Digital twins (DT) are essential tools for rapid health assessment of critical infrastructures, such as transport and energy structures, during extreme events when resources are scarce. In such situations, digital twins facilitate the prediction of dynamic response and forecasting of potential equipment failures, aiding efficient recovery planning. However, real-time surrogates are needed to address the timing constraints of current engineering modelling solutions, including finite element and seismic fragility models. In this work, we propose a hybrid approach to develop these surrogates by combining physics-based models with data- driven models [1]. This method captures the complex interactions among equipment in various configurations of electrical substations. Specifically, we integrate Graph Neural Networks (GNN) with physics-based surrogates (i.e., reduced-order models or ROMs) [2] to model individual components. We modify graph aggregation functions to ensure physical coherence and enhance interpretability. This tool has been tested through a case study focused on combinations of equipment interconnected by a busbar. The final outcome consists in a DT that incorporates monitored ground motion signals to provide real-time information about the equipment’s health and functionality, enabling swift and informed decisions regarding power restoration.