ESTIMATING S-WAVE VELOCITY MODEL FROM HORIZONTAL-TO-VERTICAL SPECTRAL RATIO BASED ON SUPERVISED MACHINE LEARNING
By Koichi Hayashi
Researcher at Geometrics/OYO Corporation, Japan
Average S-wave velocity (Vs) to 30 m depth (VS30) is indispensable information to estimate site amplification. Invasive and non-invasive methods, such as velocity loggings or active/passive surface wave methods, are generally used to directly measure the VS30. Those methods are expensive and time consuming. The VS30 is also indirectly estimated by empirical methods based on geology, geomorphology, or slope angle. Those methods are inexpensive but not accurate.
We intend to use a horizontal-to-vertical spectral ratio (H/V) to estimate the VS30. Measurement of H/V is easier and quicker compared with active surface wave methods (MASW) or microtremor array measurements (MAM). Inversion of the H/V is non-unique and it is impossible to obtain unique Vs profiles from H/V. We apply supervised machine learning to estimate the VS30 from H/V together with other information. Our machine learning consists of a neural network with several hidden layers. The pairs of the H/V spectra (input layer) and Vs profiles (output layer) are used as training data. Input layer consists of an observed H/V spectrum site coordinate, and geomorphological information. Output layer is a velocity profile obtained from the velocity loggings, surface wave measurements, or inversion of H/V. We have applied the machine learning to several different sites.
This presentation introduces studies at Napa Valley, CA, U.S. and the South Kanto Plain, Japan. We measured MASW, MAM and H/V at approximately100 and 1800 sites at the Napa Valley and South Kanto Plain respectively. The pairs of H/V spectrum together with their coordinate, geomorphological classification and Vs profile obtained from the inversion of dispersion curve and H/V compose the training data. Trained neural network predicts Vs profiles from observed H/V spectra. The VS30 calculated from predicted Vs profiles are consistent with those calculated from true Vs profiles obtained from the inversion. The results implied that the machine learning could reasonably estimate VS30 from H/V together with other information.