Development of a combined model for analyzing gas mixtures using machine learning methods

Abstract

Analysis of gas mixtures is an important task in spectroscopy, environmental monitoring, industrial control and scientific research. Accurate determination of component concentrations in complex gas environments requires advanced approaches that combine physical modeling and artificial intelligence methods. The use of neural networks in spectral analysis allows increasing the accuracy and stability of calculations under variable experimental conditions, which indicates the relevance of the work. The aim of the research is to develop a combined model for spectral light flux analysis that combines physical modeling of spectral absorption of gases with machine learning methods. This provides increased accuracy in determining the concentration of components in multicomponent gas mixtures and allows adaptive adjustment of analysis parameters depending on the measurement conditions. An integrated methodology is proposed, which includes modeling of spectral light flux based on Gaussian and Lorentzian absorption profiles, the use of the Bouguer-Lambert-Beer equations to determine gas concentrations, and training a neural network to predict the light flux. To assess the performance of the developed model, a series of numerical experiments were conducted with varying network parameters and optimizing the configuration. The results obtained confirmed the high efficiency of the model, which is reflected in the high value of the coefficient of determination and low values of the mean square error. The model was tested when changing gas concentrations and the length of the optical path, which confirmed its stability and adaptability. The study showed that the optimal configuration of the neural network includes three hidden layers with an optimal number of neurons, which provides a balance between accuracy and efficiency. A rectified linear activation function was used for stable convergence, and for weight optimization - an adaptive stochastic gradient descent method, which improves performance. The proposed method of combining physical modeling and machine learning provides high accuracy of gas mixture analysis and resistance to variations in external conditions. The scientific novelty of the study lies in the use of a combined approach, which allows adapting the model to a wide range of spectral characteristics. The practical significance of the work lies in the possibility of applying the developed methodology for industrial control, environmental monitoring, and laboratory research, providing a reliable tool for the analysis of complex gas mixtures.