Digital Cellular Twins

• Digital cellular twins are digital replicas of cellular networks that integrate physical and network-level models for real-time simulation and decision-making.
• They combine models of users, base stations, channels, and environments with data-driven synchronization and hybrid physics-data calibration to mirror network behavior.
• They enable AI-based optimization for beam selection, energy efficiency, and resource allocation, supporting enhanced planning and adaptive control in 5G/6G networks.

Searching arXiv for papers on digital twins in cellular/mobile networks. Digital cellular twins are continuously updated, high-fidelity digital replicas of cellular networks and their radio environments, constructed to mirror the behavior of physical network elements and to support monitoring, simulation, prediction, optimization, and control across the network lifecycle. In the mobile-network literature, this concept appears in closely related forms: the mobile network digital twin (MNDT) as a software system that mirrors real 4G/5G/6G networks and uses AI to optimize configurations (Li et al., 2023); real-time digital twins of the physical wireless environment and channels built from precise 3D maps and multi-modal sensing (Alkhateeb et al., 2023); and the digital twin channel (DTC) as the channel-level component that maps radio channels and associated physical-layer operations from the physical world to the digital world in real time (Wang et al., 2024). Taken together, these formulations define digital cellular twins as multi-layer, data-driven, and bi-directionally integrated systems that combine physical-level and network-level twinning to inform communication and sensing decisions across 5G/6G networks (Alkhateeb et al., 2023).

1. Definition and conceptual scope

Digital cellular twins specialize the general digital twin concept to cellular infrastructure, users, radio channels, and network services. A digital twin is a virtual replica of a physical system that is continuously synchronized with real data and used for monitoring, simulation, and optimization over the system lifecycle; in mobile networks, this becomes a mobile network digital twin (MNDT) that mirrors cells, users, radio channels, traffic, and KPIs in a high-fidelity virtual environment (Li et al., 2023). This network-level view emphasizes entities such as mobile users, base stations, wireless environments, and network performance metrics including throughput, outage, energy, QoS, and QoE (Li et al., 2023).

A complementary formulation focuses on the physical wireless environment. Real-time digital twins are described as continuously updated digital replicas of physical wireless environments, built from precise 3D maps and multi-modal sensing, and used to make operational communication and sensing decisions (Alkhateeb et al., 2023). This physical-level perspective covers buildings, streets, vehicles, pedestrians, material properties, antenna locations, and instantaneous or statistical channel information, including blockage status and path parameters (Alkhateeb et al., 2023). The literature explicitly distinguishes this physical-level twin from network-level twins that focus on traffic, services, edge computing, and cybersecurity; a digital cellular twin, in the broad sense, is their union (Alkhateeb et al., 2023).

At the radio layer, the digital twin channel (DTC) is the foundational component of a full digital cellular twin. DTC is defined as a real-time mapping of radio channels and associated physical-layer communication operations from the physical world to the digital world (Wang et al., 2024). This suggests that digital cellular twins are layered systems in which channel-level twinning supports base-station-level, UE-level, network-level, and service-level twins by providing predictive information about link quality, interference, and coverage (Wang et al., 2024).

The concept is also framed at a larger architectural scale as a digital network twin (DNT) or digital twin network (DTN), namely a live, virtual representation of an entire communication network that continuously updates from real-time data, predicts future dynamics, and informs decision-making (Liu et al., 2024). In this usage, digital cellular twins may represent a cell, a cluster of cells, an end-to-end private 5G network, or a full 5G/6G infrastructure spanning RAN, core, and edge (Costa et al., 14 Oct 2025). A plausible implication is that the term denotes not a single model class but a family of coupled replicas operating at different granularities and timescales.

2. Architectural organization and synchronization

A recurrent architectural pattern is the separation between physical network, digital replica, and control or application logic. The MNDT architecture for 6G is organized around two feedback loops: an inner loop joining simulation and optimization, and an outer loop joining physical and virtual systems (Li et al., 2023). In the inner loop, a digital mirror of network elements is simulated, an AI-based optimizer proposes configurations, the simulation engine evaluates those actions, and the optimizer is iteratively trained from the resulting performance metrics (Li et al., 2023). In the outer loop, real network data at time $t$ updates the mirror model, simulation parameters are calibrated to reduce the gap between simulated and real performance, and discrepancies observed at time $t+1$ refine both models and optimizers (Li et al., 2023).

The real-time DTN literature uses a related three-layer decomposition: physical network layer, twin layer, and network application layer (Lin et al., 2022). The twin layer includes a data domain for collection and storage, a model domain for basic and functional models, and a management domain for lifecycle and security management (Lin et al., 2022). Another formulation, aligned with ISO 23247 and ITU-T Y.3090 concepts, distinguishes an observable domain, a data collection and device control domain, a core domain, and a user domain in the implementation of a network digital twin for a 5G private network (Costa et al., 14 Oct 2025).

