Artificial intelligence is emerging as a new paradigm in materials science. This talk describes how physical intuition and (insightful) machine learning can solve the complicated task of structure recognition in materials at the nanoscale.
The Face of Nanomaterials: Insightful Classification Using Deep Learning - Angelo Ziletti
1. The Face of Nanomaterials:
Insightful Classification
Using Deep Learning
Dr. Angelo Ziletti
Deputy Group Leader in Data Science for Materials
Fritz Haber Institute of the Max Planck Society
Berlin, Germany
Berlin, July 8th
, 2018
3. 3
● Ruled by the laws of Quantum Physics
What is a nanomaterial?
International Organization for Standardization (ISO)
"Material with any external dimension in the nanoscale or having internal
structure in a size range from approximately 1 nm to 100 nm."
(A human hair is approximately 80,000- 100,000 nanometers wide)
4. 4
Why are nanomaterials important?
LEDs
Nobel Prize
Physics 2014
(blue LED)
Lasers
Nobel Prize
Physics 1964, 1981
Computers
Nobel Prize
Physics 1956
(transistor)
Levitating Trains
Nobel Prize
Physics 1972
(th. superconductivity)
… and many others...
5. 5
● Graphene:
– Single layer of graphite (carbon), 1-atom thick
– strongest material ever discovered (tensile strength= 130GPa)
– lowest known resistivity at room temperature
– better heat conductor than silver and copper
– 97% transparent
An example: two-dimensional materials
Nobel Prize
2010
Model Experiment Fabrication
7. 7
● Given an atomic arrangement in a nanomaterial, determine the (“most
similar”) prototype among the following classes:
The goal
Body-centered-tetragonal
(139)
Body-centered-tetragonal
(141)
Hexagonal
Simple cubic Face-centered-cubic Diamond Body-centered-cubic
Rhombohedral
8. 8
Structures are quite (very?) similar
Simple
cubic
Body-centered
cubic
Face-centered
cubic
9. 9
Structures are quite (very?) similar
Simple
cubic
Body-
centered
cubic
Face-
centered
cubic
Ref: B. A. Averill and P. Eldredge, Chemistry: Principles, Patterns, and Applications, Prentice Hall (2007)
10. 10
And with atom removals/deformations...
Simple
cubic
Body-centered
cubic
Face-centered
cubic
12. 12
● Nanomaterials are complex, non-rigid, three-dimensional objects with
periodically repeated structures (like the brick of a house)
● A good representation of nanomaterials must be:
– invariant with respect to system size
– stable with respect to deformations and atoms removal
Feature Engineering for periodic 3D objects
Perfect structure 25% atoms removed Random deformation
13. 13
● … and ideally:
– the representation is compact
– nanomaterials belonging to a similar class have a similar
representation
● Learning symmetries by data augmentation?
... but for each structure we would need to give:
– Nanomaterials of different sizes
– All (!) distorted configurations
→ a huge amount of data (and no learning guarantee)
Feature Engineering for periodic 3D objects
14. 14
The diffraction fingerprint: intuition
Crystal
structure to
classify
Diffraction
fingerprint
Simulated
radiation
● Rotate the crystal structure of 45°
and (-45°) about the x,y, and z axis
● Calculate the diffraction pattern
(~Fourier Transform) for each
rotation:
– around x-axis
– around y-axis
– around z-axis
● Sum the results in a RGB image
Ziletti et al., Nature Communications, in press; arXiv: 1709.02298 (2018).
15. 15
The diffraction fingerprint: results
Body-centered-tetragonal
(139)
Body-centered-tetragonal
(141)
Rhombohedral/Hexagonal
Simple cubic Face-centered-cubic Diamond Body-centered-cubic
Ziletti et al., Nature Communications, in press; ArXiv: 1709.02298 (2018).
18. 18
● A standard n-layer neural network applies to the input data a series of linear
and non-linear transformations in successions:
– non-linear operators: ReLU, sigmoid, max-pooling, softmax.
