MACHINE-LEARNING INTERATOMIC POTENTIAL FOR ULTRA-WIDE BANDGAP SEMICONDUCTOR AND X-RAY SCATTERING STUDY OF CHARGE DENSITY WAVE

Abstract

Understanding the role of native defects, impurities, and dopants in reducing the intrinsicallyhigh thermal conductivity of wurtzite AlN is of importance for effective thermal management of next-generation devices. A computational approach is pursued that implements a machine-learning interatomic potential (MLP) to transfer the accuracy of first principles methods to the lengthscales accessible by molecular dynamics (MD) simulations. MLPs are trained on a dataset obtained by density functional theory calculations to map atomic configurations to per atom energies, reconstructing the potential energy surface. In this work, a training dataset is generated and two styles of MLP, a neural network potential and a gaussian approximation potential, are trained and evaluated. The established framework for MLPs, which is limited to the consideration of short- range interactions, is not sufficient to reproduce the potential energy surface of wurtzite AlN due to the material’s piezoelectric nature. Transition metal dichalcogenides host charge density wave (CDW) states, a periodic modulation of electron density accompanied by a periodic lattice distortion that emerges below a transition temperature. A signature of the CDW state is the appearance of superlattice reflections in diffraction patterns. Synchrotron source x-ray scattering measurements of bulk 2H-TaSe2 were taken on a cooling-warming cycle over the range of the commensurate and incommensurate CDW transition temperatures. The modulation vector defining the periodicity of the CDW state exhibited thermal hysteresis, and the commensurate CDW transition shifted on the warming cycle. The persistence of superlattice reflections above the incommensurate transition temperature indicated local ordering prior to the onset of the CDW state.