UC Berkeley

Monitoring and evaluation of structural systems is essential for maintaining life safety and cost-efficient operation of infrastructure. Recent advances in sensing technologies have created a data-rich environment for structural health monitoring. Traditional structural state identification (SSI) methods are not computationally efficient to handle large volumes of input data and are limited in their ability to localize damage within a structure. Recently developed machine learning (ML)-based methods leverage large quantities of data to produce high-resolution outputs. However, these approaches are often not easily interpretable or do not explicitly enforce governing mechanical laws. This dissertation explores the potential of differentiable physics for inverse parameter estimation in structural applications, enabling identification of structural parameters from monitoring data in an interpretable and physically grounded manner. The dissertation initially presents a ML framework for identifying train classes from dynamic strain recordings. First, characterizing features are extracted from the signal. Then, the XGBoost classification algorithm is trained on the signal feature data to accurately distinguish different types of rolling stock. With train classification, further analysis of the monitoring data can be conducted after conditioning for varying load effects induced by different train types. Relationships among peak recorded strains, time, and temperature are subsequently explored to better understand long-term structural response trends. Next, a differentiable material point method (DiffMPM) framework is developed for structural parameter identification. DiffMPM leverages automatic differentiation and gradient-based optimization to inversely estimate parameters related to material properties and loading conditions. The framework is applied to several numerical examples to evaluate its effectiveness under varying levels of system observability, parameter initialization, optimizer selection, and measurement noise. Results demonstrate that DiffMPM shows promise for structural system identification while maintaining physical interpretability and computational efficiency. The DiffMPM framework is then applied to synthetic damage identification problems. In these numerical examples, the virtual model is initialized in an undamaged state, and stiffness reductions are introduced in a second data-generating model to simulate structural damage. DiffMPM is subsequently used to learn the spatial damage field of the model. To improve damage localization, shape characterization, and identification accuracy, several optimization strategies are proposed and evaluated. The results indicate that DiffMPM has strong potential for damage localization and characterization in numerical settings. Finally, the differentiable physics methodology is extended to develop a differentiable finite element method (DiffFEM) for structural parameter identification. To evaluate DiffFEM using experimental monitoring data, strain measurements from reinforced concrete prisms subjected to uniaxial tension are collected using distributed fiber optic sensors (DFOS) and camera-based digital image correlation (DIC). These strain measurements are used within DiffFEM to estimate spatial 3D damage within the structure, modeled through stiffness reduction. DiffFEM demonstrates strong strain field recovery and identifies damage in regions where cracking is observed in the reinforced concrete specimens. Overall, the findings of this research contribute new methodologies for system identification and damage detection in structural health monitoring applications. The proposed differentiable physics frameworks improve the interpretability, physical consistency, and spatial resolution of structural parameter identification, advancing capabilities for damage assessment and safe operation of structural systems.