Rapid Identification of X-ray Diffraction Spectra Based on Very Limited Data by

Interpretable Convolutional Neural Networks

Hong Wang1#, Yunchao Xie1#, Dawei Li1, Heng Deng1, Yunxin Zhao2, Ming Xin1, and Jian
Lin1,2,3*
1
Department of Mechanical and Aerospace Engineering
2
Department of Electrical Engineering and Computer Science
3
Department of Physics and Astronomy
University of Missouri, Columbia, Missouri 65211, USA
*
E-mail: [email protected] (J. L.)
#
Authors contributed equally to this work.

Abstract

Large volumes of data from material characterizations call for rapid and automatic data analysis

to accelerate materials discovery. Herein, we report a convolutional neural network (CNN) that

was trained based on theoretic data and very limited experimental data for fast identification of

experimental X-ray diffraction (XRD) spectra of metal-organic frameworks (MOFs). To augment

the data for training the model, noise was extracted from experimental spectra and shuffled, then

merged with the main peaks that were extracted from theoretical spectra to synthesize new spectra.

For the first time, one-to-one material identification was achieved. The optimized model showed

the highest identification accuracy of 96.7% for the Top 5 ranking among a dataset of 1012 MOFs.

Neighborhood components analysis (NCA) on the experimental XRD spectra shows that the

spectra from the same material are clustered in groups in the NCA map. Analysis on the class

activation maps of the last CNN layer further discloses the mechanism by which the CNN model

successfully identifies individual MOFs from the XRD spectra. This CNN model trained by the

data-augmentation technique would not only open numerous potential applications for identifying

XRD spectra for different materials, but also pave avenues to autonomously analyze data by other

characterization tools such as FTIR, Raman, and NMR.

Keywords: CNN, deep learning, interpretable, identification, X-ray diffraction

1
Introduction

High-throughput synthesis techniques have shown great potential in accelerating material

innovation.[1] Large volumes of characterization data including X-ray diffraction (XRD), Raman,

nuclear magnetic resonance (NMR), and Fourier Transform Infrared (FTIR) spectra are collected

during or after the synthesis. Among them, XRD is a powerful technique to characterize

crystallographic structures, grain size, and molecular structures.[2] Typically, experimental XRD

patterns are analyzed via comparing descriptors such as peak positions, intensities, full widths at

half maximum (FWHM) against a known database such as Crystallography Open Database and

Inorganic Crystal Structure Database, allowing scientists to identify the compounds of interest and

to map phase diagrams of combinatorial materials. However, the tedious and time-consuming

procedure due to the manual analysis at a relatively low speed severely hinders the fast decision

making.[2, 3] To fully exploit the characterization tools, it is becoming urgent to develop new data

assessment tools with automation and recommendation functions, especially with the emerging of

self-driven laboratories enabled by robots.[4-6] Despite recent progress, it has been and continues

to be a grand challenge.

Recently, machine learning (ML) models have shown great potentials in managing the large

volumes of characterization data for rapidly and automatically identifying composition-phase

maps as well as constructing composition-structure-property relationships, thereby speeding up

the materials discovery.[1, 7-15] For instance, Park et al. demonstrated well-trained convolutional

neural networks (CNNs) which exhibited satisfactory accuracy in classifying XRD spectra based

on theoretical database.[16] Oviedo and his colleagues proposed a machine learning approach to

predict crystallographic dimensionality and space groups from a limited number of thin-film XRD

spectra.[2] Angelo’s research group developed a robust CNN model to classify crystal structures

2
and also unfolded the internal behavior of the classification model through visualization.[14]

Miller’s research group implemented a CNN to determine crystallography trained on imaging and

diffraction data.[15] However, these approaches were applied to identify several classes or crystal

systems into which target materials are grouped. One-by-one identification of individual spectrum

from millions of spectrum databases is still challenging. Another big challenge for developing the

machine learning enabled methodology is the lack of experimental data for training the models.

Although a technique of Gaussian mixture was employed to augment the theoretic data,[17] it may

not fully reflect the real experiments when distinguished features arise from the experiments. It is

envisioned that directly incorporating experimental data into theoretic data is a better approach.

