DATE:

Friday, September 30, 2022

TIME:

3:30 PM

LOCATION:

GMCS 314

SPEAKER:

Dr. Yuezhi Mao, Chemistry, San Diego State University

ABSTRACT:

Electron transfer and electronic excitations are of key importance in processes harnessing light energies such as photosynthesis, photovoltaics, and photoredox catalysis. In this seminar, I will introduce our recent advances in the theoretical modeling of these processes using ab initio quantum chemistry as well as machine-learning (ML) approaches. I will first demonstrate how one can utilize density functional theory (DFT) calculations with absolutely localized molecular orbitals (ALMOs) to construct charge-localized diabatic states, the initial and final states of an electron transfer process, and to accurately evaluate the electronic coupling between them, an important parameter that governs the electron transfer rate. These ALMO-based diabatic states can be computed at a similar cost as a ground-state DFT calculation for the same system, and the nuclear forces associated with each diabatic potential energy surface can be easily obtained, making them suitable for the on-the-fly dynamics situations of electron transfer processes. We have further extended this approach to the modeling of photoinduced electron/hole transfer processes by combining ALMO-based DFT calculations with the ∆SCF method for excited states.

I will then introduce our recent development of ML models for the prediction of linear and multidimensional electronic spectra, using the example of Green Fluorescence Protein (GFP) chromophore in water. For this system, ML models fitted to excitation energies calculated by the commonly used TDDFT method turn out to severely underpredict the width of the linear absorption spectra, suggesting the necessity of using training data generated by higher-level, albeit computationally more demanding excited-state methods. Here we have developed a data-efficient approach based on transfer learning of excitation energies computed using EOM-CCSD, a high-level excited-state method, embedded in DFT environments. This transfer-learning model, when applied to the prediction of linear and 2D electronic spectra for GFP chromophore in water, was shown to capture the width of the former more accurately and yielded a more pronounced dynamical Stokes shift for the latter as compared to the corresponding predictions using models trained solely on TDDFT data. We have further revealed that these differences originate from the stronger coupling between the chromophore’s excitation energy and the hydrogen-bonding environment predicted by the higher-level EOM-CCSD method.

HOST:

Andrew Cooksy

VIDEO: