Physics-Informed Machine Learning-Driven Structural Digital Twin for Damage Identification

Abstract

With advancements in AI, data-driven methods such as autoencoders (AEs) have been widely used for damage identification through anomaly detection. However, AE-based methods primarily learn identity mappings from healthy-state data, making them less effective in detecting subtle damage. This study presents a physics-informed, machine learning-driven structural digital twin (SDT) framework for damage identification using anomaly detection. By incorporating physics-informed neural networks (PINNs), the framework reduces discrepancies between finite element model (FEM) predictions and real structural responses, enabling more accurate anomaly detection. The proposed approach is evaluated using the numerical ASCE benchmark structure. It outperforms AE and LSTM-AE baseline methods in comparatively smaller damage scenarios.