A team of researchers has developed a machine learning algorithm to predict porosity and absolute permeability of carbonate rock plugs. The researchers have applied six machine learning algorithms, including deep neural networks, to obtain averaged pore properties from 2D slices of 3D micro-CT carbonate rock images as input features. The proposed models have been trained on a dataset using methodologies such as geological analysis, laboratory methods for measuring rock porosity, and absolute permeability, and the image processing protocol. The images were processed using the Otsu localized algorithm on the watershed image segmentation technique to segment the 2D images into pores and rock matrices. The team conducted an analysis of 14 regional pore properties using the Variance Inflation Factor (VIF) to minimize multicollinearity between features.
The researchers also utilized the stacked ensemble approach, which involves combining predictions from multiple machine learning algorithms to improve the model’s accuracy. The stacked approach consists of two levels: the first level includes several machine learning and/or deep learning models trained independently on the same dataset, while the second level uses a meta-learner to combine the individual performances of the previous models. The proposed stacked approach involves stacking predictions from multiple linear and nonlinear machine learning-based models into a meta-learner linear model and stacking various predictions from multiple deep neural networks of different levels of model complexity into a meta-learner neural network.
The researchers achieved a high accuracy of the model predictions through the usage of these stacked ensemble approaches. The proposed models can predict the permeability and porosity of carbonated rock samples with varied pore-throat distributions, heterogeneous levels, and large ranges of permeability. The proposed algorithm could help scientists, researchers, and engineers in predicting permeability and porosity accurately. Such predictions can assist in accurate energy reserve estimation, thus improving the exploration operations in areas of high oil deposits. The approach could also aid in groundwater management, aiding in the assessment of aquifer properties.
The proposed model has been created using the Python platform and has been trained and tested using an Nvidia GeForce Titan graphics card with 12 GB of memory and core i7 of the 8th generation. The team used random parameter optimization (RSO) to identify the best training hyperparameters of each model and adopted the mean square error and mean absolute error functions to evaluate the model. The models have been trained on a dataset split of 80:20 and have been tested against the coefficient of determination (R), a goodness-of-fit measure of the model predictions to the actual targets.
