Prediction of Environmental Properties Using a Hybrid Group Contribution Approach

Development of predictive environmental property models is increasingly becoming crucial as stringent regulations for substances with high global warming and ozone depletion potentials are being introduced. This contribution presents new environmental property models using two group contribution (GC) based approaches for the prediction of ozone depletion potential (ODP), global warming potential (GWP) and Daphniamagna lethal concentration (LC50, 48 hr), concentration of the test chemical in water(mg/L) that causes 50% of Daphnia magna to die after 48 hours. First, the classical group contribution approach, in which a property of interest is estimated from regression models that make use of available information about the chemical structureof a given compound (i.e. functional groups), is applied to develop models for selectedproperties using robust regression with outlier treatment. Second, a hybrid approachusing only the first order GC-defined functional groups as predictors is presented todevelop a number of data-driven models (a feedforward neural network (ANN) and aradial basis function network (RBFN), regression tree, etc.). Performance of thedifferent models in predicting ODP, GWP, and LC50 is assessed for various industriallyrelevant chemicals and compared with results of the classical GC method. Theexperimental data for the selected properties is collected from the databases ofEnvironmental Protection Agency (EPA) and the fifth assessment report of theIntergovernmental Panel on Climate Change (IPCC). The results showed that hybridapproach presents significant improvement on estimation accuracy for the consideredenvironmental properties. This flexible approach builds on the basis of GC theory andextends it with nonlinear surrogate models to better describe the property of interest,which makes it a promising method to improve accuracy of property models in thewider domain of process systems engineering.