Network Digital Twins

• Network Digital Twins are real-time digital replicas of communication networks that mirror both physical and logical components for simulation and optimization.
• They integrate live telemetry with predictive models using modular, multi-domain architectures to support testing, control, and planning in diverse network environments.
• Applications span from O-RAN and 6G network optimization to urban modeling, demonstrating measurable improvements in delay reduction, efficiency, and resource scaling.

Network Digital Twins (NDTs) are virtual or digital replicas of communication networks that are designed to mirror relevant physical-network properties closely enough to support simulation, testing, optimization, monitoring, prediction, and “what-if” analysis without applying risky changes directly to production systems. Across the literature, NDTs are consistently distinguished from static offline models by their intended coupling to live telemetry and evolving operational state, although the degree of real-time synchronization, control closure, and formal validation varies widely across implementations (Almasan et al., 2022). In adjacent work, the term “Digital Network Twin” (DNT) is used for the same broad idea, especially in next-generation wireless contexts, with emphasis on life-cycle concerns such as creation, adaptation, deployment cost, and security (Liu et al., 2024). The topic has expanded from conceptual O-RAN and 6G visions toward concrete predictive, control-oriented, graph-based, and multi-domain implementations, but the field remains heterogeneous in scope, fidelity, and maturity (Mirzaei et al., 2023).

1. Conceptual scope and distinctions

NDTs are typically defined as real-time or near-real-time digital representations of communication networks that combine telemetry, structural information, and predictive or analytical models to estimate network behavior under current or hypothetical conditions. In O-RAN, an NDT is presented as a real-time digital replica of the physical network that mirrors both physical and logical components—including infrastructure, protocols, algorithms, and operating conditions—and remains connected to the live system through continuous data exchange (Mirzaei et al., 2023). In fixed-network work, the NDT is framed as a virtual model of the communication network that takes network state descriptions such as topology, traffic matrix, routing, and scheduling as input and produces outputs such as utilization, delay, or anomalies in real time (Almasan et al., 2022). In next-generation wireless surveys, the same construct is described as a real-time virtual mapping of a physical network that is built by mapping physical network objects and their internal and external properties into a digital environment and then self-evolving that environment as those properties change (Liu et al., 2024).

A persistent distinction in the literature is between NDTs and neighboring tools. Simulators are described as useful but computationally expensive and dependent on simplifying assumptions; emulators can reproduce specific components but become costly or limited for large systems; monitoring systems observe the live network but do not answer counterfactual questions; orchestration systems act on the network but do not by themselves provide a safe predictive replica (Mirzaei et al., 2023). This supports a common interpretation of NDTs as a predictive, testable, and decision-support layer rather than merely a dashboard or offline simulator. A related conceptual distinction is between “networked twins,” meaning digital twins of external physical systems that depend on communication networks, and “twins of networks,” meaning digital twins whose physical twin is the network itself (Ahmadi et al., 2021).

The literature also varies in how broad a state the twin is expected to capture. Some works model a narrow state space, such as traffic trajectories or per-flow KPIs, while others argue for holistic twins spanning topology, traffic, radio conditions, application sensitivity, and control state. This suggests that “NDT” is best treated as a family of purpose-specific architectures rather than a single canonical system. A plausible implication is that fidelity should be judged relative to the operational question being asked, not by an unattainable goal of complete network duplication (Liu et al., 2024).

2. Architectural patterns and life-cycle organization

Several recurrent architectural patterns appear across the literature. A common baseline separates the physical network, the twin domain, and the data-exchange path between them. In O-RAN, this appears explicitly as a three-pillar structure consisting of the Physical Twin (PT), the NDT, and southbound real-time data flow between them, with northbound communication between the NDT and the RIC domain (Mirzaei et al., 2023). In more general NDT formulations, the twin sits at the center of a loop that collects network state, evaluates candidate configurations, and supports optimization or control decisions before those decisions are applied back to the physical system (Almasan et al., 2022).

A second recurring pattern is life-cycle decomposition into creation, synchronization, adaptation, and operational use. Survey-style work on DNTs distinguishes “vertical mapping,” which builds the initial twin from physical-network and environment information, from “horizontal mapping,” which updates or reuses existing twins as network conditions evolve (Liu et al., 2024). In automatic twin-generation work, the emphasis is narrower: the twin is a data-driven “Unit Twin” generated from network behavior data through AutoML so that it can enter validation pipelines with much lower human effort than manual twin construction (Ding et al., 3 Oct 2025). In robot-centric and vehicular settings, the life cycle is explicitly online and incremental: a robot or vehicle-side pipeline updates the twin as new geometry, position, or traffic observations arrive, and the twin becomes progressively more accurate as exploration or sensing continues (Sanchez et al., 4 Feb 2025).

