Machine learning for robust velocity interpretation of uphole and check-shot data

Abstract Exploration of stratigraphic traps and low-relief structures strongly relies on accurate seismic imaging in the presence of near-surface complexities. Unresolved near-surface velocity anomalies adversely affect the interpretation of reservoir structures. Despite the availability of velocity modeling techniques, such as traveltime tomography and full-waveform inversion, certain geologic scenarios provide shallow complexities that necessitate direct velocity calibrations via uphole traveltime measurements. Interpretation of the uphole time-depth records in terms of velocity is typically performed through a manual operation that, in the presence of noise, may result in nonunique determinations, velocity estimations that are too coarse, or a velocity profile that is too scattered. When the operation is repeated for thousands of upholes and by multiple operators, the outcome is often an inconsistent velocity model. As machine learning (ML) algorithms continue to exhibit potential for advancing various subfields of geophysics, we introduce an ML-based approach for automatic interpretation of uphole and check-shot velocity profiles. A shallow neural network is trained with a large number of synthetically generated uphole data, with uphole traveltime as input and interval velocity as output. We test the procedure on synthetically generated uphole times contaminated with synthetic noise, where the ML approach proves to be the most robust, even when compared to manual interpretation. The tool is then developed as a software application and adopted in production for the instantaneous interpretation of thousands of upholes in several seismic exploration blocks.