Hypothesis: Predicting the surface tension (SFT)-log(c) profiles of hydrocarbon surfactants in aqueous solution is computationally non-trivial, and empirically challenging due to the diverse and complex architecture and interactions of surfactant molecules. Machine learning (ML), combining a data-based and knowledge-based approach, can provide a powerful means to relate molecular descriptors to SFT profiles.
Experiments: A dataset of SFT for 154 model hydrocarbon surfactants at 20-30 °C is fitted to the Szyszkowski equation to extract three characteristic parameters (Γmax,KL and critical micelle concentration (CMC)) which are correlated to a series of 2D and 3D molecular descriptors. Key (∼10) descriptors were selected by removing co-correlation, and employing a gradient-boosted regressor model to rank feature importance and carry out recursive feature elimination (RFE). The hyperparameters of each target-variable model were fine-tuned using a randomised cross-validated grid search, to improve predictive ability and reduce overfitting.
Findings: The ML models correlate favourably with test experimental data, with R2= 0.69-0.87, and the merits and limitations of the approach are discussed based on 'unseen' hydrocarbon surfactants. The incorporation of a knowledge-based framework provides an appropriate smoothing of the experimental data which simplifies the data-driven approach and enhances its generality. Open-source codes and a brief tutorial are provided.
Keywords: Critical micelle concentration; Machine learning; QSPR; Surface tension; Surfactant.
Copyright © 2022 The Authors. Published by Elsevier Inc. All rights reserved.