Advancing Network Digital Twin Frameworks for Generating Realistic Datasets Oscar Stenhammar, Sundeep Rangan, Gábor Fodor, Carlo Fischione
arXiv:2604.12888v1 [cs.NI] 14 Apr 2026
Abstract The integration of accurate and reproducible wireless network simulations is a key enabler for research on open, virtualized, and intelligent communication systems. Network Digital Twins (NDTs) provide a scalable alternative to costly and time-consuming measurement campaigns, while enabling controlled experimentation and data generation for data-driven network design. In this paper, we present an open and user-friendly NDT framework that integrates controllable vehicular mobility with the site-specific ray tracer Sionna and the discrete-event ns-3 network simulator, enabling virtualized end-to-end modeling of wireless networks across the radio, network, and application layers. The proposed framework is particularly well-suited for dynamic vehicular networks and urban deployments, supporting realistic mobility, traffic dynamics, and the extraction of cross-layer metrics. To promote open-source initiatives, we release both the NDT implementation and a representative dataset generated from realistic vehicular and urban scenarios. The framework and dataset facilitate reproducible experimentation and benchmarking of machine learning–based quality-of-service prediction, network optimization, and intelligent network management algorithms, lowering the entry barrier for research on virtual and open wireless network services.
I. I NTRODUCTION
T
HE advent of sixth-generation (6G) wireless communication technologies is expected to enable highly flexible, virtualized, and intelligent communication systems [1]. 6G is envisioned to support open and software-driven network architectures capable of adapting to diverse vertical applications, including intelligent transportation systems (ITS) and connected and cooperative automated mobility (CCAM) [2]. Such applications impose stringent requirements on reliability and latency, where communication failures or excessive delays may lead to severe safety and operational consequences. Consequently, future network operators must provision communication services that can sustain reliable and low-latency performance under pronounced spatiotemporal variability in traffic demand, mobility, and radio propagation conditions [3]. To address these challenges, network digital twins (NDTs) are emerging as a key enabler for open and virtualized network operation. An NDT constitutes a fine-grained, software-based replica of a communication network that mirrors the behavior of its physical counterpart and enables experimentation without disrupting live services [4]. By supporting reproducible, controllable, and scalable experimentation, NDTs facilitate data-driven network management tasks such as performance prediction, optimization, and service assurance. As such, NDTs provide a natural foundation for developing and validating machine learning (ML)–based solutions in complex and dynamic wireless environments. Achieving realism in NDTs requires several components to be fulfilled. Wireless propagation and mobility models must capture the dynamics of real environments to reflect time-varying interference and fading caused by mobility patterns. The NDT frameworks must have the ability to scale to the complexity of large and heterogeneous networks while maintaining sufficient granularity to represent distributed devices. Modeling the application layer and the behavior of end users is essential to reproduce realistic traffic patterns and capture the nature of the load experienced in operational networks. If the NDT is installed in a live network, leveraging measurements from real-time systems together is increasingly crucial to calibrate and update the NDT. The combination of these important elements, among others, enables NDTs to evolve beyond static simulations into continuously synchronized systems that can help networks adapt to dynamic network conditions and remain aligned with real-world performance [5]. Recent work has emphasized the need for realistic, controllable, and scalable NDTs to support reproducible research for vehicular and urban wireless systems [5], [6], [7]. This has sparked collective contributions to creating more realistic and accessible NDT platforms. The ms-van3t framework from the work in [8] provides an open-source, standardization-compliant simulation environment that integrates the vehicular mobility simulator SUMO and network simulator ns-3 [9] to support vehicle-to-everything (V2X) simulations. The work in [10] introduces OScar, an open and lightweight cooperative ITS (C-ITS) stack designed for affordable vehicular field tests. At the infrastructure scale, the Colosseum simulator [11] demonstrates how large radio frequency emulators and software protocol stacks can realize Open Radio Access Network (RAN) digital twins for end-to-end experimentation. Traditional discrete-event network simulators such as ns-3 [9] are well-suited for modeling protocol stacks and network traffic flows. However, they often rely on simplified channel models that cannot capture fine-grained spatial, temporal, and ©2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Oscar Stenhammar, Gabor Fodor, and Carlo Fischione are with the School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology, Stockholm, Sweden (e-mail: [email protected], [email protected], carlofi@kth.se). Oscar Stenhammar and Gabor Fodor are also with Ericsson Research, Sweden (e-mail: [email protected], [email protected]). Sundeep Rangan is with the Electrical & Computer Engineering Department, NYU Tandon ([email protected])
Fig. 1. This figure illustrates the considered scene in Sionna. Base station antennas are marked by red spheres, and blue spheres indicate the placement of the antennas of the vehicles.
