Traditional phase equilibrium models usually depend on simplified assumptions and empirical parameters, which are difficult to obtain during regular site investigations. As a result, they often under- or over-estimate soil vapor concentrations for assessing the risks of volatile organic compound (VOC)-contaminated sites. In this study, we develop several machine learning models to predict soil vapor concentrations using 2225 soil-soil vapor data pairs collected from seven contaminated sites in northern China. Compared to the classic dual equilibrium desorption model, the random forest (RF) model can provide more accurate predictions of soil vapor concentrations by at least 1-2 orders of magnitude. Among the employed covariates, soil concentration and organic carbon-water partition coefficient are two of the most significant explanatory covariates affecting soil vapor concentrations. Further examination of the developed RF model reveals the phase equilibrium behavior of VOCs in soil is that: the soil vapor concentration increases with soil concentration at different rates in the first two intervals but remains almost unchanged in the last interval; the solid-vapor partitioning interface may still exist at up to 15% mass water content in our simulations. These findings can help site investigators perform more accurate risk assessments at VOC-contaminated sites.
Keywords: Desorption; Equilibrium partitioning; Machine learning; Soil vapor concentration; Sorption.
Copyright © 2022 Elsevier Ltd. All rights reserved.