Abstract We employ a physics-based machine learning workflow that can be integrated with traditional geophysical analysis to improve estimates of stimulated rock volume in unconventional reservoirs, in addition to estimating resource density across flow units and benches, by significantly expanding the number of wells available for geomechanical analysis. Our workflow, tested in the prolific Permian Basin, USA, can accurately predict compressional-wave and shear-wave sonic data from existing log suites, expanding the number of wells available for geomechanical property calculation from hundreds to over tens of thousands; sonic logs can be predicted in wells that have only gamma ray, resistivity, and density logs available. Existing full-log-suite wells are used as a training dataset and for blind tests across all benches in the basin. The resulting predictions are used in calculations of various elastic properties and thus geomechanical properties, including minimum horizontal stress (Shmin). The geomechanical properties are then used to populate a 3D model representative of subsurface variability across flow units and benches with an aim at refining cube performance (e.g. production rate) by bridging the gap between conventional estimated resource density and the stimulated rock volume (SRV). Seismic interpretation of key horizons can then be used to guide extraction of geomechanical properties from the model across the target intervals. The 3D geomechanical model can also be used for fracture modeling and for optimizing completion design. Our machine learning workflow is physics-based and scalable, which can be applied to thousands of wells. This enables us to rapidly understand past well performance, generate development scenarios, and guide future completions decisions. Thus far, the machine learning workflow has been tested in unconventional reservoirs, where Shmin calculations are key to understanding the recoverable resource. The workflow itself has the potential to be extrapolated to other basins or scenarios to obtain a more robust understanding of subsurface properties impacting resource estimates and development plans on a large scale.
Liu et al. (Mon,) studied this question.