Prediction of the Geomechanical Properties of Anisotropic Reservoirs From Acoustic Logs Using Machine Learning

A neural network (NN) is developed to predict the elastic properties of reservoirs from acoustic data acquired during logging. The NN is applied to flexural and quadrupole slowness curves obtained from wireline and logging while drilling (LWD) acoustic instruments. Results show that the velocities are estimated with high accuracy for isotropic and vertical transversely isotropic formations vertically transverse isotropic (VTI), and that the method is stable and fast. The NN model yields accurate estimation for velocities in all directions, with the best results observed for shear velocity ( ${V}_{s}$ ) due to higher sensitivity of the slowness curves to variations in $V_{s}$ . This work demonstrates that the NN can be accurately trained with only a few thousand data points generated. Moreover, the velocities are estimated using a single slowness curve (either the flexural slowness for wireline or quadrupole slowness for LWD tools) without needing other formation parameters. We also provide recommendations and lessons learned to further improve the prediction of formation properties. Notably, we show that for VTI reservoirs, expressing the elastic properties in terms of stiffness coefficients or velocities is preferable to using Thomsen’s parameters. The methodology is fast and accurate, which facilitates real-time reservoir characterization. It is also computationally efficient since NNs are trained once and predict formation parameters from a dispersion curve with a single forward pass.