Utilizing Machine Learning for the Calculation of Vibrational Frequencies of Molecular Crystals
Conference Year
January 2021
Abstract
The vibration of atoms within a crystalline lattice can give insight into properties of the bulk material. For example, thermal properties such as heat capacity and entropy can be determined from the lattice dynamics of a crystal. Theoretical methods such as density functional theory (DFT) can be utilized to simulate the vibrational dynamics of a system with superior accuracy. But, the computational complexity of vibrational DFT simulations increases rapidly as the number of atoms increases. Less complex methods such as molecular dynamics can be utilized to simulate lattice dynamics, but sacrifice accuracy for speed due to the use of empirical potentials. High-dimensional neural network potentials (HDNNPs) are a machine learning technique developed to combine the accuracy of DFT with the speed of empirical potentials. HDNNPs are based on artificial neural networks that are trained replicate the potential energy surface of a chemical structure. This work presents the preliminary results in calculating the vibrational frequencies of various molecular crystals with HDNNPs. Furthermore, the HDNNPs utilized in this study can be employed to determine anharmonic corrections to the vibrational frequencies.
Primary Faculty Mentor Name
Michael Ruggiero
Status
Graduate
Student College
College of Arts and Sciences
Program/Major
Materials Science
Primary Research Category
Engineering & Physical Sciences
Utilizing Machine Learning for the Calculation of Vibrational Frequencies of Molecular Crystals
The vibration of atoms within a crystalline lattice can give insight into properties of the bulk material. For example, thermal properties such as heat capacity and entropy can be determined from the lattice dynamics of a crystal. Theoretical methods such as density functional theory (DFT) can be utilized to simulate the vibrational dynamics of a system with superior accuracy. But, the computational complexity of vibrational DFT simulations increases rapidly as the number of atoms increases. Less complex methods such as molecular dynamics can be utilized to simulate lattice dynamics, but sacrifice accuracy for speed due to the use of empirical potentials. High-dimensional neural network potentials (HDNNPs) are a machine learning technique developed to combine the accuracy of DFT with the speed of empirical potentials. HDNNPs are based on artificial neural networks that are trained replicate the potential energy surface of a chemical structure. This work presents the preliminary results in calculating the vibrational frequencies of various molecular crystals with HDNNPs. Furthermore, the HDNNPs utilized in this study can be employed to determine anharmonic corrections to the vibrational frequencies.