AI-Enhanced O-RAN Integration

• AI-enhanced O-RAN is a paradigm that embeds AI/ML in open, modular RAN architectures to enable automated, data-driven network control and optimization.
• It leverages multi-layer RAN Intelligent Controllers and open interfaces to perform dynamic resource allocation, network slicing, and QoE optimization across 5G, beyond-5G, and emerging 6G systems.
• The architecture emphasizes robust security, explainable AI, energy-aware orchestration, and rigorous testing to ensure scalability, real-time responsiveness, and operational resiliency.

The AI-enhanced Open Radio Access Network (O-RAN) paradigm refers to the deep integration of AI and ML into the open, modular, and disaggregated O‑RAN architecture. This integration transforms how radio access networks are controlled, managed, and optimized, leveraging open interfaces and virtualized components to enable programmable, data-driven automation, operational efficiency, robust security, and dynamic adaptability for 5G, beyond-5G, and emerging 6G wireless systems. AI-enhanced O-RAN supports closed-loop intelligence at multiple layers and timescales, introduces new testing and orchestration requirements, and enables new research in data-driven, secure, and explainable networking.

1. Architectural Integration of AI in O-RAN

AI’s role in O‑RAN is foundational, not peripheral: the O‑RAN architecture is designed specifically to support AI-driven control and management across the disaggregated RAN stack (Abdalla et al., 2021, Polese et al., 2022). Two key architectural elements enable this:

• RAN Intelligent Controllers (RICs):
  • Non–Real-Time (non-RT) RIC (part of SMO): Handles high-level orchestration, policy management, and long-term training of AI/ML models (e.g., rApps, with control timescales >1s).
  • Near–Real-Time (near-RT) RIC: Deployed at the edge, this controller executes latency-sensitive xApps (control loop latencies: 10ms–1s) using rapidly updated KPMs via the E2 interface for functions such as dynamic scheduling, slicing, beamforming, and traffic steering.
  • Proposed Real-Time (RT) RIC: For sub-millisecond control, especially at PHY, a further “RT RIC” may be co-located with O-DU/RU, executing zApps based on lightweight AI (e.g., echo state networks) and requiring hardware acceleration (Abdalla et al., 2021).
• Data and Interface Infrastructure:

The architecture’s open interfaces (E2, A1, O1, O2) provide standardized access to cross-layer metrics (including MAC/PHY data) for data-driven learning, closed-loop execution, and policy updates (Polese et al., 2022).

• AI Workflow:

A canonical AI/ML pipeline encompasses: 1. Data collection through E2/O1/A1 interfaces, 2. Feature engineering (e.g., normalization, dimensionality reduction [autoencoders]), 3. Offline training (on SMO/non-RT RIC) employing, e.g., deep reinforcement learning (DRL), with TD-loss

$$L(\theta) = \mathbb{E}_{(s,a,r,s')}[(r + \gamma\,\max_{a'}Q(s',a';\theta) - Q(s,a;\theta))^2],$$

1. Validation and publication to a model catalog,
2. Deployment as containerized xApps/rApps,
3. Online inference and adaptation,
4. Continuous monitoring and retraining (Polese et al., 2022).

This granular, hierarchical integration supports coordinated intelligence across multiple control layers and time domains.

2. Core AI-Driven Use Cases and Functionalities

AI-enhanced O‑RAN enables advanced, dynamic RAN functionalities (Abdalla et al., 2021, Polese et al., 2022, Bonati et al., 2022):

• Dynamic Resource Allocation: ML models predict demand and instantaneously adjust scheduling, RBG assignment, or MAC parameters.
• Network Slicing: xApps implement DRL policies to allocate/adjust slice resources in real time for eMBB, URLLC, and mMTC traffic; models train on observed KPIs and are validated offline before deployment (Bonati et al., 2022).
• QoE Optimization: AI/ML models leverage KPM streams to infer per-UE experience and dynamically reconfigure handover, beamforming, or rate control settings.
• Traffic Steering: Short-term demand prediction allows proactive routing/assignment of flows.
• Spectrum Sharing: AI-driven marketplace models (DDQN agents) enable operators to trade spectrum resources dynamically, with reward models such as $R = - (\alpha D + \beta S + \gamma P + T),$ where $D$ is resource deficit, $S$ surplus, $P$ monetary cost, $T$ transaction cost (Rasti et al., 19 Feb 2025).
• Security and Anomaly Detection: Continuous monitoring and cross-domain AI supports early detection of attacks by training RAN-side classifiers with high-fidelity transport network labels (Xavier et al., 17 Jan 2024).

3. Performance, Scalability, and Testing of AI-Driven Systems

• Latency Constraints:

Deterministic latency is essential, especially for 6G; dynamic functional splits (FS) adapt placement of RAN functions based on fine-grained, feedback-driven timing models, e.g., $R_W \geq T_W + \delta,$ where $R_W$ is the reception window, $T_W$ the transmission window, and $\delta$ transport variation (Abdalla et al., 2021). Dynamic scaling using latency-aware controllers (e.g., ScalO‑RAN’s QCQP formulations (Maxenti et al., 2023)) ensures that AI inference delays meet control timing:

$$t^{\mathrm{inf}}_{a,s}(y) = t^{\mathrm{exec}}_{a,s}(y) + t^{\mathrm{queue}}_{a,s}(y)$$

with constraints on aggregate inference time per app.

