Calibration-free physics-informed multi-task residual U-Net for simultaneous denoising and gas pressure retrieval from noisy voigt spectra

We present a machine learning framework based on a one-dimensional U-Net (1D U-Net) that simultaneously performs spectral denoising and pressure estimation within a unified architecture. The model is trained on simulated Voigt profiles of the P(21) CO absorption line over pressures ranging from of 1 mbar to 2 bar. To ensure realistic conditions, simulated spectra are superimposed with experimentally captured noise, making them nearly indistinguishable from real experimental scans and forcing the model to recover clean spectra from noisy inputs. Quantitative assessments indicate excellent reconstruction performance, with a Pearson correlation coefficient (PCC) approaching unity, a signal-to-noise ratio (SNR) exceeding 35 dB, and both mean absolute error (MAE) and mean squared error (MSE) remaining close to zero. The model accurately predicts pressures for 200 unseen spectra using only spectral features, bypassing traditional linewidth analysis. At 1.25 bar, the U-Net yields virtually zero error (AE ≈ 0, SE ≈ 0), demonstrating sub-percent deviation and excellent consistency with the simulated reference. Experimental validation on a difference-frequency generation (DFG) spectrometer confirms robust performance, with ≈ 70% of traces reaching a peak SNR (PSNR) above 34 dB. Pressure estimation from ramp-based scans further demonstrates high accuracy, achieving minimal errors at 539.1 mbar (AEP = 0.002, SEP = 4.0 × 10⁻⁶). These findings establish the 1D U-Net as an efficient and reliable alternative to conventional noise-reduction and pressure-estimation techniques, simplifying mid-infrared spectroscopy workflows while ensuring high fidelity and stability.