Predicting the dissolution of metal-based nanoparticles by means of QSPRs and the effect of data augmentation.

Particle dissolution is a critical process in the environmental fate assessment of metal-based nanoparticles (MNPs). Numerous attempts have been made previously to adequately quantify dissolution (kinetics), however, existing dissolution data and models are generally limited to a few nanomaterials or specific time points. Hence, they only capture phases of the process. This study aimed to develop a Quantitative Structure-Property Relationship (QSPR) model to predict the ion release (in %) of MNPs for different time points and water chemistry conditions. Furthermore, many machine learning models are frequently plagued by a lack of data and recently data augmentation has been suggested as a method to mitigate this issue. Therefore, we also investigated the effects of data augmentation on QSPRs. Following data collection from literature, QSPR models were generated and results indicate models with adequate performance (R > 0.7). Results also demonstrated significant improvements in model performance with increasing amounts of applied data augmentation. However, a deeper evaluation of the results also highlighted that data augmentation can lead to misleading and overoptimistic model evaluation. Thus, proper model assessment is necessary when evaluating QSPRs. Variable importance analysis results revealed that the "initial concentration" and features related to the size and shape of MNPs were the most critical factors in the dissolution process. The predictive models generated here for MNP dissolution can improve nanomaterial testing efficiency and guide experimental design.