Artificial Intelligence Application on Seismic Data for Automatic First-Break Arrival Picking

Abstract

Automatic arrival-picking schemes for seismic data are indispensable due to large amount of digital data recorded by wide seismic networks and the need for rapid analysis. Until now, the automation process has been applied only to seismological studies for earthquake analysis. In this study, the authors took the initiative to use the framework of automatic arrival picking and modified it for its application on Vertical Seismic Profiling (VSP) data, a class of borehole measurements used for correlation with surface seismic data to obtain higher resolution images than surface seismic images for oil and gas exploration by industries. An accurate estimation of VSP first breaks is important for VSP data processing and velocity calculation which forms the backbone of subsurface modelling for hydrocarbon exploration. The authors tested different arrival-picking algorithms on VSP data, which were based on finding changes in seismic trace attributes such as change in energy, power, amplitude or statistical properties of the seismic signal in time or frequency domain. The authors built and applied 7 arrival-picking algorithms: Short-term Average/Long-Term Average (STA/LTA) Ratio, Kurtosis, Skewness, Modified Energy Ratio (MER), Peak amplitude of a half cycle, Peak-to-Lobe Difference and Entropy Method. The authors then selected 4 best algorithms corresponding to high correlation coefficients by the method of Principal Component Analysis, which were fed to the deep learning Neural Network for its training of first-break arrival picking. The authors employed the Artificial Neural Network approach due to its adaptive and robust nature, quintessential for accurate first-break arrival picking of large volumes of complex seismic data. The authors used the trained neural network to pick first-breaks in a more realistic VSP data set. In these experiments, the authors found that the arrival picks from the trained neural network gave a very close approximation of the first-break arrival picks. The work thus provided valuable insight into the applicability of rapidly emerging Machine Learning techniques, in particular the Artificial Intelligence (AI) technology by deep learning Neural Networks, to seismic studies. The integration of AI with seismic studies for automating VSP data processing has not been explored as on date as per the author’s knowledge. The study will transform the data processing methods of the industry and will facilitate faster and more reliable seismic data processing with significant reduction in manual effort employed in the process, thus aiding in efficient and cost-effective oil and gas exploration.

AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019