Machine learning-based surrogate correlation for prediction of wax deposition rate in oil and gas pipelines

Abstract

This study developed a data-driven surrogate-box correlation to predict wax deposition rates in crude-oil pipelines and help address the limited generalizability of empirical and mechanistic models under variable flow and thermal conditions. Literature-derived dataset of 215 experimental and field observations was compiled with inputs as oil temperature , wall temperature , dynamic viscosity , wall shear stress , flow velocity , wall temperature gradient , and wall concentration gradient , and target as wax deposition rate (g/m 2 .h). After outlier control, standardization, and ANOVA-based feature engineering, Five supervised black-box models; SVR, Random Forest, MLP, Gradient Boosting, and KNN, models were trained with an 80/20 train–test split, 5-fold cross-validated, hyperparameter-tuned and evaluated using MSE, RMSE, MAE, , and AAPRE. The top performing model, KNN, was modeled into an interpretable surrogate via Elastic Net (benchmarked against polynomial Ridge degree-2/3) to yield a closed-form correlation. KNN achieved excellent test performance ( , RMSE ⁓ 0.101, AAPRE ⁓ 5.09%), outperforming alternative models. The Elastic Net surrogate preserved predictive ability with balanced generalization ( ≈ 0.61–0.65 for train/validation/test) while exposing parameter level effects and nonlinear interactions among temperature, viscosity, shear stress, velocity, and wall-scale gradients. When compared to a physics-based correlation, the surrogate exhibited tighter clustering to measurements, which indicates improved field relevance. The principal contribution of this study is a hybrid workflow that couples high-accuracy black-box learning with transparent surrogate modeling to help enable real-time monitoring and control, proactive pigging scheduling, and optimized chemical/thermal treatments of oil flow in pipelines. The resulting correlation offers a deployable, interpretable alternative to nontransparent machine learning models or assumption-heavy empirical relations, with clear value for flow-assurance planning and operational reliability.