Prediction of isobaric heat capacities of room temperature ionic liquids
Permanent link to this Item
Link to Journal
Ionic liquids (ILs) have many potential applications that require knowledge of a variety of physical and thermodynamic properties. While these properties can often be determined experimentally, this is impossible for novel, yet to be synthesised ILs; thus, property prediction from first principles is essential to unlock new developments in the rational design of ILs. The isobaric heat capacity (CP ) is an important thermodynamic property that quantifies the amount of heat needed to increase the temperature of a material and is thus of great importance in engineering applications involving the design of heat-transfer systems. From a theoretical viewpoint, the heat capacity is a fundamental quantity that expresses the temperature dependence of enthalpy and entropy. Several models for the prediction of CP have been developed to date; however, these are often trained on limited data sets and published model performance is largely dependent on the judicious choice of the testing data. Moreover, popular techniques such as group contribution methods (GCMs) cannot always be applied to structurally novel ILs and quantitative structure property relationships (QSPRs) are highly dependent on the diversity of training data. In this work, predictive models for CP have been developed using linear and nonlinear machine-learning methods. A large data set of 2463 temperature-dependent CP values, spanning 208 ILs, was obtained from the ILThermo database. Molecular volumes, features based on the electrostatic potential (ESP) and other molecular descriptors were calculated for each cation and anion in the data set. Following this, several multiple linear regression models were developed, for which Lasso regression was used reduce the number of features, where necessary. The models were developed using a methodology that attempts to reduce the dependency of the results on the identity of the specific species in the training set. The complexity of these models was gradually increased from a simple volume-based model (inspired by the success of the Volume Based Thermodynamics (VBT) approach of Glasser and Jenkins [L. Glasser and H. D. B. Jenkins, Chem. Soc. Rev., 2005, 34, 866], which was applied to ionic liquids and augmented by Krossing and co-workers [W. Beichel et al., J. Mol. Liq., 2014, 192, 3]), to the addition of electrostatic potential surface areas and finally including the General Interaction Properties Functions (GIPFs) of Murray and Politzer [J. S. Murray et al., J. Mol. Struct. (THEOCHEM), 1994, 307, 55], which are statistically well-defined quantities derived from ESP data, and a Feed Forward Neural Network (FFNN) was developed using the most effective of the aforementioned feature sets. In addition to reporting test-set errors, an external data set was carefully compiled, containing ILs with components (either the cation or anion) not present in the training data, and structurally distinct. This was done to assess the general applicability and flexibility of the final models, and to allow for a fair comparison of model performance. Of the linear models developed, that using interacting features consisting of molecular volumes and GIPFs produced the lowest errors; this is likely due to the ability of the interaction features to describe intermolecular interactions between cations and anions. Consequently, molecular volumes and GIPFs features were also used to develop a nonlinear FFNN. Finally, the linear interacting GIPFs model and FFNN also produced the lowest errors of 3.2 1.5% and 3.8 2.4%, respectively, when applied to the external data set.