Predicting Oil Sludge Formation During Crude Oil Production in the Niger Delta Using Machine Learning Models

Abstract Sludge formation during crude oil production presents significant flow assurance challenges, including reduced throughput, equipment fouling, and increased maintenance costs. Accurate prediction of sludge mass is therefore essential for proactive mitigation and optimized production management. This study develops and validates a machine learning-based predictive framework for estimating sludge formation using production data from OML 4, 38, 41, 53, and 55. An experimental dataset comprising 144 crude oil samples was analyzed, incorporating compositional, thermophysical, and operational parameters, including a wide range of SARA fractions, BS&W, flowing temperature and pressure, API gravity, viscosity, Wax Appearance Temperature (WAT), and wax content. A structured modeling workflow was implemented using leakage-free preprocessing, hyperparameter optimization, learning curve diagnostics, and nested cross-validation to ensure statistical robustness. Five regression algorithms were evaluated: Random Forest, Gradient Boosting, XGBoost, LightGBM, and Polynomial Support Vector Regression (PolySVR). Nested cross-validation results demonstrated that PolySVR achieved superior generalization performance, with an average R2 of 0.962 ± 0.005. Independent test evaluation confirmed strong predictive accuracy (R2 = 0.976; RMSE = 0.00876 g sludge/g oil). Ensemble boosting methods also performed competitively but exhibited higher variance under limited data conditions. Feature importance analysis identified wax content as the dominant predictor of sludge mass, followed by API gravity, BS&W, and WAT, consistent with established wax crystallization and aggregation mechanisms. The findings indicate that sludge formation in the evaluated assets exhibits smooth nonlinear behavior primarily governed by compositional and thermodynamic factors. The proposed framework is computationally efficient and suitable for integration into real-time production surveillance systems, providing a practical tool for early sludge risk detection and improved flow assurance management.