24th International Colloquium Tribology - January 2024 163 A Machine Learning approach to Tribological Performance Prediction of New Lubricant Formulations Wahyu Wijanarko 1* , Nuria Espallargas 1* 1 Norwegian University of Science and Technology * Corresponding author: wahyu.wijanarko@ntnu.no, nuria.espallargas@ntnu.no 1. Introduction Research on finding new Environmentally Acceptable Lubricants (EALs) has become more critical due to new regulations requiring lubricants to be non-toxic, readily biodegradable, and non-bioaccumulative [1]. New regulations put limit on the usage of petroleum-based oils, especially for their use in the maritime industry, where there is a potential continuous leakage from different parts of the vessel to the sea. A fully formulated lubricant consists of 70% to 99% base lubricant and 30% to 1% lubricant additives. To be considered as an EAL, both base lubricant and additives need to be environmentally acceptable. Examples of environmentally acceptable base lubricants are vegetable oils, synthetic esters, glycols, low viscosity polyalphaolefins, and water [2]. If those are used as base lubricants to formulate an EAL, the toxicity of the fully formulated lubricants will be controlled by the additives used in the formulation. Finding the right additives for environmentally acceptable base lubricants is challenging due to the limited types of chemicals that are both, compatible with the environmentally acceptable base lubricants, and can perform as lubricant additives. Our research focuses mostly on water-based lubricants as a family of EALs. Water-based lubricants are interesting lubricants due to their readily environmentally acceptability [1,2]; however, they have poor performance compared to mineral or synthetic oils. Therefore, finding new additives for water-based lubricants becomes our main goal. One family of chemicals that is attracting attention due to their potential benefits as lubricants is ionic liquids. Ionic liquids are organic salt consisting of cation and anion moieties [3]. Ionic liquids are environmentally acceptable chemicals due to their low volatility, non-flammability, and high thermal stability [4]. In addition, ionic liquids are highly polar substances, meaning they are highly surface active with a tendency to adsorb on the metallic surfaces, making them suitable as additives [5]. There are large numbers of cations and anions precursors available for ionic liquids synthesis, therefore over 1 million possible combinations of ionic liquids can be produced. Nowadays, around 300 to 400 ionic liquids are commercially available. Such large availability of ionic liquids will take huge number of experiments to study their tribological performance, therefore machine learning can be a great tool to predict the tribological performance of water-based lubricants formulated with ionic liquids using small experimental datasets saving time in formulating new lubricants. In the future and with enough results, machine learning can also be used to predict the relationship between structure and performance to propose new ionic liquid chemistries. 2. Methodology In this work, the coefficient of friction of water-based lubricants formulated with ionic liquids was predicted using machine learning. The dataset consisted of water (W), four glycols, four water-glycols, and two polyalphaolefins as the base lubricants. Glycols that were used are monoethylene glycol (MEG), diethylene glycol (DEG), monopropylene glycol (MPG), and dipropylene glycol (DPG). Polyalphaolefins that were used are polyalphaolefin with 2 cSt viscosity (PAO2) and polyalphaolefin with 8 cSt viscosity (PAO8). Water-glycols were formulated by mixing the water with each glycol in a ratio of 1: 1, namely, water-monoethylene glycol (WMEG), water-diethylene glycol (WDEG), watermonopropylene glycol (WMPG), and water-dipropylene glycol (WDPG). Each base lubricant was formulated with six ionic liquids, namely, 1,3-dimethylimidazolium dimethylphosphate (IM), (2-hydroxyethyl)trimethylammonium dimethylphosphate (AM), tributylmethylphosphonium dimethylphosphate (PP), 1-butyl-1-methylpyrrolidinium tris(pentafluoroethyl)trifluorophosphate (BMP), trihexyltetradecylphosphonium bis(2,4,4-trimethylpentyl)phosphinate (PB), and trihexyltetradecylphosphonium decanoate (PC). One weight percent (1 wt.