Model updating of structures by combining reduced order modelling and deep reinforcement learning

This paper presents an automated framework for structural model updating that integrates reduced-order modelling (ROM) with deep reinforcement learning (DRL). Continuous structural health monitoring (SHM) data are used to calibrate numerical models acting as digital twins of real structures, a task that becomes computationally prohibitive when high-fidelity finite element models are employed for highly uncertain systems. To address this challenge, a ROM is constructed from a physically consistent dataset generated by a full-order model and embedded within a DRL environment. An intelligent agent autonomously explores the parameter space and identifies optimal values by minimizing discrepancies between simulated and measured modal quantities. The proposed methodology, termed DRUM-Av2.0, is validated through two fictitious case studies of increasing complexity and a real prestressed concrete bridge. Results demonstrate up to two orders of magnitude reduction in computational time compared to FOM-based updating, while maintaining high accuracy in parameter identification. The framework provides a scalable and automated solution suitable for practical SHM applications and continuous digital twin calibration.