Approximation of Inverse Models for Temperature-Concentration Dependences of the Transmission Function of a Single-Component Homogeneous Gas Medium by Artificial Neural Networks

摘要

The problem of application of artificial neural networks for approximation of inverse models of temperature-concentration dependences of the transmission function of a single-component homogeneous gas medium is considered on the example of carbon monoxide. The gas transmission function is calculated using the line-byline method for five spectral centers at partial pressures 0.1-1 atm and temperatures 300-2500 K. The inverse models are approximated using a multilayered perceptron with three hidden layers. The artificial neural network is learned using the Levenberg-Marquardt algorithm with Bayesian regularization. The errors of the obtained inverse models are analyzed depending on the number of the employed spectral centers and the leaning sample size. A tendency toward a decrease in error values with increase of these parameters is demonstrated. Maximal steps of the uniform concentration-temperature grid required for correct approximation of the inverse models by the artificial neural networks are determined. The inverse model of the temperature-concentration dependence of the carbon monoxide transmission function, providing a solution of the inverse optical problem on the determination of its partial pressure and temperature, is obtained with relative errors less than 3% in the examined ranges of their variations.