A third pattern is modularization. Rather than assuming a monolithic end-to-end twin, recent work advocates composition of multiple domain- or task-specific twins. The “Unified Twin Transformation” (UTT) framework treats transfer, merging, and splitting of task-oriented NDTs as first-class operations, allowing a trajectory twin, position twin, and tracking twin to be transformed into each other through joint multimodal encoding, fusion, and distributed model aggregation (Zhang et al., 2 Sep 2025). In multi-domain 6G management, NDT capabilities are exposed as a specialized service domain, and a DT Orchestrator dynamically selects and composes domain-specific DT modules and simulators in response to predictive or prescriptive “what-if” queries (Buzcu et al., 11 Jun 2026). This supports the view that large-scale NDT systems may be ecosystems of coordinated twins rather than single global models.

Deployment placement is also treated as an architectural variable. O-RAN work presents edge deployment as suitable for delay-sensitive functions such as real-time monitoring and fault detection, while cloud deployment is more appropriate for network-wide planning and optimization (Mirzaei et al., 2023). In adaptive vehicular NDTs, a hierarchical cloud-edge split is used: cloud-side calibration periodically benchmarks fidelity choices, and the edge executes the strict real-time predictive loop every transmission interval (Makvandi et al., 21 May 2026). This suggests that NDT architecture is increasingly shaped by latency budgets and by the computational asymmetry between calibration and runtime control.

3. Modeling paradigms and algorithmic formulations

NDT modeling spans several distinct paradigms. One major class treats the twin as a learned surrogate mapping from network state to KPIs. In fixed IP networks, RouteNet-E predicts end-to-end delay from topology, traffic matrix, and routing policy, functioning as a graph-based surrogate for packet-level simulation (Almasan et al., 2022). PLAN-Net extends this idea by integrating path, link, and node embeddings so that the network configuration can be mapped to per-flow KPIs—delay, jitter, throughput, and drops—in a single forward pass (Li et al., 2023). M3Net further broadens the predictive scope to multiple per-flow metrics and uses a mixture-of-experts readout on top of a GNN backbone, reporting delay MAPE reduction from 20.06% to 17.39% and classification accuracies of 66.47% for jitter and 78.7% for packets dropped on a real testbed dataset (Guda et al., 10 Dec 2025).

A second class emphasizes graph-native representations for topology, state, and application semantics. An AI-NDT built on a multi-layered knowledge graph uses three coordinated graphs: the Network Topology Knowledge Graph (NTKG), the Network State Knowledge Graph (NSKG), and the Application State Knowledge Graph (ASKG). The NTKG captures structure such as router/switch interconnections, degree, centrality, and connectivity score; the NSKG stores RTT, jitter, packet loss, and bandwidth utilization; and the ASKG links these to QoE-sensitive application classes such as web browsing, video streaming, VoIP, gaming, and file transfer (Zacarias et al., 4 Aug 2025). On a graph with 989 nodes and 908,752 edges derived from RIPE Atlas data, this work compares GraphSAGE, ChebNet, ResGatedGCN, and GraphTransformer and reports GraphTransformer as the strongest model with $R^2 = 0.9763$ and MAE $= 0.0750$ (Zacarias et al., 4 Aug 2025).

A third class incorporates explicit control-theoretic synchronization. In real-time traffic synchronization for mobile/heterogeneous networks, the twin predicts traffic $\hat{T}(t)$ from a windowed observation and computes synchronization error

$e(t) = T_{\text{actual}(t)} - \hat{T}(t),$

followed by PID correction

$$u(t) = K_p e(t) + K_i \int e(t)\,dt + K_d \frac{de(t)}{dt}.$$

This creates a closed loop of PT measurement, DT prediction, error computation, correction, DT update, and visualization (Sengendo et al., 23 Oct 2025). The practical contribution is clear, but the paper also shows a recurring limitation in the field: the “adaptive PID” label is stronger than the formal specification provided, because no explicit self-tuning law or stability proof is given (Sengendo et al., 23 Oct 2025).

A fourth paradigm is hybrid-system modeling of wireless network state. In a multi-cell 5G testbed, the NDT is defined as

$\tilde N:(t,x)\mapsto(Q,c),$

where $Q(t,x)\in\mathbb{R}^l$ is a communication-quality vector and $c(t,x)\in\{1,\dots,m\}$ is the serving-cell identity (Mavridis et al., 31 Oct 2025). The twin is modeled as a hybrid system whose modes correspond to base stations and whose sub-modes correspond to regions $\Sigma_j(t)$ of the workspace with similar network behavior. Online deterministic annealing is then used to learn region prototypes, local communication-quality values, and class assignments from online user measurements (Mavridis et al., 31 Oct 2025). This is notable because it yields a compact, explainable NDT rather than a monolithic black-box regressor.

