Improving δ13C Measurement Accuracy of Off-Axis Integrated Cavity Output Spectroscopy with a Weighted Feedforward Neural Network.

Despite the advantages of laser absorption spectroscopy for high-sensitivity and continuous isotope measurements, isotope analyzers often suffer from concentration dependence and nonlinear errors, particularly over wide concentration ranges and at low signal-to-noise ratios, which severely limit measurement accuracy. Here, we developed a dual-gas carbon isotope analyzer for methane (CH4) and carbon dioxide (CO2) based on off-axis integrated cavity output spectroscopy (OA-ICOS) and proposed a feature-enhanced weighted feedforward neural network (FNN) for correcting concentration-dependent effects. By constructing a 9-dimensional feature space containing high-order interaction terms and combining boundary-constrained data augmentation with a region-of-interest weighting strategy, nonlinear instrument responses were effectively decoupled from true isotope values. After correction, concentration dependence was eliminated across the measurement range, with residual errors within ±0.4‰. Excellent agreement with reference values was achieved across different isotopic compositions (R2 > 0.999). Long-term stability tests indicated that at 2 ppm of CH4 and 500 ppm of CO2, the δ13C measurement precision (1σ) reached 0.19‰ and 0.075‰, respectively, with a 5 min integration time. Furthermore, a continuous field comparison with a commercial analyzer showed high consistency in capturing dynamic atmospheric isotope variations. These results demonstrate that neural network-based correction provides an effective solution for mitigating concentration dependence in high-precision laser isotope spectroscopy.