Machine learning moves to the heart of 6G networks in new systematic review

A systematic review published in Discover Networks has mapped how machine learning is shifting from an optional optimization tool in wireless networks to a foundational design element of sixth-generation (6G) communication systems. The review, conducted under the PRISMA 2020 framework, screened 5,635 records from major scientific databases published between January 2020 and April 2026 and retained 81 studies that met predefined quality criteria for detailed analysis. From this synthesis, the authors propose a new conceptual framework, the Integrated AI-Native 6G Communication Architecture (IACA), intended to organize the technologies, services and trade-offs that will shape intelligent wireless networks.

The review’s central observation is an evolution in how artificial intelligence is applied to communications. Early studies relied on supervised, unsupervised and reinforcement learning to solve isolated tasks such as channel estimation, resource allocation, traffic prediction and spectrum management. As networks grew more complex, deep learning architectures including convolutional neural networks, recurrent networks, graph neural networks and deep reinforcement learning enabled more adaptive optimization across diverse wireless environments. The most recent work, the authors report, moves toward AI-native systems in which learning, inference and decision-making are embedded within the network architecture itself rather than bolted on as external optimization modules.

Reinforcement learning and deep reinforcement learning emerged as the dominant approaches for autonomous radio resource management, learning adaptive allocation policies directly from network interactions. According to the reviewed studies, these methods achieve better throughput, spectrum utilization and latency than conventional optimization by continuously adapting without requiring explicit mathematical models. Supervised learning contributes through traffic prediction and demand forecasting, allowing resources to be allocated before congestion degrades performance. Yet the synthesis also flags limitations: deep reinforcement learning remains computationally intensive and prone to convergence instability, which constrains its practicality in large-scale real-time deployments.

In the physical layer, deep learning has become a powerful alternative for beamforming optimization and channel estimation in millimeter-wave and terahertz systems, where severe path loss and rapid channel fluctuations limit conventional analytical techniques. The review finds that different architectures serve complementary roles: convolutional networks excel at spatial feature extraction for beamforming and signal classification, recurrent and long short-term memory networks model temporal channel variations and mobility, and transformer-based architectures capture long-range contextual dependencies for beam prediction and channel state estimation. No single architecture is universally superior, however, and the authors emphasize that choices must weigh computation efficiency, latency and deployment requirements alongside accuracy.

Several emerging 6G technologies feature prominently in the synthesis. Reconfigurable intelligent surfaces, which control electromagnetic wave propagation through programmable reflections, benefit from machine learning methods that learn optimal reflection patterns directly from communication data, improving coverage, throughput and energy efficiency in challenging propagation scenarios. Semantic communication, which aims to transmit meaningful information rather than raw bit streams, is enhanced by deep learning, transformers and large language models that support context reasoning and knowledge-driven information exchange. Integrated sensing and communication frameworks combine wireless transmission with environmental perception, and machine learning improves their joint optimization of sensing and communication resources. The review is careful to note that semantic communication remains a complementary paradigm rather than a replacement for existing communication frameworks, with challenges in semantic synchronization, interoperability and standardization unresolved.

Federated learning and edge intelligence receive particular attention as architectural paradigms for distributed intelligence. Federated learning allows distributed devices to collaboratively train models without sending raw data to central servers, supporting privacy protection in applications such as healthcare systems, industrial internet of things deployments and autonomous vehicles. Edge intelligence moves inference and decision-making closer to end-user devices, reducing latency for applications like the tactile internet and immersive environments. The review nonetheless identifies significant trade-offs: federated learning faces heterogeneous data distributions, communication overhead, synchronization instability and model poisoning attacks, while edge intelligence is constrained by limited computational resources, energy availability and the complexity of coordinating distributed devices.

Security emerges as a double-edged theme. Machine learning strengthens intrusion detection, anomaly detection, adaptive authentication, trust evaluation and physical-layer security, with deep learning models outperforming rule-based methods in detecting distributed denial of service attacks, spoofing, jamming and unauthorized access in large heterogeneous networks. At the same time, AI-driven security systems themselves face threats from adversarial machine learning, model poisoning, data manipulation and privacy leakage, particularly in federated setups. The authors conclude that improving detection accuracy alone is insufficient without ensuring the robustness, explainability and trustworthiness of the security mechanisms themselves.

Sustainability presents a notable contradiction in the literature. While machine learning improves network efficiency through intelligent resource scheduling, adaptive power allocation, traffic-aware networking and sleep-mode management, large-scale deep learning models, transformers and foundation models carry substantial computational and energy costs of their own. The review states that the energy required by AI models can partially cancel out the communication efficiency gains they enable, suggesting that communication intelligence should be evaluated not only by network performance but also by its overall computational and environmental footprint.

The proposed IACA framework organizes these findings into four hierarchical layers: communication infrastructure, machine learning intelligence, AI-native communication services and intelligent applications, supported by cross-cutting trust and sustainability requirements. The communication infrastructure layer encompasses reconfigurable intelligent surfaces, terahertz communications, satellite and non-terrestrial networks, UAV-assisted networking, edge computing and massive IoT. The intelligence layer combines supervised, unsupervised and reinforcement learning with deep architectures, federated learning and foundation models. The services layer embeds intelligence into core functions such as resource management, beamforming, semantic communication, sensing and security, while the applications layer covers smart cities, intelligent transportation, connected healthcare, extended reality and large-scale IoT ecosystems. The authors position IACA as a technology-agnostic reference abstraction that complements rather than substitutes standardized ITU and 3GPP architectures.

Beyond the architecture, the review identifies deployment trade-offs that will shape practical realization: intelligence versus computational complexity, autonomy versus trustworthiness, performance versus sustainability, heterogeneity versus interoperability, and adaptability versus scalability. It closes with a research agenda calling for continual and lifelong learning, self-supervised and online learning, scalable federated intelligence, multimodal foundation models, explainable and adversarially robust AI, and privacy-preserving machine learning. Because the synthesis is qualitative rather than a meta-analysis, given the diversity of datasets and evaluation metrics across the included studies, its conclusions rest on narrative synthesis of the 81 retained studies. The authors state that no datasets were generated or analysed during the study, and that the framework is intended to guide future research toward autonomous, trustworthy, scalable and sustainable AI-native 6G communication ecosystems.

Subject of Research: A systematic review of machine learning technologies and architectures for AI-native 6G communication systems

Article Title: A systematic review of machine learning technologies and intelligent architectures toward AI-native 6G communication ecosystems

Article References: Favanza Kwaha, I., & Bulus, B. (2026). A systematic review of machine learning technologies and intelligent architectures toward AI-native 6G communication ecosystems. Discover Networks, 2(1), Article 42. https://doi.org/10.1007/s44354-026-00056-9

Image Credits: AI Generated

DOI: 10.1007/s44354-026-00056-9

Keywords: 6G, machine learning, artificial intelligence, systematic review, PRISMA, federated learning, semantic communication, ISAC, reconfigurable intelligent surfaces, edge intelligence, energy efficiency, network architecture