Different modelling approaches for predicting titanium dioxide nanoparticles mobility in intact soil media

Understanding the transport behaviour of new and emerging materials such as engineered nanoparticles (ENPs) is vital for the accurate assessment of their functionality and fate in environmental systems. Predicting ENP mobility in soil systems based on common attributes of either soil or ENPs is of significant interest as an alternative to conducting laborious and time consuming column experiments. Thus this study investigates the importance of different soil properties and experimental conditions on titanium dioxide nanoparticles (nTiO2) mobility in real soil media and also evaluates four different modelling approaches including Multiple Linear Regression (MLR), Classification and Regression Tree (CART), Random Forest (RF) and Artificial Neural Network (ANN) for predicting nTiO2 mobility in soil media. The performance of both ANN and RF models were good for predicting nTiO2 transport in soil media, with ANN predictions being slightly superior to RF with less generalization errors. However, RF had the advantage of requiring less input predictors. In comparison the MLR model exhibited poor performance in both calibration and validation datasets, and while the validity of CART was almost acceptable in the calibration dataset, its efficiency was poor for the validation dataset. In addition to soil solution chemistry and hydraulic properties, other important factors having a major contribution to nTiO2 transport through soil included soil fracture associated properties and the existence of preferential flows. [Display omitted] •ANN and RF models can properly predict nTiO2 transport in soil media.•ANN predictive model had the least generalization errors amongst all models.•MLR model exhibited poor performance for predict nTiO2 transport in soil media.•Preferential flow influences nTiO2 transport in soil media.