Improved model development and feature ranking for rock permeability prediction by coupling petrophysical log data and ensemble machine learning techniques

Permeability is a crucial petrophysical rock parameter to assess the reservoir flow ability to capture the reservoir quality and reservoir characterization during petroleum field development. Due to the importance of permeability to minimize the risk of exploration and production forecasting, geoscientists and petroleum engineers should be able to make reliable permeability for sedimentary rock. The study objectives are to examine the effects of petrophysical parameters and parametric sensitivity analysis to develop an improved model while predicting formation permeability (K). The predictive models are constructed by coupling petrophysical log parameters and ensemble machine learning approaches, including random forest and extreme gradient boosting (XGB). To verify the predictive model outcomes for feature ranking, the least square support vector machine with a global approach of coupled simulated annealing is applied, and assessed the accuracy of models using statistical performance indicators. It is found that the XGB-based model achieved better accuracy than other ensemble machine learning approaches. The primary log variables such as bulk density and compressional acoustic travel time are adopted to obtain improved correlation for K estimation. The improved model of K is developed with the least absolute residuals (LAR) technique and it is compared and verified with existing correlation using real field log data which performs a high correlation coefficient (99%) and minimal average error. The improved model of rock K can be adopted to assess the reservoir rock quality in simulation studies for the oil and gas field development stages. The studied model development strategies can be applied for further assessment of reservoir rock and fluid properties in reservoir studies and management with low costs on time.