A Molecular Dynamics and Interpretable Machine Learning Framework for Quantitative Viscosity-Contribution Analysis of Heavy Oil

The molecular origin of heavy-oil viscosity is difficult to quantify because compositional, structural, and interaction effects are strongly coupled. Here we propose a molecular dynamics and interpretable machine learning framework for the quantitative viscosity-contribution analysis of a single parent heavy-oil system, without introducing additional chemical variables. Starting from one digital-oil model, the SARA composition was systematically varied within physically meaningful bounds to generate 147 composition-variant systems sharing the same chemical description. Non-equilibrium molecular dynamics provided the zero-shear viscosity of each system and, based on a motif partition (aromatic core, side chain, and heteroatom), 60 pairwise interaction-energy features (ten SARA pairs × six motif pairs), of which ten non-redundant features were retained for modeling. A gradient boosting regressor predicted the base-10 logarithmic viscosity with a held-out test R2 of 0.972 (RMSE0.064), outperforming the linear baseline and other tested regressors. SHAP analysis ranked the aromatic-core AC–AC interactions among the heavy fractions, led by the asphaltene–resin pair, as the dominant microscopic driver of viscosity, while polar heteroatom association and asphaltene aromatic self-stacking form a second tier. Case studies tie the statistics to the π–π stacking networks observed in the simulations. The framework links viscosity predictions to molecular interactions and association structures through quantitative contribution analysis, providing guidance for the molecular design of viscosity reducers.