Molecular Structure - Properties Mapping
Mapping
Molecular weight 29.06 g/mol
Boiling Point -154.8 °F
Std enthalpy of formation
+52.47 kJ/mol
Atomization energy
...
Structure PropertiesFigure 1.[1] Structure of ethylene
Motivation
* …
Future
Machine Learning > Solve Schrödinger equation (SE)
❏https://nanohub.org/tools/cnn2d/status❏Tool name: ‘Molecular Quantum Machine’❏Input: a string that includes the Coulomb
matrix of a molecule❏Output: the predicted atomization energy of
that molecule
Online Simulation Tool
QM7 dataset (a subset of GDB-13) 23 by 23 matriceshttp://quantum-machine.org/datasets/#qm77165 organic molecules Up to seven ‘heavy’ atoms including C, N, O, S
Data Source
Input Data representation
Figure 2. [2] Coulomb matrix of an ethylene molecule
1.Convolutional Neural Network (CNN)2.Fully-connected Neural Network
TensorFlow based network ● An open source library developed by
Google● Designed for Machine Learning
Two Machine Learning networks
Figure 3. Fully connected neural network vs. Convolutional Neural NetworkMarc’Aurelio Ranzato: Neural Nets for Vision https://www.slideshare.net/zukun/p03-neural-networks-cvpr2012-deep-learning-methods-for-vision
Fully connected neural network Convolutional neural network
1. Convolutional Neural Network● Parallel permutations● 4 convolutional layers (2*2 kernel size)● Seperable convolutions● Max pooling
Input Coulomb Matrix X
Permutation X1
TF
Prediction Y1
Permutation X2
TF
Prediction Y2
Permutation X3
TF
Prediction Y3
Permutation X4
TF
Prediction Y4
Permutation X5
TF
Prediction Y5
Mean of all predictions Y
Optimize all the TF layers
2. Fully-connected (Dense) Neural Network●6 dense layers in total●4 of them are residual layers●Use ReLu activation functions instead of
sigmoid activation functions
Results & ConclusionConvolutional
Neural NetworkFully-connected Neural Network
Training time 2~3 hours 3~4 hours
Mean Absolute Error (kcal/mol)
11.5 12.8
Root Mean Squared Error
(kcal/mol)
16.48 17.9
Error ~0.72% ~0.81%
Fully-connected Neural Network
Convolutional Neural Network
1. https://en.wikipedia.org/wiki/Ethylene2. Hansen, Montavon, Biegler, Fazli, Rupp, Scheffler, Von Lilienfeld,
Tkatchenko, and Müller. "Assessment and Validation of Machine Learning Methods for Predicting Molecular Atomization Energies." Journal of Chemical Theory and Computation 9.8, 2013. 3404-19. Web.
Reference