Prediction of ionic liquids’ speed of sound and isothermal compressibility by chemical structure based machine learning model

Abstract

The speed of sound (u) and isothermal compressibility coefficient (K_T) are important thermodynamic parameters of ionic liquids (ILs), crucial in describing their behavior, deriving additional thermodynamic properties, and developing the advanced equations of state. In this work, we developed an artificial neural network (ANN) model, integrated with the group contribution method (GCM), to predict the u and K_T of pure ILs. The model leverages a newly comprehensive dataset. GCM was employed to divide molecules of ILs into constituent groups and use these groups as input features for the ANN algorithm. The model offers simple and reliable predictions of u and K_T of ILs without relying on other properties. To achieve higher model generalizability, cross-validation was performed and two distinct dataset division strategies were applied: IL-division and datapoint-division. The model demonstrates exceptional predictive accuracy across both strategies. For the u-test set, the IL-division and datapoint-division achieve an average absolute relative deviation (AARD) of 0.9083 % and 0.4134 %, respectively. Similarly, for K_T, the IL-division and datapoint-division methods for the test set obtain AARD of 4.2679 % and 1.1651 %, respectively. In the datapoint-division method, the same IL was perhaps included in both training, validation, and test sets, yielding better results. However, the IL-division approach allows prediction on completely new ILs with no available experimental data. Furthermore, correlation analysis was conducted to explore the influence of molecular group occurrences on the model's predictions, offering deeper insights into the structure-property relationships of ILs.