A Modular Network Digital Twin for Radio Coverage Prediction: From Theory to Practice

Network Digital Twins (NDTs) offer a structured framework for modeling, predicting, and optimizing wireless networks. This paper presents a modular NDT implementation based on the GreyCat platform, integrating graph-based data models and external functional algorithms for indoor radio coverage prediction. For the first time, we implement an NDT system aligned with ITU-T Recommendation Y.3090, covering both basic and functional model instantiation from modular and interoperable abstract structures. We generated a practical dataset using a software-defined radio (SDR)-based OpenAirInterface5G setup, with a gNB and commercial UE deployed in a controlled environment. This real-world dataset was used to benchmark Gaussian Process Regression (GPR) and Convolutional Neural Network (CNN) models for predicting RSRP-based radio coverage. Our results show that CNN outperforms GPR in under-sampled conditions, and we demonstrate how the modular architecture supports flexible model integration and benchmarking. This work represents a significant step toward practical, data-driven NDT deployments for wireless systems.