Colloquium: Dr. Kamal Choudhary, NIST

ABSTRACT:

In this talk, we’ll discuss deep learning methods 1) Graph neural network (GNN) for improved atomistic material property predictions of solids and molecules, 2) Convolutional neural network for STM and STEM image related tasks, and 3) quantum algorithm method: Variational Quantum Eigensolver (VQE) for predicting electron and phonon properties. Many GNN models for atomistic property predictions are based on bond-distances mainly. We developed Atomistic Line Graph Neural Network (ALIGNN) that performs message passing on both the bond-distances as well as bond-angles. We apply ALIGNN to train 60 material property models in the Materials Project, JARVIS-DFT, hMOF and QM9 datasets leading to up to 44 % improved performance compared to previously known GNN methods. Next, we’ll discuss the AtomVision package which can be used to generate scanning tunneling microscope (STM) and scanning transmission electron microscope (STEM) datasets. Then we apply deep learning frameworks for image classification and defects detection tasks for 2D materials. Currently, the application of quantum algorithms such as VQE is mainly limited to molecules. We’ll show using tight-binding approaches for electrons and phonons, quantum circuit-based methods can be applied for solids using AtomQC package. All of the above projects are part of the NIST-JARVIS infrastructure (https://jarvis.nist.gov/).