Interpretable Viscosity Prediction of Waxy Crude Oils via Physics-guided Machine Learning

• Physics-guided explainable ML framework for robust viscosity modeling. • XGBoost predicts viscosity using four parameters under dynamic conditions. • Interpretability reveals key factors and interactions governing viscosity Dynamic and accurate prediction of waxy crude oil viscosity is pivotal for ensuring flow assurance in pipeline transportation. However, the viscosity of pipeline crude oil is influenced by multifarious factors, and the inherent uncertainty of these factors often constrains the performance of conventional data-driven modeling. To address this challenge, this study proposes a highly interpretable data-driven framework for predicting waxy crude oil viscosity under practical engineering conditions. First, physical laws governing crude oil viscosity, including suspension theory, the Arrhenius equation, and the power-law model, are integrated into feature engineering. Through mathematical transformations, nonlinear physical relationships are converted into linearized features, such as reciprocal temperature (from Arrhenius), logarithmic shear rate (from power-law), and combined resin-asphaltene content (from suspension theory), to construct an input set tailored for uncertain operating conditions. Subsequently, an extreme gradient boosting model (XGBoost) is developed based on these physics-guided features to achieve high-precision estimation. Furthermore, a tree-specific interpretation method is employed to elucidate the estimating mechanisms and variation trends. Experimental results based on 2609 data points, including density, pour point, resin content, asphaltene content, shear rate, and temperature, showed that the baseline model achieved an average absolute relative error (AARE) of 5.38% and an R² of 0.9622. With the introduction of physics-guided features, the average absolute relative error further improved to 4.31%. Moreover, SHAP analysis demonstrated that the model's key influencing factors are consistent with established rheological mechanisms. The proposed framework provides a robust strategy for predicting waxy crude oil viscosity with high accuracy, enhanced interpretability, and significant practical applicability.