Hybrid machine learning–geochemical modeling for predictive assessment of barium sulfate scaling in oilfield fluids: a case study from the Haoud Berkaoui formation

Barite scaling is a critical flow-assurance problem in oilfield water systems, and rapid estimation of saturation state from routine water chemistry can improve monitoring and mitigation. This study aimed to develop and evaluate an interpretable thermodynamic–machine-learning framework for predicting barite saturation index in produced water from the Haoud Berkaoui field, Algeria. A thermodynamically calculated saturation index derived from measured water chemistry and temperature was used as the reference target for 1202 produced-water records. Four routinely measured variables—pH, temperature, salinity, and total dissolved Fe—were used to train and compare Linear, Ridge, Lasso, Random Forest, and Gradient Boosting regression models. The ensemble models clearly outperformed the linear models. Gradient Boosting achieved the best test performance (R2 = 0.9984, RMSE = 0.0356, MAE = 0.0154), followed by Random Forest (R2 = 0.9976, RMSE = 0.0437, MAE = 0.0144), whereas the linear models showed weak predictive capability. Shapley Additive Explanations (SHAP) and response-surface analyses showed that salinity was the dominant predictor, while temperature and Fe had secondary effects and pH had only a minor contribution. By integrating thermodynamic saturation-index calculation, interpretable machine learning, and deployable digital screening, the proposed framework supports near-real-time identification of undersaturated, near-equilibrium, and supersaturated conditions in oilfield water-management operations.