Earthquake Ground Motion Simulation using Novel Machine Learning Tools

Abstract

A novel method of model-independent probabilistic seismic hazard analysis(PSHA) and ground motion simulation is presented and verified using previously recorded data and machine learning. The concept of “eigenquakes” is introduced as an orthonormal set of basis vectors that represent characteristic earthquake records in a large database. Our proposed procedure consists of three phases, (1) estimation of the anticipated level of shaking for a scenario earthquake at a site using Gaussian Process regression, (2) extraction of the eigenquakes from Principal Component Analysis (PCA) of data, and (3) optimal combination of the eigenquakes to generate time-series of ground acceleration with spectral ordinates obtained in phase (1). The benefits of using a model-independent method of PSHA and ground motion simulation, particularly in large urban areas where dense instrumentation is available or expected, are argued. The effectiveness of the proposed methodology is exhibited using eight scenario examples for downtown areas of Los Angeles and San Francisco where it is shown that no dependency on specific ground motion prediction equations or processes of selection and scaling would be needed in our procedure. Furthermore, PCA allows systematic analysis of large databases of ground motion records that are otherwise very difficult to handle by conventional methods of data analysis. Advantages, disadvantages, and future research needs are highlighted at the end.