Formation lithology predictions based on measurement while drilling (MWD) using gradient boosting algorithms

Abstract

Many subsurface engineering applications require petrophysical and geomechanical properties as inputs for estimating the recoverable hydrocarbon volume and the reservoir deformation and properties' change during the life cycle of the production. Well log data plays a key role in estimating these reservoir properties, including volumes of minerals, Young's Modulus, and Poisson's Ratio. A conventional workflow using a multi-mineral petrophysical inversion model based on core-log integration requires an a priori rock-fluid model; therefore, it is often unavailable in reservoirs of complex minerals, such as carbonate or unconventional fields. In addition, most unconventional wells do not acquire Logging While Drilling (LWD) and/or wireline logs due to economic constraints or borehole instability reasons. Available measurements for these wells are typically Measurement While Drilling (MWD) Gamma Ray (GRMA) logs along with drilling dynamics measurements including Surface Weight-On-Bit (SWOB), Rate Of Penetration (ROP), Torque (TQA), Revolutions Per Minute (RPM), and the differential pressure. Thus, in this study, we demonstrate a machine learning approach, which does not require an a priori model, to predict formation lithologies based on conventional MWD and wireline well logs.We applied the Extreme Gradient Boosting Regression (XGBR) method to simultaneously predict the weight fractions of 5 minerals: QFM (Quartz-Feldspar-Mica), carbonate, clays, pyrite, and siderite from drilling dynamics measurements. The XGBR algorithm was tested with datasets from four adjacent wells in the Volve Field, located in the North Sea. A total of 10,575 data points were used to train and validate the model. The trained model was then applied to the blind test dataset of 10,446 data points from the same field. Lithology predictions from the trained model yield excellent results of weight concentrations with Root-Mean-Square Errors (RMSE) of 6.01 and 7.44% for training and test datasets, respectively. The prediction results agree with the lithology obtained from the Elemental Capture Spectroscopy (ECS) data as well as the dominant lithology record from rock cuttings. With the developed model, lithologic information and sonic logs of the formation could be predicted with MWD data, which could ultimately help optimize real-time drilling and completion operations in the field.