An accurate and transferable machine learning potential for carbon

Patrick RoweVolker L. DeringerPiero GasparottoGábor Csányi, and  Angelos Michaelides

Home >The Journal of Chemical Physics >Volume 153, Issue 3 >10.1063/5.0005084

The same characteristics which make carbon a fascinating element for study also make it challenging to model computationally. It exhibits some of the greatest structural diversity - and associated diversity of properties - of any of the elements. This variety of structures and properties makes the design of an accurate empirical model for carbon quite challenging. Such a model must be robust enough to treat high-temperature liquid phases, while remaining accurate enough to model weak van der Waals in layered graphitic structures. Machine learning approaches for the construction of interatomic potentials offer this flexibility, while also being able to achieve accuracies which surpass those of traditional empirical models. In this work, we present just such a machine learning potential for elemental carbon, trained using the Gaussian approximation potential (GAP) approach. We show that it is accurate for a wide range of important material properties, including bulk crystalline phases, defects and surfaces. We go on to demonstrate the model's flexibility and potential for scientific discovery by performing simulations on structures not included in the training - including a GAP driven random structure search, which identifies some interesting crystal structures. 


Follow @tyc_london for updates from the Thomas Young Centre.