Predicting soil sorption coefficients of organic chemicals using a neural network model

Abstract

AbstractThe soil/sediment adsorption partition coefficient normalized to organic carbon (Koc) is extensively used to assess the fate of organic chemicals in hazardous waste sites. Several attempts have been made to estimate the value of Koc from chemical structure or its parameters. The primary purpose of this study was to develop a nonlinear model for estimating Koc applicable to polar and nonpolar organics based on artificial neural networks using the octanol/water partition coefficient (Kow) and water solubility (S). An analytic equation was obtained by starting with a neural network, converging the bias and weight values using the available data on water solubility, octanol/water partition coefficient, and the normalized soil/sediment adsorption partition coefficient, and then combining the equations for each node in the final neural network. For the 119 chemicals in the training set, estimates using the neural network equation lie outside the 2σ region (the standard deviation for the training set, σ = 0.52) for only five chemicals, while all the chemicals in the test set lie within the 2σ region. It was concluded that the neural network equation outperforms the linear models in fitting the Koc values for the training set and predicting them for the test set.