Permeability Prediction of Carbonate Cores With Gaussian Process Regression Model

Abstract

Abstract Machine leaning (ML) methods are widely adopted in predictions affected by various factors. This paper presents a step-by-step workflow of applying a ML approach to develop a heterogeneous permeability prediction model from the CT images of core samples. In this work, over ten thousand 3-D sub-image were randomly extracted from the CT images of two heterogeneous carbonate core samples. The permeability of each sub-image is simulated using pore network modeling (PNM) method. Ten features including porosity, pore size, surface area, specific surface area and connection coefficient etc. are extracted from sub-image by a statistical method. Three training datasets were built with features and permeability. Each set of training data is input into a ML model pool, which contains 19 regression models of 5 types including linear regression models, regression trees, support vector machines, Gaussian process regression models and ensembles of trees. Then, regression models are trained to identify the one that can yield the best permeability prediction. The trained model with the highest R-Squared value is selected for permeability prediction from binary CT images. Overall, comparing the training outputs indicate that Gaussian Process Regression models (GPR) correlate features and permeability well. For the tested heterogeneous core plugs, the exponential Gaussian Process model performs the best. The R-Squared values of the three sets of training data are 0.88, 0.87 and 0.91 respectively. Afterwards, the selected ML model was tested with additional data, and the R-squared value of each test dataset was greater than 0.85, confirming a strong predictive performance. The trained model based on ML method eliminates the conventional time-consuming operations including distance transformation and watershed segmentation. It also avoids excessive memory consumption, which makes the method suitable for images with large size. The paper provides a way to develop an alternative approach of PNM simulation method for permeability prediction from CT images.