ICAMS / Interdisciplinary Centre for Advanced Materials Simulation


Neural network potentials for atomistic simulations

Date: 15.12.2008
Place: ICAMS, UHW Raum 1102

Jörg Behler, Department of Theoretical Chemistry, Ruhr-Universität Bochum, Bochum, Germany

The reliability of the results obtained in theoretical simulations strongly depends on the quality of the employed interatomic potentials. While electronic structure methods like density-functional theory (DFT) provide an accurate description of many materials, the high computational costs severely limit the system size that can be studied. In the past decade artificial Neural Networks (NN) have become a promising new tool for the construction of accurate potential-energy surfaces (PES) [1,2]. NN potentials are based on electronic structure calculations, but once they have been constructed, they are several orders of magnitude faster to evaluate than the underlying electronic structure methods, while the accuracy is essentially maintained. Thus they are ideally suited for large-scale molecular dynamics simulations, but so far NN potentials have been applicable only to rather low dimensional PESs. In this talk an extension of the NN methodology to high-dimensional PESs of condensed systems is presented [3], now enabling applications in the field of materials science. The capabilities of the method are demonstrated for silicon as a benchmark system. We show that by a series of NN-based metadynamics simulations the full sequence of pressure-induced phase transitions can be obtained in excellent agreement with experiment [4]. Further, the prospects for an extension of the methodology to more complex systems are discussed.

[1] T.B. Blank, S.D. Brown, A.W. Calhoun and D.J. Doren, J. Chem. Phys. 103, 4129 (1995).
[2] S. Lorenz, A. Gro, and M. Scheffer, Chem. Phys. Lett. 395, 210 (2004).
[3] J. Behler, and M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007).
[4] J. Behler, R. Martok, D. Donadio, and M. Parrinello, Phys. Rev. Lett. 100, 185501 (2008).

Supporting information:

« back