Application of Machine Learning to Predict Shale Wettability

Abstract CO2 wettability in shale formations is an important parameter for different applications including, CO2 EOR, CO2 sequestration in saline aquifers where the shale formations are the seal cap rock, CO2 sequestration in the shale formation, and hydraulic fracturing process in shale. Different experimental work can be used to estimate the wettability including quantitative and qualitative methods such as contact angle, Amott method, NMR, flotation methods, relative permeability, and recovery curves. In addition to the difficult surface preparation processes, laboratory experiments take a lot of time, money, and effort. Therefore, this paper seeks to use various machine-learning tools to calculate the contact angle which is an indication of the shale wettability. A collection of 200 data points was gathered for various shale samples under varying conditions. Machine learning models such as linear regression (LR) and Random forests (RF) were employed to forecast the wettability of shale-water-CO2 as a function of shale characteristics, pressure, temperature, and water salinity. The data was randomly divided into two parts with a 70:30 training-testing ratio. A separate, unseen set of data was used to validate the predictive models. The results indicated that the most significant factors impacting shale wettability are, among others, operating pressure and temperature, total organic content (TOC), and mineral matter. The linear regression (LR) model was employed to evaluate the linear dependence of contact angle values on the input parameters, but it failed to accurately predict the contact angle for several points with an R2 value lower than 0.8. In contrast, the Random Forest (RF) model accurately forecasted the contact angle in the shale-water-CO2 system based on shale properties and system conditions with a high R2 of 0.99 for the training dataset and 0.95 for the testing dataset. The root mean square error (RMSE) was less than 6 degrees for both training and testing datasets in both models. The developed model was validated using unseen data and the correlation coefficient between the actual and predicted contact angle was found to be above 0.94. This study demonstrates the dependability of the suggested models in determining the contact angle in the shale-water-CO2 system based on shale properties, pressure and temperature, and water salinity, eliminating the requirement for intricate measurements or calculations through experimentation.