Physics-guided spectral attention for baseline-free concentration regression from laser absorption spectra.

Tunable diode laser absorption spectroscopy (TDLAS) is widely used for quantitative trace gas analysis. However, direct concentration regression from raw transmitted-intensity spectra remains challenging because weak absorption signals are entangled with baselines, optical-intensity fluctuations, and photodetector noise. Conventional absorbance-based inversion methods rely on baseline extraction and may introduce distortions, while end-to-end models often treat spectra as ordinary one-dimensional sequences and insufficiently account for the multiplicative coupling between optical background and gas absorption. To address this problem, this paper proposes a physics-guided dual-perspective spectral attention network for path-integrated concentration regression directly from raw TDLAS transmitted-intensity spectra without explicit baseline fitting or absorbance conversion. A physics-guided feature enhancement module introduces a second-derivative correlation prior into attention scores, while representative spectral cross-attention retrieves similar patterns using representative spectra selected from the training set. Experiments on concentration-disjoint methane and nitrous oxide datasets show that the proposed model achieves RMSE = 5.1708 ± 1.2274 and R2 = 0.9993 ± 0.0003 on methane, reducing RMSE by 24.2% compared with the strongest Transformer baseline. On nitrous oxide, it achieves RMSE = 3.5669 ± 0.6032 and R2 = 0.9947 ± 0.0045, reducing RMSE by 9.1% compared with the strongest CNN baseline. These results demonstrate that the proposed model provides an effective chemometric modeling strategy for quantitative analysis of laser absorption spectra under complex background interference.