SCEC Award Number 17239 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Application of Machine Learning in Deterministic Ground Motion Simulation
Investigator(s)
Name Organization
Ricardo Taborda University of Memphis
Other Participants Naeem Khoshnevis
SCEC Priorities 4a, 4c, 2b SCEC Groups GM, EEII, CS
Report Due Date 06/15/2018 Date Report Submitted 06/30/2018
Project Abstract
 This project focused on investigating and testing the potential use of machine learning methods in three-dimensional (physics-based) ground motion simulation, as a means to optimize data analysis and modeling parameters. We concentrated efforts in two particular problems: (i) prioritizing ground motion validation metrics; and (ii) evaluating attenuation models. For the first of these problems, we analyzed a dataset of ground motion validation results with eleven goodness-of-fit metrics using semi-supervised and supervised learning methods. We performed a clustering analysis on the dataset and identified multi-dimensional patterns in order to label the data samples, which allowed us to produce decision trees that with a prioritized and narrowed choice of metrics. In the end, we identified the response spectrum, total energy and peak ground acceleration as the most relevant metrics, followed by the Arias intensity and peak ground velocity. For the second problem, we used ma-chine learning techniques to create surrogate simulators that could predict the peak velocity, peak acceleration, area under the envelope of seismograms and response spectrum ordinates, to then feed an optimization method to invert the parameters associated with an attenuation (Q-Vs) relationship. The procedure was tested for both idealized models and realistic models, and it was found to effectively identify the parameters for the assumed Q-Vs relationship when using synthetic data as reference. Initial tests with real data showed promise but require additional optimization steps and user supervision.
Intellectual Merit This study contributes to exploring new alternatives that could potentially accelerate the time-to-completion of routine data analysis and computing intensive processes. In particular, the activities carried out contribute to the goals of SCEC in areas of the ground motion, scientific computing, and the earthquake engineering implementation interface disciplinary and interdisciplinary groups, and the community modeling environment group. In terms of developing creative and original concepts, the research carried out contributed two methodologies not available before. On the one hand, we developed an algorithm for validation of ground motion synthetics that can provide an accurate assessment of the level of accuracy of simulated seismograms when compared to data using only a few metrics (2 or 3), as opposed to the more elaborate methods in use today employing a larger family of goodness-of-fit metrics (11). And on the other hand, we developed a procedure to invert for the parameters defining a Q-Vs relationship in a simulation domain based on an optimization process using training data to predict ground motion metrics.
Broader Impacts From a broader impacts perspective, the main contribution of this project was to advance the use of machine learning techniques in areas of earth sciences and earthquake engineering. Machine learning is an area of growing interest and in this project we show that there is potential for its use in two types of problems, data analysis and synthesis, and prediction of ground motion characteristics. Furthermore, the project provided a particular learning opportunity for a graduate research assistant. The original proposal for this project was prepared by PhD Geophysics student Naeem Khoshnevis in the Center for Earthquake Research and Information at the University of Memphis. Khoshnevis also carried out the research as his own with very little input from the PI of record, Ricardo Taborda.
Exemplary Figure Marked as exemplary figure in the report.