Machine Learning-Based Prediction of Pore Types in Carbonate Rocks Using Elastic Properties

Abstract

This paper explores the innovative application of machine learning and neural network algorithms to predict pore types in carbonate rocks using experimental acoustic properties under ambient pressure conditions. Carbonate reservoirs, crucial for hydrocarbon storage and extraction, present a challenge due to their complex pore structures influenced by diverse depositional environments and diagenetic processes. Traditional petrographic methods for identifying pore types, though accurate, are time-consuming and destructive. Recent approaches leverage log and core-measured compressional wave velocities and porosity, yet variability in data remains an issue. Addressing the challenge, this study distinguishes itself by employing high-resolution physical rock samples from the early Miocene dam formation, eastern province of Saudi Arabia. Through meticulous data preparation, feature engineering, and the evaluation of logistic regression, random forest classifier, gradient boosting classifier, and support vector classifier models, we have developed an advanced model capable of predicting pore types with significant accuracy. Our findings reveal that logistic regression achieves the highest accuracy (71%) among the models, effectively capturing the inherent patterns within our dataset. A detailed analysis using principal component analysis underscored the discriminative power of these models, particularly in identifying interparticle–intraparticle and moldic pore types. This study’s innovative approach, leveraging experimental measurements and machine learning techniques, offers a robust framework for accurately predicting pore types in carbonate rocks. While challenges such as data size and feature limitations persist, the potential implications of our findings for reservoir modeling and efficient hydrocarbon extraction are significant, providing a foundation for future research to build upon.