Problem

The transition to Sixth Generation (6G) mobile networks demands exceptionally stringent key performance indicators (KPIs) such as ultra-reliable low-latency communications (URLLC), enhanced mobile broadband (eMBB), and massive machine-type communications (mMTC). Traditional reactive network management struggles to meet these demands, especially with the use of high-frequency bands like Millimeter Wave (mmWave) and Terahertz (THz) spectrums and denser network topologies, which are highly susceptible to environmental blockages. The Digital Twin Network (DTN) paradigm has emerged as a promising foundational technology to enable proactive and deterministic network orchestration. However, the existing literature predominantly explores DTNs conceptually, lacking a formal classification of architectures and a comprehensive evaluation of their technical and computational feasibility, leaving a significant gap in understanding their practical implementation and scalability.

Method

This survey addresses the existing gaps by formally categorizing state-of-the-art DTN architectures into two primary types: passive monitoring twins and active control twins. It provides an in-depth evaluation of the key enabling technologies essential for DTNs, which include Ray-Tracing (RT) for high-fidelity 3D environmental modeling and channel simulation, Reconfigurable Intelligent Surfaces (RIS) for dynamically manipulating the electromagnetic environment, Artificial Intelligence (AI) (specifically Deep Reinforcement Learning, Supervised Deep Learning, and Federated Learning) for autonomous orchestration and predictive capabilities, and Mobile Edge Computing (MEC) for localized, low-latency processing. A crucial aspect of the methodology involves a detailed mathematical and computational complexity analysis of state-of-the-art solutions, assessing hardware scalability and inference bottlenecks. To facilitate fair comparisons across diverse system assumptions, the paper introduces a normalized classification framework that evaluates each architecture based on its latency class, memory requirements, hardware dependence, and scalability trends.

Results

The analysis of passive monitoring twins highlights that high-fidelity electromagnetic (EM) reconstruction using Ray-Tracing (RT) is feasible, but CPU-based RT faces significant computational bottlenecks, particularly with large 3D models and real-time convolutions. Solutions often involve shifting computational load to GPU-accelerated neural network (NN) inference, which is then bounded by GPU memory. For AI-driven channel prediction, managing model drift is critical, with methods like transfer learning and Age of Information (AoI)-triggered retraining proving effective. However, increasing fidelity (e.g., input resolution, beam counts) often introduces performance ceilings due to mounting computational overhead. For active control twins focused on edge computing and task offloading, Deep Reinforcement Learning (DRL) is widely used, but its scalability is highly sensitive to the action space dimensions, necessitating complex hybrid DRL architectures. RIS provides substantial benefits in resource-constrained environments, yet its utility diminishes beyond a saturation point or when spectrum resources are abundant. Dynamic RIS deployment, such as on UAVs, consistently outperforms static placements. In real-time mobility tracking, centralized multi-agent DRL for vehicular networks demonstrates linear scaling, while decentralized neuromorphic computing for pedestrians shows promising low-latency, low-energy, and sub-linear scaling, though current neuromorphic hardware remains immature. Across both twin types, a consistent trade-off emerges between emulation fidelity, computational complexity, and capital expenditures (CAPEX). Scalability often mandates segmenting large digital twins and addressing hardware limitations, with model drift and real-time synchronization posing persistent challenges.

Implications

Digital Twin Networks (DTNs) are crucial for overcoming the physical electromagnetic challenges and resource constraints that hinder 6G service deployment across various domains, including smart cities, industries, healthcare, smart grids, and the unified Internet of Everything (IoE). For smart cities, DTNs enable predictive planning and blockage mitigation; for industries, they facilitate predictive maintenance and robust wireless connectivity; in healthcare, they optimize infrastructure and support human digital twins; and for smart grids, they provide intelligent load balancing and dynamic aerial infrastructure. Ultimately, for the IoE, DTNs serve as central orchestrators, resolving cross-domain resource disputes. The widespread deployment of DTNs will necessitate fluid, non-terrestrial networks, leveraging satellites and UAVs for on-demand resource provisioning. Cybersecurity is paramount, requiring a shift to zero-trust, decentralized architectures that integrate Federated Learning (FL) and blockchain technologies to mitigate state desynchronization attacks and data poisoning, thereby ensuring public safety and data integrity. Future research directions should focus on advancements such as active RIS, Integrated Sensing and Communication (ISAC), energy-efficient neuromorphic hardware, predictive actuation to overcome latency, Explainable AI (XAI) for transparency, semantic communications for data efficiency, and quantum computing for intractable optimization problems. Furthermore, establishing a universally standardized, multi-dimensional taxonomy for DTNs is essential for comprehensive evaluation and future development.