Development of machine learning based prediction models for hazardous properties of chemical mixtures

QSPR models to predict the physical hazards of mixtures: a state of art

Physical hazards of chemical mixtures, associated for example with their fire or explosion risks, are generally characterized using experimental tools. These tests can be expensive, complex, long to perform and even dangerous for operators. Therefore, for several years and especially with the implementation of the REACH regulation, predictive methods like quantitative structure-property relationships have been encouraged as alternatives tests to determine (eco)toxicological but also physical hazards of chemical substances. Initially, these approaches were intended for pure products, by considering a molecular similarity principle. However, additional to those for pure products, QSPR models for mixtures recently appeared and represent an increasing field of research. This study proposes a state of the art of existing QSPR models specifically dedicated to the prediction of the physical hazards of mixtures. Identified models have been analysed on the key elements of model development (experimental data and fields of application, descriptors used, development and validation methods). It draws up an overview of the potential and limitations of current models as well as areas of progress towards enlarged deployment as a complement to experimental characterizations, for example in the search for safer substances (according to safety-by-design concepts).

Prediction of attachment efficiency using machine learning on a comprehensive database and its validation

Colloidal particles can attach to surfaces during transport, but the attachment depends on particle size, hydrodynamics, solid and water chemistry, and particulate matter. The attachment is quantified in filtration theory by measuring attachment or sticking efficiency (Alpha). A comprehensive Alpha database (2538 records) was built from experiments in the literature and used to develop a machine learning (ML) model to predict Alpha. The training (r-squared: 0.86) was performed using two random forests capable of handling missing data. A holdout dataset was used to validate the training (r-squared: 0.98), and the variable importance was explored for training and validation. Finally, an additional validation dataset was built from quartz crystal microbalance experiments using surface-modified polystyrene, poly (methyl methacrylate), and polyethylene. The experiments were performed in the absence or presence of humic acid. Full database regression (r-squared: 0.90) predicted Alpha for the additional validation with an r-squared of 0.23. Nevertheless, when the original database and the additional validation dataset were combined into a new database, both the training (r-squared: 0.95) and validation (r-squared: 0.70) increased. The developed ML model provides a data-driven prediction of Alpha over a big database and evaluates the significance of 22 input variables.