skip to content

Yusuf Hamied Department of Chemistry

Image of Dr Christoph Schran sitting at his desk

Dr Christoph Schran at his desk, courtesy ICE group

A new strategy to enable molecular simulations of complex systems has opened the door to a better understanding of complex materials.

The powerful machine-learning framework, developed by researchers from this Department, University College London, Imperial College London and Charles University in Prague, is reported in the journal PNAS.  

Understanding complex materials, in particular those with solid-liquid interfaces, such as water on surfaces or under confinement, is a key prerequisite in tackling some of the fundamental issues of our time, such as climate change or the lack of clean water as it enables rational material design of technologies like water nano-filtration, electrical energy storage, and heterogeneous catalyses.

Here, the researchers show how the simple and automated machine-learning procedure can provide accurate models of these complex systems, overcoming previous limitations. To show how the models can be applied in exhaustive simulations, the researchers successfully used the methodology on a wide set of aqueous systems with increasing degrees of complexity.  

"The key to the success of the procedure is that we’ve automated crucial steps such as the development and validation of the machine learning model," said first author Dr Christoph Schran from this Department. “The process we’ve created is simple and user-friendly, where the machine learning potential is constructed with minimum human effort through a data-driven active learning protocol.”

The procedure will enable more thorough and accurate investigation and understanding of complex aqueous processes such as water structuring in contact with interfaces and wetting or ice formation on surfaces.

Because it enables the fast screening of different materials at high accuracy, it should also prove useful for other materials and liquids in contact with solids, as well as general solvation phenomena.

"Informally, we keep calling our models 'disposable machine-learning potentials' as they are very easy to create for a given purpose, essentially like a single-use item. But the method works so well, that we have found we do not throw the models away afterwards, so we need a new name!" said Schran.

Schran said the approach is particularly useful for situations where long timescales are involved as for the exploration of absorption phenomena, or dynamical properties, such as the friction or viscosity of liquids in contact with interfaces.

“We’ve essentially outlined a straightforward strategy for the uncomplicated yet accurate molecular simulation of many complex systems,” said Schran, who noted that the team have made all their methods openly accessible so the new strategy can be used by  researchers everywhere. 

Figure explaining procedure

Machine learning potentials for complex aqueous systems made simple, C. Schran, F. L. Thiemann, P. Rowe, E. A. Müller, O. Marsalek, A. Michaelides, PNAS (Sept 2021).