(592d) Prediction of CO2 Solubility in Ionic Liquids Using Artificial Neural Networks and COSMO σ-Moments

CO₂ separation using Ionic Liquids(IL) is a highly active research topic both
from experimental and theoretical perspectives. Some of the most interesting
properties of ILs that make them promising for separation tasks are their low
volatility and flammability, wide liquid range, stability, and tunability by
careful selection of anion and cation.

The need for a predictive model accurate enough to guide the search in a huge
chemical space like the one spanned by ILs requires a more sound theoretical
basis than the one provided by models like UNIFAC and other group contribution
models; but also more functional flexibility than what can be obtained by
statistical associating fluid theory(SAFT) or screening models such as COSMO-RS.
Activity coefficient models are also know to be unsuitable for high pressure
descriptions, and equation of state approaches are usually targeted at
regression and not prediction.

In addition, the recent availability of large experimental databases of
vapor/liquid equilibrium (VLE) such as ILThermo and review papers with the most
up-to-date relevant measurements in a curated form provides the opportunity of
new machine learning models to test and challenge the supremacy of
group-contribution and other fixed-form models, both in regression and
prediction.

Here, a novel and easy to use tool for the screening and prediction of CO₂
solubility in ionic liquids on a wide range of pressures and temperatures of
practical interest is presented. The method is based on Artificial Neural
networks taking as inputs COSMO Ï?-moments¹Â and is tuned for physical
absorption based ILs. The network is implemented as a feed-forward multilayer
perceptron (MLP) with Levenberg-Marquardt as the learning algorithm and
log-sigmoid transfer functions in the hidden and output layers.

The selection of the moment descriptors is justified on the basis of physical
insight, prior art², and avoiding simultaneous selection of highly correlated
descriptors. The final set chosen for this work contains 4 parameters for each
cation and anion, for a total of 10 values, when considering temperature and
pressure.

With respect to the neural network itself, the architecture (size and number of
layers), activation functions, and training method were established to maximize
the predictive power of the network.

We apply the method to a recent, large, and heterogeneous dataset³.
Our results show that the neural networks generated are capable of:
a) representing the full dataset for interpolation on the known ionic liquids
b) reproduce homologous series effects(such as chain length effect, fluorination
degree) on ionic liquids which show this behavior
c) predict solubility in unknown ILs with state-of-the-art performance

Of special interest is the fact that the method is shown to perform adequately
even on ILs where both cation and anion are only part of the external test set,
demonstrating that the descriptors capture the relevant entropic and enthalpic
information. The quality of prediction is competitive with current methods such
as COSMO-SAC and UNIFAC in terms of R²of prediction.

References

[1] A. Klamt, COSMO-RS: From Quantum Chemistry to Fluid Phase Thermodynamics and
Drug Design, Elsevier Science,2005

[2] J. Palomar, J. S. Torrecilla, V. R. Ferro, F. Rodríguez , Development of an
a Priori Ionic Liquid Design Tool. 1. Integration of a Novel COSMO-RS Molecular
Descriptor on Neural Networks, Industrial & Engineering Chemistry Research 47
(13) (2008), 4523-4532, DOI: 10.1021/ie800056q

[3] Z. Lei, C. Dai, B. Chen, Gas solubility in ionic liquids, Chemical Reviews
114 (2) (2014) 1289â??1326, DOI:10.1021/cr300497a

Acknowledgements

The authors acknowledge The Research Council of Norway NFR project no 209337

and The Faculty of Natural Science and Technology, Norwegian University of Science

and Technology (NTNU) for financial support.