Data-informed simulation of temperature effects in finite element models for structural health monitoring using inverse surrogate modeling

This work presents a data-informed framework to simulate temperature effects in a finite element model (FEM) for structural health monitoring (SHM), enabling realistic generation of synthetic modal data for both healthy and damaged states. A challenge in vibration-based SHM is that real damaged data are rarely available, so damage detection and classification methods are often validated using FEM simulations. However, temperature strongly affects modal data through changes in stiffness, connections, and boundary conditions, and these effects are difficult to reproduce realistically using standard thermo-mechanical assumptions, especially when only few temperature measurements are available. The proposed approach combines field monitoring data with an updated FEM through inverse surrogate modeling. First, an inverse surrogate model is trained on FEM simulations to recover temperature-sensitive mechanical parameters (e.g., Young’s modulus of elasticity) from modal frequencies. Then, this surrogate is applied to long-term monitoring data to estimate how these parameters evolve with temperature. A second regression model then learns a direct mapping from temperature to FEM parameters. Injecting these temperature-dependent parameters into the FEM allows simulation of realistic temperature-frequency trends, for both healthy and damage scenarios. The method is demonstrated on a numerical truss and on the KW51 railway bridge. Results show high accuracy in modal frequency prediction and consistent anomaly detection performance, indicating that the framework reproduces in-situ thermal effects without requiring explicit thermo-mechanical modeling, while preserving damage detectability under temperature variations.