Can machine learning predict the phase behavior of surfactants?

We explore the prediction of surfactant phase behavior using state-of-the-art machine- learning methods, using a data set for twenty-three non-ionic surfactants modified from Bell, Phil. Trans. R. Soc. A 2016, 374, 20150137. We recover Bell’s observation that most machine learning methods are capable to some degree of filling in missing data in partially-complete phase diagrams. However, strong data bias and a lack of chemical space information generally lead to poor results for de novo phase diagram prediction. Although some machine learning methods perform better than others, these observations are largely robust to the particular choice of algorithm. For the gap-filling challenge we additionally examine how to sample state points effectively in the context of typical experimental protocols, to make the most of machine-learning interpolation. Our results indicate what factors should be considered when preparing for machine- learning prediction of surfactant phase behavior in future studies.