Rock Physics and Machine Learning Analysis of a High-Porosity Gas Sand in the Gulf of Mexico

Abstract

Rock physics transforms established on the well data play an important role in predicting seismic rock properties. However, a data-driven approach, such as machine learning, can also estimate the targeted outputs from the well data. This study aims at comparing the accuracy of rock physics and machine learning analyses for the prediction of the P-wave velocity of porous rocks at the well log scale by employing the well data from the Mississippi Canyon, Gulf of Mexico. Rock physics diagnostics (RPD) was used as a physics-driven methodology for predicting the P-wave velocity, while artificial neural network (ANN) was used as a machine learning approach. To train the neural network, the well data were divided into two sections where the ANN model was optimized on the upper well data interval and tested in the lower interval. During the rock physics analysis, the lower interval was employed to compare the obtained results from the physics-driven and data-driven approaches in the same well interval. Based on the results from RPD, the constant cement model with a high coordination number describes the well data under examination. The established rock physics model is used for predicting elastic properties of rocks, including the P-wave velocity from measured petrophysical properties, namely porosity, mineralogy, and the pore fluid. However, the mineralogy input, such as the clay content, was missing in the well data. Therefore, the clay content was calculated from the gamma ray log and used in the rock physics model established. On the other hand, the ANN model was developed and tested using well log inputs such as porosity, gamma ray, and resistivity logs. Results showed that the accuracy of the machine learning model outperforms that of the rock physics model in the prediction of the P-wave velocity. In particular, a correlation coefficient (R) of 0.84 and absolute average percentage error (AAPE) of 2.71 were obtained by the ANN model, while the constant cement model reached CC of 0.65 and AAPE of 4.07. However, one should be aware that the computed clay content from the gamma ray log was a major factor in obtaining low CC compared to the ANN model as it significantly introduced uncertainty in our computations.