will they interoperate? This is the "neural codec" interoperability problem.

• Robustness and worst-case guarantees: ML models trained on data from environment A may fail catastrophically in environment B (distribution shift). Safety-critical applications (V2X, URLLC) may require provable performance bounds that current ML cannot provide.

• Computational cost: Running transformer-based neural receivers at sub-millisecond latency on a mobile chipset is beyond current hardware. Model compression, quantisation-aware training, and hardware-software co-design are active research areas.

Integrating NTN into 6G poses unique challenges beyond the link budget:

• Delay and Doppler: LEO round-trip delays of 4 -- 25 ms break the HARQ timing assumptions of terrestrial NR. 3GPP NTN uses disabling of HARQ feedback or extended HARQ timers, sacrificing latency for coverage. 6G aims for protocol designs that natively accommodate variable and large delays.

• Handover storm: With LEO satellite footprints sweeping the ground at $\sim$7 km/s, a stationary UE experiences frequent handovers. Predictive handover using satellite ephemeris data (known hours in advance) is essential.

• Inter-system interference: NTN and terrestrial systems sharing spectrum (e.g., at S-band or L-band) must manage inter-system interference through coordination zones, power control, and beamforming null placement.

• Mega-constellation scalability: Routing traffic across thousands of interconnected LEO satellites with inter-satellite links (ISLs) requires distributed, low-latency routing algorithms — a space networking problem with no terrestrial analogue.

The three architectural pillars — disaggregation (Open RAN), intelligence (AI-native), and expanded coverage (NTN) — converge toward a 6G network that is far more than a faster data pipe. The ITU-R IMT-2030 usage scenarios reflect this convergence:

• Immersive communication: XR, holographic telepresence (peak rates $> 100$ Gbps, latency $< 1$ ms).
• Hyper-reliable low-latency: Industrial control, remote surgery ($10^{-7}$ reliability, $0.1$ ms latency).
• Massive communication: $10^7$+ devices/km² for IoT.
• Ubiquitous connectivity: Global coverage via NTN.
• Integrated AI and communication: Network-as-a-service for distributed ML training and inference.
• Integrated sensing and communication (ISAC): Using the wireless signal for environment sensing (covered in Section 33.6).

The remaining sections of this chapter dive into specific enabling technologies: near-field XL-MIMO, full-duplex, over-the-air computation, and digital twins / ISAC.