– : weight matrices and bias vectors
● Neural networks have been extremely successful in a large variety of task
(computer vision, speech recognition, machine translation, etc)
● For image recognition: Convolutional Neural Network (ConvNets)[1]
Prediction model: neural network
[1] LeCun et al., Neural Comput. 1, 541 (1989)
19. 19
How do we (humans) subconsciously classify an image?
Looking for identifiable (pre-learned) features (e.g. for dogs: paws, 4 legs)
ConvNets: human analogy
How does a computer classify an image?
Looking at low level features (edges and curves), and then build more
abstract concepts though a series of (convolutional) layers.
20. 20
Computing a convolution
Ref: V. Dumoulin, F. Visin, A guide to convolution arithmetic for deep learning, https://arxiv.org/abs/1603.07285 (2016)
● Slide kernel throughout the image
● For each position in the image:
– Element-wise multiplication between
image and kernel
– Sum of all elements (within the region)
Output
Input
Kernel
21. 21
Computing a convolution: example
Ref: V. Dumoulin, F. Visin, A guide to convolution arithmetic for deep learning, https://arxiv.org/abs/1603.07285 (2016)
24. 24
Convolutional layer recap
● Convolution is spatial filtering
● Different filters (weights) extract different
characteristics of the input → multiple filters
● Complexity of the filters increases layer by layer
● Filters learned minimizing the training error
● Multiple conv. Layers:
– 1st
layer: input=image → low-level filters (e.g. curve or straight edges)
– 2nd
layer: input=activation map → higher level filters (e.g. semicircles:
curve+straight edges, squares)
– nth
layer: high level filters (e.g. face)
25. 25
Pooling layer
● Replaces the output at a certain location with a summary statistic of
the nearby outputs
● Makes the representations smaller (downsampling)
● Different poolings: e.g. max pooling, average pooling
● It is not crucial and can be avoided
Images from Stanford CS231n: Convolutional Neural Networks for Visual Recognition (http://cs231n.github.io/convolutional-networks/)
28. 28
● Dataset 1:
– Includes ~90 chemical elements
– Different nanomaterials’ sizes
● Dataset numbers:
– 10,517 images; 7 classes
– 90% training, 10% validation (randomly)
– ConvNet runtime: train: ~80min, pred. ~70 ms @img
The pristine dataset
Training accuracy [%] Validation accuracy [%]
100.0 100.0
29. 29
● Dataset 2: dataset 1 with added defects
– Random displacements: up to st. dev. 0.06 Å
– Random vacancies: up to 25%
– Substitutions (randomly change the type of
atom: e.g. C -> H)
● Dataset numbers:
– 105,170 images
– 7 classes
The defective dataset (test set)
Training accuracy [%] Test accuracy [%]
No Training 100.0
32. 32
Comparison with materials science state-of-the-art
● Our deep learning-based method outperforms the
state-of-the-art approach
● “Fairness” note: smaller number of materials
classes (so far), need correctly labeled (!) training
data
Spglib: Grosse-Kunstleve, Acta Crystallographica A, 55, pp. 383 (1999); A. Togo, https://atztogo.github.io/spglib/ (2009)
Deep learning-based: Ziletti et al., Nature Communications, in press, arXiv: 1709.02298 (2018)
34. 34
Back-projection to image space
Method: Zeiler and Fergus, European Conf. on Computer Vision, Springer, 2014.
● Project feature activities back to the input pixel space
35. 35
“Going backwards” in a convolutional layer
Method: Zeiler and Fergus, European Conf. on Computer Vision, Springer, 2014.