• Testing and Verification:

Rigorous validation of AI-based functions is critical due to potential network destabilization (e.g., conflicting/erroneous AI decisions). Automated, specification-aware frameworks (e.g., AI5GTest: Gen-LLM, Val-LLM, Debug-LLM pipeline (Ganiyu et al., 11 Jun 2025)) and distributed, AI-enabled testbeds (AI-enhanced fuzzing, RL-based exploration, adversarial stress-testing (Tang et al., 2022)) are essential for compliance and security evaluation.

Component	Function	Typical Timescale
Non-RT RIC (rApps)	Policy/orchestration/training	>1 s
Near-RT RIC (xApps)	Real-time closed-loop control	10 ms–1 s
RT RIC (zApps)	Physical layer, sub-ms reaction	<1 ms

4. Security, Trust, and Explainability

• Security Risks and Mitigation:
  • Mutual authentication,
  • Dynamic cryptography with secure key management,
  • Adoption of a zero-trust architecture,
  • Real-time, AI-driven anomaly detection and adaptive security policies (research direction).
• Explainable AI (XAI) Needs:
  • Attribute AI decisions to input features/KPIs,
  • Quantify model confidence, robustness, and consistency,
  • Resolve multi-vendor policy conflicts and drive intent-based decision making in resource allocation.
• Automation and Standards:

Support for MLOps/DevOps-style pipelines, continuous validation, and the CI/CD automation of both models and their explanation layers is required to maintain system robustness and facilitate adaptivity under dynamic channel/network conditions.

5. Energy Efficiency, Scalability, and Resource Management

• Energy-Aware Orchestration:
  • Latency-aware scaling (ScalO‑RAN),
  • Lossless migration (CORMO-RAN via stateful or shared data layer migration with explicit optimization of downtime vs. resource overhead).
• AI-RAN Convergence:
  • Soft Actor-Critic RL (SAC) agents are used for predictive, adaptive resource allocation, balancing RAN/AI task completion rates:

  $$R_t = w_\text{RAN}(t) \tfrac{C_\text{RAN}(t)}{d_\text{RAN}(t)} + w_\text{AI}(t)\tfrac{C_\text{AI}(t)}{d_\text{AI}(t)} - \lambda(1-\tfrac{r_\text{RAN}(t)+r_\text{AI}(t)}{R_\text{max}})$$

The orchestrator uses Y1 interface for real-time radio analytics and integrates anomaly detection and LSTM-based forecasting to achieve >99% RAN demand fulfiLLMent in dynamic traffic conditions (Shah et al., 12 Jul 2025).

6. Emerging Domains and Future Research

• Spatio-Temporal Spectrum Markets:

Multi-granularity spectrum marketplaces within O‑RAN can be operated using discriminative and generative AI, allowing dynamic, fine-grained spectrum trading among operators (Rasti et al., 19 Feb 2025).

• Non-Terrestrial O-RAN:

The extension of O‑RAN into satellite and non-terrestrial domains (Space-O-RAN) is enabled via hierarchical, closed-loop control where AI-driven dApps on satellites provide real-time, in-orbit decision making for scheduling, beam management, and inter-satellite coordination, with periodic policy and model updates from terrestrial digital twins (Baena et al., 21 Feb 2025).

• Autonomous Operation and Resilience:

Edge agentic AI frameworks leverage persona-based, multi-tools architectures and context-aware, multimodal data fusion (including external event/weather intelligence) to preempt outages and ensure operational safety, achieving reported 0% outage rates even under high-stress conditions (Salama et al., 29 Jul 2025).

• Hardware Acceleration:

Designs such as HeartStream demonstrate custom AI/baseband shared-memory clusters (64 RISC-V cores, 410 GFLOP/s, 204.8GBps L1 bandwidth) with systolic execution, providing energy and compute density to support demanding AI-driven PHY and MAC functions under B5G/6G constraints (Zhang et al., 10 Sep 2025).

Research Theme	Architectural/AI Innovation
AI-Driven Closed-Loop Control	Multi-layer RIC (non-RT, near-RT, RT), xApps/rApps
Proactive Resource Allocation	RL-based orchestration, predictive LSTM modules
Secure, Explainable Networking	XAI overlays, AI-driven anomaly/security detection
Energy-Aware Compute Scaling	Latency-aware optimization, lossless xApp migration
Non-Terrestrial/Spectrum Sharing	Dynamic interfaces/mapping, AI spectrum broker

7. Challenges and Directions

• Trade-offs:

Achieving the right balance among real-time control, security, explainability, and energy efficiency requires fine-grained coordination between architectural, AI/ML, and operational domains.

• Standards and Interoperability:

Ongoing standardization and best-practice development for MLOps pipelines, XAI interfaces, and secure, zero-trust frameworks are foundational for deploying robust, vendor-agnostic, and future-proof AI-enhanced O‑RAN networks (Abdalla et al., 2021, Brik et al., 2023).

• Testing and Validation:

Automated, LLM-powered platforms are now essential for validating conformance, interoperability, and functional robustness at the scale and heterogeneity inherent to O‑RAN deployments (Ganiyu et al., 11 Jun 2025).

In sum, AI-enhanced O-RAN represents the confluence of open, modular RAN principles with multi-layer, closed-loop, data- and learning-driven control, offering adaptive, intelligent, and secure wireless systems architecture foundational for operators and industry as the sector advances toward 6G (Abdalla et al., 2021, Polese et al., 2022, Polese et al., 9 Jul 2025, Shah et al., 12 Jul 2025).