In more operational settings, synchronization is implemented through periodic ingestion of configuration and telemetry. In the proof-of-concept for a 5G private network, configuration data from the physical network are converted into structured JSON and YAML for the twin, while packet captures are replayed to reproduce traffic behavior; synchronization delay is described as approximately $T + t_{\text{comm}}$, where $T$ is the capture duration and $t_{\text{comm}}$ is transfer and processing time (Costa et al., 14 Oct 2025). The same work discusses Twin Alignment Ratio, State Consistency Index, and data freshness as conceptual metrics for synchronization quality (Costa et al., 14 Oct 2025). This suggests that temporal coherence between physical and digital domains is a first-order systems problem, not merely a data-ingestion task.

Cellular-twin architectures also differ in where intelligence is placed. Real-time digital twins for 6G distinguish three levels: local information, individual decision; shared information, individual decision; and shared information, joint/cooperative decision (Alkhateeb et al., 2023). VH-Twin proposes a hierarchy of local BS twins, cluster twins, and a global twin, using vertical twinning for synchronous hierarchical initialization and horizontal twinning for asynchronous maintenance (Zhang et al., 2024). Tiny-Twin, by contrast, is a CPU-native, full-stack digital twin that integrates a complete 5G stack with per-UE channel emulation and near-real-time RIC integration on commodity CPUs (Mamaghani et al., 13 Jan 2026). These formulations collectively indicate that digital cellular twins may be centralized, hierarchical, distributed, or multi-fidelity depending on latency, scale, and fidelity requirements.

3. Core models: users, base stations, channels, and environments

Digital cellular twins rely on explicit models of mobile users, base stations, wireless environments, and network state. In MNDT, user twins capture both mobility and traffic demand (Li et al., 2023). Outdoor trajectories are generated from origin–destination schedules, routes are computed via the A* algorithm,

$$\text{route} = \text{A*}( \text{origin}, \text{destination}, \text{road graph} ),$$

and movement along routes evolves according to the Krauss model (Li et al., 2023). Indoor movement is modeled with repulsive forces from obstacles, attractive forces toward destinations, and a physics-infused neural network to generate realistic crowd trajectories (Li et al., 2023). User traffic demand is generated by a multi-scale hierarchical generator that reflects mixtures of behavioral patterns, with instantaneous demand $D_i(t)$ used as a constraint for rate allocation (Li et al., 2023).

Base-station twins encode both static parameters and functional modules. In the MNDT formulation, each base station includes GPS location, carrier frequency $f$, maximum transmit power $P_{\text{max}}$, and antenna geometry and pattern, together with user association, scheduling, multiple access, and mobility management functions (Li et al., 2023). The achievable error-free transmission rate for a BS–UE pair is approximated by Shannon’s equation,

$C = B \log_2(1 + \text{SINR}),$

with SINR determined by channel modeling and power allocation (Li et al., 2023). These models support metrics such as BS sum rate, per-user rate, overall throughput, outage ratio, and power consumption, which later serve as rewards and costs in optimization (Li et al., 2023).

Channel and environment twinning are particularly central in 6G-oriented work. In real-time digital twins of wireless environments, a 3D map $t+1$0 encodes static geometry and materials, while dynamic objects are overlaid based on sensing data (Alkhateeb et al., 2023). A ray-tracing engine computes propagation paths with path lengths, delays, angles of departure and arrival, and complex gains; a generic channel model is written as

$t+1$1

or, in narrowband form,

$t+1$2

Path loss may be modeled as

$t+1$3

(Alkhateeb et al., 2023). In MNDT, overall path loss is expressed as

$t+1$4

with free-space loss, shadowing, and small-scale fading components; ray tracing combined with real map data is used to provide per-link path loss and hence SINR (Li et al., 2023).

The DTC literature systematizes channel twinning as a progression from analytic and statistical models to site-specific, real-time, and autonomous channel twins (Wang et al., 2024). At the highest level, autonomous DTC uses multi-modal sensing, “all data,” closed-loop interaction, and data-model dual-driven (DMDD) AI with a Radio Environment Knowledge Pool (REKP) (Wang et al., 2024). In a concrete site-specific implementation, a two-step channel twin first tunes a ray-traced environment model and then fine-tunes CSI using an AI model trained to map ray-traced CIRs to measured CIRs; the reported NMSE improves from $t+1$5 dB for the tuned ray-tracing twin to $t+1$6 dB for the AI-driven twin (Haider et al., 27 Jan 2025). This suggests that high-fidelity digital cellular twins increasingly rely on hybrid physics-plus-data calibration rather than pure analytical abstraction.

