Comparative of machine learning classification strategies for electron energy loss spectroscopy: Support vector machines and artificial neural networks

Abstract

Machine Learning (ML) strategies applied to Scanning and conventional Transmission Electron Microscopy have become a valuable tool for analyzing the large volumes of data generated by various S/TEM techniques. In this work, we focus on Electron Energy Loss Spectroscopy (EELS) and study two ML techniques for classifying spectra in detail: Support Vector Machines (SVM) and Artificial Neural Networks (ANN). Firstly, we systematically analyze the optimal configurations and architectures for ANN classifiers using random search and the treestructured Parzen estimator methods. Secondly, a new kernel strategy is introduced for the soft-margin SVMs, the cosine kernel, which offers a significant advantage over the previously studied kernels and other ML classification strategies. This kernel allows us to bypass the normalization of EEL spectra, achieving accurate classification. This result is highly relevant for the EELS community since we also assess the impact of common normalization techniques on our spectra using Uniform Manifold Approximation and Projection (UMAP), revealing a strong bias introduced in the spectra once normalized. In order to evaluate and study both classification strategies, we focus on determining the oxidation state of transition metals through their EEL spectra, examining which feature is more suitable for oxidation state classification: the oxygen K peak or the transition metal white lines. Subsequently, we compare the resistance to energy loss shifts for both classifiers and present a strategy to improve their resistance. The results of this study suggest the use of soft-margin SVMs for simpler EELS classification tasks with a limited number of spectra, as they provide performance comparable to ANNs while requiring lower computational resources and reduced training times. Conversely, ANNs are better suited for handling complex classification problems with extensive training data.