AI-based Network Digital Twin (AI-NDT)
- AI-based Network Digital Twin (AI-NDT) is a virtual replica of physical networks that continuously synchronizes with real-world data for simulation, training, and optimized control.
- It integrates layered architectures and diverse AI methods—such as reinforcement, meta, federated, and graph learning—to enable closed-loop operations in 6G and beyond.
- Key applications include mobile network optimization, energy-aware control, vehicular adaptation, and secure network management, with challenges in synchronization and scalability.
AI-based Network Digital Twin (AI-NDT) denotes a network digital twin in which artificial intelligence is embedded into the twin’s construction, synchronization, prediction, optimization, and control loops. Across the literature, the concept is framed as a virtual representation of physical networks that is continuously synchronized with real-world data, while also functioning as a safe environment for simulation, synthetic data generation, AI training, and decision validation. In 6G-oriented work, AI-NDT is presented not as a single algorithmic pattern but as a class of closed-loop systems that combine digital twinning with machine learning, deep learning, reinforcement learning, federated learning, graph learning, and, more recently, generative and agentic AI for optimization, forecasting, security, and network operations (Bariah et al., 2022).
1. Definition and conceptual scope
AI-NDT emerged from the convergence of two complementary developments: the use of digital twins as real-time replicas of physical networks, and the use of AI to address modeling, control, and adaptation problems that become intractable in large-scale, dynamic, heterogeneous networks. In this formulation, the digital twin is not merely a visualization or simulation artifact; it is a cyber representation that ingests telemetry, mirrors network entities and environments, and participates in bi-directional feedback with the physical network. AI is then used to build surrogate models, forecast network states, abstract patterns, learn policies, and automate model life-cycle management (Lin et al., 2022).
A central conceptual claim in the literature is that AI and digital twins are enablers for each other. AI supports representation, inference, coordination, and control inside the twin, while the twin supports AI through synthetic data generation, safe training, scenario exploration, and continual validation. This bidirectional relation is explicitly articulated as a bridge between model-driven and data-driven methodologies, especially for 6G systems where purely analytical modeling is often insufficient and purely data-driven methods may lack interpretability or robustness (Bariah et al., 2022).
The term also covers different lifecycle roles of twins. The literature distinguishes planning twins, training twins, and operational twins. Planning twins are used before physical deployment; training twins host AI training in the cyber domain; operational twins act as the “network brain,” supporting real-time inference, retraining, and control. This suggests that AI-NDT is best understood as an architectural paradigm spanning design-time, training-time, and run-time phases rather than as a narrowly defined runtime control system (Bariah et al., 2022).
A common misconception is that AI-NDT is equivalent to conventional network simulation. The surveyed work consistently presents a broader construct: real-time synchronization, closed-loop interaction, and deployment-facing decision support are treated as defining properties. Simulation is only one function among modeling, verification, emulation, forecasting, and AI-assisted control (Lin et al., 2022).
2. Architectural patterns and closed-loop organization
A recurrent architectural pattern is a layered or hierarchical cyber–physical system with explicit feedback loops. In the mobile-network formulation of "How AI-driven Digital Twins Can Empower Mobile Networks," the Mobile Network Digital Twin (MNDT) uses an integration model with a simulation-optimization paradigm organized around inner and outer loops. The mirror model maintains up-to-date virtual representations of users, base stations, and environments; the simulation model evaluates configurations and what-if scenarios; and the integration model maintains bidirectional coupling with the physical network. The inner loop connects an AI-based optimizer to a simulation engine, whereas the outer loop uses real-world network performance to recalibrate the twin and improve optimizer robustness (Li et al., 2023).
Other architectures adopt hierarchical distribution across cloud and edge. In vehicular applications, a two-tier learning framework is organized over a three-layer vehicular network of end devices, edge servers, and cloud. Hierarchical digital twins are deployed at the edge and cloud servers: the cloud DT maintains high-tier meta models and coordinates global updates, while each edge DT acts as a local virtual copy of a physical local vehicular network and supports both emulation and local adaptation. This structure is explicitly designed for automated life-cycle management of intelligent network management functions under spatio-temporal nonstationarity (Qu et al., 2024).
The AI-native 6G formulation introduces differentiated twin types: User Digital Twin (UDT), Infrastructure Digital Twin (IDT), and Slice Digital Twin (SDT). UDTs track profile, behavioral, and networking status of end users; IDTs mirror base stations, routers, edge servers, and related infrastructure; SDTs represent logical slices together with service requirements, user groups, slicing configuration, and performance metrics. These twins interact so that user and infrastructure states inform slice-level decisions, which are validated in the twin before being enacted in the physical network (Wu et al., 2024).
