Machine learning for seismic exploration: Where are we and how far are we from the holy grail?

ABSTRACT Machine-learning (ML) applications in seismic exploration are growing faster than applications in other industry fields, mainly due to the large amount of acquired data for the exploration industry. The ML algorithms are constantly being implemented for almost all the steps involved in seismic processing and interpretation workflow, mainly for automation, processing time reduction, efficiency, and in some cases for improving the results. We carry out a literature-based analysis of existing ML-based seismic processing and interpretation published in SEG and EAGE literature repositories and derive a detailed overview of the main ML thrusts in different seismic applications. For each publication, we extract various metadata about ML implementations and performances. The data indicate that current ML implementations in seismic exploration are focused on individual tasks rather than a disruptive change in processing and interpretation workflows. The metadata indicate that the main targets of ML applications for seismic processing are denoising, velocity model building, and first-break picking, whereas, for seismic interpretation, they are fault detection, lithofacies classification, and geobody identification. Through the metadata available in publications, we obtain indices related to computational power efficiency, data preparation simplicity, real data test rate of the ML model, diversity of ML methods, etc., and we use them to approximate the level of efficiency, effectivity, and applicability of the current ML-based seismic processing and interpretation tasks. The indices of ML-based processing tasks indicate that current ML-based denoising and frequency extrapolation have higher efficiency, whereas ML-based quality control is more effective and applicable compared with other processing tasks. Among the interpretation tasks, ML-based impedance inversion indicates high efficiency, whereas high effectivity is depicted for fault detection. ML-based lithofacies classification, stratigraphic sequence identification, and petro/rock properties inversion exhibit high applicability among other interpretation tasks.