Prediction of organic additive-induced rheological enhancement of CTAB solutions using machine learning methods

Abstract

Various organic additives can induce structural rearrangement of surfactant micelles in aqueous solutions, leading to enhanced rheological properties, namely an increase in viscosity, appearance of viscoelasticity, and non-Newtonian behavior. This phenomenon can be exploited in several industries to obtain strong gels. The selection of proper additives and their concentrations remains largely empirical owing to the high number of factors influencing structural rearrangement. This paper aims to employ machine learning (ML) methods to predict whether an enhancement of rheological behavior will occur in cetyltrimethylammonium bromide solutions (CTAB) upon addition of an organic compound. Five different shallow ML algorithms were tested along with graph neural networks (GNNs), using a dataset of 605 data points collected from the literature. Two types of descriptors (RDKit and PaDEL) were applied. Shallow ML algorithms achieved a maximum precision of 0.85 under cross-validation with a random fold split, and this performance did not deteriorate significantly when a structure-based split was applied (0.82). The random forest algorithm was selected based on its comparatively stable performance. GNN achieved a precision of 0.85 (random split) and 0.80 (structure-based split). Analysis of relative contribution of system properties and molecular features to the predictions of shallow and GNN models is consistent with the known physicochemical basis. Experimental validation of ML methods on four prospective compounds unseen during training was carried out. Comparison with ML predictions demonstrated that both models can correctly identify compounds that promote rheological enhancement, but tend to overpredict positive outcomes. However, this drawback may be partially mitigated by applying a pessimistic consensus, and then the accuracy of 0.84 and the precision of 0.73 are achieved for evaluation on the experimental dataset. The present work demonstrates that ML methods can be applied to address the problem of surfactant self-aggregation in solution upon addition of organic additives with room remaining for performance improvement.