Predictive modeling of wax appearance temperature (WAT) using advanced machine learning algorithms: Application to oil recovery and production

Heavy oil is characterized by its high density and viscosity, frequently exhibiting higher sulfur and metal contents, with organic and inorganic deposits. This type of oil includes notable amounts of wax and asphaltenes in its composition. A critical parameter, Wax appearance temperature (WAT), governs wax crystallization in oil samples, directly influencing flow assurance and operability of pipelines. In this research, four advanced machine learning (ML) models, namely adaptive boosting (AdaBoost), categorical boosting (CatBoost), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM) are developed to predict WAT. To do this, a comprehensive databank consisting of 313 experimental measurements is employed. Wax content, cooling rate, solvent molecular weight, and wax type (macrocrystalline and microcrystalline) are considered as the input variables, enabling estimation of both paraffinic and naphthenic waxes. Results demonstrated that AdaBoost achieved the highest accuracy with an average absolute percent relative error (AAPRE) of 1.78% and a coefficient of determination (R 2 ) of 0.9947. Moreover, sensitivity and trend analyses confirmed the physical consistency of the predictions, while SHAP results identified wax type as the most influential parameter. Furthermore, outlier detection using the William’s plot showed that over 98% of data points fall within the valid region. Although the proposed models demonstrate strong predictive capability, their real‑time applicability is limited by the offline nature of several input variables. Integrating real‑time measurements and hybrid ML-thermodynamic frameworks can be applied in future work to enhance operational deployment.