Leveraging Machine Learning for Dynamic System Modeling - Application to Shake Table Data

Machine learning stands at the forefront of technological progress, proving invaluable across various disciplines and is expanding in the realm of natural hazard analysis and modeling. This project serves as a comprehensive case study and workflow, showcasing the adept application of machine learning algorithms for precise regression analysis on experimental data. In the domain of structural engineering, where reliability is paramount, physics-based models offer the most robust foundation for integrating machine learning methodologies. The initial approach employs the widely recognized linear regression algorithm, tailored specifically for this structural engineering application. We've implemented specialized tools that guide the development of features, ensuring both model stability and accurate fitting of nonlinear terms. The second set of algorithms forms a deep neural network providing increased model accuracy at the expense of model interpretability. Further work is in place for expanding these to a physics informed neural network to increase interpretability and generalization. The dataset underpinning this case study comprises dynamic tests conducted at the UCSD Caltrans Seismic Response Modification Device Testing Facility. While some functionalities are tailored to this specific dataset, our notebooks provide clear instructions for customization, allowing seamless adaptation to any regression analysis dataset featuring input-output pairs.