Integrated detection and investigation of bad borehole section in the Wolfcamp Formation in the Midland Basin using machine learning, petrophysics, and core characterization

The quality of formation evaluation and reservoir es-timates depends a lot on the quality of open hole logs, which is intricately tied to the borehole geometry and stand-offs. Bad borehole sections cause interesting pat-terns on wireline logs that can result in multiple interpre-tations. We have conducted an automated identification of bad borehole sections (>50 ft [15 m] thick) in multiple wells in the Wolfcamp Formation in the Midland Basin of the United States and investigated the causalities behind the unstable borehole section. The wells impacted by bad hole conditions appear to be within a particular re-gion of the northern Midland Basin, and the bad hole condition impacts a specific zone in the Wolfcamp D. We consider well logs as multivariate time series (or depth series), where they share interdependencies to a large extent. We use this concept and unsupervised multivariate time series clustering to automatically iden-tify bad borehole sections. Our proposed workflow si-multaneously clusters the wireline logs, resulting in a large labeled regional-scale data set that can be further used for log quality assurance, processing, and petro-physical modeling. More importantly, we investigate the reasons behind the borehole washout, integrating sedimentology and geochemical data. Thick borehole washout sections are related to the presence of a signifi-cant amount of clay (>73 wt%, a mixture of smectite and illite) in the Wolfcamp D. The results also indicate that the manual editing of wireline logs and even supervised machine learning-based log reconstruction (commonly known as a missing log prediction problem) can lead to erroneous and inconsistent interpretation of subsur-face characteristics. Such exercise can be meaningless if derived models are not validated with the ground truth (preserved subsurface rocks).