Contents 
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import plot_confusion_matrix

Material classification

Neutron scattering can give complementary information about materials, such as the structural information from diffraction and electronic information from spectroscopy. In this exercise, we will look at using a random forest classification of structural and electronic information to create a model that can guess the space group for new materials.

The first thing that we need is some data, I have obtained data for lithium, sodium, and potassium containing materials of the four most common space groups (space group numbers: 2, 14, 15, 19). This data can be found in the file materials_data.csv.

data = pd.read_csv('materials_data.csv')

This data is read in as a pandas.DataFrame so it can be visualised as so.

data
pretty_formula spacegroup.number band_gap density volume formation_energy_per_atom nsites
0 KAl9O14 14 4.1426 3.142646 534.648960 -3.364229 48
1 K2Se2O 15 0.0975 3.303683 253.443517 -0.940536 10
2 K2AgF4 14 0.6151 3.343987 260.263084 -2.371859 14
3 Na2TeS3 14 1.7919 2.904619 616.908734 -1.000993 24
4 K3Zn2Cl7 2 3.9024 2.345705 702.644832 -1.909423 24
... ... ... ... ... ... ... ...
537 Na2Cr2O7 2 2.5253 2.654095 655.601817 -1.869509 44
538 K4SnO4 2 2.3423 2.977324 378.252471 -1.839865 18
539 LiTa3O8 15 3.3841 7.889695 285.303789 -3.263699 24
540 LiSmO2 14 3.7655 6.485182 193.881819 -3.185751 16
541 K4UO5 2 2.1910 4.207670 374.454975 -2.645721 20

542 rows × 7 columns

The first thing to do is separate our data into training and validation data (this process is common in machine learning methods). For this we will use the scikit-learn function train_test_split.

train, validate = train_test_split(data, test_size=0.2)

We have split the data, which consisted of data about 542 materials, so that 80 % will be used to train and the remaining 20 % is reserved for validation.

train
pretty_formula spacegroup.number band_gap density volume formation_energy_per_atom nsites
244 NaInI4 14 2.4484 3.715767 1153.737028 -0.785825 24
205 Li2SnS3 15 1.4195 3.435021 442.395841 -1.213020 24
99 Na3DyCl6 14 5.4269 2.831208 521.042810 -2.353868 20
378 KC2N3 14 4.5026 1.611816 433.272097 -0.278986 24
33 K5SnBi3 14 0.0145 4.095423 1526.391127 -0.349438 36
... ... ... ... ... ... ... ...
533 Na2SnSe3 14 1.3283 3.803238 701.320138 -0.774873 24
303 K4SiO4 14 3.2937 2.492700 662.100652 -2.329418 36
76 KTaTe3 14 0.4181 5.607999 3570.078081 -0.572082 100
213 Li12Mo5O17 2 1.3118 4.182842 662.955856 -2.220343 68
204 Li2PtF6 14 2.5122 5.110123 209.889942 -2.437627 18

433 rows × 7 columns

validate
pretty_formula spacegroup.number band_gap density volume formation_energy_per_atom nsites
366 Na5Ga3F14 14 5.4154 3.416719 573.577513 -3.034718 44
291 KNb2Cl11 2 2.1275 2.391408 853.936652 -1.816086 28
340 NaYO2 15 4.1319 4.251371 224.814378 -3.141605 16
1 K2Se2O 15 0.0975 3.303683 253.443517 -0.940536 10
132 Li2PbO3 15 1.0712 6.684437 133.689100 -1.796154 12
... ... ... ... ... ... ... ...
208 KMoO4 2 0.6734 3.181245 415.569254 -1.824456 24
362 KPO3 14 5.0464 2.339255 1005.757685 -2.695846 60
207 Na2TeSe3 14 1.2319 3.848153 708.479141 -0.628667 24
372 NaC2N3 14 3.5806 1.874699 946.327099 -0.350918 72
112 KAsO3 2 3.1488 2.902151 556.216456 -1.947867 30

109 rows × 7 columns

We then define the columns of interest and segment our data into the X and y elements.

columns = ['band_gap', 'formation_energy_per_atom']
X_train = train[columns]
y_train = train['spacegroup.number']
X_validate = validate[columns]
y_validate = validate['spacegroup.number']

With the data split up, we can define our random forest model.

model = RandomForestClassifier(n_estimators=100).fit(X_train, y_train)

Let’s see how well our model managed to classify the validation data?

plot_confusion_matrix(model, X_validate, y_validate, normalize='true')
plt.show()
_images/classification_exercise_15_0.png

The confusion matrix allows us to compare the actual space group of the validation data, with that from the classification. For a perfect classification, you would have an identity matrix with the dimension of the number of labels.

Exercise

It is clear from the above confusion matrix, that our current selection of columns of interest is not sufficient to accurately classify the data. In this exercise, you should try different combinations of the columns of interest to see which give the best classification of the space group for these materials.

Random forests Clustering