A fifth modeling direction is site-specific geometry-aware twins. For LoS/NLoS identification, a 3D digital twin of the urban area of Milan is combined with Sionna ray tracing to generate labeled channel data for supervised learning, with the NDT serving as a high-fidelity digital representation of the physical environment (Zhu et al., 21 May 2025). For 5G-enabled mobile robots, the twin is built from ROS/ROS2 data, SLAM, OctoMap, Open3D, and Wireless InSite ray tracing to infer radio quality over a progressively discovered indoor environment (Sanchez et al., 4 Feb 2025). In adaptive vehicular control, AdaPTwin combines trajectory prediction, updated 3D scenes, multi-fidelity ray tracing, and beamforming/association optimization under a strict deadline (Makvandi et al., 21 May 2026). These works treat geometry and propagation not as peripheral features but as primary state variables of the NDT.

4. Synchronization, validation, and trustworthiness

Synchronization is a defining property of NDTs, but the literature shows that it is also one of the least mature. The broad requirement is continuous alignment between the physical twin and digital representation through telemetry ingestion, state updating, and model refresh (Mirzaei et al., 2023). In practice, synchronization mechanisms range from simple model retraining or online correction to more elaborate event-triggered adaptation. The hybrid wireless NDT explicitly uses regression-error and classification-error triggers to “heat” the annealing process when the twin no longer matches observed network conditions, and it adds temporary mode-specific corrections for localized cell malfunction without retraining the whole model (Mavridis et al., 31 Oct 2025). In traffic-synchronization work, the operational loop is online, but the adaptation law for controller gains remains qualitative (Sengendo et al., 23 Oct 2025). This suggests that many current NDTs are operationally adaptive without yet being formally adaptive in the control-theoretic sense.

Validation practices are equally varied. Some predictive twins are validated primarily against simulation outputs or held-out testbed traces using accuracy metrics such as MAPE, MAE, RMSE, $R^2$, or F1-score (Li et al., 2023). Others emphasize visual or qualitative runtime tracking (Sengendo et al., 23 Oct 2025). A more recent development is explicit trustworthiness validation for counterfactual “what-if” analysis. In cloud-edge resource management, a 6G-TWIN-based framework extends telemetry collection and semantic harmonization and then evaluates predictive models not only by regression accuracy but also by directional correctness under regime changes. The paper introduces Sign Agreement ($= 0.0750$0) and Directional Sensitivity as validation criteria and shows that although both DNN and XGBoost achieve $= 0.0750$1, XGBoost is much more trustworthy for proactive scaling because it achieves $= 0.0750$2 while the DNN underreacts to perturbations (Agudelo et al., 16 Apr 2026). This is important because it demonstrates that high interpolation accuracy is insufficient when the twin is used for action selection.

Trustworthiness also depends on model and measurement quality. A power-network digital twin built on an operational UK microgrid shows that imperfect telemetry semantics, sparse instrumentation, wrong transformer tap assumptions, and low observability can make a twin analytically elegant but operationally unreliable (Deakin et al., 2023). This is a strong reminder that NDT quality is not only a modeling problem but also a data-assimilation problem. A plausible implication is that many communication-network NDTs will face analogous issues when telemetry semantics, sensor placement, or hidden state are mismatched to the twin’s estimation objectives.

Backward maintenance is emerging as another aspect of trustworthiness. Network digital untwinning introduces targeted removal of deprecated contributions from a distributed NDT so that the resulting model is $= 0.0750$3-indistinguishable from a scratch-rebuilt twin. The framework uses connectivity-aware untwinning sets, rollback checkpoint selection, Gaussian perturbation, and remapping, with both single-request and parallel-request versions (Zhang et al., 30 Apr 2026). This extends the NDT life cycle beyond creation and synchronization to include deletion, privacy, compliance, and model integrity.

5. Applications and operational roles

The applications of NDTs now span planning, validation, optimization, observability, and cross-domain orchestration. In O-RAN, NDTs are proposed for prior-deployment validation, post-deployment optimization, interoperability testing, traffic steering, energy efficiency, anomaly detection, and support for AI/ML training (Mirzaei et al., 2023). In fixed networks, they act as fast surrogates for simulator-in-the-loop routing optimization (Almasan et al., 2022). In testing pipelines, automatically generated twins can serve as efficient validation tools aligned with the ITU-T experimentation subsystem, reducing end-to-end testing time from over 900 hours to about 3.4 hours in a simple Mininet use case (Ding et al., 3 Oct 2025).

