Due to its ability to solve numerous problems associated with traditional water-based fracturing, CO2 foam utilization in unconventional reservoir fracturing has garnered a great deal of interest owing to its ability to overcome various issues connected with traditional water-based fracturing. However, the instability of CO2 foam under severe reservoir conditions necessitated more study into foam stabilization using other compounds, including nanoparticles. The rheology of nanoparticle-stabilized-CO2-foam (NP-CO2 foam) determines how successful the fracturing operation would be to a great extent. A variety of parameters, including temperature, foam quality, salinity, nanoparticle concentration, and shear rate affect this rheology. The design of the optimal injection strategy is based on the quantification of these parameters under tough reservoir conditions that are too expensive and/or time-consuming to achieve in laboratory studies and with mechanistic foam models. Empirical correlations are unable to make rapid and accurate predictions when different combinations of all these parameters are considered. Therefore, this paper explores the use of simple-to-use, relatively fast, and versatile machine learning (ML) methods in a predictive study. Apparent viscosity for NP-CO2 foam is determined while considering all input parameters.

In this work, we compare the performance of four data-driven non-linear ML algorithms: Multilayer Perceptron Neural Network, Support Vector Regression, K-Nearest Neighbor, and Multivariate Polynomial Regression. We apply these algorithms to estimate apparent viscosity, a key rheological property of NP-CO2 foams when used as fracturing fluid. A dataset containing 262 experimental data records were utilized in the training, optimization, and testing of resultant models.

When evaluated on the test data, all the models had a prediction accuracy of at least 97 percent, with the radial basis function-kernelized Support Vector Regression (SVR) model providing the most accurate estimation efficiency (MSE: 0.17, RMSE: 0.41cp, MAE: 0.23, R2:0.99). The SVR model's algorithm also took the shortest amount of time to run out of all the models. For the NP-CO2 foam, the SVR model showed that temperature was the most important factor in predicting its apparent viscosity. This ML approach allows a thorough understanding of the nonlinear relationship between the investigated factors and apparent viscosity. Therefore, in cases where precise laboratory experimental measurements are not available or too expensive to obtain, the SVR algorithm can be considered as a complementary tool for the efficient assessment of the effects of NP-CO2 foam fracturing fluid.

This content is only available via PDF.
You can access this article if you purchase or spend a download.