High-accuracy photoacoustic gas sensor enabled by deep learning-based concentration inversion from full waveforms

In photoacoustic spectroscopy (PAS) gas sensors, existing concentration inversion methods typically rely on either a single scalar or a feature vector. In principle, the complete signal waveform contains richer concentration-related information, and its utilization is expected to further improve inversion accuracy, which remains unverified to date. To achieve this, this study proposes an end-to-end method based on a one-dimensional convolutional neural network (1D-CNN) that directly maps the full second-harmonic (2f) waveform to gas concentration. Experimental validation with chloroform (CHCl₃) showed strong predictive robustness under stochastic waveform perturbations, achieving a coefficient of determination (R²) of 0.9999; among the noise-augmented evaluation samples, 95.95% exhibited absolute errors below 1 ppm, and 61.8% exhibited relative errors below 1%. This method autonomously extracts concentration information from the full waveform and establishes a complex nonlinear mapping to gas concentration. Furthermore, this framework may provide a strategy for concentration inversion in other gas sensors employing second-harmonic wavelength modulation spectroscopy (WMS-2f) detection.