Finally, the deep learning models like CNN are usually treated as a “black-box”. Interpreting the

underlying mechanism of such as black-box for decision-making or obtaining the final desired

results is still an open problem. Therefore, developing a procedure that can better interpret the deep

learning models when they are applied to material research has recently seen a resurgence.

In this paper, we propose a CNN model that was trained for rapid one-to-one identification of

experimental XRD spectra of metal-organic frameworks (MOFs). To increase robustness of the

CNN model, noise was extracted from the experimental spectra to augment the theoretic spectra

for training. In the cases of very noisy experimental spectra, the fast Fourier transform (FFT) was

applied to reduce the noise before they were input into the CNN for improving the prediction

accuracy. The optimized CNN model showed the highest identification accuracy of 96.7% for the

Top 5 ranking among a dataset of 1012 MOFs. Data dimension reduction analysis on the

experimental XRD spectra by the neighborhood components analysis (NCA) shows that the

spectra from the same MOF are clustered in individual groups in the NCA map, while the XRD

spectra from different MOFs but with very similar characteristics may have overlapping. Further

3
analysis on the class activation maps (CAMs) of the fourteenth layer of the CNN model shows that

the grouped spectra that are highly distinguishable in the NCA map exhibit very different

activation characteristics. This observation can well explain why the CNN can identify individual

spectra from the library.

The novelty of this work can be summarized as follows. First, to the best of our knowledge,

this is the first demonstration that a CNN enables one-by-one identification of XRD spectrum for

individual materials. The previous reported machine learning algorithms only classify several

classes or crystal systems into which target materials are grouped. Second, the model was trained

by theoretical spectra combined with very limited experimental data. Third, the noise-based data

augmentation technique is very easy and straightforward to implement, but results in very effective

outcomes. Fourth, the trained CNN model can successfully and robustly perform one-by-one

classification with the help of noise filtering procedure even though the experimental XRD spectra

exhibit peak shift, scaling in peak intensities, or FWHM broadening compared to the theoretic

spectra. Lastly but not the least, the study on the CAMs of the convolutional layers discloses the

mechanism of how the CNN model makes the decision, thus shedding new light on the

interpretable deep learning for materials characterization data analysis. Consequently, the

proposed solution is of great interest and appears to be very promising, not only because of the

applications of XRD to characterize different types of materials, but also the possible extension to

spectra collected by other characterization techniques including Raman, NMR, and FTIR.

Results and Discussion

The flowchart showing the procedure for rapid identification of XRD spectra enabled by the

CNN is illustrated in Fig. 1a. First, theoretical CIF files of MOFs downloaded from Cambridge

4
Crystallographic Data Centre (CCDC)[18] were converted into theoretical XRD spectra.

Experimental XRD spectra were collected from as-synthesized MOFs powder in Bruker D8

Advance. Detailed synthesis and characterization can be found in Experimental Section. Then, the

noise was extracted from experimental spectra and shuffled, then merged with the main peaks that

were extracted from theoretical spectra to obtain new synthesized spectra. By this data

augmentation method, sufficient training datasets were realized. The experimental XRD spectra

that were used as the testing datasets were filtered to reduce the noise level if needed. The detailed

procedure of data augmentation and noise reduction is described in the following paragraph. A

CNN was built from scratch based on Lenet5 and VGG16.[19, 20] Its architecture is shown in Fig.

1b with hyperparameters summarized in Table S1. Basically, the CNN consists of one input layer,

four convolutional layers, three fully-connected layers, and one output layer. The first layer inputs

synthesized XRD spectra with 2θ ranging from 5 degrees to 50 degrees. Then the data is fed

subsequently into four convolution layers with kernel filters followed by a max pooling layer. The

kernels for these convolutional layers are six, sixteen, thirty-two, and sixty-four filters with a size

of 5×1 and a stride of one. The max pooling layer has two filters with a stride of two and a dropout

with a dropout rate of 0.2.[21-23] All convolutional layers were activated by the function of the

rectified linear unit (Relu). After one flatten layer, the data was fed to three dense layers with filters

of 120, 84, and 1012, respectively. The Adam optimizer was implemented to minimize the

categorical cross-entropy loss function.[24, 25]

5
Figure 1. (a) Flowchart showing the process of XRD spectrum identification. (b) Architecture of

proposed convolutional neural network.

