Estimation of Adsorbed Amounts in Organoclay by Machine Learning

Adsorption properties of organoclay have been investigated for decades focusing on the morphology and physicochemical properties of two-dimensional interlayers. Experimental studies have previously revealed that the adsorption mechanisms depend on the molecular species of the organocation and adsorbate, making it difficult to estimate the adsorbed amount without experiments. Considering that the adsorption of aromatic compounds has been reported by using various clays, organocations, and adsorbates, machine learning is a promising method to overcome the difficulty. In the present study, we collected adsorption data from the literature and constructed models to estimate the adsorbed amount of the organoclay by random forest regression. The composition of the clay, molecular descriptors of the organocation and adsorbate obtained by the RDKit, and experimental conditions were used as the explanatory variables. Simple model construction by using all the experimental data resulted in low R2 and a mean absolute error. This problem was solved by the correction of the adsorbed amount data by the Langmuir or Freundlich equation and the following model construction at various equilibrium concentrations. The plots of the adsorbed amount estimated by the latter model were located close to the corresponding adsorption isotherm, while that by the former was not. Thus, it was revealed that the adsorbed amount was estimated quantitatively without understanding the adsorption mechanisms individually. Feature importance analysis also revealed that the combination of the organocation and adsorbate is important at high equilibrium concentrations, while the clay should be selected carefully as the concentration gets lower. Our results give an insight into the rational design of the organoclay including the synthesis and adsorption properties.