Machine learning − based shale-alkane-brine contact angle prediction at in-situ reservoir conditions

Abstract

Wettability is a pivotal parameter to determine hydrocarbon reserves and production in shale reservoirs, typically characterized quantitatively by shale-oil-brine contact angles. This parameter is a complex function of shale composition, fluid properties, temperature etc. and is therefore very difficult for conventional theoretical methods to predict accurately and efficiently.

Therefore, herein machine learning methods [eXtreme gradient boosting (XGBoost) and the Shapley additive explanations (SHAP)] were integrated to predict this parameter and analyze its sensitivity. To make up the shortage of available data in literature (only 162 data points), another 100 data points about shale/mineral-alkane-water contact angles were measured under in-situ reservoir conditions using the sessile droplet method, thus totally 262 data points were used for machine learning.

Experimental results showed that shale, quartz, mica and albite became more hydrophilic with increasing temperature from 25 ℃ to 70 ℃, while K-feldspar and dolomite demonstrated the opposite trend; shale-alkane-water contact angle increased from 60° to 149° (thus wettability shifted from water-wet to oil-wet) with increasing TOC content from 1.9 wt% to 10 wt%. The XGBoost model demonstrated superior predictive accuracy than the gradient boosting regressor and support vector machine models (e.g. R² is 0.913, 0.876 and 0.623, respectively). The SHAP sensitivity analysis revealed that brine ionic strength, TOC content, calcite content and quartz content were the four most influential factors affecting wettability.

This work presents an efficient artificial intelligence method for shale wettability prediction, which is beneficial for hydrocarbon reserves estimation and production in shale reservoirs.