Dr. Michele Pavanello, Rutgers University
Seminar Title: "Making and breaking electronic structures: lessons from embedding and machine learning"
Host: John Herbert, Herbert.44@osu.edu
ABSTRACT: The most enthusiastic modeler claims to accurately predict chemical reaction thermodynamics, kinetics, and nonequilibrium dynamics. Unfortunately, current models, while more robust and predictive than in past years, are often either too approximate to provide a faithful representation of reality or too computationally expensive to yield answers within a reasonable time. The talk argues that it is imperative to develop new-generation electronic structure methods to aid experiments, as these face different yet similarly difficult circumstances.
The talk introduces models based on orbital-free density-functional theory (OF-DFT), machine learning and density embedding to tackle large molecular condensed-phase systems that are too extensive for conventional electronic structure methods. To achieve chemical accuracy, nonstandard workflows are presented. These involve the dynamic combination of OF-DFT and conventional DFT methods, resulting in black-box-like adaptive embedding methods where molecular fragments are merged and split dynamically along a Born-Oppenheimer dynamics trajectory.
The presented methods are available to the broader community as open-source Python implementations, such as Quantum ESPRESSO in Python (QEpy), OF-DFT software (DFTpy), and density embedding software (eDFTpy). The argument is made that with such a software arsenal at our disposal, we are ready to tackle the most difficult and timely electronic structure challenges available today.