%) of additives is utilized for the lubricant formulation. In addition, the base lubricant alone and the individual ionic liquids were also tested. Each test was performed twice to check the repeatability. In total, 83 lubricants were tested to generate 166 samples for building up the experimental dataset. Each sample in the dataset needs to be transformed to a machine-readable format to be able to be used in the machine learning programming. The transformation required Simplified Molecular Input Line Entry System (SMILES) to represent each chemical structure in the lubricant. As explained in the previous paragraph, seven compounds were used to formulate the base lubricants and six ionic liquids were used as lubricant additives, so in total 13 chemicals were involved in the lubricant formulation. The SMILES for each chemical can be obtained from the database webpage, for example PubChem from National Library of Medicine - National Institutes of Health (NIH) or CompTox from United States Environmental Protection Agency (EPA) [6,7]. Once the SMILES is obtained, the molecular descriptors were calculated using AlvaDesc software [8]. The software calculates 4179 molecular descriptors for each chemical based on its physical and chemical properties. Cleaning (curing) the molecular descriptors was done by removing the molecular descriptors that had constant and missing values throughout the 13 chemicals, resulting in 912 molecular descriptors that could be used for machine learning. In addition to the molecular descriptors generated A Machine Learning approach to Tribological Performance Prediction of New Lubricant Formulations 164 24th International Colloquium Tribology - January 2024 by AlvaDesc, nine experimental molecular descriptors were also added to the dataset, namely, pH, electrical conductivity, molecular weight, number of C, H, O, N, P, and F atoms. So, in total 921 molecular descriptors for the 13 chemicals were generated. As explained before, 83 lubricants were formulated using 13 chemicals, thus the molecular descriptors for each lubricant were calculated based on the volumetric ratio of the chemicals in the lubricant. For the machine learning part of the work, Python version 3.9.13 combined with Jupyter Notebook version 6.4.12 were used as the program language. Scikit-learn free online package tool version 1.2.2 together with Random Forest Regressor model were utilized to perform the coefficient of friction prediction [9]. The number of estimators used for Random Forest Regressor were 1000 decision tress. 3. Results and discussion The coefficient of friction of the water-glycols and polyalphaolefins based lubricants formulated with IM were predicted. This prediction was made by splitting the dataset into a training and a testing sub-dataset. The training dataset contains samples or lubricants with no IM in the lubricant (142 samples). The testing samples were PAO2-IM, PAO8-IM, WMEG-1IM, WDEG- 1IM, WMPG-1IM, and WDPG-1IM (6 samples). The results of the coefficient of friction predictions are shown in Figure 1. For each prediction, twenty-five iterations were performed, and the maximum, median, and minimum values were reported in the graph along with root mean square error (RMSE) and accuracy calculations. The true value (from experimental result) was also reported. The prediction results showed that Random Forest Regressor performed very well in prediction of water-glycols and polyalphaolefins formulated with IM ionic liquids with accuracy in the range of 83.26% to 99.47%. Figure 1. Prediction of coefficient of friction for water-glycols and polyalphaolefins formulated with IM ionic liquid. 4. Conclusion The application of machine learning for tribological performance prediction of new lubricant formulation was investigated in water-based and synthetic oil-based lubricants. From this study Random Forest Regressor could predict the coefficient of friction of lubricants formulated with ionic liquids with high accuracy. This high accuracy could be due to similar ionic liquids present in the training dataset, hence, predicting completely a new structure of chemical could be challenging. Therefore, for future work, the dataset will be expanded using more base lubricants and additives to increase the variation and the learning input. In addition, important molecular descriptors that influence the prediction accuracy will be analyzed. References [1] Agency, U.S.E.P., Environmentally Acceptable Lubricants. 2011, United States Environmental Protection Agency [2] Das R. Eco-friendly Lubricants for Tribological Application. Handbook of Ecomaterials. Cham (Switzerland): Springer, 2017 [3] Welton T. Room-Temperature Ionic Liquids. Solvents for Synthesis and Catalysis. Chemical Reviews 99: 2071-2084 (1999) [4] Earle M J, Seddon K R. Ionic liquids. Green solvents for the future. Pure and Applied Chemistry 72: 1391- 1398 (2000) [5] Zhao H. Innovative applications of ionic liquids as ‘green’ engineering liquids. Chemical Engineering Communications 193: 1660-1677 (2006) [6] Explore Chemistry. Quickly find chemical information from authoritative sources. https: / / pubchem.ncbi.nlm. nih.gov [7] Computational Toxicology Chemicals Dashboard (CompTox Dashboard). https: / / comptox.epa.gov/ dashboard/ advanced-search [8] Mauri A. AlvaDesc: A Tool to Calculate and Analyze Molecular Descriptors and Fingerprints. Ecotoxicological QSARs. Methods in Pharmacology and Toxicology. New York (US): Humana, 2020 [9] Pedregosa et al. Scikit-learn: Machine Learning in Python. The Journal of Machine Learning Research 12: 2825-2830 (2011)