Prediction of ionic liquid thermophysical properties by group contribution method using random forest regression

• Main Author: Nuqui, Jesus Patrick E.
• Format: text
• Language: English
• Published: Animo Repository 2023
• Subjects: Materials—Thermal properties; Materials—Thermal conductivity; Chemical Engineering; Engineering; Thermodynamics
• Online Access: https://animorepository.dlsu.edu.ph/etdm_chemeng/20; https://animorepository.dlsu.edu.ph/context/etdm_chemeng/article/1020/viewcontent/Prediction_of_ionic_liquid_thermophysical_properties_by_group_contribution_method_using_random_forest_regression.pdf
• Institution: De La Salle University

Ionic liquids are tunable, there are vast combinations of cations and anions and attached substituent groups that can be combined together to achieve a target property. Mathematical equations in which the contributions of each component groups of an ionic liquid are considered in a predictive model are common in literature, however, it encounters limitations in terms of number of input variables that can be accommodated. In this study, the random forest (RF) regression algorithm was used to create a model to predict thermophysical properties, particularly the isobaric heat capacity, viscosity, thermal conductivity, and surface tension of ionic liquids using the group contribution of the main cation, main anion, and their substituent groups as input variables. Hyperparameter tuning was used to identify the optimal parameters for the random forest model for each thermophysical property. It was found that the algorithm is capable of handling multiple input variables and provide accurate predictions based on the obtained R2 and %AAD values. The following were the obtained R2 values of the training and testing data sets at 80-20 split of the RF models for the isobaric heat capacity, viscosity (range limited), surface tension, and thermal conductivity respectively: 0.9964 and 0.9985; 0.9628 and 0.8141; 0.9981 and 0.9906; and 0.9960 and 0.9863. Additionally, the %AAD of the of the training and testing data sets at 80-20 split of the four thermophysical properties following the same sequence as above are as follows: 0.28% and 0.63%; 7.03% and 15.94%; 0.67% and 1.47%; and 0.56% and 1.21%. These favorable performance statistical metrics on top of being able to handle large number of input variables and data points confirmed the competitiveness of the RF regression algorithm in predicting the properties of complex substances such as ILs.