Integrating DFG spectroscopy with machine learning for direct and accurate gas pressure diagnosis from spectral...

We introduce what we believe to be a novel machine learning (ML)-based ResNet algorithm for predicting gas pressure from spectral imagery, eliminating the need for traditional peak fitting. Evaluated using simulated and experimental carbon monoxide (CO) spectra, the model accurately predicts pressures across a wide range (1 mbar - 2 bar), even with noisy data, outperforming conventional methods like PeakFit. The ResNet model demonstrates minimal discrepancies between predicted and actual pressures, achieving a mean absolute error (MAE) of 0.095 and mean squared error (MSE) of 0.009 in simulations, and maximum MAE of 1.2×10^-2 and MSE of 1.46×10^-4 experimentally below 94 mbar. This approach significantly enhances quantitative spectroscopy by focusing on line shape imagery, showing promising applications in atmospheric science, industrial monitoring, and environmental research. This work is a substantial improvement over our previous models.