If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact us at firstname.lastname@example.org.
Towards Gaussian Process Regression Modeling of Simulation-Based Ground Motion Coherency Functions
AuthorWaldvogel, Stephen Eugene
AdvisorSeylabi, Elnaz E
Civil and Environmental Engineering
AltmetricsView Usage Statistics
As the design of long-span structures will be subjected to different seismic loading across the entirety of its span (a.k.a multi-support excitation), determining how those ground motions will change is critical for seismic analysis of distributed infrastructure. Due to the lack of densely recorded ground motion databases, ground motions are commonly interpreted from known to unknown points along the span of the structure using ground motion coherency. The coherency is a statistical approach for measuring the similarity of ground motions between two geospatial points. Several methods for approximating coherency have been developed using empirical and semi-empirical functions. These models are developed mainly based on a limited number of dense arrays and simplified wave propagation theories and, therefore, lack the capabilities of modeling source, path, and site complexities and fail to provide consistently accurate values. In the past, many studies have been performed to see how physical parameters may influence coherency loss of ground motions, and many have found importance from numerous site, path, and source effects. This research aims to utilize the results of broadband deterministic earthquake simulations to explore how different physical parameters influence the loss of ground motion coherency as a function of frequency and how machine learning can be used to provide forward progress towards developing non-ergodic and physics-based coherency models.SW4, a fourth-order finite difference code, is used for broadband deterministic earthquake simulations in the Bay Area region and for generating dense arrays of ground motions for coherency analysis. Three of the SW4 simulations investigated are magnitude 4.5 point-source events, and the fourth is a simulated magnitude 6.5 fault rupture event. A machine learning technique based on Gaussian Process Regression (GPR) is used to train non-parametric coherency models and measure the correlation between different physical parameters including frequency, spacing, VS30, and station and source positions on coherency. In total, eight GPR models were developed to measure coherency across the region of interest. The first model looks at the first SW4 point source simulation and forms coherency predictions using input features longitude, latitude, separation distance, and frequency, conditioned against a training set composed of observed data from the same simulation. Models 2 through 5 investigate 5-feature model accuracy for all three point-source events and the simulated magnitude 6.5 fault rupture event, respectively. These five features include station longitude and latitude, separation distance, frequency, and VS30. Like model 1, models 2 through 5 are conditioned against training sets composed of observed data from their respective simulations. Models 6 through 8 investigate the 6-feature model accuracy of each point source event and includes input features of station longitude and latitude, separation distance, frequency, VS30, and station distance from the point-source epicenter. Unlike the first 5 GPR models, models 6 through 8 are conditioned against a training set composed of data from all three point sources to see how well they performed when the training set consisted of greater uncertainty. Results found accurate fitting in all eight models, with a better fit observed on the 5-feature models when compared to 4-feature and 6-feature predictions.