Prediction of partition coefficient in high-pressure carbon dioxide–water systems using machine learning

Abstract

A method of predicting the partition coefficients (log K) of organic compounds in high-pressure carbon dioxide–water systems using machine learning was investigated. Using the collected literature data of log K, several linear and non-linear regression models were constructed. A cross-validation using these models indicated that log K can be approximately predicted with a root-mean-squared error of 0.6–1.1. Overall, the non-linear model predicted log K better, but linear models such as lasso regression exhibited comparable performances when features describing compounds were reduced. Parity plots indicated several outliers, most of which contain several polar functional groups. The analysis of feature importance revealed that the constructed model primarily consisted of the feature of log P (the 1-octanol–water partition coefficient) and was modified using process parameters and features related to the charges of compounds. The machine-learning approach yields a physicochemically reasonable model.