A computational approach for prediction of viscosity of chemical compounds based on molecular structures

• Use of SMILES notation to generate molecular descriptors of 320 chemical compounds and train machine learning models to predict viscosity of the compounds based on the data. • Comparative analysis of three supervised machine learning models (RF, GB and CatBoost) used to predict viscosity of the 320 compounds. • Identification of the most important features for each model which contributed the most in predicting the viscosity. • Comprehensive analysis of the best and worst performances across all three models is conducted. • CatBoost model outperformed the RF and GB models in terms of R 2 and MSE parameters and the features- Number of Hydrogen atoms and TPSA were the highest contributing factors in the viscosity prediction using the CatBoost model. The research paper explores the feasibility of predicting the viscosity of a diverse chemical compound by using molecular structures at 25 °C through supervised machine learning methods. In this paper, Random Forest, Gradient Boosting and CatBoost supervised algorithms were implemented. The dataset consists of the Simplified Molecular Input Line Entry System (SMILES) notation of 320 chemical compounds and their corresponding viscosities at 25 °C. The study generated ten features from the compounds to correlate and predict the viscosity values. The results suggest that Catboost algorithm (R 2 = 0.94 and MSE = 0.64) performs better than Gradient Boosting (R 2 = 0.90 and MSE = 1.05) and Random Forest algorithm (R 2 = 0.86 and MSE = 1.39), while measuring the viscosity. Although the Random Forest model lags behind, the overall results may be useful for predicting viscosity with accuracy. The findings of the research work also reveal which feature contributes the most to a given model. The results of the proposed model can be used in several industries including food and beverage, cosmetics, medicine, pharmaceuticals etc.