Digital Wellbore Stability Prediction with Machine Learning

Abstract

Abstract One of the main challenges during drilling is wellbore instability. Traditionally, geomechanical model construction and wellbore stability (WBS) analysis are manually executed by geomechanics experts for well planning and drilling. The procedures are usually complicated and time-consuming due to subsurface complexity, and the results highly depend on the executor's expertise. This makes WBS analysis far from ideal and automatic. In this study, we present a physics-incorporated machine learning method that performs WBS analysis in a simple and automatic way. First, it characterizes and digitalizes subsurface geostructures geometry by labeling formations and its lithology. Then, it trains a digital geomechanics model using a series of machine learning algorithms with existing data, such as geology, well logs, drilling data, and geomechanical data. The rock mechanical properties, including rock elastic modulus and rock strength, are trained as formation material property models which describe the changing patterns in each formation. The formation pore pressure and in-situ earth stresses are trained using a physics-based hybrid algorithms, taking into account formation compaction and tectonic settings. Lastly, wellbore stability along any planned well trajectories can be predicted using this digital geomechanics model to identify drilling risks, optimize safe mud weight, and hence improve drilling practices. This digital approach was tested and validated in a shale oil field in Ordos Basin, China. In this field, horizonal wells are drilled targeting a shale oil reservoir, this requires pre-drill WBS analysis, which usually takes several weeks following a manual methodology. With the developed new method, the digital geomechanical model was trained with seven surfaces representing different geological formations and well data from six existing vertical wells. The digital model and WBS results, including formation collapse pressure, mud loss pressure and breakdown pressure, were then compared against manual results calculated by geomechanics experts using traditional methods. The digital results matched well with manual results. The comparison demonstrated the applicability and reliability with a learning accuracy of over 99%. With this digital model, the geomechanical properties and WBS analysis of five planned horizontal wells were accurately predicted and proved consistent with actual drilling results. Another significant advantage is the high computational efficiency and reduced need for supervision. In this case, the digital machine-learning method reduced the WBS analysis time for five wells from weeks to hours. This field case confirms the effectiveness and efficiency of transferring domain knowledge and data into digital models, it enables support for massive cluster horizontal drilling activities on well pad and field scale.