Multidomain simulations of aluminum nitride with machine-learned force fields

D Behrendt and V Nascimento and AM Rappe, PHYSICAL REVIEW B, 110, 035204 (2024).

DOI: 10.1103/PhysRevB.110.035204

Aluminum nitride (AlN) and other wurtzite materials are widely used in piezoelectric microelectromechanical systems and are of great interest for future thin-film ferroelectric devices. Much progress has been made by modeling these materials with quantum mechanical methods such as density functional theory (DFT). However, there are very few existing methods that can model AlN on a larger scale, and none that can model multiple phases and domain walls with the accuracy of DFT. In this work, we present a machine-learned molecular dynamics force field (MLFF) for AlN constructed by fitting an artificial neural network to an underlying DFT dataset. Using our trained MLFF, we can predict the energies, forces, and phonon dispersions of AlN with the accuracy of DFT at dramatically lower computational cost. Accordingly, our MLFF can simulate systems orders of magnitude larger than DFT, enabling the study of emergent and long-range effects, such as the frequencydependent dielectric function and multiple ferroelectric domains. This method can easily be expanded to other wurtzite nitrides, oxides, and solid solutions.

Return to Publications page