Mar 20 – 22, 2019
Nordita, Stockholm
Europe/Stockholm timezone

Computational experiments with machine learning-based interatomic potentials: explaining the growth mechanism in amorphous carbon

Mar 22, 2019, 12:15 PM
132:028 (Nordita, Stockholm)


Nordita, Stockholm


Miguel A Caro (Aalto)


Molecular dynamics (MD) simulations are a useful tool to understand the interactions between atoms and to get insight into the processes that take place at the nanoscale and give rise to the observed properties of materials. "Classical" interatomic potentials, based on i) harmonic description of bonds, ii) partial electrostatic charges and iii) Lennard-Jones approximations for dispersion interactions, are computationally efficient but do not grant accurate representation of the real underlying physics/chemistry. They tend to fail at flexibly describing molecules in changing environments, especially when there is bond rearrangement, i.e., when chemical reactions take place. Density functional theory (DFT), on the other hand, offers a satisfactory description of interatomic interactions and can be used to characterize bond formation and annihilation. Unfortunately, DFT becomes prohibitively expensive when running MD of systems beyond a few hundreds of atoms or for time scales longer than a nanosecond. To bridge this gap between computational efficiency and accuracy, algorithmic developments that make use of machine learning techniques are being adopted by the community. In this seminar, I will briefly introduce one of such approaches, the Gaussian approximation potential (GAP) framework [1]. Then I will go on to discuss two applications of this approach. In the first part of the seminar, I will present GAP simulations of amorphous carbon depositions which allowed us to explain, for the first time, how the "diamond-like" properties of dense amorphous carbon arise [2]. In the second part, I will introduce a new method that we have developed to predict adsorption energies, with application to amorphous carbon surfaces [3]. I will also present a new type of atomic descriptor which allows us to improve the predictive ability of GAP models and therefore bring them closer to full DFT accuracy [3]. [1] A.P. Bartók, M.C. Payne, R. Kondor, G. Csányi. Phys. Rev. Lett. 104, 136403 (2010). [2] M.A. Caro, V.L. Deringer, J. Koskinen, T. Laurila, and G. Csányi. Phys. Rev. Let. 120, 166101 (2018). [3] M.A. Caro, A. Aarva, V.L. Deringer, G. Csányi, and T. Laurila. Chem. Mater. 30, 7446 (2018).

Presentation materials

There are no materials yet.