بررسي امكان استفاده از روش يادگيري ماشين در استخراج فشار گاز مونواكسيد كربن از داده¬هاي تجربي بيناب¬نمايي در فروسرخ ميانه

This study aims to develop a novel framework for predicting the pressure of the P(21) spectral line of CO gas by leveraging precise analysis of the Voigt profile an‎d machine learning algorithms. The Voigt profile, which combines Gaussian broadening (due to the Doppler effect) an‎d Lorentzian broadening (resulting from molecular collisions), has long posed computational challenges in accurate modeling. In this work, an intelligent system was designed an‎d implemented using the advanced deep neural network architecture ResNet50 to predict key spectroscopic parameters—including gas pressure, noise type, an‎d signal-to-noise ratio—by focusing on the spectral images of CO gas at various pressures. The training data for the model was generated from simulated Voigt profiles. The initial data were extracted from the well-known HITRAN database [1], focusing on the CO P(21) absorption lines near the wavenumber 2055.4 cm⁻¹. The proposed model achieved a mean absolute error (MAE) of 0.095 an‎d a mean squared error (MSE) of 0.009, demonstrating higher accuracy compared to conventional curve-fitting methods. One of the key achievements of this project was the significant reduction in processing time an‎d result generation compared to traditional numerical algorithms, as well as the elimination of human error in such fitting-based computations. Additionally, the results reveal the model’s notable robustness against various noise types applied to the spectrum. The main innovation of this research lies in the substantial improvement in quantitative spectroscopy accuracy an‎d the ability to simultaneously learn hidden physical parameters from raw spectroscopic data using deep learning. This advancement paves the way for the development of next-generation systems in atmospheric pollutant monitoring, pharmaceutical compound analysis, an‎d laser-based medical diagnostics. The findings demonstrate the proposed framework’s high potential in automating spectroscopic analysis an‎d reducing reliance on manual parameter tuning, marking a significant step toward the design of intelligent spectroscopic systems of the future. Keywords: Deep learning, machine learning, ResNet network, absorption spectroscopy, Voigt profile, pressure prediction, spectral line modeling.

4