Deep spectral deconvolution for image-based broadband spectral data analysis

Spectroscopy has become a routine in modern science. However, its everyday use in industry still lags behind academia. One major obstacle is spectral deconvolution, a critical but persistent bottleneck in data analysis. The conventional method of spectral interpretation demands expertise from trained spectroscopists, as manual curve-fitting is tedious and prone to error. Here, we present a deep learning model that enhances spectroscopy by automatically identifying and quantifying chemical species from the corresponding absorption spectra, without relying on database searches, rule-based heuristics, or peak labelling. The deep spectral deconvolution (DSD) model employs convolutional neural networks capable of extracting chemical information directly from raw spectrum images, automating the routine process of cross-validating numerical results with visual interpretability. Two complementary loss functions are used to simultaneously train the model, balancing numerical accuracy with graphical fidelity. Applied to optical absorption spectroscopy of an air-plasma chemical reactor, the trained DSD model achieves higher accuracy than conventional mathematical methods. By enabling fully autonomous spectral deconvolution and interpretation from raw images, the DSD streamlines practical applications and can be readily extended to optical plasma diagnostics and other fields. • Image-based AI model explains chemical spectra in a robust and consistent manner. • Accurate, reproducible, bias-free quantification from optical broadband spectral data. • Outperforms standard curve-fitting on noisy, overlapping peaks. • General tool for plasma chemical diagnostics and beyond.