Robust Method for Property Prediction via Artificial Neural Networks: Incorporating Key Structural Features for Carbon Dioxide–Ionic Liquid Mixtures

This
study addresses a critical gap in the existing literature
on carbon dioxide and ionic liquid (IL) mixtures, where fragmented
and incomplete data, particularly for flow properties, hinder practical
applications. Therefore, this work aimed to establish a robust and
efficient method for predicting the density of the CO2–IL
mixtures across diverse operating conditions and IL families using
novel validation techniques. Both linear and symbolic regression models
provided relevant insights but failed to accurately capture the IL–CO2 interactions in a mixture that determine the molar volume
of CO2 at infinite dilution when solubilized by a given
IL. Therefore, more mathematically flexible artificial neural networks
(ANN) were trained based on three different sets of features: (1)
IL critical properties, (2) IL structural descriptors, and (3) a selective
combination of (1) and (2). While all models showed relative deviations
consistently below 3% for the testing data, combining critical and
structural data significantly improved accuracy (R2 = 0.986, testing data set). A postprocessing outlier-handling
method enhanced model performance, removing a minimal fraction (below
0.2%) of unphysical data points. Furthermore, molecular dynamics simulations
validated the robust generalization of all ANN models, with the combined
model exhibiting remarkable accuracy over operating conditions outside
the training ranges for ILs in the training set and even for ILs that
are not included in this data set. This computational approach provides
a significantly faster and broader alternative to other thermodynamical
tools, establishing a solid method for future machine learning (ML)-based
property prediction augmented by external validation from cross-comparison
tests and statistical thermodynamics models.