Fig. 2 exhibits the detailed procedure of preprocessing and augmenting theoretic spectra,

reducing the noise level of experimental spectra, and training the CNN model. The spectra used

for training were synthesized by merging extracted peaks from the theoretic spectra and baseline

noise from the experimental spectra. The main peaks were extracted from a theoretic spectrum

containing the largest 400 points. The noise was collected from the baseline of raw experimental

spectra after the main peaks were removed. Then these two components were concatenated to form

a new spectrum. Since the noise can be randomly sampled and shuffled, the spectra can be largely

augmented for training the CNN model. These training spectra were synthesized from a library of

a total of 832 theoretical spectra. Next, FFT and inverse FFT (iFFT) were applied to smooth the

6
experimental spectra for the purpose of reducing their noise level. FFT is a popular algorithm that

converts the signal in an original domain to a representation in a frequency domain.[26] After the

raw XRD spectra were converted to FFT spectra in the frequency range of 0 ~ 200 Hz, the data

beyond the frequency of 200 is the white noise. Here, we use the mean peak value of the white

noise as a criterion to determine whether iFFT should be applied to reduce the noise of

experimental XRD spectra. If the mean value is larger than 1, it means that the noise has high

intensity and the iFFT is applied to first 200 Hz. Otherwise, the original data can be directly used

as testing data. A total of 24 experimental spectra collected from eight types of MOFs after noise

filtering were used as the testing datasets. Compared to previous reports on the procedure of

preprocessing XRD spectra, no much human intervention steps such as background removal,

smoothing and interpolation, region exclusion, and peak shifting were involved,[3, 27] thereby

significantly improving autonomy of the model for data analysis.

Figure 2. Flowchart showing process of augmenting theoretical spectra and filtering noise of

experimental spectra.
7
 After the data preprocessing, the synthesized spectra and noise-reduced experimental spectra

were used as the training and test datasets, respectively, to train and test various supervised ML

algorithms for evaluating one-by-one identification performance. We first tested five classical ML

algorithms such as Naïve Bayes (NB), k-Nearest Neighbors (KNN), Logistic Regression (LR),

Random Forest (RF), and Support Vector Machine (SVM) and their classification accuracies are

shown in Fig. 3a and summarized in Table S2. Here, we define Top 1 to Top 5 as the ranking

positions of identification results of the testing spectra among the library consisting of total 823

MOFs (Table S3). For example, Top 1 means that a ML algorithm can successfully rank a MOF

sample at the first position. Different from previous study which maps the spectra into 7 crystal

systems or 230 space groups among thousands of materials,[2, 16] it is much more challenging to

reach the goal of one-by-one material classification. One-by-one classification mission is similar

to the large-scale image-classification challenge attempted by the ImageNet, which classifies high-

resolution images into 1,000 different categories. It can be seen that all five classical ML

algorithms exhibit < 50% identification accuracy for Top 1-to-Top 5 rankings. In comparison, the

best result of CNN performed much better with 56.7%, 76.7%, 90%, and 93.3% accuracy for Top

1, Top 2, Top 3, and Top 4 ranking, respectively (Table S4), and reached 96.7 % accuracy for Top

5 ranking. It demonstrates that high-level hidden and meaningful features learned by the CNN help

identify XRD spectra with a much higher accuracy than the classical ML algorithms.[28] In

addition, the classification accuracy of Top 5 over the classical ML algorithms and CNN model

under different theoretical data is also investigated and shown in Fig. 3b. It is obvious that the

classification accuracy of the classical ML algorithms decreases sharply when the data size

increases, whereas our CNN model is robust enough to deliver predictive accuracy of 96.7% even

8
the theoretical spectra in the library increases from 189 to 1012 (Fig. 3b). It can be expected that

as the library size further increases, the prediction accuracy would be well maintained.

