A convolutional neural network (CNN) classifier for determining the oxidation state of Manganese in electron energy-loss spectroscopy. This 1D CNN takes signals describing the bonding enviroment in transition metals and it accurately classifies the valence of the material exceeding 99% test accuracy. The CNN has proven to be robust to chemical shifts (the real-world signal is capable of sliding temporally) and is deemed temporally invariant across the energy spectrum. It is robust to noise after using PCA as a form of data augmentation. PCA is effective at modeling the real noise found in EELS signals.

Just a note. This is not a maintained repository. It is a collection of scripts.

• HyperSpy 1.3
• Keras 1.2.1
• Tensorflow 0.12
• numpy 1.13.1
• scikit-learn 1.18
• pandas 0.18.1
• matplotlib 1.5.3

Training a CNN with name "cnn_model1" with 1x translation training augmentation. python train_crossval.py cnn 1 cnn_model1

The default settings are for 500 epochs, batch size of 512, 10-fold stratified cross-validation and no PCA data augmentation. These params are all adjustable in train_crossval.py.

This script will make a new directory titled cnn_model1 and save all 10 weights in 10 separate directories. The argument cnn can be replaced with mlp or custom to use other graph architectures. These architectures are found in the build_graph_architecture helper method in the training script.

In addition to using train_crossval.py there is also a Jupyter Notebook file which walks through what happens during training for a single fold (no cross validation).

python test_digitized_spectra.py

This will run the models specified in the script ( neural_network_name ) and use the the file of spectra found in data_path everything will be evaluated using 10-fold cross validation.

Test a CNN named highest_val_acc_weights_cnn.h5 :

python test_model_accuracy.py cnn

Compare a CNN and MLP named highest_val_acc_weights_cnn.h5 and highest_val_acc_weights_cnn.h5

python test_model_accuracy.py both

For rapid prototyping and model development, it is advantageous sometimes to test the most recently trained set of weights in the weights folder

python test_model_accuracy.py last

The above line will test the most recent set of weights. To test the second last, or third last etc. set of weights, additional numbers can be specified after last. python test_model_accuracy.py last-1, python test_model_accuracy.py last-2, python test_model_accuracy.py last-3

The usage of using an unknown spectra is as follows:

python test_unknown_spectra.py

The path to the weights can be specified in the script. The path to the EELS spectra can be specified in the script. HyperSpy is required to load the EELS file.