Developing Machine Learning Models for Predicting Multiple Physical Properties of Ionic Liquids through a Combined Constitution-Structure-Interaction Descriptor

Ionic
liquids (ILs) have garnered significant research interest
due to their wide-range applications in separation, catalysis, and
synthesis. However, the combination of an extensive quantity of diverse
anions and cations makes it a grand challenge for traditional methods
(theoretical calculations and experiments) to analyze the properties
of unexplored ILs. The emerging machine learning (ML) technique can
uncover complex relationships within large data sets. However, the
prediction accuracy of ML models is highly dependent on the selection
of universal descriptors that are both physically meaningful and broadly
applicable. In this work, we designed a composite descriptor incorporating
the constitution, structure, and interaction. Then we utilized this
descriptor in combination with five conventional ML models to predict
six important properties of ILs including density, surface tension,
heat capacity, viscosity, electrical conductivity, and thermal conductivity.
Detailed comparison results show that the gradient boosting regression
tree (GBRT) model with respect to other ML models exhibits superior
accuracy and generalizability in predicting all six properties of
ILs. The results of feature importance and Shapley additive explanations
(SHAP) values further reveal that the structure descriptor in our
proposed composite descriptor plays a key role in predicting the properties
of ILs. Our study demonstrates that the integration of a constitution-structure-interaction
descriptor with an ML model yields remarkable accuracy in predicting
properties of ILs, and we anticipate that this can serve as a powerful
tool for predicting properties of other materials beyond ILs.