CNN-based spectral image analysis for advanced plasma optical diagnostics

Spectroscopy is an indispensable analytical tool in fields ranging from biomedicine and materials processing to fusion plasma research. As one of the passive diagnostics, optical emission spectroscopy (OES) is especially well suited to plasma characterization. However, its industrial adoption lags behind laboratory success because spectral interpretation still relies on expert-driven curve-fitting and modeling. In this contribution, we introduce an AI-based platform that automates and standardizes spectroscopic interpretation, delivering accurate and consistent results independent of user expertise. The core of this platform is Deep Spectral Deconvolution (DSD) model that is based on a deep learning-based deconvolution strategy with a multi-objective loss function. This architecture learns both numerical features (e.g., peak positions and widths) and graphical trends directly from 2D dynamic spectral images, eliminating empirical fitting processes and post-data processing including manual noise masking. The model is validated across three representative spectroscopic modalities: (i) broadband optical absorption spectroscopy for ozone and nitrogen oxides in air discharges, (ii) laser absorption spectroscopy for helium metastable atoms, and (iii) optical emission spectroscopy for temperature estimation of CN radical. In all cases, our model reproduces reference values within experimental uncertainty while reducing analysis time to milliseconds. The proposed platform is expected to support automated diagnostics in fields less familiar with spectroscopy, bridging the gap between scientific spectroscopy and high-throughput industrial applications.