Seismological geothermometer. Part I: Neural network modeling of the temperature prediction from seismic velocity data at borehole depth scale

Abstract

A feasibility study is carried out to assess the possibility of predicting the subsurface temperatures from seismic tomography data. To this end, we used seismic velocity data along the W-E profile in the Northern Tien Shan and geotherms from three boreholes located at different distances from the profile. To model the temperature prediction from seismic velocity data, we used a supervised artificial neural network (ANN) approach. Estimation of the temperature prediction accuracy was fulfilled in two modes: (1) extrapolation in depth, and (2) assessment in another geological environment. The accuracy estimates of the temperature predictions at depth have shown that the relative errors depend on the ratio between the depth for which the data are available and the target depth. In particular, temperature prediction from seismic S-wave velocity data is markedly more accurate than the prediction from P-wave velocity if the target depth exceeds the source depth 5–10 times. On the other hand, in extrapolation to a depth less than twice the initial depth the average relative errors are 1 % and 2 %, accordingly. The accuracy analysis of temperature predictions in different geological environments showed that the accuracy of temperature prediction practically does not depend on the spacing between the locations of the “source” and “target” boreholes. On average, the accuracy of the prediction from P- and S-waves is approximately similar, with average relative errors of 4.8 and 4.7 %, respectively. It can be concluded that neural network prediction of the subsurface temperature from seismic velocity data can be performed with acceptable accuracy (at least, at borehole depth scale) and can be used as a seismological geothermometer.