A different architectural trajectory appears in operations-oriented work. Aether integrates Generative Agentic AI with a unified Network Digital Twin for network change validation. Its digital twin combines modeling, simulation, and emulation, while five specialized network operations AI agents collaborate across intent analysis, impact assessment, test planning, execution, and diagnostics. The Network Digital Map acts as a temporal, multi-layered knowledge graph that functions as a single source of truth for pre- and post-change snapshots (Auge et al., 20 Apr 2026).
The following table summarizes representative architectural motifs already described in the literature.
| Architecture | Core organization | Representative source |
|---|---|---|
| Simulation-optimization twin | Mirror model, simulation model, integration model, inner/outer loops | (Li et al., 2023) |
| Hierarchical cloud-edge twin | Cloud DT for meta learning, edge DT for local adaptation | (Qu et al., 2024) |
| AI-native multi-twin system | UDT, IDT, SDT supporting prediction, abstraction, and management | (Wu et al., 2024) |
| Agentic operational twin | Multi-agent AI over a unified modeling/simulation/emulation NDT | (Auge et al., 20 Apr 2026) |
Across these architectures, the closed loop is not optional. Real-time data collection, synchronization, policy validation, and feedback-driven recalibration are treated as necessary conditions for fidelity and safe deployment. This suggests that architectural quality in AI-NDT is determined not only by model accuracy but also by synchronization discipline, model update mechanisms, and the placement of twin functions across cloud, edge, and physical domains.
3. AI mechanisms embedded in network digital twins
Reinforcement learning is one of the most prominent AI mechanisms in AI-NDT. In MNDT, dynamic resource allocation is formulated as a constrained Markov decision process, with state variables including user demands, base station power, and channel states; actions including user association, resource block assignment, and power allocation; reward defined as total throughput; and costs encoding demand satisfaction constraints. Penalized Proximal Policy Optimization and DDPG-like procedures are described for mixed discrete-continuous control, while Mean Field RL is used for base-station sleep and cell on/sleep/off decisions under QoS constraints (Li et al., 2023).
Meta learning is used where network conditions are explicitly nonstationary. In the vehicular framework, high-tier cloud learning extracts generalizable features from multiple local vehicular networks, and low-tier edge learning rapidly personalizes models for local contexts. Hierarchical meta models are trained for different categories such as time of day, city, or road type, and updates are triggered when distribution drifts are detected through adaptation loss changes. In the cooperative-perception case study, PPO-based meta learning is compared with transfer learning and training from scratch, with the meta-learning variant reported to converge faster while matching asymptotic performance (Qu et al., 2024).
Federated learning appears chiefly in forecasting and distributed intelligence. A 2025 synthesis redesigns forecasting as a joint data-scenario prediction task and proposes Hybrid FL with synchronous and asynchronous modes. Clustering and region virtualization form local DNTs, which contribute to a global model. The same work also describes federated reinforcement learning pipelines for secure vehicular settings, where DNT-generated adversarial scenarios and server-side filtering are used to resist Byzantine or poisoning attacks (Zhang et al., 8 Mar 2025).
Graph-based learning plays two distinct roles. First, graph neural networks are used as predictive models for network evaluation and KPI estimation. Real-data-driven DTN work proposes the autoencoder-based skip connected message passing neural network (AE-SMPN), combining autoencoding, message passing GNNs, and recurrent models to capture spatiotemporal features from the BNN-UPC dataset. Second, graph structures are used as the organizing substrate of the twin itself. A multi-layered knowledge-graph AI-NDT trained on RIPE Atlas data uses Network Topology, Network State, and Application State Knowledge Graphs, and compares GraphSAGE, ChebNet, ResGatedGCN, and GraphTransformer, reporting that GraphTransformer attains the best predictive performance while shorter training time may favor other architectures (Shin et al., 2024); (Zacarias et al., 4 Aug 2025).
Generative models serve both data augmentation and operational scenario generation. GANs and VAEs are used to synthesize traffic or fault scenarios and reconstruct incomplete or noisy telemetry for predictive maintenance and anomaly detection. In tactical networks, diffusion models and transformers generate diverse and adversarial scenarios for agent training inside synchronized digital twins. More recently, agentic generative AI has been integrated directly into network-change validation workflows, combining LLM-based planning and tool use with a knowledge-graph-backed twin (Muhammad et al., 2024); (Ray, 28 Jul 2025); (Auge et al., 20 Apr 2026).
A common misconception is that AI-NDT is identical with RL-driven optimization. The literature is broader: RL is important, but AI-NDT also includes supervised forecasting, graph learning, meta learning, federated learning, generative modeling, and hybrid-system identification. The choice of AI method depends on whether the twin is being used for control, prediction, calibration, verification, or scenario synthesis.
