Robust-tuning machine learning algorithms for precise prediction of permeability impairment due to CaCO3 deposition

• Published in: Scientific reports Vol. 15; no. 1; pp. 29967 - 24
• Main Authors: Khodabakhshi, Mohammad Javad, Bijani, Masoud, Hasani, Masoud
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 15.08.2025; Nature Publishing Group; Nature Portfolio
• Subjects: 639/4077/4082; 639/4077/4082/4061; 639/4077/4082/4090; Algorithms; Artificial intelligence; Barium; CaCO3 deposition; Calcium carbonate; Calcium permeability; Carbon dioxide; Chemical reactions; Decreased permeability; Deep learning; Enhanced oil recovery; Flow velocity; Humanities and Social Sciences; Hydrochloric acid; Hyperparameter tuning; Investigations; Learning algorithms; Machine learning; Machine learning methods; multidisciplinary; Neural networks; Oil fields; Oil recovery; Permeability; Predictions; Regression analysis; Reservoirs; Salinity; Scale; Science; Science (multidisciplinary); Sedimentation & deposition; Sediments; Temperature; Water flooding; deposition; Decreased permeability; Scale; CaCO; Machine learning methods; Hyperparameter tuning
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-11267-5

Scale buildup, especially calcium carbonate (CaCO₃), is a common problem in Enhanced Oil Recovery (EOR) operations, often caused by injecting incompatible water or by changes in pressure and temperature that trigger chemical reactions. This buildup can clog reservoirs, damage wells, and affect surface equipment by reducing permeability. This study explores how factors like temperature, pressure, pH, and ion concentration influence CaCO₃ deposition and how it affects reservoir performance. Using machine learning models—Support Vector Regression (SVR), Extra Trees (ET), and Extreme Gradient Boosting (XGB)—the research aims to predict how much permeability is lost due to scaling. With proper tuning of these models, prediction accuracy significantly improved: SVR rose from 92 to 99.88%, and XGB reached 99.87%, while ET remained consistently high at around 99.98%. The real value of this work lies in building a fine-tuned, practical machine learning approach that applies proven models to real-world EOR challenges. Instead of creating new algorithms, the study focuses on refining existing ones to make them more effective for the field. These accurate predictions can help engineers make smarter decisions about maintaining wells and reservoirs, ultimately improving efficiency and cutting operational costs.