4. Optimization, control, and AI-enabled decision making

Digital cellular twins are not merely representational; they are optimization substrates. In MNDT, the simulation–optimization module is the “brain” of the twin (Li et al., 2023). At each time step, the simulator computes rates, throughput, energy consumption, and outage ratios from current UE positions, demands, BS configurations, and channel losses, and these become reward and cost signals for AI optimizers (Li et al., 2023). The RL state includes timestep, remaining user demands, traffic history, BS and cell status, and channel decay factors, while actions include user association, resource-block assignment, power levels, and cell on/sleep/off decisions (Li et al., 2023).

Two RL-based optimizers are described explicitly. The dynamical wireless network resource allocation (DWRA) problem is formulated as a constrained Markov decision process,

$t+1$7

with reward

$t+1$8

and a cumulative demand-satisfaction constraint (Li et al., 2023). The formulation includes constraints such as at-most-one-BS association per user, one RB per user, and per-BS power constraints (Li et al., 2023). The paper describes a safe RL method, for example Penalized Proximal Policy Optimization (P3O), together with bipartite matching via the Hungarian algorithm for association and RB assignment, and continuous-action power allocation with a DDPG-like method (Li et al., 2023).

The second optimizer addresses sleep mechanisms for energy saving through multi-agent RL. Each cell is controlled by an individual agent sharing a common policy and value network, and the reward combines negative total power consumption and negative switching cost (Li et al., 2023). In experiments using one week of Nanchang traffic traces, the RL method shows lower energy consumption than both the “all cells always on” baseline and a traffic-driven heuristic, while tracking traffic trends more closely (Li et al., 2023). The paper states that this implies better adaptation to temporal traffic variation and controlled switching behavior (Li et al., 2023).

At a broader level, DTC introduces a five-module closed loop: multi-modal data collection, data processing and modeling, channel prediction with DMDD AI, action decision, and real-time interaction and feedback (Wang et al., 2024). The predicted channel

$t+1$9

is mapped to control actions

$T + t_{\text{comm}}$0

for beam selection, MCS, RB assignment, handover, and power allocation (Wang et al., 2024). This layered action mapping makes clear that digital cellular twins can host both predictive models and explicit decision policies across PHY, MAC, and network layers.

What-if analysis is itself formalized in DTN-based wireless management. One framework uses a physical twin in NS-3, a cloud-hosted digital twin layer in Azure Digital Twins, and a service layer with ML and RL for optimizing carrier sensitivity threshold and transmit power control (Ak et al., 2024). Synthetic scenarios are generated with a conditional tabular GAN, and four normalized KPIs—throughput, latency, packet loss, and coverage—are aggregated into a scenario score and then an overall effectiveness score $T + t_{\text{comm}}$1 (Ak et al., 2024). The paper reports that integrating all scenarios via CTGAN yields about $T + t_{\text{comm}}$2 improvement in $T + t_{\text{comm}}$3 versus designing only for the current-behavior scenario, although with a $T + t_{\text{comm}}$4 wider confidence interval (Ak et al., 2024). This suggests that robust digital cellular twins are increasingly evaluated under distributions of possible futures rather than single nominal conditions.

5. Construction methodologies, data pipelines, and prototype systems

A major line of work addresses how digital cellular twins are built and maintained from real data rather than how they are used once available. VH-Twin is explicitly a mapping framework for wireless networks into digital reality, organized around Dynamic Connectivity Segmentation, vertical twinning, and horizontal twinning (Zhang et al., 2024). Each base station $T + t_{\text{comm}}$5 is associated with a time-series dataset

$T + t_{\text{comm}}$6

and the global twin-training objective is

$T + t_{\text{comm}}$7

with quadratic loss (Zhang et al., 2024). Base stations are clustered using geographic distance, backhaul similarity, coverage overlap, and traffic-distribution similarity, combined into a relationship metric

$T + t_{\text{comm}}$8

(Zhang et al., 2024).

Within each cluster, local twins are synchronously aggregated into cluster twins and then into a global twin: $T + t_{\text{comm}}$9 (Zhang et al., 2024). Horizontal twinning then updates the global twin asynchronously when cluster-level divergence exceeds a threshold $T$0 (Zhang et al., 2024). In experiments on Telecom Italia Milan traffic data, VH-Twin yields about $T$1 lower synchronization overhead in H-twinning compared with a single-level twin, while maintaining strong predictive performance (Zhang et al., 2024). This is a concrete example of cellular twinning as a distributed systems and federated learning problem.

