Predicting Mud Weight Window from Well Logs by Integrating Deep Neural Networks and Physics Models

Abstract

The selection of an appropriate mud weight is important in drilling operations, as it plays a pivotal role in mitigating the potential for costly wellbore instability issues. The safe mud weight window is typically computed through analytical solutions that necessitate detailed rock properties as integral inputs. Conventionally, these rock properties are estimated based on well logs through empirical correlations. This paper introduces a wellbore stability analysis workflow that makes two changes to the conventional methodology. First, microporomechanics models are used to upscale the nano and micro properties of the mineral constituents to the macro rock properties. Unlike the correlation methods, this scientific approach can explain the origin of the rock properties. To help get the mineral composition data, a deep neural network (DNN) is trained on 15,979 data points to predict the volume fractions of silt inclusions, clay, and kerogen from gamma ray, resistivity, density, neutron porosity, and photoelectric logs. Second, another DNN is used in the workflow to speed-up the analytical solution for mud weight window computation. This DNN is trained to predict the mud weight window from in-situ stresses, pore pressure, well trajectory, and the rock properties. Its prediction is used as the starting point in the analytical wellbore stability solution to quickly determine the correct mud weight window. To demonstrate the practical application of this workflow, evaluations were conducted using a 480-foot shale well segment comprising 961 depth intervals. The results show that the hybrid approach can calculate 961 mud weight windows 5 times faster than the purely analytical solution.