AI-Enhanced O-RAN

• AI-enhanced O-RAN is a network architecture that integrates AI/ML into disaggregated RAN modules with open interfaces to enable dynamic, real-time control.
• It employs advanced techniques such as reinforcement learning and supervised models for precise resource allocation, network slicing, and anomaly detection.
• Deployment testbeds using containerized xApps and standardized interfaces validate significant improvements in throughput, cost efficiency, and adaptive service provisioning.

AI-enhanced Open Radio Access Network (O-RAN) refers to the integration of advanced AI and ML techniques into the architecture and operational workflows of Open RAN systems. O-RAN fundamentally transforms wireless Radio Access Network deployments by disaggregating legacy base stations into virtualized, modular components (RU/DU/CU), connecting them via open interfaces (E2, A1, O1), and orchestrating them using layered RAN Intelligent Controllers (RICs). By embedding AI-driven applications (xApps/rApps) at multiple control levels, O-RAN aims to achieve real-time, adaptive, and autonomous resource management, improved Quality of Experience (QoE), rapid service agility, and operational efficiency across diverse network scenarios (Kouchaki et al., 2022, Abdalla et al., 2021, Baena et al., 21 Feb 2025, Tang et al., 15 Mar 2025, Zhao et al., 12 Jan 2025, Navidan et al., 15 Feb 2026, Masur et al., 2021, Polese et al., 2022, Bonati et al., 2022, Malakalapalli et al., 25 Apr 2025, Tang et al., 2022, Salama et al., 29 Jul 2025, Parada et al., 25 Feb 2025, Niknam et al., 2020, Shokouhi et al., 26 Feb 2026, Habibi et al., 2024, Feraudo et al., 3 Feb 2026, Shah et al., 12 Jul 2025).

1. O-RAN Architecture and Functional Decomposition

O-RAN re-engineers RAN infrastructure by decoupling legacy base stations into standardized virtual network functions:

• Radio Unit (RU): Physical-layer (PHY) components for RF, FFT/IFFT, and digital beamforming.
• Distributed Unit (DU): MAC, RLC, HARQ; responsible for lower-PHY functions and resource scheduling.
• Centralized Unit (CU): SDAP, PDCP, RRC; supports both control-plane and user-plane separation.
• RAN Intelligent Controllers (RICs):
  • Non-Real-Time RIC (Non-RT RIC): Resides within the Service Management and Orchestration (SMO); responsible for policy management, AI/ML model training, and long-term analytics (>1 s timescale).
  • Near-Real-Time RIC (Near-RT RIC): Edge-deployed, runs xApps for resource control and optimization in 10 ms–1 s loops.
• Open Interfaces: E2 (Near-RT RIC ↔ DU/CU), A1 (Non-RT RIC → Near-RT RIC), O1 (Non-RT RIC → all nodes).

This disaggregated design, coupled with containerized deployment and open interfaces, enables modular, vendor-neutral AI integration, supporting closed-loop control and facilitating interoperability across heterogeneous deployments (Kouchaki et al., 2022, Abdalla et al., 2021, Polese et al., 2022, Feraudo et al., 3 Feb 2026).

2. AI/ML Workflows and Control Loops

AI-enhanced O-RAN features a multi-scale closed-loop pipeline for data-driven optimization:

• Data Collection: Continuous extraction of RAN KPIs (e.g., used/available PRBs, UE count, CQI, throughput) via E2 Indication or O1 reporting (Kouchaki et al., 2022, Habibi et al., 2024).
• Model Training and Deployment:
  • Offline Training (Non-RT RIC): Supervised, unsupervised, or reinforcement-learning models trained on historical KPI logs; policy frameworks formalize optimization goals (e.g., log-utility throughput maximization, fairness constraints, energy efficiency) (Tang et al., 15 Mar 2025, Zhao et al., 12 Jan 2025, Shokouhi et al., 26 Feb 2026).
  • Distribution (A1): Trained model artifacts are provisioned to Near-RT RIC for inference.
• Online Inference and Control (Near-RT RIC): xApps subscribe to live telemetry, run AI models for scheduling, slicing, or mobility management, and issue control directives (E2 Control) within strict latency budgets (<100 ms for TTI-scale loops) (Kouchaki et al., 2022, Bonati et al., 2022).
• Feedback and Policy Adaptation: Post-execution KPIs are fed back for continuous retraining, policy tuning, or anomaly detection; service assurance is closed via A1 policy revisions and runtime monitoring (Abdalla et al., 2021, Tang et al., 15 Mar 2025, Navidan et al., 15 Feb 2026).

