Reconstructing spectral shapes with GAN models: A data-driven approach for high-resolution spectra from low-resolution spectrometers
Journal
Chemometrics and Intelligent Laboratory Systems
Journal Volume
258
Start Page
105333
ISSN
0169-7439
Date Issued
2025-03-15
Author(s)
Min-Hsu Tai
Abstract
This study presents the development of a generative adversarial network (GAN) to generate high-resolution (HR) spectra from low-resolution (LR) spectra. Plasma emissions with second positive system of nitrogen are used for demonstration. Specair™ is used to generate HR and LR spectra pairs as the training data covering the range of rotational temperatures (Trot) and vibrational temperatures (Tvib) ranging from 300 to 1200 K and 2000 to 6500 K, respectively. Optical emission spectra from low-pressure and atmospheric-pressure plasmas are used as the testing data to show the feasibility of the model for generating HR spectra with spectra acquired using LR spectrometers. Feature matching is used during the training stage to tackle the instability issues. The distributions of the discriminator scores are used as an initial criterion to monitor the training procedure. The results show a weighted coefficient of determination (R‾2) greater than 0.9999 between the simulated and generated HR spectra. The fitting errors for Trot and Tvib between generated HR spectra and experimental HR spectra acquired from an HR spectrometer are mostly below 5 %. The results indicate that this GAN serves as an efficient approach to obtain HR spectra when HR spectrometers are not available.
Subjects
Deep learning
Feature matching
Generative adversarial network (GAN)
Machine learning
Optical emission spectroscopy (OES)
Publisher
Elsevier BV
Type
journal article