Predicting multi-component oil viscosity using machine learning methods

Chemical Engineering Sciences. 2027. Vol. 337. Article 124616.

Deshchenya V., Nikolay D. Kondratyuk, Sivakova A., Sushkov O., Kozhevnikov D., Petrov V., Aliev T., Muravev A., Michael G. Medvedev, Skorb E.

Accurate prediction of complex formulation viscosity is crucial for the lubricant industry and engineering processes. However, developing generalizable models remains challenging due to the complexity of industrial mixtures and the scarcity of public experimental data, often resulting from corporate proprietary restrictions. This study presents a machine-learning workflow for viscosity estimation from a proprietary pseudonymized dataset, showing that robust models can be trained under data-protection limitations. We evaluated multiple machine learning architectures, including Decision Tree, Random Forest, Support Vector Regression, and Gradient Boosting, employing feature selection strategies adapted to pseudonymized data. Additionally, the Walther formula was used as a physics-based reference. While this formula required deanonymized multi-temperature viscosity data, it proved valuable for identifying experimental outliers. Among the evaluated models, Gradient Boosting demonstrated superior performance, achieving a median absolute percentage error of 16.2% on the held-out test set. These results indicate that machine learning can support viscosity screening for complex industrial fluids without requiring full compositional transparency.

Keywords: machine learning, materials science, Kinematic viscosity, Lubricating oils