(140e) Data-Driven Modeling to Predict the Physical Properties of the Lubricant

Lubricants are essential to reduce friction on machines or vehicles, and the required properties of lubricants vary widely depending on the type or specification of the applied device. Also, they have various physical properties by blending many different ingredients¹. For this reason, it is crucial to find the recipe of the specific lubricant with the required physical properties. However, it takes much time and cost to find the specific recipe for producing the target lubricant, because the proper recipe for the target lubricant has been found by numerous experiments with trial and error. Furthermore, lots of ingredient types and their combinations cause more repeated experiments. Hence, this study suggests data-driven modeling to predict the physical properties of lubricants to solve the time-and-cost-consuming. First, the lubricant dataset, which consists of 830 recipes with 55 ingredients and 3 main properties (viscosities at 40℃ and 100℃, and density)², is preprocessed using categorization³. Second, multiple linear regression, random forest, and catboost^4–6 are applied to data-driven modeling and the developed models are evaluated with R². As a result, each catboost-based prediction model for viscosities at 40℃ and 100℃, and density shows the highest R² with 0.9977, 0.9962, and 0.9596, respectively. Therefore, the catboost-based models have the best prediction performance and they are expected to solve the time-and-cost-consuming problem in the target lubricant production.

Literature cited:

[1] Martini A, Ramasamy US, Len M. Review of Viscosity Modifier Lubricant Additives. Tribol Lett. 2018;66(2):1-14. doi:10.1007/s11249-018-1007-0

[2] Marinović S, Jukić A, Doležal D, Špechar B, Kritović M. Prediction of used lubricating oils properties by infrared spectroscopy using multivariate analysis. Goriva I Maz. 2012;51(3):205-215. http://hrcak.srce.hr/index.php?show=clanak&id_clanak_jezik=132656

[3] Joo C, Park H, Kim J. Development of physical property prediction models for polypropylene composites with optimizing random forest hyperparameters. 2021;(January). doi:10.1002/int.22700

[4] Liu W, Li M, Zhang M, et al. Estimating leaf mercury content in Phragmites australis based on leaf hyperspectral reflectance. Ecosyst Heal Sustain. 2020;6(1). doi:10.1080/20964129.2020.1726211

[5] Bentéjac C, Csörgő A, Martínez-Muñoz G. A Comparative Analysis of Gradient Boosting Algorithms. Vol 54. Springer Netherlands; 2021. doi:10.1007/s10462-020-09896-5

[6] Zhang Y, Zhao Z, Zheng J. CatBoost: A new approach for estimating daily reference crop evapotranspiration in arid and semi-arid regions of Northern China. J Hydrol. 2020;588:125087. doi:10.1016/J.JHYDROL.2020.125087