Prediction of Sanding Potential in Unconsolidated Sandstones Using Selected Machine-Learning Techniques

Abstract Extracting hydrocarbons from unconsolidated sandstone formations poses significant operational challenges, particularly sand production, which can affect well integrity, equipment durability, and overall production efficiency. Prior research has primarily focused on empirical and analytical models requiring extensive production data, which hinders their use for proactive measures. This study overcomes these limitations by leveraging advanced machine learning techniques to predict sand production in a more accurate and efficient manner, thereby enhancing sand control strategies in oil and gas extraction. This research utilized a dataset from Triassic sandstones, characterized by oil presence and a thin gas cap, located at a depth of 1100 meters below the seabed. The dataset which was derived from sonic logs, included parameters such as compressional wave transit time (DTC), shear wave transit time (DTS), pore pressure, overburden pressure, and horizontal stress. The methodology involved developing and validating several machine learning models to predict the unconfined compressive strength (UCS) of the sandstone—a key indicator of its susceptibility to sand production. The models, trained using a supervised learning approach, included linear regression, Elasticnet Regression, and Lasso Regression, which were evaluated based on their ability to predict sand production rates and conditions. The models demonstrated that Lasso Regression was the most effective, achieving an R2 score of 0.977 and an RMSE of 133.45, indicating high predictive accuracy and minimal error. The success of the Lasso Regression model underscores its utility in predicting sand production, making it a valuable tool for proactive sand management in oil and gas operations. The findings of this study not only advance academic understanding of predictive modeling in geotechnical engineering but also offer practical solutions that can enhance the efficiency and sustainability of extraction practices. The results advocate for a broader adoption of data-driven, machine learning-based methodologies within the oil and gas industry.