TransposedConvolution: Im et al., Generating images with recurrent adversarial networks, arXiv: 1602.05110 (2016)
Input to layer
Convolution
Pooling
Next layer
Nonlinearity
Reconstruction
Fractionally strided
convolution
Unpooling
Layer above reconstruction
Nonlinearity
Forward pass Going backwards: reconstruction
Also called:
- Transposed convolution
- Backward strided convolution
- Deconvolution
In Tensorflow:
tf.nn.conv2d_transpose
36. 36
Attentive response maps: forward pass
● Forward pass of the image
– for each pooling layer: store pool switches
– for conv. layer of interest (e.g. last):
● calculate filters’ activations
● order filters by activation value
– select the top most-activated filters
Method: Zeiler and Fergus, European Conf. on Computer Vision, Springer, 2014.
Application to anatomy classification: Kumar et al., IEEE Int. Symp. on Biomed. Imaging, arXiv: 1611.06284 (2018)
Application to materials science: Ziletti et al., Nature Communications, in press, arXiv: 1709.02298 (2018)
Input image
ClassificationConv
Layer 1
Conv
Layer 2
Last Conv
Layer
FC
Layers
...
37. 37
Attentive response maps: back-projection
Input image
Conv
Layer 1
Conv
Layer 2
Last Conv
Layer
...
● Back-propagate to image space the top most-activated filters
– for max-pooling layers→ unpooling
– for convolutional layers→ fractionally strided convolution
Method: Zeiler and Fergus, European Conf. on Computer Vision, Springer, 2014.
Application to anatomy classification: Kumar et al., IEEE Int. Symp. on Biomed. Imaging, arXiv: 1611.06284 (2018)
Application to materials science: Ziletti et al., Nature Communications, in press, arXiv: 1709.02298 (2018)
38. 38
Attentive response maps: per-pixel max
Input image
Conv
Layer 1
Conv
Layer 2
Last Conv
Layer
...
● Compute the per-pixel max of
these back-projected maps
Max
Individual response maps
Attentive response map
Method: Zeiler and Fergus, European Conf. on Computer Vision, Springer, 2014.
Application to anatomy classification: Kumar et al., IEEE Int. Symp. on Biomed. Imaging, arXiv: 1611.06284 (2018)
Application to materials science: Ziletti et al., Nature Communications, in press, arXiv: 1709.02298 (2018)
39. 39
Attentive response maps (Summary)
● Forward pass of the image
– for each pooling layer: store pool switches
– for conv. layer of interest (e.g. last):
● calculate filters’ activations
● order filters by activation value
– select the top most-activated filters
● Back-propagate to image space the top most-activated filters
– for max-pooling layers→ unpooling
– for convolutional layers→ fractionally strided convolution
● Compute the per-pixel max of these back-projected maps
Method: Zeiler and Fergus, European Conf. on Computer Vision, Springer, 2014.
Application to anatomy classification: Kumar et al., IEEE Int. Symp. on Biomed. Imaging, arXiv: 1611.06284 (2018)
Application to materials science: Ziletti et al., Nature Communications, in press, arXiv: 1709.02298 (2018)
40. 40
Understanding ConvNets
Devinder Kumar
(University of
Waterloo, Canada)
Attentive response maps
Input
image
Layer 1 Layer 2 Layer 3 Layer 4 Layer 5 Layer 6
green
red
Ziletti et al., Nature Communications, in press; ArXiv: 1709.02298 (2018).
42. 42
What did the ConvNet learn?
● Sum of the last convolutional layer attentive response maps:
● has learned nanomaterials templates automatically from the data
● uses the same landmarks a materials scientist would use
although never explicitly instructed to do so
Our ConvNet:
43. 43
● The challenge
● How to represent a nanomaterial
● Convolutional Networks
● Opening the black-box
Summary
46. Dr. Angelo Ziletti
Fritz Haber Institute of the Max Planck Society, Berlin, Germany
Insightful Classification of Crystal Structures
Using Deep Learning
Ziletti et al., Nature Communications, in press (2018).
Online: https://arxiv.org/abs/1709.02298
ziletti@fhi-berlin.mpg.de