environmental variations. Ray-tracing-based frameworks such as Sionna RT [12] can produce detailed channel realizations, but lack the integration of full protocol stack behaviour. The work in [13] integrates ns-3 with Sionna RT to create the first full-stack open-source NDT, which enables deterministic ray-traced channel modeling. The study in [14] shows that real network data is subject to concept drift, meaning that traffic patterns, channel conditions, and user behavior can change significantly over time. This highlights the need for an NDT capable of generating heterogeneous and evolving network scenarios so that ML models can be trained and validated under realistic, non-stationary conditions. Realistic NDTs have the ability to empower wireless network research and evaluation of ML algorithms. However, there is a gap for open and accessible NDT frameworks that unify ray-tracing, user mobility, and full network stack simulations for scalable ML-oriented research. We extend prior work in [13], which integrates Sionna RT with ns-3 to support fullstack network data flows over channels computed by ray-tracing, by developing a framework that incorporates user mobility and varying network load into the model. This framework significantly enhances the simulation model by supporting a high number of controllable mobile devices and vehicles. Although SUMO generates realistic vehicle mobility, our framework allows for greater controllability. The network load is implemented to follow a realistic pattern over the hours of the day, while maintaining a degree of stochasticity to reflect authentic network behaviors. This user-friendly NDT provides a rigorous baseline implementation, but can be easily configured to accommodate the desired experimental setup. The simulation framework supports dynamic vehicular mobility with a heterogeneous environment and trajectory selection in real urban grids. A comprehensive logging system is integrated, enabling data sources from the entire network stack. This produces an accessible and reproducible environment towards an NDT ideal for research into dynamic vehicular networks, adaptive communication strategies, and ML-based algorithms. To demonstrate its utility, we provide an open-source repository providing the NDT tool1 that includes extensive examples. To make this tool even more accessible, a representative dataset is published [15], generated from network load and vehicular mobility reflecting reality. The resulting dataset includes spatial, temporal, and network data, and serves as a rich foundation for ML tasks such as predictive quality of service (QoS), anomaly detection, and adaptive resource management. The remainder of this paper is organized as follows. The simulation setup is described in Section II. A data analysis is conducted based on the generated data, and a simple QoS prediction example is provided, in Section III. We present a discussion of the presented work and potential studies with the generated data for future work in Section IV. Finally, we conclude our work with a conclusion in Section V. II. S IMULATION S ETUP To obtain a realistic NDT, we simulate a wireless network by combining the ns-3 discrete-event simulator with the Sionna link-level ray-tracer based on the work in [13]. Ray-traced propagation data is generated in the predefined 500 × 500 m Munich scenario in Sionna and imported directly into ns-3, replacing standard stochastic channel models and enabling site-specific modeling of path loss, shadowing, and multipath effects. A fraction of the scene is visualized in Fig. 1. Vehicular mobility of the users is modeled by simulating connected cars that traverse the road network of the Munich scene. The number of active vehicles is an input parameter. We set probabilities that define crowded and less-populated regions of 1 https://github.com/osst3224/ns3-rt-mobility
Fig. 2. A map of the signal strength in the virtual scenario. Base station antennas are marked by red spheres. The black lines indicate the streets along which vehicles can travel.
the map and spawn vehicles accordingly. We introduce a random number generator that assigns each vehicle a trajectory, with a higher likelihood of selecting routes through crowded areas and a lower likelihood of entering sparsely populated ones. Within each region, we apply area-specific speed limits between 30 km/h to 100 km/h to reflect normal traffic conditions. This approach allows us to generate heterogeneous and realistic mobility patterns throughout the simulated urban environment. To emulate realistic network dynamics, we model background traffic in the network using a 24-hour diurnal load profile; low during early morning hours, peaking during rush hours, and moderate throughout the day. We place 12 base stations with one antenna unit each, and set inter-site distances to ensure coverage of the simulated area. The signal strength is plotted as a heatmap in the considered scenario in Fig. 2. We set the network to operate at 3.5 GHz with 20 MHz bandwidth and a maximum transmit power of 30 dBm. Within this environment, ns-3 instantiates the network stack and applies the time-varying traffic load, while Sionna calculates the link-level radio conditions. The user mobility that we implement is managed in the ns-3 part of the code. To collect multilayer data, we incorporate a FlowMonitor from ns3 with a sampling period of one second. Logged features include flow identifiers, serving cell load, UE position, velocity, and direction, as well as packet-level statistics such as transmitted and received packet error rate, packet sizes, end-to-end latency, throughput, and jitter. Radio-level measurements capture SINR, RSRP, and line of sight (LoS) status. By running repeated simulations across varying traffic and mobility conditions, we generate a large-scale dataset [15] suitable for training and evaluating ML models for applications that require a heterogeneous dataset. III. DATA A NALYSIS A. Simulation Setup We provide an analysis of the generated dataset that we provide in [15], focusing on diversity among the different cells and the interplay between radio and traffic data. The correlation matrix in Fig. 3 summarizes the linear dependencies among the key QoS metrics in the dataset. As expected, the channel-quality metrics SINR and RSRP are positively correlated. Both SINR and RSRP show a moderate negative correlation with latency and packet error rate, reflecting their role as primary drivers of link reliability. LoS conditions also correlate positively with SINR and RSRP, indicating that LoS propagation is associated with stronger received power and improved signal quality. All simulation parameters used for this simulation are included in the provided GitHub repository. On the traffic side, packet error rate shows a strong negative correlation with received packets and a positive correlation with latency, consistent with congestion- or interference-driven performance degradation. The datarate correlates positively with