TSU COLLOQUIUM
Speaker: Dr. Bikash Kanungo
Affiliation: Research Scientist, Dept. of Mechanical Engineering, University of Michigan
Title: Towards large-scale ground-state and time-dependent density functional theory at quantum accuracy.
Date and Time: 21 February 2023 (Tuesday) at 02:30 pm (Tea/Coffee: 02:15 pm)
Venue: Nevill Mott Hall, JNCASR
Abstract:
Density functional theory (DFT) and time-dependent density functional theory (TDDFT), owing to their great balance of speed and accuracy, have remained essential tools to understand all manners of nanoscale processes and materials behavior. Although, in principle, an exact theory, in practice, DFT (TDDFT) requires approximations to the exact exchange-correlation (XC) functionals to encapsulate the quantum many-electron interactions into a mean-field of the electron density. The existing XC approximations, despite their successes, exhibit several notable deficiencies—inaccurate band-gaps, bond-dissociation curves, reaction barriers, to name a few. These deficiencies of the existing XC approximations severely limit the reliability of DFT (TDDFT) in predictions of material properties. Additionally, the high computational demands of DFT (TDDFT) limits their routine usage to length-scales of few hundred atoms and timescales of few tens of picoseconds for ab initio molecular dynamics (AIMD) (few tens of femtoseconds for TDDFT). This, in turn, renders a wide array of problems—energetics of dislocation in crystalline materials, dynamics of water splitting, surface plasmons in metal nanoclusters, to name a few—inaccessible to DFT (TDDFT). In this talk, I will present different strategies to address the above accuracy and efficiency challenges in DFT/TDDFT. First, I will introduce a data-driven approach to model the XC approximation. In particular, I will present an accurate and robust solution to the inverse DFT problem that connects DFT to the wavefunction based methods (e.g., configuration interaction, quantum Monte Carlo), and hence, is crucial to the generation of training data needed to model the XC approximation. Subsequently, I will discuss various machine-learning approaches to construct the XC approximation, using the training data from inverse DFT. Lastly, I will present various efficient spatial and temporal discretization schemes, ranging from mixed basis formulation to exponential time-integrators, that can enable large-scale DFT/TDDFT calculations than possible, heretofore.
,
,
