SmallTalks: Designing better soaps with machine learning: Predicting surfactant properties from molecular structure

Abstract 

Surfactants – the active ingredients in soaps, detergents, foams, and emulsions – are everywhere in industry and daily life, yet predicting their behaviour from chemical structure remains stubbornly hard. Measuring their properties demands either careful wet-lab experiments or expensive molecular simulations, both of which limit how quickly new candidates can be screened.

In this talk, I will present SurfPro-MD, a unified dataset that combines curated experimental data with molecular-dynamics-derived properties (diffusion, viscosity, interfacial surface tension) across 1436 surfactant systems and nine target properties.

Building on this, we trained machine learning surrogate models that predict surfactant behaviour directly from molecular structure, using a rigorous cluster-based evaluation to test how well they generalize to chemically novel compounds. A feature analysis reveals a small set of molecular descriptors that recur across properties, hinting at shared structural determinants of surfactant transport. All data and models are openly released.