SchNet: A continuous-filter
convolutional neural
network for modeling quantum
interactions
Jin-Woo Jeong
Network Science Lab
Dept. of Mathematics
The Catholic University of Korea
E-mail: zeus0208b@catholic.ac.kr
Schütt, Kristof, et al. "Schnet: A continuous-filter convolutional neural network
for modeling quantum interactions." Advances in neural information
processing systems 30 (2017).
2
 Introduction
 Continuous-filter convolutions
 SchNet
 Experiments and results
 Conclusions
Q/A
3
Introduction
Introduction
• In recent years, there have been increased efforts to use machine learning for the accelerated discovery
of molecules and materials with desired properties.
• However, these methods are only applied to stable systems in so-called equilibrium.
• Data sets such as the established QM9 benchmark contain only equilibrium molecules.
• Predicting stable atom arrangements is in itself an important challenge in quantum chemistry and
material science.
4
Introduction
Contributions
• Propose continuous-filter .
• Propose : a neural network specifically designed to respect essential quantum chemical constraints.
• Present a new, challenging benchmark – ISO17 – including both chemical and conformational changes.
5
Continuous-filter convolutions
Continuous-filter convolutions
• Normal convolutional layers are not sufficient for unevenly spaced inputs such as the atom positions of a
molecule
• They propose to use continuous filters that are able to handle unevenly spaced data, in particular, atoms at
arbitrary positions.
6
Continuous-filter convolutions
Continuous-filter convolutions
• Given the feature representations of n objects with at locations with the continuous-filter convolutional
layer requires a filter-generating function
7
SchNet
Architecture
• SchNet is designed to learn a representation for the prediction of molecular energies and atomic forces.
• : non-linear
8
SchNet
Architecture
9
SchNet
Training with energies and forces
• the interatomic forces are related to the molecular energy, so that we can obtain an energy-conserving
force model by differentiating the energy model w.r.t. the atom positions
• Loss fuction:
10
Experiments and results
Datasets
• QM9 – chemical degrees of freedom
• Task: Predicting chemical properties
• Contains only equilibrium molecules
• Diverse chemical property data
• Molecule size: up to 9 heavy atoms (C, O, N, F included)
• Includes energy, dipole moment, HOMO/LUMO, etc.
• MD17 – conformational degrees of freedom
• Task: Molecular dynamics
• Predicting conformational changes of single molecules
• Includes molecular trajectories
• Provides energy and force data
• Based on small organic molecules
• ISO17 – chemical and conformational degrees of freedom
• Task: Predicting chemical + conformational variations
• Includes isomers ()
• Data on both chemical and structural variations
• 645,000 labeled examples
11
Experiments and results
Results
12
Experiments and results
Results
13
Experiments and results
Results
14
Conclusion
Conclusion
• SchNet introduces continuous-filter convolutional layers to model quantum interactions in molecules,
respecting fundamental quantum-chemical principles like energy conservation and rotational invariance.
• It achieves state-of-the-art performance on QM9 and MD17 datasets and establishes ISO17 as a
challenging benchmark for modeling both chemical and conformational variations, paving the way for
advancements in machine learning-driven quantum chemistry.
15
Q & A
Q / A

250106_JW_labseminar[SchNet: A continuous-filter convolutional neural network for modeling quantum interactions].pptx

  • 1.
    SchNet: A continuous-filter convolutionalneural network for modeling quantum interactions Jin-Woo Jeong Network Science Lab Dept. of Mathematics The Catholic University of Korea E-mail: zeus0208b@catholic.ac.kr Schütt, Kristof, et al. "Schnet: A continuous-filter convolutional neural network for modeling quantum interactions." Advances in neural information processing systems 30 (2017).
  • 2.
    2  Introduction  Continuous-filterconvolutions  SchNet  Experiments and results  Conclusions Q/A
  • 3.
    3 Introduction Introduction • In recentyears, there have been increased efforts to use machine learning for the accelerated discovery of molecules and materials with desired properties. • However, these methods are only applied to stable systems in so-called equilibrium. • Data sets such as the established QM9 benchmark contain only equilibrium molecules. • Predicting stable atom arrangements is in itself an important challenge in quantum chemistry and material science.
  • 4.
    4 Introduction Contributions • Propose continuous-filter. • Propose : a neural network specifically designed to respect essential quantum chemical constraints. • Present a new, challenging benchmark – ISO17 – including both chemical and conformational changes.
  • 5.
    5 Continuous-filter convolutions Continuous-filter convolutions •Normal convolutional layers are not sufficient for unevenly spaced inputs such as the atom positions of a molecule • They propose to use continuous filters that are able to handle unevenly spaced data, in particular, atoms at arbitrary positions.
  • 6.
    6 Continuous-filter convolutions Continuous-filter convolutions •Given the feature representations of n objects with at locations with the continuous-filter convolutional layer requires a filter-generating function
  • 7.
    7 SchNet Architecture • SchNet isdesigned to learn a representation for the prediction of molecular energies and atomic forces. • : non-linear
  • 8.
  • 9.
    9 SchNet Training with energiesand forces • the interatomic forces are related to the molecular energy, so that we can obtain an energy-conserving force model by differentiating the energy model w.r.t. the atom positions • Loss fuction:
  • 10.
    10 Experiments and results Datasets •QM9 – chemical degrees of freedom • Task: Predicting chemical properties • Contains only equilibrium molecules • Diverse chemical property data • Molecule size: up to 9 heavy atoms (C, O, N, F included) • Includes energy, dipole moment, HOMO/LUMO, etc. • MD17 – conformational degrees of freedom • Task: Molecular dynamics • Predicting conformational changes of single molecules • Includes molecular trajectories • Provides energy and force data • Based on small organic molecules • ISO17 – chemical and conformational degrees of freedom • Task: Predicting chemical + conformational variations • Includes isomers () • Data on both chemical and structural variations • 645,000 labeled examples
  • 11.
  • 12.
  • 13.
  • 14.
    14 Conclusion Conclusion • SchNet introducescontinuous-filter convolutional layers to model quantum interactions in molecules, respecting fundamental quantum-chemical principles like energy conservation and rotational invariance. • It achieves state-of-the-art performance on QM9 and MD17 datasets and establishes ISO17 as a challenging benchmark for modeling both chemical and conformational variations, paving the way for advancements in machine learning-driven quantum chemistry.
  • 15.

Editor's Notes

  • #15 thank you, the presentation is concluded