In In the realm of Structural Health Monitoring (SHM), much research has been devoted in recent years to automated methods allowing real-time unsupervised monitoring of structures. When damage occurs, extended monitoring of a structure should include detection, localization, quantification, and prognosis of the residual life. While there is a broad consensus on the use of control charts for damage detection purposes, various techniques are found in the literature concerning damage identification [1], encompassing both localization and quantification. When monitoring is conducted through vibration-based techniques, damage identification typically involves the inverse calibration of finite element models through non-linear optimization, presenting computational challenges incompatible with real-time SHM [2-3]. Therefore, the identification of high-fidelity surrogate models for real-time model updating based on continuous monitoring data is a challenging topic in structural system identification. A surrogate model is a mathematical function or algorithm that approximates the behaviour of a structure based on collected data from the actual structure (i.e. accelerations, deformations, displacements) in an inexpensive computational manner. Subsequently, surrogate models can be employed to make predictions about the future structural health of the system based on current and past observations [4]. While surrogate models are widely applied to heritage masonry structures, there is no evidence in the literature regarding the application of these powerful techniques in the case of infrastructures like bridges. With the aim of addressing this gap in the literature, the effectiveness of surrogate models is demonstrated in the case study of a real in-operation multi-span Italian roadway bridge.