TY - JOUR
T1 - Machine learning coupled with group contribution for predicting the density of deep eutectic solvents
AU - Roosta, Ahmadreza
AU - Haghbakhsh, Reza
AU - Duarte, Ana Rita C.
AU - Raeissi, Sona
N1 - Funding Information:
The parts of the research which were carried out at Universidade Nova de Lisboa were funded by the European Union Horizon 2020 Program , under the grant agreement ERC-2016-CoG 725034 (ERC Consolidator Grant Des.solve), and furthermore, supported by the Associate Laboratory for Green Chemistry- LAQV which is financed by the national funds from FCT/MCTES ( UID/QUI/50006/2019 ). The authors also wish to thank Shiraz University, University of Isfahan and Universidade Nova de Lisboa for the facilities provided.
Publisher Copyright:
© 2022
PY - 2023/2
Y1 - 2023/2
N2 - Deep Eutectic Solvents (DESs) are a recently introduced class of green solvents with unique and favorable characteristics. Despite their recent debut, the scientific community has begun to place greater emphasis on them as alternatives to ionic liquids (ILs). Knowledge of the various physical properties of DESs is essential for various applications in the chemical industries and related fields. In this study, a comprehensive database including 1410 density data points, from 166 different DESs at various temperatures and atmospheric pressure, were retrieved from open literature to develop models to increase the accuracy of density predictions. The densities of DESs were used to develop two commonly used machine learning models, namely Multilayer Perceptron Artificial Neural Network (MLPANN) and Least Square Support Vector Machine (LSSVM), in conjunction with the group contribution (GC) method. Based on the GC method, each fragment of a compound contributes a specific amount to the physical property's value. By considering this, the prediction ability was improved by applying the GC method in the model development procedure. Both models predict the DES densities by taking into account the effect of 35 functional groups, the temperature, and the HBA/HBD molar ratios. The optimum MLPANN model structure consists of a single hidden layer with five neurons and a logarithmic sigmoid transfer function. By employing this MLPANN-GC model, the values of the squared correlation coefficient, R2, and absolute average relative deviation percent, AARD%, were 0.99 and 0.61%, respectively, while for the LSSVM-GC model (with the radial basis function (RBF) kernel), they were 0.99 and 0.56%, respectively. Also, K-fold cross-validation was used to assess the performance of the LSSVM-GC model. The presented machine learning models in this study were found to perform more accurately than those obtained using the best current correlations and GC models for DES densities in the open literature. The more accurate results, in addition to the enhanced predictability behavior of the developed models, give these models a preference for use in industrial and academic applications.
AB - Deep Eutectic Solvents (DESs) are a recently introduced class of green solvents with unique and favorable characteristics. Despite their recent debut, the scientific community has begun to place greater emphasis on them as alternatives to ionic liquids (ILs). Knowledge of the various physical properties of DESs is essential for various applications in the chemical industries and related fields. In this study, a comprehensive database including 1410 density data points, from 166 different DESs at various temperatures and atmospheric pressure, were retrieved from open literature to develop models to increase the accuracy of density predictions. The densities of DESs were used to develop two commonly used machine learning models, namely Multilayer Perceptron Artificial Neural Network (MLPANN) and Least Square Support Vector Machine (LSSVM), in conjunction with the group contribution (GC) method. Based on the GC method, each fragment of a compound contributes a specific amount to the physical property's value. By considering this, the prediction ability was improved by applying the GC method in the model development procedure. Both models predict the DES densities by taking into account the effect of 35 functional groups, the temperature, and the HBA/HBD molar ratios. The optimum MLPANN model structure consists of a single hidden layer with five neurons and a logarithmic sigmoid transfer function. By employing this MLPANN-GC model, the values of the squared correlation coefficient, R2, and absolute average relative deviation percent, AARD%, were 0.99 and 0.61%, respectively, while for the LSSVM-GC model (with the radial basis function (RBF) kernel), they were 0.99 and 0.56%, respectively. Also, K-fold cross-validation was used to assess the performance of the LSSVM-GC model. The presented machine learning models in this study were found to perform more accurately than those obtained using the best current correlations and GC models for DES densities in the open literature. The more accurate results, in addition to the enhanced predictability behavior of the developed models, give these models a preference for use in industrial and academic applications.
KW - Artificial neural network
KW - Density
KW - DES
KW - Group contribution
KW - Machine learning
KW - Physical property
KW - Support vector machine
UR - https://www.scopus.com/pages/publications/85142326265
U2 - 10.1016/j.fluid.2022.113672
DO - 10.1016/j.fluid.2022.113672
M3 - Article
AN - SCOPUS:85142326265
SN - 0378-3812
VL - 565
JO - Fluid Phase Equilibria
JF - Fluid Phase Equilibria
M1 - 113672
ER -