Another implementation track uses open-source emulation to twin an actual private network. In the 5G private-network proof of concept, Amarisoft configuration is parsed and mapped into Open5GS and UERANSIM running inside ComNetsEmu, while live traffic is mirrored by replaying packet captures (Costa et al., 14 Oct 2025). Measurements on the physical infrastructure and the related digital twin are reported to demonstrate high accuracy in reproducing the state and behavior of the actual 5G system (Costa et al., 14 Oct 2025). The work frames the NDT as more than an emulator because it adds continuous data collection, dynamic mapping, and synchronization logic between NPT and NDT (Costa et al., 14 Oct 2025).

Tiny-Twin represents a different practical emphasis: full-stack, repeatable 5G experimentation on commodity CPUs (Mamaghani et al., 13 Jan 2026). It integrates time-varying multi-tap convolution

$T$2

with a complete 5G stack, plug-and-play replay of channel traces, and real-time RIC integration (Mamaghani et al., 13 Jan 2026). The evaluation reports 90th-percentile TTI computation of about $T$3 ms with up to 10 UEs and around 20 taps per UE, preserving protocol timing and end-to-end behavior (Mamaghani et al., 13 Jan 2026). A plausible implication is that digital cellular twins now span a wide implementation spectrum from cloud-native network mirrors to waveform-level full-stack emulators.

6. Applications, trade-offs, and open problems

The application space of digital cellular twins is broad but converges around planning, closed-loop optimization, predictive management, and safe experimentation. The MNDT literature lists network planning, real-time optimization, fault prediction and anomaly detection, and autonomous network operation as key applications (Li et al., 2023). Real-time digital twins of wireless environments extend this toward beamforming, initial access, beam management, predictive handover, blockage mitigation, JCAS/ISAC, and URLLC support (Alkhateeb et al., 2023). DTC further emphasizes adaptive air-interface design, immersive communication, massive communication and resource allocation, space–air–ground connectivity, and HRLLC (Wang et al., 2024).

Digital cellular twins are also increasingly used as data engines. Real-time digital twins built from matched real and synthetic environments can train models in the digital world and transfer them to reality with limited calibration (Alkhateeb et al., 2023). In the beam-prediction example, training on 200 synthetic data points yields $T$4 top-2 beam prediction accuracy on real data, and with fewer than 20 real-world data points for transfer learning the model approaches near-optimal beams; with more than 20 real-world data points, an explicit mismatch in beam codebooks no longer causes a performance gap (Alkhateeb et al., 2023). The Multiverse framework generalizes this into a multi-fidelity family of twins, dynamically selecting among them according to latency and compute constraints for vehicular mmWave beam selection (Salehi et al., 2023).

The literature consistently highlights trade-offs rather than presenting digital cellular twins as frictionless abstractions. Major open problems include scalability to full-city or nationwide networks, data quality and availability, privacy and security, model fidelity and calibration, and interoperability with O-RAN, 3GPP, and existing OSS/BSS systems (Li et al., 2023, Alkhateeb et al., 2023, Wang et al., 2024, Liu et al., 2024). Resource trade-offs are often explicit: twin accuracy versus latency versus communication cost in real-time sensing and fusion (Alkhateeb et al., 2023), fidelity versus compute and latency in multi-fidelity twin selection (Salehi et al., 2023), or synchronization overhead versus prediction accuracy in VH-Twin via the threshold $T$5, where roughly a $T$6 reduction in overhead corresponds to about a $T$7 loss in accuracy (Zhang et al., 2024).

A common misconception is that a digital cellular twin is simply a simulator. Multiple sources reject this equivalence. MNDT is described as an integration-model digital twin that interacts bi-directionally with the network rather than only mirroring it (Li et al., 2023). The 5G private-network proof of concept argues that an NDT is not just an emulator because it includes continuous synchronization with a physical twin and structured data exchange (Costa et al., 14 Oct 2025). Conversely, another misconception is that a network-level twin alone suffices; the physical-environment and channel-twin literature argues that radio-layer fidelity is foundational, especially for 6G where propagation, blockage, sensing, and environmental geometry directly influence control decisions (Alkhateeb et al., 2023, Wang et al., 2024).

A plausible synthesis is that digital cellular twins are evolving toward modular, multi-layer, closed-loop systems: channel twins providing the radio substrate; user, base-station, and service twins capturing operational dynamics; hierarchical or federated network twins coordinating at scale; and AI-driven optimizers or controllers using these replicas to evaluate, train, and deploy network policies under realistic constraints. That trajectory is visible across network-level architectures (Li et al., 2023), real-time environment twins (Alkhateeb et al., 2023), channel-twin frameworks (Wang et al., 2024), mapping methodologies (Zhang et al., 2024), what-if management systems (Ak et al., 2024), and practical full-stack prototypes (Mamaghani et al., 13 Jan 2026).