Machine Learning Data Augmentation Strategy for Electron Energy Loss Spectroscopy: Generative Adversarial Networks.

Area of Science:

• Materials Science
• Data Science
• Spectroscopy

Background:

• Supervised machine learning (ML) requires substantial high-quality data for effective algorithm training.
• Electron energy loss spectroscopy (EELS) data is often limited, posing a challenge for ML model development.

Purpose of the Study:

• To develop a novel data augmentation (DA) strategy for EELS data using generative adversarial networks (GANs).
• To enable the training of ML classifiers with limited EELS spectral data.

Main Methods:

• Implementation of a data augmentation generative adversarial network (DAG) approach.
• Exploration of optimal GAN configurations for generating realistic EELS spectra.
• Utilizing generated spectra to train artificial neural networks (ANNs) and support vector machines (SVMs).

Main Results:

• The DAG successfully generated realistic EELS spectra from a small dataset (around 100 spectra).
• Classifiers trained on DAG-generated data achieved success in classifying experimental EEL spectra.

Conclusions:

• The developed DAG strategy effectively addresses the data scarcity issue in EELS.
• Generated EELS spectra are viable for training robust ML classifiers for real-world applications.