Condensed Matter Seminars
Thursday, September 30, 2021
Physics Building, Room 204
Aravind Krishnamoorthy , University of Southern California
[Host: Utpal Chatterjee]
The accelerated discovery and design of new quantum materials requires atomic-level information about chemical reactions, phase transformations, mechanical deformations and other collective and emergent quantum phenomena. Several techniques have been developed recently that can learn the potential energy surface (PES) of complex materials. Machine Learning (ML) models, particularly deep neural networks, have proven capable of learning highly complex non-linear relationships between atomic structure and properties and theory and experiments. In this talk, I will describe two examples of ML-driven MD called neural-network quantum molecular dynamics (NNQMD) to tackle problems related to large systems and long trajectories that cannot be investigated by Quantum Molecular Dynamics (QMD).
First, we use NNQMD for quantitatively characterizing the intermediate range order, manifested as first sharp diffraction peak in GeSe2. In the second example, we compute the dielectric constant, ε0, and its temperature dependence for liquid water using fluctuations in macroscopic polarization using two coupled neural network models. The first network, NNQMD, learns the PES of liquid water from QMD training data. The second network, neural-network maximally localized Wannier functions, NNMLWF, is trained to predict dipole moments.
I will also briefly discuss applications of ML to discovery of new dielectric polymer materials with high breakdown strengths and to optimization of chemical vapor deposition synthesis of quantum materials.
To add a speaker, send an email to phys-speakers@Virginia.EDU. Please include the seminar type (e.g. Condensed Matter Seminars), date, name of the speaker, title of talk, and an abstract (if available).