Developing a Novel Machine Learning-Based Petrophysical Rock Typing (PRT) Classification: Applied on Heterogenous Carbonate Reservoirs

Abstract Petrophysical Rock Typing (PRT) from core and well log data is a valuable tool for reservoir discrimination and recoverable reserve estimation in heterogeneous multimodal pore carbonate systems. However, the traditional PRT methods are often ineffective in these systems because of the complex pore geometry and distribution. This work introduces an innovative PRT technique that utilizes cutting-edge machine learning algorithms to categorize petrophysical rock types in a supergiant oil and gas field located in southern Iraq. The study initially employs Ward's hierarchical clustering method to define PRT from mercury injection capillary pressure (MICP) data. Ward's algorithm methodology discovers five separate groups, each of which corresponds to a different petrophysical rock type with a distinct pore geometry and system. The probabilistic neural network (PNN) approach is then used to classify PRTs from well-logging data in a cored well. The PNN algorithm is able to learn the non-linear complex relationships between well logging data and characterized PRTs, and it is able to predict PRTs with high accuracy. The confusion matrix and the total correct percent (TCP) were used as additional criteria to analyze the effectiveness of the PNN classification algorithm's performance. The study's results showed that the predicted and observed PRT had an exceptional match by attaining 94% of TCP based on the testing subset and 95 % of TCP for the entire dataset. The predicted PRT distribution is utilized as an extra independent variable to improve the modelling and prediction of core permeability and saturation height function that are essential to achieve a realistic geological distribution. The study's findings demonstrate that the suggested PRT approach can accurately categorize PRT in heterogeneous multimodal pore carbonate systems. The strategy can be easily applied to different reservoirs and has the potential to enhance the overall reservoir characterization.