Modelling Si–O nanostructures with active machine learning

sio

Modelling Si–O nanostructures with active machine learning

The silicon–oxygen system is central to chemistry. Beyond the widely-known mineral quartz, there are many other solid-state forms of silicon dioxide, including some that form only under very high pressure. There are glassy phases, nanostructured composites, and an amorphous material with an uncommon composition (“silicon monoxide”). Understanding the seemingly simple binary Si–O system on the atomic scale has therefore remained a major research challenge.

In a study published in Nature Communications, researchers at Oxford Chemistry and the Technical University of Darmstadt (Germany) now describe a machine-learning model that can capture atomic structures and dynamics in the full Si–O system. Their work enables the realistic modelling of nanostructured silica and silicon monoxide on the length scale of tens of nanometres.

To meet the complex modelling challenges – high-pressure structures, surfaces and pores, and phases of mixed stoichiometric composition – the team developed an ML model for the atomic interactions. In particular, they used a technique called “active learning” to detect structures that the model has not yet accounted for, to select the most important new structural fragments, and to feed them back into the training of the model.

The research was carried out by Linus Erhard, a senior doctoral student in the group of Professor Karsten Albe at Darmstadt. Linus visited Oxford twice: in early 2020, and again in 2022 when he spent 6 months working with the Deringer group in the Inorganic Chemistry Laboratory. His research stays at Oxford were supported by the Erasmus+ programme and the German Academic Exchange Service (DAAD). A first joint paper described an ML potential for silica, the widely known phase with 1:2 stoichiometric composition.

But the team did not stop there: their new paper now enables the computational study of the entire binary system, with a special focus on SiO. The image above shows a close-up of a nanoscale segregated phase: SiO2-rich regions (oxygen atoms in red) coexisting with amorphous silicon inclusions (blue).

Linus said:

This is the first time that we can model the interface between silicon and silicon dioxide on a scale of millions of atoms with high accuracy. Therefore, our model is highly promising for solving and understanding future materials problems in this system on the atomic scale.”

The work is expected to be of interest to the materials modelling community, but also to experimental chemists. The research data, including the parameters for running simulations, are openly available.

Volker Deringer, an Associate Professor in the Department who led the Oxford side of the work, commented:

We are seeing an unprecedented degree of realism in materials modelling these days, made possible by many years of methodological developments in atomistic machine learning. Linus’ work is an outstanding example of this (and has been a fantastic collaboration with the Darmstadt team!). It is an extremely exciting time to be working on computational solid-state and materials chemistry.”

The work is available via Open Access at https://doi.org/10.1038/s41467-024-45840-9.