This architecture enables real-time, fine-grained resource control, e.g., dynamic PRB allocations and adaptive slicing under fluctuating traffic and mobility patterns (Zhao et al., 12 Jan 2025).

3. AI/ML Methodologies and Optimization Formulations

AI-enhanced O-RAN leverages a diverse set of modeling techniques, subject to architectural and operational constraints:

• Reinforcement Learning (RL): Employed at both fine (per-TTI) and coarse (strategic) time scales. Core formulations:
  • MDP state $s_t$ includes RAN KPIs such as MCS, resource requests, per-UE fairness.
  • Action $a_t$ is typically resource assignment (e.g., scheduling blocks, PRB splits).
  • Reward $r(s_t, a_t)$ captures QoE, throughput, and fairness, often structured as:

  $$r_{t}^u = R_{t}^u(C_t, f_t) \cdot \left(\min_u f_t^u / \max_u f_t^u\right)$$ - Algorithms: Advantage Actor-Critic (A2C), Proximal Policy Optimization (PPO), Bayesian learning with GP surrogates and ADMM coordination for resource slicing (Kouchaki et al., 2022, Zhao et al., 12 Jan 2025, Abdalla et al., 2021, Bonati et al., 2022).

• Supervised/Regression Models: For tasks such as energy consumption prediction, LSTM-based traffic/latency forecasting, and anomaly detection. Ensemble regressors (Random Forest, Gradient Boosting, XGBoost) are evaluated with MSE metrics (Malakalapalli et al., 25 Apr 2025, Parada et al., 25 Feb 2025, Tang et al., 15 Mar 2025).

• Explainable AI (XAI): SHAP and LIME-based explanation modules analyze feature attributions (e.g., airtime, buffer status report, goodput) for model transparency and energy efficiency (Malakalapalli et al., 25 Apr 2025).

• Agentic and Hierarchical AI: Large and Small LLM (LLM/SLM) agents interpret operator intents, generate policies, orchestrate xApp lifecycle, and coordinate with physical-layer foundation models (WPFM) for low-latency tasks (Navidan et al., 15 Feb 2026, Shokouhi et al., 26 Feb 2026, Salama et al., 29 Jul 2025).

• Edge Adaptive and Continual Learning: Continual updating of slice controllers and traffic predictors (e.g., AdaOrch, LSTMs) supports adaptation under non-stationary and high-mobility regimes (Zhao et al., 12 Jan 2025, Salama et al., 29 Jul 2025).

The architectural integration of these models is enabled by portable frameworks (e.g., xDevSM, OpenRAN Gym), which expose high-level SDKs, KPM-streaming, and unified control abstractions (Feraudo et al., 3 Feb 2026, Bonati et al., 2022).

4. Deployment, Interoperability, and Real-World Validation

Multiple research testbeds and frameworks support experimentation and real-world deployment of AI-enhanced O-RAN:

• Testbeds: OpenAirInterface, FlexRIC, srsRAN, Colosseum, PAWR, Arena, USRP B210-based SDRs (Kouchaki et al., 2022, Bonati et al., 2022, Zhao et al., 12 Jan 2025).

• Containerization and Orchestration: xApps/rApps packaged as Docker containers; deployment via Kubernetes Pod specs; hot-swapped on Near-RT RIC clusters (Kouchaki et al., 2022, Tang et al., 15 Mar 2025).

• Interoperability: Cross-stack support achieved via normalized service models (E2SM-KPM/RC), ASN.1-encoding, and generic SDK bindings; xDevSM provides an abstraction layer across OAI, srsRAN, ARC-OTA (Feraudo et al., 3 Feb 2026).

• Performance Metrics and Benchmarks:

  • Throughput, PRB utilization, normalized reward, convergence time (TTIs to policy saturation), latency, and energy prediction MSE.
  • Demonstrated improvements: up to 40% throughput gain with DRL xApps, 64.2% cost reduction and 45.5% performance boost with AdaSlicing, and sub-10 ms inference times on commodity hardware (Kouchaki et al., 2022, Zhao et al., 12 Jan 2025, Bonati et al., 2022).
• Edge and Distributed Architectures: Persona-based Edge Agentic frameworks and hierarchical agent organization allow robust, zero-outage operation with strict timing. Space-O-RAN extends these principles to non-terrestrial networks, partitioning AI pipelines between onboard dApps, cluster-level SPACERICs, and terrestrial SMO (Salama et al., 29 Jul 2025, Baena et al., 21 Feb 2025).

