GEC Early Career Award: Empowering Optical Plasma Diagnostics with Advanced Spectral Analysis via Multi-Objective Deep Learning

Optical diagnostics have been pivotal in characterizing plasma in both basic science and industrial applications. They make it possible, for instance, to determine electron density and temperature from continuum emission or line-ratio analysis, and to retrieve electron- or ion-energy distribution functions with laser-based techniques. Yet converting raw spectra into quantitative plasma parameters still depends heavily on manual curve-fitting and expert intuition based on domain knowledge, rendering spectroscopy time-consuming and prone to user bias. Here, we present the application of a deep neural network to extract plasma information directly from spectrum images, which is called deep spectral deconvolution (DSD). Specifically, DSD is trained with a multi-objective loss that combines (i) a regression term, which forces agreement between predicted and true species concentrations, and (ii) a reconstruction term, which ensures the recovered optical-depth spectrum faithfully reproduces the measured data. By learning jointly from numerical values and their graphical context, the network overcomes the ill-posedness that hampers purely numerical deconvolution. Details on the trained model will be presented. As a case study, this method is demonstrated on broadband UV-visible absorption spectroscopy, of which results are largely overlapped spectra by multiple absorbing molecules in plasma. Considering chemicals of interest (O₃, nitrogen oxides, and relevant species), training datasets of synthetic spectra is prepared to train and optimize the proposed AI model. Synthetic spectra spanning realistic concentration ranges serve as the training set, and independent synthetic and experimental spectra are used for validation. Across repeated trials, DSD consistently outperforms conventional linear-regression methods in accuracy, fidelity, and robustness. We believe that this pioneering application of AI-powered optical data analysis will offer promising alternatives to empower conventional optical plasma diagnostics.