Wavelength calibration with deep-learning using long short-term memory architectures

Abstract

In this paper, the long-term short-term memory (LSTM) architecture is investigated as a tool for wavelength calibration of a spectrometer. Polynomial fitting is the most common method of wavelength calibration, whereby wavelength standards, such as neon and krypton are recorded and the position of the spectral lines on the detector together with the known wavelengths are used for fitting. The method performs poorly when only a small number of lines appear within the bandwidth recorded by the spectrometer. We demonstrate how the basic encoderdecoder LSTM architecture can be used to provide superior wavelength calibration accuracy when five or less lines are present. We believe with further development, machine learning could outperform the traditional methods in all cases.