Figure 3. (a) Comparison of various ML models for identifying XRD spectra among the library

with 1012 MOFs. NB: Navie Bayes, kNN: k-Nearest Neighbors, LOG: Logistic Regression, RF:

Random Forest, SVM: Support Vector Machine, CNN: Convolutional Neural Network. (b) The

Top 5 accuracy of various ML models trained with different number of theoretical spectra.

It is well-known that the size of the datasets may affect the performance of a ML algorithm.[29]

Thus, we investigated the effect of the library size of the MOFs on the prediction accuracy. First,

the ratio of augmented training data to corresponding theoretical data also affects the prediction

accuracy. As it increases, one can identify that the accuracy quickly reaches the highest number

(96.7%), while it gradually drops to 79.4 % when the ratio of augmentation is further increased to

300 (Fig. S1a). Other than the dataset size, the number of training epochs also affects the

performance (Fig. S1b). With the increase of the epoch number, the identification accuracy first

increases to 96.7% from 75%, then decreases to 75% again if it further increases. At the epoch

number of 40, the accuracy reaches the highest value of 96.7%. Thus, the early stopping was set

to 40 epochs. The identification accuracy of the convolutional layer number also exhibits the

similar trend (Fig. S1c), i.e., first increases to 96.7% with the increase of the layer number to 7,

9
then decreases to 77.8% as the number of the layers increases to 9. Hence, the layer number is

fixed to be 7. It is because a larger number of layers introduce more parameters, which may lead

to overfitting, while a small number of layers usually have inadequate parameters, which may

result in underfitting.

It is usually difficult for the deep neural networks to afford insights toward interpreting

mechanism since they introduce the complexity of interactions and nonlinearities.[30] Thus, it is

also known as “black boxes” for a long time. To reduce the dimensionality of the data for better

understanding how the CNN makes the decision, neighborhood components analysis (NCA) was

employed to analyze the experimental XRD spectra.[31] As shown in the NCA map (Fig. 4), these

XRD spectra from the same MOFs are clustered into ten separate groups, while the XRD spectra

from different MOFs but with very similar spectra characteristics may result in overlapped or very

close groups in the component map. It indicates that characteristics of main peaks such as position,

intensity, and full width at half maximum—usually the main criteria to distinguish the XRD

spectra—are reduced to show distinguished features as shown in the LLE component map.

Figure 4. Neighborhood components analysis (NCA) map for clustering of XRD spectra of all

MOFs samples.

10
In order to understand the mechanism of how the CNN model distinguishes individual

experimental spectra, their CAMs in the fourteenth layers and corresponding XRD spectra were

compared and shown in Fig. 5 and Fig. S3. In an image classification task, CAMs are usually used

to reflect the main discriminative features of images, which helps interpret and improve

classification accuracy.[2, 32] In our case, CAMs may afford a clear and direct impression on the

characteristics of the XRD spectra that were the most relevant to the specific class. It is found that

the red regions in the CAMs correspond to the main peaks of the XRD spectra. Hence, it is deduced

that the CNN model can well distinguish XRD spectra according to the main peaks, i.e., the most

important peaks. Further observation shows that MOF-74 (Fig. 5a) only exhibits six colorful

regions including two dominated red regions, and ZIF-8 (Fig. 5b) exhibited much more colorful

regions. This is similar to the way that a professional material scientist analyzes the spectra data.

