Abstract
The physical hazards of chemical mixtures are typically characterized using experimental tools that could benefit to be prioritized by using predictive methods. Indeed, experimental tests can be costly, complex, time-consuming, and potentially dangerous for the operator. In the last decades, particularly with the implementation of the REACH regulation, predictive methods such as QSAR/QSPR (Quantitative Structure-Activity/Property Relationships) have been encouraged and utilized as rapid and economical alternatives to experimental testing for determining (eco)toxicological and physical hazards of chemical substances.
Initially designed for pure compounds, adaptations of the QSPR approach were proposed to predict the properties of mixtures even if their development, in particular for physical hazards, is still an emerging field. Indeed, existing QSPR models still present some limitations to complement mixing rules and experimental approaches, and there is a need for new and more reliable models to extend applicability and improve prediction accuracy. A possible orientation could be using advanced machine learning approaches, taking advantage of scientific progress in artificial intelligence beyond classical multilinear regressions. More complex non-linear approaches (such as neural networks or random forests) have recently been used with the hope of better accounting for mixture complexity in QSPR models for mixtures.
This research aims to investigate if integrating advanced AI analytical methods can enhance the performance and applicability of QSPR models for predicting the physical hazards of chemical mixtures. To this end, applications of different machine learning methods were tested to evidence the advantages and limits of these Advanced AI algorithms compared to the more classical MLR approach when developing models for the flammability of liquid mixtures.
