A neural network approach for line detection in complex atomic emission spectra measured by high-resolution Fourier transform spectroscopy

TL;DR

This paper introduces a neural network method using LSTM and fully connected layers to detect spectral lines in complex atomic emission spectra, significantly improving accuracy over traditional techniques especially in noisy or blended conditions.

Contribution

The study presents a novel neural network approach for line detection in atomic spectra, enabling automated analysis and identification of spectral lines with higher precision.

Findings

Enhanced detection of spectral lines in noisy and blended spectra

Successful identification of previously unidentifiable Ni II energy levels

Outperforms conventional line detection methods in complex spectra

Abstract

The atomic spectra and structure of the open d- and f-shell elements are extremely complex, where tens of thousands of transitions between fine structure energy levels can be observed as spectral lines across the infrared and UV per species. Energy level quantum properties and transition wavenumbers of these elements underpins almost all spectroscopic plasma diagnostic investigations, with prominent demands from astronomy and fusion research. Despite their importance, these fundamental data are incomplete for many species. A major limitation for the analyses of emission spectra of the open d- and f-shell elements is the amount of time and human resource required to extract transition wavenumbers and intensities from the spectra. Here, the spectral line detection problem is approached by encoding the spectrum point-wise using bidirectional Long Short-Term Memory networks, where transition wavenumber positions are decoded by a fully connected neural network. The model was trained using simulated atomic spectra and evaluated against experimental Fourier transform spectra of Ni ($Z=28$) covering 1800-70,000 cm$^{-1}$ (5555-143 nm) and Nd ($Z=60$) covering 25,369-32,485 cm$^{-1}$ (394-308 nm), measured under a variety of experimental set-ups. Improvements over conventional methods in line detection were evident, particularly for spectral lines that are noisy, blended, and/or distorted by instrumental spectral resolution-limited ringing. In evaluating model performance, a brief energy level analysis of Ni II using lines newly detected by the neural networks has led to the confident identification of two Ni II levels, $3\text{d}^8$$(^3\text{F}_4)6\text{f} [2]_{3/2}$ at 134,261.8946 $\pm$ 0.0081 cm$^{-1}$ and $3\text{d}^8$$(^3\text{F}_4)6\text{f} [1]_{3/2}$ at 134,249.5264 $\pm$ 0.0054 cm$^{-1}$, previously concluded to be unidentifiable using previously analysed Ni spectra.