- The paper presents an open-source framework that combines site-specific ray-tracing with ns-3 simulation to realistically model urban vehicular network environments.
- The paper demonstrates substantial improvements in predictive QoS modeling with cell-specific machine learning models reducing mean squared error compared to global approaches.
- The paper emphasizes the practical and theoretical impact of unified simulation techniques for advancing reproducible, data-driven wireless research and AI optimization.
Advancing Network Digital Twin Frameworks for Generating Realistic Datasets
Introduction
This paper introduces an open-source Network Digital Twin (NDT) framework that unifies vehicular mobility, site-specific ray-tracing, and end-to-end network stack simulation. The framework addresses critical gaps in existing NDT platforms, specifically the lack of realistic propagation environments coupled with controllable mobility and application/traffic dynamics—features essential for modern wireless network research, particularly within urban and vehicular communication contexts. By systematically integrating the Sionna ray-tracer and the ns-3 discrete-event network simulator, the architecture enables scalable, reproducible experimentation across spatial, temporal, and protocol-layer dimensions. The authors further substantiate their approach with an open-access software repository and a representative dataset, thereby promoting reproducibility and accessibility in data-driven wireless research.
Framework Architecture and Simulation Methodology
The proposed NDT framework combines the deterministic, site-specific channel modeling capabilities of Sionna RT with the flexible protocol stack emulation offered by ns-3. The Sionna-based radio channel simulations account for multipath, fading, and environmental effects by employing a ray-traced Munich urban scenario covering a 500 × 500 m area. Mobility is realized via a highly configurable vehicular model, where vehicles are spawned according to spatial density probabilities and traverse the road network guided by both randomized and biased trajectory assignments, reflecting realistic traffic distributions.
Temporal network dynamics are emulated through a diurnal load profile, synthesizing stochastic yet smoothly varying traffic consistent with real-world diurnal fluctuations. The simulation environment comprises 12 base stations operating at 3.5 GHz with 20 MHz bandwidth and integrates a sampling infrastructure (via FlowMonitor) for collecting fine-grained measurements at one-second intervals. Key logged parameters include user equipment (UE) position, velocity, serving cell load, SINR, RSRP, latency, packet error rate, and throughput, structured to facilitate cross-layer analysis.
Dataset Features and Statistical Analysis
The dataset generated encompasses approximately 90,000 samples over a 24-hour simulation, capturing extensive heterogeneity in both radio and traffic conditions. Analysis reveals expected statistical relationships: SINR and RSRP display strong positive correlation, and both metrics are inversely related to latency and packet error rate, confirming their relevance as primary link reliability drivers. LoS status strongly co-varies with better radio conditions, and increased cell traffic is associated with higher latency and packet errors. The framework's ability to create spatially and temporally diverse latency distributions across cells is highlighted, showcasing the capacity for studying localized QoS variations and the impact of both environmental and traffic-induced non-stationarities.
Machine Learning Experiment: Predictive QoS Modeling
A key use-case explored is predictive modeling of QoS metrics, specifically one-hour-ahead prediction of cell-level latency distribution parameters. The authors benchmark three approaches: a naive persistence model, a global model aggregated across all cells, and localized (per-cell) models trained solely on local data. A relatively standard MLP architecture is employed, optimized using Adam and regularized via dropout. Numerical results underscore a substantial reduction in mean squared error using cell-specific models (0.16 ms) relative to the global (0.59 ms) and naive (0.61 ms) baselines, establishing the criticality of spatially localized model adaptation due to pronounced concept drift and cell-specific correlation structures. The performance differential supports the claim that heterogeneity and local dynamics in urban vehicular networks cannot be adequately captured by global model aggregation, directly motivating future research into federated and adaptive ML schemes.
Practical and Theoretical Implications
The presented NDT platform addresses reproducibility and accessibility barriers common in wireless network research. By incorporating ray-traced propagation, dynamic mobility, realistic traffic, and end-to-end protocol emulation, the framework supports rigorous study of ML-driven network optimization, predictive QoS, and adaptive resource management under faithfully modeled urban conditions. From a practical perspective, it lowers the resource burden of data collection campaigns and supports comparative benchmarking of learning algorithms under repeatable conditions.
Theoretically, the findings on spatial and temporal heterogeneity and concept drift—amplified by strong empirical evidence—imply that future AI for communications will increasingly need to adopt localized, adaptive, or meta-learning paradigms. The open dissemination of both the toolchain and datasets will likely accelerate innovation in areas such as online learning robustness, anomaly detection in non-stationary environments, and transfer learning between urban morphologies.
Limitations and Future Directions
While the NDT platform achieves high realism by integrating environmental awareness and realistic mobility, direct fidelity comparison with real deployed networks remains infeasible without access to comprehensive, synchronized cross-layer real-world measurements, which are obstructed by privacy and commercial constraints. The computational complexity associated with ray-tracing also presents practical limits for very large-scale simulation, although GPU acceleration and parallelization are recommended.
Promising future work includes systematic validation via targeted measurement campaigns, scalable anomaly detection benchmarks, and studying transferability of learned models across urban scenarios with distinct topology and user distributions. The availability of open-source tools and datasets signals a shift toward more transparent and standardized methodology in wireless machine learning research.
Conclusion
This work presents a hybrid, open-source NDT framework merging ray-tracing-based channel modeling with discrete-event vehicular network simulation for the generation of realistic, reproducible datasets. By facilitating controlled experimentation over spatially and temporally diverse scenarios, the system enables robust evaluation of ML-based network management and QoS prediction algorithms, as substantiated by strong local model numerical results. These contributions lower the barrier for data-driven research and catalyze further AI innovation in next-generation wireless systems.