Digital twins of engineered systems allow for enhanced decision-making in real-world problems through the use of advanced computational methods, which enable both modelling and forecasting under uncertainty (e.g., Reduced Order Modeling, Reinforcement Learning (RL)). Previous work by Arcieri et al. [1] has demonstrated the successful inference of a simulation environment of a deteriorating railway track and applied RL techniques to plan optimal maintenance actions that outperform current human policies. However, the inference of a reliable simulation environment for complex systems, such as an entire railway network, still poses severe challenges that harm a successful implementation of these mathematical models. A railway network presents system-level interactions that need to be accounted for. These are often based on the network topology, such as the correlation in the deterioration of interlinked track sections and the economies of scale that arise from coordinated maintenance of different tracks. The ability to model features intrinsically linked to topological configurations could therefore be of significant value. In this work, we infer a simulation environment of a deteriorating railway network via Markov Chain Monte Carlo sampling of a hierarchical Bayesian model based on the Gaussian Process on Graphs [2] technique. The inference directly exploits real-world monitoring data from a portion of the Swiss railway network around the Zurich metropolitan area. The resulting digital twinned railway network is thus composed of 100 components (track sections). Multi-agent RL is then used to achieve maintenance policies that jointly and cooperatively optimize the entire railway network. In particular, in this work we explore the combination of multi-agent RL with graph-based deep learning models to derive agents that are aware of the network topology allowing them to account for the complex system-level interactions between different components.