A traditional way to identify the compounds of interest and to map phase diagrams of

combinatorial materials is to match descriptors of their XRD patterns such as peak positions,

intensities, FWHM with a known database. Thus, it is straightforward to train machine learning

models via data augmentation by peak scaling, peak elimination, and peak shift.[2, 16] The models

successfully identify several crystal systems into which target materials are grouped. However, it

is much more challenging to reach the goal of one-by-one classification, i.e., assign a correct label

to individual XRD spectrum instead of seven crystal systems or 230 space groups among

thousands of datasets. In our case, the trained CNN model can successfully and robustly perform

one-by-one classification even though the experimental XRD spectra exhibit peak shift, scaling in

peak intensities, or FWHM broadening compared to the theoretic spectra. As evident in Fig. 5 and

Fig. S3, all XRD spectra of the MOFs exhibit peak shift, the existence of noise and peak intensity

11
scaling compared to their theoretical spectra. Our model can tackle these abnormal phenomena

and reach 96.7 % one-by-one classification accuracy. In addition, amazingly, the CNN can still

identify the spectra even though when intensities of main peaks are largely changed. Take MOF-

199 (Fig. 5c) and MOF-5 (Fig. 5d) as examples. The ratios of (220) to (222) peaks for experimental

spectra of three MOF-199 samples are varied from 0.82 to 0.45. The trained model can all identify

them in Top 3 (Table S4). The CAM on MOF-199 agrees well with this result, which shows that

the dominated red region shifts to (222) from (220). Similarly, the ratios of (200) to (220) for

MOF-5 are 0.29, 2.40, and 0.99 for M1, M2 and M3 of MOF-5, respectively. The model also can

rank MOF-5 in Top 3 (Table S4). For MOF-5 M2 sample, the peak corresponding to the (400)

plane increases to be the second highest peak. This result agrees well with the CAM analysis result

and it shows that the two dominated regions correspond to the (200) and (400) peaks. In addition,

we also observed that the lower crystallinity and existence of noise greatly affect the classification

accuracy. Here, ZIF-9 was chosen as an example (Fig. S2f). The CNN model can classify the ZIF-

9 M2 and ZIF-9 M3 in Top 2 and Top 4 rankings, but cannot distinguish ZIF-9 M1. The CAM

analysis shows only one red dominated region, corresponding to the first highest peak. Due to the

lower intensity and the existence of noise, the CNN could not learn the main features which can

effectively assign all three samples to ZIF-9.

12
Figure 5. CAMs of the 14th layer output from the CNN model and corresponding XRD spectra: (a)

MOF-74; (b) ZIF-8; (c) MOF-199; (d) MOF-5.

Conclusion

In summary, we demonstrated a CNN model trained by the theoretical XRD spectra

augmented by noise from limited experimental spectra for rapid one-to-one identification of

13
individual MOFs. CNN model is employed to identify XRD spectrum instead of categorizing them

into groups or crystallinity systems. The optimized CNN model showed the highest identification

accuracy of 96.7% for the Top 5 rankings among a dataset of 1012 XRD spectra. The advantages

of the proposed CNN model can be summarized as follows: 1) it is a one-by-one identification

instead of predicting several crystal groups; 2) the model was trained based on very limited

theoretical data; 3) simple and straightforward noise-based data augmentation—not like the past

technique that employed multi-step operations (peak scaling, elimination and shifting)—was

deployed, thus it is easy to operate and requires less hyperparameter tuning; 4) the procedure of

noise filtering can greatly increase the classification accuracy of CNN model. Finally, the proposed

CNN model has the potential in not only numerous applications of XRD in materials science, but

also the possible expansion of the solution to several other characterization techniques such as

Raman, NMR, and FTIR.

Methods

Chemicals. Zn(NO3)2·6H2O (Sigma Aldrich), ZnCl2·6H2O (Fluka), Zn(CH3COO)2·2H2O (Fisher),

Co(NO3)2·6H2O (Sigma Aldrich), CoCl2·6H2O (Sigma Aldrich), Co(CH3COO)2·4H2O (Fisher), 2-

methylimidazole (Fisher), benzimidazole (Fisher), 4,5-dicholorimidazole (Fisher), 2-

imidazolecarboxaldehyde (Fisher), terephthalic acid (BDC, Fisher), 1,3,5-benzentricarboxylic

acid (BTC, Fisher), cetyltrimethylammonium bromide (CTAB, Sigma Aldrich), diethylamine

(DEA, Fisher), triethylamine (TEA, Fisher), polyvinylpyrrolidone (PVP, Mw = 360000 or 40000,

Sigma Aldrich), N, N-dimethylformamide (DMF, Fisher), methanol (Fisher), and ethanol (Fisher)

were used without any further purification.