4. Modeling fidelity, learning substrates, and real-time digital twinning
High-fidelity modeling is a defining requirement because AI policies and forecasts are only as useful as the twin states on which they are trained and validated. Multiple papers therefore focus on learnable or adaptive surrogate models that replace or augment conventional simulation.
One line of work develops learnable KPI surrogates. PLAN-Net is a learning-based NDT for network simulators that integrates node, edge, and path embeddings and maps a network configuration to KPIs in a single forward pass. The architecture uses iterative message passing with RNNs, MLPs, and edge-conditioned GCNs, and is reported to outperform baseline learning models while achieving comparable performance to simulators with substantially higher computational efficiency (Li et al., 2023).
A second line concentrates on real-data-driven evaluation. The AE-SMPN model uses real DTN data rather than simulation-only data. Experimental results on the BNN-UPC dataset report test MAPE values of 42.42 for AE-MPNN and 42.82 for AE-SMPN3, compared with a baseline test MAPE of 100.52, for average delay prediction on held-out topologies. The paper attributes the gains to AE-based feature extraction, message passing, LSTM-based spatiotemporal modeling, and skip-connected readout (Shin et al., 2024).
A third line addresses radio-environment fidelity. For real-time propagation modeling, a U-Net-like deep learning model combines a 3D elevation map with a rough propagation estimate and outputs full-area path-gain heatmaps. Over a 37,210 square meter area, the reported normalized RMSE is less than 0.035 dB, with inference times of 46 ms on a GPU and 183 ms on a CPU, whereas high-fidelity ray tracing requires approximately three orders of magnitude more time. The model can also be calibrated with measurement data, further reducing median error to 0.0113 dB in the reported setting (Saeizadeh et al., 2024).
Environment-specific digital twins are also used to generate labeled training data. For line-of-sight identification, a city-scale NDT of Milan together with deterministic ray tracing is used to produce angle-delay channel power matrices for Deep Learning models. The proposed training strategies outperform the state of the art 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 stated baseline solutions (Zhu et al., 21 May 2025).
Another trajectory is hybrid-system identification. "Learning a Network Digital Twin as a Hybrid System" models the NDT with modes corresponding to base stations and sub-modes corresponding to workspace partitions with similar network characteristics, learned through an online deterministic annealing algorithm driven by measurements from moving users. On a two-cell 5G testbed at Ericsson, Stockholm, the hybrid NDT is reported to converge in approximately 20 seconds using approximately 500 training samples and 38 codevectors, with lower mean squared error and faster post-change adaptation than a neural-network regressor (Mavridis et al., 31 Oct 2025).
These efforts show that fidelity in AI-NDT is achieved through different substrates: simulator surrogates, GNN-based evaluators, deep propagation models, environment-specific ray-tracing twins, and hybrid-system models. A plausible implication is that no single modeling substrate is sufficient across all tasks; instead, AI-NDT research is converging on a stack of specialized twin components tuned to topology, radio, workload, and control requirements.
5. Principal application domains and empirical results
Mobile-network optimization is a canonical AI-NDT application. In the MNDT prototype, the emulated network spans an area of with 13,000 users, 184 base stations, 1,850 lanes, and 1,417 areas of interest. In the dynamic resource-allocation experiment with 8,000 moving users and 184 base stations, the reported throughput is with a satisfaction ratio of 95.28%, compared with 4.12e5 and 99.02% for equally dividing demand, and 5.18e5 and 91.45% for a throughput-maximizing baseline that ignores demand satisfaction. The paper interprets this as achieving high throughput while maintaining the demand-satisfaction threshold (Li et al., 2023).
Energy-aware control is another established use case. In the same MNDT framework, a Mean Field RL sleep mechanism is trained using one day of traffic data and evaluated over a week of real traffic data from Nanchang, with decisions every 30 minutes. The reported result is that the RL agent’s energy consumption closely tracks actual traffic and outperforms both always-on operation and the minimal-cell-fit heuristic of Peng et al. 2011 in energy saving (Li et al., 2023).
Vehicular AI-NDT focuses on rapid adaptation under nonstationarity. In the cooperative-perception case study, the high-tier/low-tier twin-assisted meta-learning framework uses PPO for switching between Standalone Perception and Cooperative Perception in response to workload, radio resources, and channel conditions. The reported outcome is faster convergence and better adaptation than transfer learning, together with higher policy entropy during adaptation, which is interpreted as improving exploration and reducing suboptimal convergence risk (Qu et al., 2024).
