Exploring for Wolfcamp Reservoirs, Eastern Shelf of the Permian Basin, USA, Using a Machine Learning Approach

Abstract

The objective was to leverage prestack and poststack seismic data in order to reconstruct 3D images of thin, discontinuous, oil-filled packstone pay facies of the Upper and Lower Wolfcamp formation (Sakmarian time: 293-296 Ma). The well-to-seismic tie was carefully established using synthetic seismograms, which enabled the facies log to be properly associated with the corresponding seismic samples. The seismic data were all resampled from 2 ms to 0.5 ms in anticipation of being able to recover facies thicknesses on the order of 2 m. Six neural networks with diverse learning strategies were trained to recognize the nine facies classes in the high-resolution seismic stack: Instantaneous Frequency, Instantaneous Q Factor, Inversion (P-Impedance), Semblance, Dominant Frequency, Most Negative Curvature, and eight Angle Stacks, using a two-stage learning and voting process. At the wells, the nine facies were reconstructed from seismic at a 97% accuracy rate. The bootstrap classification rate, a proxy for blind well testing, was over 80%, which indicates a high-quality modeling process. The pay facies was described with no false positives or false negatives. In the 3D seismic volume between the wells, the procedure produced a Most Likely Facies volume (unsmoothed and smoothed), and nine individual Facies Probability volumes. The pay facies was visualized in a 3D voxel visualization canvas using opacity, and also in a two-way time thickness map. The usable vertical and horizontal resolution was much greater than that of the original seismic. Based on these classification results, additional drilling locations were chosen to further target the oil-filled packstones. The classification results were created by neural networks, which can be used as a substitute for traditional AVO, inversion and cross-plotting techniques for seismic reservoir characterization. The time need to create the Machine Learning results for this small dataset was on the order of ten minutes.