Prediction of seismic P-wave velocity using machine learning

Abstract

Measurements of seismic velocity as a function of depth are generally restricted to borehole locations and are therefore sparse in the world's oceans. Consequently, in the absence of measurements or suitable seismic data, studies requiring knowledge of seismic velocities often obtain these from simple empirical relationships. However, empirically derived velocities may be inaccurate, as they are typically limited to certain geological settings, and other parameters potentially influencing seismic velocities, such as depth to basement, crustal age, or heat flow, are not taken into account. Here, we present a machine learning approach to predict the overall trend of seismic P-wave velocity (vp) as a function of depth (z) for any marine location. Based on a training dataset consisting of vp(z) data from 333 boreholes and 38 geological and spatial predictors obtained from publicly available global datasets, a prediction model was created using the random forests method. In 60 % of the tested locations, the predicted seismic velocities were superior to those calculated empirically. The results indicate a promising potential for global prediction of vp(z) data, which will allow the improvement of geophysical models in areas lacking first-hand velocity data.