A neural network approach for line detection in complex atomic emission spectra measured by high-resolution Fourier transform spectroscopy
Milan Ding, Sean Z. J. Lim, Xiaoran Yu, Christian P. Clear, Juliet C. Pickering · 2025
Read A neural network approach for line detection in complex atomic emission spectra measured by high-resolution Fourier transform spectroscopy by Milan Ding, Sean Z. J. Lim, Xiaoran Yu, Christian P. Clear, Juliet C. Pickering (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-12276.txt.
Introduction
Milan Ding and co-authors teach a neural net to find lines in dense atomic light spectra in a 2025 paper. Some atoms give off tens of thousands of lines, and hand work to list them eats years. The net reads the spectrum point by point with a two-way memory design, then calls line spots through a plain output layer.
Trained on fake spectra, the net was tried on true nickel and neodymium reads across wide ranges. It beat classic ways most on noisy, merged, or ringing lines. Lines it newly found let the team pin two nickel energy levels that had long been marked as out of reach.
Catalog record
| CREATOR | Milan Ding, Sean Z. J. Lim, Xiaoran Yu, Christian P. Clear, Juliet C. Pickering |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-12276
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A neural network approach for line detection in complex atomic emission spectra measured by high-resolution Fourier transform spectroscopy Milan Ding, Sean Z. J. Lim, Xiaoran Yu, Christian P. Clear, Juliet C. Pickering 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 underpin 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 Z=28 ) covering 1800–70 000 cm -1 (5555–143 nm) and Nd ( Z = 60 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, 3d 8 ( 3 F 4 ) 6 f [ 2 ] 3 / 2 (^{3}\text{F}_{4})6\text{f}\,\,[2]_{3/2} at 134 261.8946 ± \pm 0.0081 cm -1 and 3d 8 ( 3 F 4 ) 6 f [ 1 ] 3 / 2 (^{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 the Ni spectra. January 2025 Keywords : heavy elements, atomic structure, Fourier transform spectroscopy, high resolution, spectrum simulation, peak-finding, neural networks 1 Introduction Atoms and ions with the most complex spectra contain electrons in the open d- and f-shells [ 1 ] , where interactions between the multiple valence electrons lead to many thousands or tens of thousands of observable electronic transitions. The experimental determination of fundamental atomic data such as fine structure level energies and angular momenta, and energy level transition wavenumbers is crucial in almost any high to moderate resolution spectroscopic plasma investigations of these species. These data for the open d- and f-shell elements are of interest to a wide variety of fields, such as in the lighting and metal manufacturing industries, in fusion research [ 2 ] , stellar astronomy [ 3 ] , Galactic spectroscopic surveys [ 4 ] , and stellar nucleosynthesis investigations [ 5 ] . Additionally, empirical knowledge of atomic energy levels constrains and improves theoretical atomic structure models (e.g. the Cowan codes [ 6 ] ), offering improvements over ab initio estimates. High-resolution Fourier transform (FT) spectroscopy is one of the primary experimental methods for the empirical determination of atomic structure of open d- and f- shell species. Recent examples include the open 3d-shell Ni II [ 7 , 8 ] , the open 4d-shell Zr I-II [ 9 ] , the open 5d-shell Hf I-II [ 10 ] , and the open 4f-shell Nd III [ 11 , 12 ] . The resolving power of FT spectrometers can be up to a few million, corresponding to wavenumber accuracies that are many orders of magnitude above those achievable with theoretical calculations. Spectral line Doppler and pressure broadening widths of the plasma sources are commonly the factors limiting spectral resolution. With the wide spectral range of FT spectrometers, and with the complexity of atomic structure of the open d- and f-shell species, tens of thousands of transitions between fine structure energy levels of each species are measurable across the mid-infrared (IR) to vacuum-ultraviolet (UV) range. The immense quantities of transition wavenumbers and intensities of the open d- or f-shell elements are extracted from observed FT spectra into line lists through means of least-squares fitting of model spectral line profiles. Existing methods involve initialising fitting parameters using peak-finding algorithms, optimising the parameters, and then manually including undetected lines or correcting poor fits. This analysis is widely carried out using the FT atomic spectra analysis program Xgremlin [ 13 ] , with a peak-finding algorithm that averages the peak positions and second derivative peak positions determined using a set of specialised kernels. The greatest challenges arise from unresolved line blending and weak lines that are not discernible from the noise level, causing incorrect line wavenumbers and incomplete extraction of the list of transitions, which hinders the spectrum analyses. Moreover, a relatively high signal-to-noise ratio (S/N) threshold parameter, smin , of around 5 is typically set in Xgremlin to avoid noise from being detected as lines [ 13 ] . The manual adjustments to the line list typically take months for a set of spectra of just one element across the IR-UV. There is a demand for more efficient methodologies. In addition to dramatically improving the efficiency of the FT atomic spectra line list making process, we also aimed to develop a method to increase the accuracy and completeness of the wavenumbers of weak and blended lines in the line lists. This is important because, the reliable observation of one transition of an energy level could be the key to determine that level and then finding other levels connected through other transitions. As far as we know, there are no records on the application of neural networks (NNs) for line detection in high-resolution FT complex atomic spectra in the literature. In this paper, we present our simplest and highest performing NN model inspired by similar investigations for molecular nuclear magnetic resonance spectra [ 14 , 15 ] . Our NN model encodes each spectrum point using bidirectional Long Short-Term Memory (LSTM [ 16 ] ) network layers, and classifies each spectrum point for spectral line detection using a fully connected neural network (FCNN) as the decoder. The LSTM-FCNN model was trained against simulated FT spectra using focal loss [ 17 ] and evaluated on FT spectra obtained for Ni ( Z=28 ) covering 1800–70 000 cm -1 (5555–143 nm) [ 7 , 8 ] and Nd ( Z=60 ) covering 25 369–32 485 cm -1 (394–308 nm) [ 11 , 12 ] from a variety of FT spectrometers and atomic emission spectrum light sources. Results show that the LSTM-FCNN model outperforms the peak-finding algorithms of Xgremlin [ 13 ] in detecting weak lines, blends, and lines affected by instrumental spectral resolution-limited ringing. Around 95% of the spectral lines in previously human-made line lists were detected by the LSTM-FCNN model. A significant fraction of the NN-detected lines that had not been included previously in the human-made line lists could be confirmed by identification in energy level analyses. As an example, we present the identification of two energy levels of Ni II that were previously concluded to be unidentifiable from