Preprint: XNDT: Towards Explainable Network Digital Twins using a High-Fidelity Hardware Testbed

Abstract

Growing traffic volumes and dynamic operating conditions complicate the management of modern communication networks. Machine Learning (ML)-based Network Digital Twins (NDTs) can support performance prediction and what-if analysis, with Graph Neural Networks(GNNs) capturing topology-dependent network behavior. Yet existing explainable AI (XAI) methods typically return attribution scores, feature rankings or graph structures that remain difficult to interpret operationally. We present XNDT, a research vision for operator-facing explainability in GNN-based NDTs. XNDT links model-level attribution evidence to the active network state and independent telemetry to produce traceable network-semantic explanations of predicted performance degradations. We instantiate this explanation chain in a controlled congestion scenario on a reconfigurable hardware testbed. The proof of concept traces predicted packet loss to the relevant traffic demands, their paths through a shared bottleneck and independent hardware evidence from interface throughput and queue-drop counters. The results provide initial feasibility evidence and motivate broader evaluation across network scenarios and explanation methods, including technical evaluation of operator-facing explanations.