CO/NH₃ concentration retrieval under pressure-induced spectral overlap using a CNN-FiLM-Transformer

Retrieval of CO and NH₃ from adjacent absorption features is complicated by pressure broadening under variable-pressure conditions, which increases spectral overlap and alters the measured spectral response. A CNN-FiLM-Transformer (CFT) framework was developed to estimate CO and NH₃ concentrations from measured wavelength-modulation second-harmonic (WMS-2f) spectra. The model combines convolutional neural networks for local spectral-feature extraction, a Transformer encoder for modeling dependencies within overlapped spectra, and a FiLM-inspired pressure-conditioned residual feature-scaling mechanism. The method was evaluated using controlled CO/NH₃/N₂ mixtures at pressures from 1.0 to 3.0 atm. Under condition-wise grouped evaluation, the scan-level RMSE values were 21.43 ± 5.44 ppm for NH₃ and 14.70 ± 5.18 ppm for CO. For reference, the regression-stratified scan-level test yielded RMSE values of 5.83 ppm for NH₃ and 4.91 ppm for CO, respectively, and these values are reported as secondary indicators of within-distribution performance. Among the evaluated chemometric and neural-network methods, the CFT model yielded the lowest within-distribution retrieval errors. These results indicate the feasibility of CO/NH₃ retrieval under the investigated controlled laboratory conditions. Application to flue-gas measurements requires further performance evaluation.