Advancing Wax Appearance Temperature Prediction in Crude Oil Through Physics Informed Neural Networks

Abstract Wax deposition remains one of the most persistent flow assurance challenges in crude oil production and transportation, occurring when fluid temperature falls below the Wax Appearance Temperature (WAT), leading to paraffin crystallization, increased viscosity, and potential pipeline blockage. Accurate prediction of WAT is therefore essential for pipeline design, thermal management, and mitigation strategies. Conventional laboratory methods such as Differential Scanning Calorimetry and viscometry are reliable but often slow, costly, and limited in scalability, while empirical correlations and data-driven models including Artificial Neural Networks, Support Vector Machines, and ensemble regressors provide strong predictive performance but lack physical consistency and may generalize poorly under unseen conditions. This study presents a hybrid predictive framework that combines a feed-forward neural network with a Physics-Informed Neural Network (PINN) to estimate WAT from crude oil composition data from a field in Niger Delta. The PINN formulation incorporates thermodynamic phase-equilibrium constraints governing wax precipitation into the training process, ensuring that predictions remain physically consistent, while transport-related considerations are included only in relation to measurement and system behaviour. The baseline regressor achieved strong statistical performance with R2 = 0.956, MAE = 1.85 K, and RMSE = 2.39 K. The PINN improved the predictive accuracy to R2 = 0.983, reduced the mean absolute error to 0.884 K, and lowered RMSE to 1.77 K, corresponding to approximately 34% reduction in error magnitude and 26% improvement in predictive precision. The low physics residual (~1.6 × 10-3) further confirms convergence toward physically consistent solutions. Compared to other machine learning models such as Support Vector Machines (R2 = 0.980) and Random Forest (R2 = 0.978), the developed PINN in this study demonstrated greater interpretability, stability, and extrapolative strength under unseen operating conditions. The results show that embedding thermodynamic constraints into neural network training enhances both predictive reliability and physical consistency, particularly in data-sparse conditions. This framework provides a practical and scalable tool for improving flow assurance decision-making in wax-prone crude oil systems, especially in regions where experimental data are limited.