The a priori screening of potential organic solvents using artificial neural networks

Abstract

A QSPR model using artificial neural networks was constructed to estimate the binary interaction parameters for the temperature-dependant form of the NRTL model with the objective of using it as a supplement to assist limitations of group contribution methods in the screening of potential solvents for liquid-liquid extraction processes. Parameters were regressed using experimental LLE and VLE data and checked for consistency. Molecule structures were drawn and descriptors determined with the use of Materials Studio. The QSPR model uses 31 descriptors as input and produced absolute average deviations of 0.23 and 0.19 for each pair of binary interaction parameters respectively. The development of this model is shown to be effective in improving the robustness of solvent screening processes.