Design of Gold Extraction Solvents Using Machine Learning Models

Abstract

Dibutyl carbitol (DBC) has been conventionally used for solvent extraction of Au(III), which exists as tetrachloroauric acid (HAuCl4), leached by chlorination. However, the water solubility of DBC (3 g/dm3) leads to organic contamination of the aqueous phase. To identify solvents with both high gold extractability and low water solubility, a model Y = f(X) was constructed to predict extractability and water solubility Y from molecular descriptors X of chemical structures using experimental data for various compounds. The charge density profile calculated with COSMO-RS was used as X in the extractability prediction model and genetic algorithm-based wavelength selection, which allows variable selection in regions, was applied. The constructed models showed high prediction accuracy, with correlation coefficients exceeding 0.9 for both extractability and water solubility. The inverse analysis of these models was used to search for solvents with both high extractability and low water solubility. Solvents including (−)-fenchone and 4-phenyl-2-butanone were identified as promising solvents exhibiting comparable extractability to DBC but lower water solubility.