Thesis Defense: Neng Xiong, Ph.D. Candidate
Location: KWGL 123
Department: Earth, Environmental and Planetary Sciences
Defense Date: April 12, 2022
Time: 2:30 p.m.
Data driven and machine learning based methods for Coulomb failure stress calculation, velocity model evaluation and automated receiver function selection
In recent years, machine learning has become a very popular field of study in geophysics and seismology. As a data intensive field, seismology provides a great opportunity for the development of data driven method. Machine learning, with its great ability to find and extract hidden information from data, gives a new perspective to solve existing problems in seismology in a statistical way. Utilizing the great power of data science, my projects create data driven and machine learning based methods to tackle existing problems in Coulomb stress calculation, velocity model evaluation, and receiver function selection, in the purpose of providing more precise, robust, and efficient solutions.
My first project aims at providing a more accurate estimation of Coulomb failure stress changes (ΔCFS) in regions with poor understanding of the structure and seismic station coverage. Applied to the 2015 Mw 7.8 Gorkha earthquake, my designed method uses ΔCFS calculated from aftershocks, with known moment tensor solutions, as true value, to constrain the predicted value of maximum Coulomb failure stress changes occurred with the optimum focal mechanisms and provide a better ΔCFS estimation for areas with no receiver fault information. Based on the calculation result, this study confirm that a detailed rupture model, that fits the realistic geometry of the fault, is crucial to ΔCFS calculation of thrust fault earthquakes. My result also indicates a significant stress accumulation in the unruptured area south and west to the Gorkha earthquake.
Next, my thesis shifts the focus to traditional machine learning method, using unsupervised clustering analysis to tackle the problem of velocity model evaluation. Comparing to the forward 3-D waveform simulation, which is often used for velocity model validation, my second project provides a much more efficient solution using K-means clustering and applies to two velocity models in Southern California (CVM-H15.1 and CVM-S4.26). This is done by first calculating synthetic surface wave velocity dispersion curves input velocity models, and then clustering the synthetic and observed velocity dispersion curves independently into certain number of groups. The velocity model is rated by estimating the similarity between spatial patterns obtained from the synthetic and observed dispersion data. Comparing to the forward 3-D waveform simulation, this new evaluation scheme is extremely efficient and no longer limited by the source-receiver configuration. My studies also suggests that the CVM-S4.26 fits the observed dispersion maps better than the CVM-H15.1 in term of the clustering pattern.
Last, my thesis targets on solving the problem of receiver function (RF) selection. Traditional human handpicked receiver function lacks established picking criteria and requires significant amount of time. My third project incorporates the feature engineering process to extract meaningful features from the RF data sets, in the purpose of understanding what are the characteristics that separate the good and bad RFs. Build on the selected features, I apply the fuzzy C-means clustering method to automatically picking the RFs and achieves a high F1 score of ~90% for the testing data sets. My study identifies 4 features with great separation between good and bad RFs, which can be used as the guide metrics for RF selection. The clustering based classifier could further eliminate the need for human picking, saving significant amount of time for large seismic studies.