14
MOFs synthesis. Here, all MOFs were synthesized by three different methods according to

reported literatures. The detailed synthesis methods were described in Supporting Information.

Characterization. Powder X-ray diffraction (XRD) were obtained on a Bruker D8 Discover

diffractometer (Cu Kα, λ=0.15406 nm).

Machine learning models.

Five classical machine learning algorithms, i.e., Naïve Bayes (NB), k-Nearest Neighbors (KNN),

Logistic Regression (LOG), Random Forest (RF), and Support Vector Machine (SVM), were well-

trained and employed for identifying XRD spectra. Feed-forward convolutional neural networks

were constructed using Keras software library with the TensorFlow backend.[33] The machine

learning models were performed using Python with Scikit-Learn on a high-performance computer

with Intel i7-9700k CPU and Nvidia EVGA GeForce RTX 2070 GPU.

Data availability

All original XRD spectrum including theoretical and experimental data and Python scripts for

preprocessing, augmentation and classification are available at https://github.com/javenlin/MOFs.

Acknowledgement

This work was supported by grants from US Department of Energy (Award number: DE-

FE0031645) with Program Manager Karol K. Schrems, and National Science Foundation (Award

number: 1825352) with Program Manager Khershed P. Cooper.

Author contributions

15
H. W proposed data augmentation method. He designed and trained CNN as well as implementing

MLLE and activation mapping analysis. Y. X. synthesized and characterized MOFs. He also

downloaded and processed the theoretic data. D. L. initially explored the idea by testing different

types of machine learning models. Y. Z offered valuable suggestions in developing CNNs. M. X.

provided valuable discussions and inspired solutions to the problems. J. L. conceived the idea,

organized the research scopes, and oversaw all phases of the project. Y. X. wrote the manuscript

which was edited by J. L. All authors commented on the manuscript.

Competing interests

The authors declare no competing interests.

Additional information

See supplemental information.

References

1. Juan-Pablo Correa-Baena, Kedar Hippalgaonkar, Jeroen van Duren, et al. Accelerating

Materials Development via Automation, Machine Learning, and High-Performance
Computing. Joule 2018, 2, 1410.
2. Felipe Oviedo, Zekun Ren, Shijing Sun, et al. Fast and interpretable classification of small
X-ray diffraction datasets using data augmentation and deep neural networks. npj Comput.
Mater. 2019, 5, 60.
3. Pascal Marc Vecsei, Kenny Choo, Johan Chang, et al. Neural network based classification
of crystal symmetries from x-ray diffraction patterns. Phys. Rev. B 2019, 99, 245120.
4. Connor W. Coley, Dale A. Thomas, Justin A. M. Lummiss, et al. A robotic platform for
flow synthesis of organic compounds informed by AI planning. Science 2019, 365,
eaax1566.
5. Sebastian Steiner, Jakob Wolf, Stefan Glatzel, et al. Organic synthesis in a modular robotic
system driven by a chemical programming language. Science 2019, 363, eaav2211.
6. Jarosław M. Granda, Liva Donina, Vincenza Dragone, et al. Controlling an organic
synthesis robot with machine learning to search for new reactivity. Nature 2018, 559, 377.

