Harry is working on machine learning-based interatomic potentials, aimed at unraveling the functional properties of complex materials such as SiC.

Silicon carbide (SiC) is a prototypical material for high temperature applications (e.g aerospace, [1] automotive [2] and thermoelectric [3,4]) involving complex microscopic processes typically inaccessible to experiments. [5] To gain insight into the functional properties of SiC e.g. nanostructures, computationally expensive quantum mechanical methods such as density functional theory (DFT) must be employed. These methods are expensive due to the size of the systems that must be simulated to probe the macroscopic properties we are intreted in; driving up the computational time, by at least N 3 with regard to system size in the case of DFT. Why not use less computationally demanding methods?

Unfortunately, they are almost always not accurate enough. In fact, similar to silicon (Si) and carbon (C) alone, various empirical interatomic potentials have been developed for SiC, such as Tersoff [6] or Stillinger-Weber. [7] These potentials are designed to reproduce specific features of the material, at the expense of transferability to a wider range of functional properties. As DFT is unusable for systems of this size, and the current empirical potentials are not accurate enough, machine learning (ML) is the best approach. This will enable the "best of both worlds", the accuracy of DFT combined with the speed of empirical potentials.

The aim of this project is to build a general purpose interatomic potential for SiC (e.g using machine-learning regression starting from a DFT dataset of representative configurations [8,9]), and to use this to probe the macroscopic properties of SiC in the realworld applications listed above. As a first step towards achieving this goal we here report a comparative study of the performance of currently available models for SiC, [6,7,10] and two machine learned interatomic potentials (MLIPs): a Gaussian approximation potential (GAP) and a neural network potential (NNP). Focusing in particular on the structural properties of the liquid/amorphous phase.

[1] D. G. Senesky, B. Jamshidi, K. B. Cheng and A. P. Pisano, IEEE Sensors Journal, 2009, 9, 1472–1478.
[2] A. Zeb and S. J. Milne, Journal of Materials Science: Materials in Electronics, 2015, 26, 9243–9255.
[3] C.-H. Pai, Y. Sasaki, K. Koumoto and H. Yanagida, Journal of the American Ceramic Society, 1991, 74, 2922–2924.
[4] L. A. Valentín, J. Betancourt, L. F. Fonseca, M. T. Pettes, L. Shi, M. Soszyński and A. Huczko, Journal of Applied Physics, 2013, 114, 184301.
[5] I. Aharonovich and M. Toth, Nature Physics, 2014, 10, 93–94.
[6] J. Tersoff, Physical Review B, 1989, 39, 5566–5568.
[7] P. Vashishta, R. K. Kalia, A. Nakano and J. P. Rino, Journal of Applied Physics, 2007, 101, 1–12.
[8] A. P. Bartok, J. Kermode, N. Bernstein and G. Csanyi, Physical Review X, 2018, 8, 41048.
[9] G. C. Sosso, V. L. Deringer, S. R. Elliott and G. Csányi, Molecular Simulation, 2018, 44, 866–880.
[10] L. Pastewka, A. Klemenz, P. Gumbsch and M. Moseler, Physical Review B - Condensed Matter and Materials Physics, 2013, 87, 1–12.