BROWSE

ITEM VIEW & DOWNLOAD

MACHINE LEARNING - BASED EXCITED STATE MOLECULAR DYNAMICS

Cited 0 times inthomson ciCited 0 times inthomson ci
Title
MACHINE LEARNING - BASED EXCITED STATE MOLECULAR DYNAMICS
Author
Kim, Kicheol
Advisor
Min, Seung Kyu
Issue Date
2019-08
Publisher
Graduate School of UNIST
Abstract
We present a new methodology for excited state molecular dynamics (ESMD) (or nonadiabatic molecular dynamics) based on machine learning (ML) technique. The most time consuming process in conventional on-the-fly ESMD simulations is an electronic structure calculation including analytic gradients of potential energy surfaces (PESs). Our study proposes that we can bypass this by exploiting ML. We consider ensemble density functional theory, especially state-interaction state-averaged spin-restricted ensemble-referenced Kohn-Sham (SI-SA-REKS, or SSR for brevity) method as electronic structure calculation for the ML model since SSR(2,2) provides two diabatic electronic states, namely perfectly spin-paired singlet (PPS) and open-shell singlet (OSS), and their analytic gradients as well as interstate couplings ( SA). In this study, we exploit the SchNetPack ML python library for ML procedure. For compatibility, the decoherence-induced surface hopping based on exact factorization (DISH-XF) program is implemented in Python (pyDISH-XF) for nuclear dynamics. Some part of pyDISHXF is written in C programming language to minimize the slowdown. We investigate ESMD of an ethylene molecule to benchmark pyDISH-XF. The performance of pyDISH-XF is better than UNI-xMD (a reference program is Fortran90). For overall ML-based ESMD, we focus on photo-induced cis-trans isomerization of trans-Penta-2,4-dieniminium cation (PSB3), which is a typical model molecule for rhodopsin. The SchNet model is trained with 7500 and 50000 PSB3 geometries getting from previous ESMD studies. We get the poor result with 7500 training set, while the result with 50000 training set is reliable. Mean absolute error (MAE) energy and force of PPS is evaluated 0.01001 eV and 0.01750 eV/Å, MAE energy and force of OSS is evaluated 0.00888 eV and 0.01785 eV/Å, and MAE energy and force of SA is evaluated 0.00948 eV and 0.03720 eV/Å. We investigate ESMD of PSB3 with ML and compare the result with conventional ESMD of PSB3 with SSR method. The result is comparable to the conventional result. Furthermore, it takes 21 minutes using 1 cpu thread, while the conventional result takes 1.7 days using 2 gpu for propagating one trajectory.
Description
Department of Chemistry
URI
Go to Link
Appears in Collections:
PHY_Theses_Master
Files in This Item:
MACHINE LEARNING - BASED EXCITED STATE MOLECULAR DYNAMICS.pdf Download

find_unist can give you direct access to the published full text of this article. (UNISTARs only)

Show full item record

qrcode

  • mendeley

    citeulike

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

MENU