On the use of recurrent neural networks for fast and accurate non-uniform gas radiation modeling

Abstract

This paper focuses on the construction of a fast though accurate gas radiation model based on a Recurrent neural network (RNN) formulation. The model is founded on recent works in which a solution to a non-uniform technique proposed by Godson in the 50s was derived explicitly. The method uses a non-linear transformation of a set of physical / geometrical paths, directly related to a non-uniform path which is first discretized into uniform sub-layers, into a sequence of equivalent absorption lengths. This process, studied thoroughly within the frame of the development of the ℓ-distribution approach, can be naturally handled using an algorithm that takes the form of an RNN model. The method is assessed against LBL calculations in non-uniform high temperature gaseous media and found to provide more accurate results than a CKD (Correlated K-Distribution) model with 16 gray gases. This paper is the first to suggest an RNN to treat radiative transfer in non-uniform gaseous media. Moreover, all the weights involved in the RNN have a clear physical meaning so that the structure of the RNN can be readily interpreted, avoiding the black box disadvantage of most brute force machine learning strategies.