Short-cut models for environmental impact metrics of chemical production

Abstract

This work has investigated the ability of molecular structure-based (MSB) models to predict the ReCiPe indicators for environmental impact assessment. A dataset of 189 observations and 28 molecular descriptors (MDs) has been used to predict four endpoint indicators and 18 midpoint indicators. The endpoint indicators were: Ecosystem quality (EQ), human health (HH), resource depletion (R) and the total ReCiPe score (T). Linear models in form of a partial least squares (PLS) regression and nonlinear radial basis function artificial neural networks (ANNs) have been compared. It has been found that ANNs perform significantly better than linear models. The human health (HH) indicator as well as the total (T) ReCiPe indicator could be predicted with a satisfactory precision with a coefficient of determination of 0.52 and 0.44 and model size of 15 and 13 molecular descriptors (MDs) respectively. The structure of the ANN and as well as the most important MDs has been analysed. It has been found that there is a tendency to include some oxygen related functional groups, nitrogen and the molecular weight for HH and T. The results were compared with results for the EI99 indicator from literature to investigate whether it is more useful to predict the total ReCiPe indicator directly, or to correlate it with a good prediction of the total EI99 indicator. A correlation of r2 = 0.92 between EI99 and ReCiPe has been found. This correlation is useful, provided there is a good prediction of the EI99 indicator. The dataset that has been used in this work predicts the ReCiPe indicator with a higher precision than the EI99 indicator, which makes is more convenient to model the ReCiPe indicator for this particular case directly. The analysis of the results has also indicated weaknesses in the modelling procedure, suggesting improvements for future applications.