Machine learning approach to predict variations of compressional wave velocity as function of pressure in carbonates

Understanding how the elastic wave velocities change with pressure is essential for better interpretation and modelling of time lapse seismic, as well as for geomechanics related applications. Several studies have attempted to develop empirical relations to predict velocity as function of pressure based on the measured velocities at low pressure or ambient conditions. Nevertheless, the developed relations are based on empirically determined parameters where the link to petrophysical properties and microstructure of the rock is not clear. In this study, we present an artificial neural network (ANN) model to predict compressional velocity as function of pressure in carbonate rocks. The data set consists of petrophysical properties and compressional velocities measured as function of confining pressure for 145 reservoir carbonate samples. We investigated the significance of various rock properties to select the appropriate input parameters (i.e., features) that impact the velocity-pressure relationship. In this work, we considered five different parameters: porosity, permeability, pore stiffness, dominant pore type (inter versus intra particle porosity), and pressure. Based on the results, the prediction of the ANN model outperforms that of the traditional regression approach. The ANN approach could predict the velocities with a correlation coefficient of 0.99 and average RMSE of 70 m/s.