Forecasting and proactive resource management form a separate domain. The AI-enabled DT framework based on a three-layer BiLSTM stack with $128,128,64$ units, dropout 0.2, and batch normalization predicts one-step-ahead network demand from four KPIs sampled every 5 minutes over more than a month. On the test set, the AI-enabled DT reports MAE approximately 25 and RMSE approximately 40. For resource management, median efficiency is approximately 0.987, wastage is less than 0.05, utilization is approximately 1, and over-provisioning is much less than 100 units, whereas the two static baselines show lower efficiency and substantially higher over-provisioning (Sengendo et al., 23 Oct 2025).
Security- and safety-critical applications are increasingly prominent. In edge caching, a DNT plus safe RL pipeline with intervention modules reports a cache hit rate of approximately 0.88, a maximum base-station load reduced to approximately 0.35, and a minimum load increased to 0.12 under the full DNT+RL+intervention configuration. In secure vehicular networks, a DNT plus Byzantine-robust federated RL pipeline maintains a 100% no-collision rate under model and data poisoning attacks, even with up to 40 agents, after training over more than 50,000 HighwayDT scenarios (Zhang et al., 8 Mar 2025).
Network operations is a newer empirical frontier. Aether evaluates automated change validation on eight synthetic scenarios and two real-world ISP incidents. The paper reports 100% error detection on past incidents, diagnostic coverage of 92–96%, and validation times of 6–7 minutes, contrasting these outcomes with scattered traditional testing workflows and partial coverage (Auge et al., 20 Apr 2026).
Industrial and IoT deployments provide another applied strand. A DT-native IoV architecture that combines a TCP-based data-flow pipeline and a DDPG learner reports approximately 30% processing time-saving relative to manual data-flow triggering, and identifies the actor/critic learning-rate pair as the most successful among the tested combinations over 135 episodes (Duran et al., 2023).
6. Challenges, limitations, and research directions
The literature is unusually explicit about unresolved issues. Synchronization fidelity remains fundamental: exact cyber–physical alignment, freshness of data, periodic refresh, and low-latency interfaces are repeatedly identified as preconditions for reliable inference and control. Without them, prediction accuracy, validation soundness, and sim-to-real transfer all degrade (Bariah et al., 2022).
Scalability is a second persistent challenge. Real-time AI-NDTs must operate from cell-site scope to city-scale or larger topologies, often with heterogeneous vendors and cross-domain data. Papers therefore emphasize distributed twins, federated learning, graph representations, GPU acceleration, and cloud-edge partitioning as practical necessities rather than optional enhancements (Lin et al., 2022).
Computation and memory overhead also recur. The BiLSTM-based traffic-forecasting twin reports a model size of about 2.7 MB, larger than its static baselines, and explicitly notes a computational trade-off. The propagation-modeling literature similarly frames real-time capability as contingent on replacing or augmenting expensive ray tracing with learnable surrogates (Sengendo et al., 23 Oct 2025); (Saeizadeh et al., 2024).
Security and trustworthiness remain contested areas. Federated and multi-agent settings are vulnerable to data poisoning and model poisoning; digital twins themselves become part of the attack surface because corrupted synchronization or simulated scenarios can propagate unsafe policies. Proposed responses include robust aggregation, anomaly detection, secure filtering, and adversarial scenario generation inside the twin to harden learned policies (Zhang et al., 8 Mar 2025); (Ray, 28 Jul 2025).
Interpretability is not fully resolved. Several papers stress explainability, visualization, or knowledge-graph structures, yet many performance-leading components remain deep or hybrid black boxes. The broader survey literature therefore ties AI-NDT development to responsible AI requirements, including transparency, privacy preservation, fairness, accountability, and documentation of automated decisions (Al-Shareeda et al., 2024).
Open research directions are correspondingly broad. The AI-native 6G framework highlights hierarchical digital-twin deployment, hybrid data-model driven decision-making, and efficient collaboration among twins as central problems. The broader AI–DT interplay literature adds mathematical interpretability, ultra-fast reliable AI algorithms, and principled integration of model-driven and data-driven methods. More recent operational work suggests that network digital twins are expanding from optimization platforms into verification, testing, and intent-aware orchestration systems (Wu et al., 2024); (Bariah et al., 2022); (Auge et al., 20 Apr 2026).
Taken together, these directions indicate that AI-NDT is evolving from a simulation-assisted optimization paradigm into a general cyber–physical intelligence layer for future networks. The field’s distinctive feature is not merely the use of AI inside a digital twin, but the systematic coupling of synchronized virtual replicas, data-driven and model-driven learning, and deployment-facing feedback loops across planning, adaptation, and operations.