A hybrid data-physics framework with conformal GNN for enhanced damage identification

Abstract

[EN] Structural damage identification is crucial for ensuring safety, yet existing data-driven and physics-based methods often suffer from accuracy and computational limitations. To address these issues, we propose a hybrid framework that integrates Graph Neural Networks (GNNs) with a physics-based Finite Element (FE) model updating approach. The first module employs a GNN trained on modal data from FE simulations to estimate the location and severity of structural damage, with an evolutionary AutoML framework optimizing the GNN's architecture and hyperparameters. In the second module, a conformal prediction technique quantifies uncertainty in the GNN's predictions, ensuring robust confidence bounds for damage estimations. These uncertainty-aware predictions initialize a warm-started FE model updating workflow, where the Water Strider Algorithm (WSA) efficiently minimizes a cost function based on limited modal data. The proposed methodology has been validated on benchmark structures, including the Louisville bridge, IASC-ASCE building and a dome structure, demonstrating a remarkable increase in damage identification accuracy compared to conventional approaches. Unlike pure data-driven and physics-based methods, this hybrid framework leverages their strengths while integrating uncertainty quantification, enhancing their efficiency. This hybrid approach is scalable to various structural configurations, making it a promising solution for enhanced structural health monitoring.