Towards calibration-invariant spectroscopy using deep learning.

The interaction between matter and electromagnetic radiation provides a rich understanding of what the matter is composed of and how it can be quantified using spectrometers. In many cases, however, the calibration of the spectrometer changes as a function of time (such as in electron spectrometers), or the absolute calibration may be different between different instruments. Calibration differences cause difficulties in comparing the absolute position of measured emission or absorption peaks between different instruments and even different measurements taken at different times on the same instrument. Present methods of avoiding this issue involve manual feature extraction of the original signal or qualitative analysis. Here we propose automated feature extraction using deep convolutional neural networks to determine the class of compound given only the shape of the spectrum. We classify three unique electronic environments of manganese (being relevant to many battery materials applications) in electron energy loss spectroscopy using 2001 spectra we collected in addition to testing on spectra from different instruments. We test a variety of commonly used neural network architectures found in the literature and propose a new fully convolutional architecture with improved translation-invariance which is immune to calibration differences.