5. Security, Robustness, and Explainability

The open, programmable nature of O-RAN increases the attack surface, particularly for AI-driven controllers:

• Security Threats: Misconfiguration, interface spoofing, poisoning attacks on AI models/rApps/xApps, adversarial input manipulation, and model extraction (Abdalla et al., 2021, Rahman et al., 2021).
• Countermeasures:
  • Mutual authentication (TLS, X.509), zero-trust networking, secure enclave-based key management.
  • Robust ML training: adversarial training, randomized smoothing, and blockchain-anchored data/model provenance.
  • Container-level secure attestation (e.g., TPM quotes) at deployment.
• Testing and Validation Frameworks: Distributed, automated, AI-driven test frameworks exercise AI models in both simulation and hardware. Techniques include sensitivity analysis, fuzzing, adversarial generation, and RL-based exploration of decision spaces. Standard metrics: throughput, latency, loss ratio, action success rates, and security robustness (Tang et al., 2022).
• XAI Integration: Visualization of feature contributions (SHAP, LIME) to promote interpretability of AI operator decisions, particularly for energy efficiency and anomaly detection (Malakalapalli et al., 25 Apr 2025).

6. Research Challenges and Future Directions

AI-enhanced O-RAN faces significant open challenges, especially as the vision advances toward 6G:

• Extreme Latency/Real-Time Control: While 10–100 ms Near-RT RIC loops suffice for resource scheduling, sub-ms PHY control remains an open challenge; emerging solutions involve RT RIC/zApps and hardware acceleration (FPGA/GPU, DPUs) (Abdalla et al., 2021, Navidan et al., 15 Feb 2026).
• Robust Multi-Timescale Orchestration: Coordinating cross-layer, multi-timescale controllers (Non-RT/AI for strategic policies, Near-RT for tactical, RT for PHY) and resolving conflicts between independently developed xApps/rApps (Navidan et al., 15 Feb 2026).
• Scalability and Model Lifecycle: Federated and continual learning, efficient policy/ML transfer, standardized MLOps pipelines for safe, explainable, and adaptive operation in large-scale, multi-vendor environments (Habibi et al., 2024, Polese et al., 2022).
• Security and Trust: End-to-end, provable security for AI pipelines, including post-quantum hardening of A1/E2, secure supply chain, and runtime attestation (Abdalla et al., 2021, Rahman et al., 2021).
• Testbeds and Digital Twins: Reproducible, scalable testbed support (digital twins, federated field trials) for validation of autonomous, AI-augmented O-RAN protocols and service models (Abdalla et al., 2021, Bonati et al., 2022, Tang et al., 2022).
• Non-Terrestrial Extensions: O-RAN architectural principles adapted to LEO constellations (Space-O-RAN), including interface mapping (dynamic aggregation of E2/A1/O1 over ISL/feeder/GSL links), onboard lightweight dApps, and digital-twin–assisted SMO orchestration (Baena et al., 21 Feb 2025).

7. Representative AI-Driven Use Cases and Performance Outcomes

• Resource Allocation and Scheduling: DRL xApps (A2C/PPO) in Near-RT RIC outperform heuristics by up to 30% in convergence speed and show lower variance in reward and throughput (Kouchaki et al., 2022).
• Adaptive Network Slicing: AdaSlicing with Bayesian learning agents and ADMM achieves 45.5% normalized performance improvement and 64.2% cost reduction on a live testbed (Zhao et al., 12 Jan 2025).
• Intent-Driven Edge AI Service Provisioning: LLM-powered rApps automate service deployment with measured end-to-end provisioning under 3 minutes and 98.7% QoS compliance for sub-50 ms latency (Tang et al., 15 Mar 2025).
• Cell-Free O-RAN Optimization: Agentic LLM-based multi-agent frameworks in cell-free architectures yield 41.93% fewer active O-RUs and 92% memory savings via QLoRA-adapter sharing, compared to naive or separate LLM agent baselines (Shokouhi et al., 26 Feb 2026).
• Energy Efficiency and Explainability: XAI-integrated models (SHAP, LIME) inform policy tuning and achieve 8–10% instantaneous energy savings while reducing model MSE over naive baselines by over 40% (Malakalapalli et al., 25 Apr 2025).
• Edge Agentic AI: A persona-based edge agent framework achieves zero measured network outages under high-stress compared to 8.4% (fixed) and 3.3% (LLM-only) outages, maintaining CV(SINR)<0.15 under peak load (Salama et al., 29 Jul 2025).

---

The integration of advanced AI/ML algorithms into the open, modular O-RAN architecture enables adaptive, scalable radio resource management, accelerates network service innovation, and offers a path toward fully autonomous, ultra-reliable, and efficient 6G RAN operations. Continued advances in AI model security, hardware acceleration, orchestration, and cross-domain adaptability remain key enablers and open challenges in the evolution of AI-enhanced O-RAN.