16
7. Olexandr Isayev, Corey Oses, Cormac Toher, et al. Universal fragment descriptors for
predicting properties of inorganic crystals. Nat. Commun. 2017, 8, 15679.
8. Juan J. de Pablo, Nicholas E. Jackson, Michael A. Webb, et al. New frontiers for the
materials genome initiative. npj Comput. Mater. 2019, 5, 41.
9. Keith T. Butler, Daniel W. Davies, Hugh Cartwright, et al. Machine learning for molecular
and materials science. Nature 2018, 559, 547.
10. Anton O. OliynykArthur Mar. Discovery of Intermetallic Compounds from Traditional to
Machine-Learning Approaches. Acc. Chem. Res. 2018, 51, 59.
11. Fei Li, Jinsong Han, Tian Cao, et al. Design of self-assembly dipeptide hydrogels and
machine learning via their chemical features. PNAS 2019, 116, 11259.
12. Logan Ward, Ankit Agrawal, Alok Choudhary, et al. A general-purpose machine learning
framework for predicting properties of inorganic materials. npj Comput. Mater. 2016, 2,
16028.
13. Zachary Wu, S. B. Jennifer Kan, Russell D. Lewis, et al. Machine learning-assisted
directed protein evolution with combinatorial libraries. PNAS 2019, 116, 8852.
14. Angelo Ziletti, Devinder Kumar, Matthias Scheffler, et al. Insightful classification of
crystal structures using deep learning. Nat. Commun. 2018, 9, 2775.
15. J. A. Aguiar, M. L. Gong, R. R. Unocic, et al. Decoding crystallography from high-
resolution electron imaging and diffraction datasets with deep learning. Sci. Adv. 2019, 5,
eaaw1949.
16. Woon Bae Park, Jiyong Chung, Jaeyoung Jung, et al. Classification of crystal structure
using a convolutional neural network. IUCrJ 2017, 4, 486.
17. Yuta Suzuki, Hideitsu Hino, Masato Kotsugi, et al. Automated estimation of materials
parameter from X-ray absorption and electron energy-loss spectra with similarity measures.
npj Comput. Mater. 2019, 5, 39.
18. Peyman Z. Moghadam, Aurelia Li, Seth B. Wiggin, et al. Development of a Cambridge
Structural Database Subset: A Collection of Metal–Organic Frameworks for Past, Present,
and Future. Chem. Mater. 2017, 29, 2618.
19. Karen SimonyanAndrew Zisserman Very Deep Convolutional Networks for Large-Scale
Image Recognition. arXiv 2014.
20. Y. Lecun, L. Bottou, Y. Bengio, et al. Gradient-based learning applied to document
recognition. Proc. IEEE 1998, 86, 2278.
21. Yotaro Kubo, George Tucker, and Simon Wiesler Compacting Neural Network Classifiers
via Dropout Training. arXiv 2016.
22. Yarin GalZoubin Ghahramani A Theoretically Grounded Application of Dropout in
Recurrent Neural Networks. arXiv 2015.
23. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, et al. Dropout: a simple way to
prevent neural networks from overfitting. J. Mach. Learn. Res. 2014, 15, 1929.
24. Diederik P. KingmaJimmy Ba Adam: A Method for Stochastic Optimization. arXiv 2014.
25. Ilya Sutskever, James Martens, George Dahl, et al. On the importance of initialization and
momentum in deep learning. International conference on machine learning 2013.
26. Raymond D Kent, Charles Read, and Ray D Kent, The acoustic analysis of speech. Vol.
58. 1992: Singular Publishing Group San Diego.
27. Valentin Stanev, Velimir V. Vesselinov, A. Gilad Kusne, et al. Unsupervised phase
mapping of X-ray diffraction data by nonnegative matrix factorization integrated with
custom clustering. npj Comput. Mater. 2018, 4, 43.
17
28. Jindong Wang, Yiqiang Chen, Shuji Hao, et al. Deep learning for sensor-based activity
recognition: A survey. Pattern Recog. Lett. 2019, 119, 3.
29. Joseph Lemley, Shabab Bazrafkan, and Peter Corcoran Smart Augmentation - Learning an
Optimal Data Augmentation Strategy. arXiv 2017.
30. Jason Yosinski, Jeff Clune, Anh Nguyen, et al. Understanding Neural Networks Through
Deep Visualization. arXiv 2015.
31. Jacob Goldberger, Geoffrey E Hinton, Sam T Roweis, et al. Neighbourhood components
analysis. Advances in neural information processing systems 2005.
32. B. Zhou, A. Khosla, A. Lapedriza, et al. Learning Deep Features for Discriminative
Localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
2016.
33. Martín Abadi, Ashish Agarwal, Paul Barham, et al. TensorFlow: Large-Scale Machine
Learning on Heterogeneous Distributed Systems. arXiv 2016.

18