Deep Learning Spectroscopy

Neural Networks for Molecular Excitation Spectra

Kunal Ghosh, Annika Stuke, Milica Todorović, Peter Bjørn Jørgensen, Mikkel N. Schmidt, Aki Vehtari, Patrick Rinke*

*Corresponding author for this work

Research output: Contribution to journalJournal articleResearchpeer-review

159 Downloads (Pure)

Abstract

Deep learning methods for the prediction of molecular excitation spectra are presented. For the example of the electronic density of states of 132k organic molecules, three different neural network architectures: multilayer perceptron (MLP), convolutional neural network (CNN), and deep tensor neural network (DTNN) are trained and assessed. The inputs for the neural networks are the coordinates and charges of the constituent atoms of each molecule. Already, the MLP is able to learn spectra, but the root mean square error (RMSE) is still as high as 0.3 eV. The learning quality improves significantly for the CNN (RMSE = 0.23 eV) and reaches its best performance for the DTNN (RMSE = 0.19 eV). Both CNN and DTNN capture even small nuances in the spectral shape. In a showcase application of this method, the structures of 10k previously unseen organic molecules are scanned and instant spectra predictions are obtained to identify molecules for potential applications.

Original languageEnglish
Article number1801367
JournalAdvanced Science
Volume6
Number of pages7
ISSN2198-3844
DOIs
Publication statusPublished - 2019

Keywords

  • Artificial intelligence
  • DFT calculations
  • Excitation spectra
  • Neural networks
  • Organic molecules

Cite this

Ghosh, Kunal ; Stuke, Annika ; Todorović, Milica ; Jørgensen, Peter Bjørn ; Schmidt, Mikkel N. ; Vehtari, Aki ; Rinke, Patrick. / Deep Learning Spectroscopy : Neural Networks for Molecular Excitation Spectra. In: Advanced Science. 2019 ; Vol. 6.
@article{1340797fa49e46cb966f7272a1621780,
title = "Deep Learning Spectroscopy: Neural Networks for Molecular Excitation Spectra",
abstract = "Deep learning methods for the prediction of molecular excitation spectra are presented. For the example of the electronic density of states of 132k organic molecules, three different neural network architectures: multilayer perceptron (MLP), convolutional neural network (CNN), and deep tensor neural network (DTNN) are trained and assessed. The inputs for the neural networks are the coordinates and charges of the constituent atoms of each molecule. Already, the MLP is able to learn spectra, but the root mean square error (RMSE) is still as high as 0.3 eV. The learning quality improves significantly for the CNN (RMSE = 0.23 eV) and reaches its best performance for the DTNN (RMSE = 0.19 eV). Both CNN and DTNN capture even small nuances in the spectral shape. In a showcase application of this method, the structures of 10k previously unseen organic molecules are scanned and instant spectra predictions are obtained to identify molecules for potential applications.",
keywords = "Artificial intelligence, DFT calculations, Excitation spectra, Neural networks, Organic molecules",
author = "Kunal Ghosh and Annika Stuke and Milica Todorović and J{\o}rgensen, {Peter Bj{\o}rn} and Schmidt, {Mikkel N.} and Aki Vehtari and Patrick Rinke",
year = "2019",
doi = "10.1002/advs.201801367",
language = "English",
volume = "6",
journal = "Advanced Science",
issn = "2198-3844",
publisher = "Wiley",

}

Deep Learning Spectroscopy : Neural Networks for Molecular Excitation Spectra. / Ghosh, Kunal; Stuke, Annika; Todorović, Milica; Jørgensen, Peter Bjørn; Schmidt, Mikkel N.; Vehtari, Aki; Rinke, Patrick.

In: Advanced Science, Vol. 6, 1801367, 2019.

Research output: Contribution to journalJournal articleResearchpeer-review

TY - JOUR

T1 - Deep Learning Spectroscopy

T2 - Neural Networks for Molecular Excitation Spectra

AU - Ghosh, Kunal

AU - Stuke, Annika

AU - Todorović, Milica

AU - Jørgensen, Peter Bjørn

AU - Schmidt, Mikkel N.

AU - Vehtari, Aki

AU - Rinke, Patrick

PY - 2019

Y1 - 2019

N2 - Deep learning methods for the prediction of molecular excitation spectra are presented. For the example of the electronic density of states of 132k organic molecules, three different neural network architectures: multilayer perceptron (MLP), convolutional neural network (CNN), and deep tensor neural network (DTNN) are trained and assessed. The inputs for the neural networks are the coordinates and charges of the constituent atoms of each molecule. Already, the MLP is able to learn spectra, but the root mean square error (RMSE) is still as high as 0.3 eV. The learning quality improves significantly for the CNN (RMSE = 0.23 eV) and reaches its best performance for the DTNN (RMSE = 0.19 eV). Both CNN and DTNN capture even small nuances in the spectral shape. In a showcase application of this method, the structures of 10k previously unseen organic molecules are scanned and instant spectra predictions are obtained to identify molecules for potential applications.

AB - Deep learning methods for the prediction of molecular excitation spectra are presented. For the example of the electronic density of states of 132k organic molecules, three different neural network architectures: multilayer perceptron (MLP), convolutional neural network (CNN), and deep tensor neural network (DTNN) are trained and assessed. The inputs for the neural networks are the coordinates and charges of the constituent atoms of each molecule. Already, the MLP is able to learn spectra, but the root mean square error (RMSE) is still as high as 0.3 eV. The learning quality improves significantly for the CNN (RMSE = 0.23 eV) and reaches its best performance for the DTNN (RMSE = 0.19 eV). Both CNN and DTNN capture even small nuances in the spectral shape. In a showcase application of this method, the structures of 10k previously unseen organic molecules are scanned and instant spectra predictions are obtained to identify molecules for potential applications.

KW - Artificial intelligence

KW - DFT calculations

KW - Excitation spectra

KW - Neural networks

KW - Organic molecules

U2 - 10.1002/advs.201801367

DO - 10.1002/advs.201801367

M3 - Journal article

VL - 6

JO - Advanced Science

JF - Advanced Science

SN - 2198-3844

M1 - 1801367

ER -