APPLICATIONS OF STATISTICAL LEARNING ALGORITHMS IN ELECTRON SPECTROSCOPY

Abstract

Building on the recent advances in computer vision with convolutional neural networks, we have built SpectralNet, a spectroscopy-optimized convolutional neural network architecture capable of classifying spectra despite large temporal (i.e. translational, chemical, calibration) shifts. Present methods of measuring the local chemical environment of atoms at the nano-scale involve manual feature extraction and dimensionality reduction of the original signal such as: using the peak onset, the ratio of peaks, or the full-width half maximum of peaks. Convolutional neural networks like SpectralNet are able to automatically find parts of the spectra (i.e. features) of the spectra which maximally discriminate between the classes without requiring manual feature extraction. The advantage of such a process is to remove bias and qualitative interpretation in spectroscopy analysis which occurs during manual feature extraction. Because of this automated feature extraction process, this method of spectroscopy analysis is also immune to instrument calibration differences since it performs classification based on the shape of the spectra. Convolutional neural networks are an ideal statistical classifier for spectroscopy data (i.e. time-series data) due to its shared weighting scheme in neural network weights which is ideal for identifying local correlations between adjacent dimensions of the time-series data. Over 2000 electron energy loss spectra were collected using a scanning transmission electron microscope of three oxidation states of Mn. SpectralNet was trained to learn the differences between them. We prove generalizability by training SpectralNet on electron energy loss spectroscopy data from one instrument, and test it on a variety of reference spectra found in the literature with perfect accuracy. We also test SpectralNet against a wide variety of high noise samples which a trained human spectroscopist would find incomprehensible. We also compare other neural network architectures used in the literature and determine that SpectralNet, a dense-layer free neural network, is immune to calibration differences whereas other styles of network are not.