Radio and wireless applications are especially prominent. Geometry-aware NDTs support LoS/NLoS classification, where NDT-generated data improve classification accuracy by 5% in very low SNR conditions and by approximately 10% in medium-to-high SNR scenarios while reducing inference FLOPs by 98.55% relative to the state-of-the-art SegNet-based solution (Zhu et al., 21 May 2025). For mobile robots, an NDT built from robot exploration traces and ray-traced propagation maps supports radio-aware navigation, with route comparisons showing stronger received-signal distributions than shortest-path navigation (Sanchez et al., 4 Feb 2025). For vehicular networks, AdaPTwin uses predictive ray tracing and adaptive fidelity to support proactive beamforming and association, achieving up to 90% sum-rate gain and 80% outage-probability reduction compared to non-adaptive NDTs while maintaining real-time performance (Makvandi et al., 21 May 2026).

Network management and observability applications are also expanding. In B5G core networks, a GFT-MPNN operating across real and NDT domains classifies 16 failure types with domain-level validation accuracies of 94.27% on the NDT domain, 97.83% on real-network training, and 98.05% on real-network testing (Isah et al., 2024). In graph-based Internet-scale prediction, an AI-NDT predicts RTT and packet loss from RIPE Atlas measurements, linking network-state estimation to QoE-aware management (Zacarias et al., 4 Aug 2025). In cloud-edge orchestration, trustworthy what-if NDTs are used for proactive resource scaling under unseen high-load regimes (Agudelo et al., 16 Apr 2026).

Cross-domain and sustainability-oriented uses indicate a further expansion of scope. A modular 6G architecture composes an O-RAN DT module with a two-stage solar-allocation simulator and, over a 105-base-station deployment in Poznan, identifies 17 base stations as both coverage-active and high-priority solar candidates. The resulting joint optimization reduces daily grid consumption by 28.5% with 32 solar panels at the diminishing-returns threshold (Buzcu et al., 11 Jun 2026). This suggests that future NDTs may increasingly function as cross-domain analytical mediators, coupling telecom state to energy, compute, mobility, or urban-context models.

6. Open challenges and research directions

Despite rapid progress, the literature repeatedly identifies unresolved issues in fidelity, scalability, standardization, uncertainty, and operational integration. Standardization is still immature: O-RAN and 3GPP provide adjacent interface and testing specifications, but a fully standardized O-RAN-native NDT architecture remains open (Mirzaei et al., 2023). Survey work on DNTs likewise emphasizes that the full life cycle—especially fine-grained creation, real-time adaptation, resource-efficient deployment, and security protection—has not yet been comprehensively solved (Liu et al., 2024).

Scalability remains a recurring limitation. Many demonstrations are small or medium scale: six-host traffic emulators, 989-node graph datasets, 5–8-node router testbeds, robot-scale environments, or two-cell 5G trials. Even when claims of scalability are promising, they are often demonstrated only at one scale rather than supported by asymptotic studies (Zacarias et al., 4 Aug 2025). A plausible implication is that large-scale carrier-grade NDTs will require hierarchical decomposition, selective fidelity, modular orchestration, and aggressive abstraction rather than direct extension of single-domain prototypes.

Generalization and transfer are also unsettled. Several papers explicitly note uncertainty about whether results transfer across network domains, topologies, or traffic regimes (Zacarias et al., 4 Aug 2025). Task-oriented twin transfer, merging, and splitting have been proposed as one answer to this reuse problem (Zhang et al., 2 Sep 2025), but the robustness of these operations under strong domain shift remains open. Similarly, predictive twins trained on synthetic or emulated data often improve practical deployment readiness, yet full sim-to-field validation is still limited (Zhu et al., 21 May 2025).

Trust, privacy, and governance are becoming more central as twins absorb richer telemetry and move closer to orchestration loops. Work on digital untwinning highlights deletion and compliance. Work on distributed DTs over Named Data Networking points to freshness, consistency, naming, and stale-cache risks in data-centric dissemination architectures (Chen et al., 7 May 2025). Conceptual 6G discussions also raise ownership, smart contracts, and federated DT coordination as open governance problems (Ahmadi et al., 2021). This suggests that trustworthy NDT deployment will require not only better prediction and synchronization but also explicit lifecycle governance.

Finally, the field still lacks a universally accepted criterion for “full-fledged” NDT capability. Some works focus on KPI prediction, others on synchronization, others on safe experimentation, others on cross-domain decision support. The cumulative literature suggests that a mature NDT will likely combine several properties: synchronized telemetry ingestion, purpose-specific fidelity, predictive and prescriptive what-if analysis, explicit validation beyond raw accuracy, modular composition across domains, and a clearly bounded role in human-supervised or automated control (Agudelo et al., 16 Apr 2026). The current research trajectory indicates movement toward such systems, but also makes clear that NDTs remain an active systems problem rather than a finished architectural commodity.