Abstract:While neural networks have been successfully applied to the full-spectrum k-distribution (FSCK) method at a large range of thermodynamics with k-values predicted by a trained multilayer perceptron (MLP) model, the required a-values still need to be calculated on-the-fly, which theoretically degrades the FSCK method and may lead to errors. On the other hand, too complicated structure of the current MLP model inevitably slows down the calculation efficiency. Therefore, to compensate among accuracy, efficiency and storage, the simple MLP designed based on the nature of FSCK method are developed, i.e., the simple FSCK MLP (SFM) model, from which those correlated k-values and corresponding ka-values can be efficiently obtained. Several test cases have been carried out to compare the developed SFM model and other FSCK tools including look-up tables and traditional FSCK MLP (TFM) model. Results show that the SFM model can achieve excellent accuracy that is even better than look-up tables at a tiny computational cost that is far less than that of TFM model. Considering accuracy, efficiency and portability, the SFM model is not only an excellent tool for the prediction of spectral properties, but also provides a method to reduce the errors due to nonlinear effects.

What problem does this paper attempt to address?

This paper mainly explores how to use Multilayer Perceptron (MLP) neural network model to improve the efficiency and accuracy of predicting the radiation spectral characteristics in high thermodynamic states in Full Spectrum Correlated k Distribution (FSCK) method. Although the current FSCK method achieves high accuracy through pre-computed lookup tables, it has a large data storage requirement and is not suitable for engineering applications. On the other hand, the traditional FSCK MLP (TFM) model reduces the storage requirement but may introduce errors when computing a values in real-time. To address these issues, the paper proposes a simplified FSCK MLP (SFM) model that can efficiently retrieve associated k and a values directly from input parameters, reducing computational costs while maintaining high accuracy. The SFM model optimizes the structure of the input and output layers, taking the cumulative k distribution (g) as one of the inputs and reducing the number of output layer nodes to only predict k values and ka values, avoiding the direct prediction of fluctuating a values. By expanding the training dataset to include more gas states with smaller mole fractions, the SFM model can mitigate non-linear effects. Experimental results show that the SFM model significantly reduces computation time and required storage space while maintaining the same or even better prediction accuracy as the lookup table. Compared to the TFM model, the CPU time of the SFM model is reduced to around one-tenth, and the model size is only about 0.43MB, making it more suitable for implementation and application on various Computational Fluid Dynamics (CFD) platforms. In summary, this paper addresses the trade-off between efficiency, accuracy, and storage in the FSCK method and proposes a new SFM model that is applicable to a wide range of thermodynamic states, providing a more efficient, compact, and portable solution for radiation modeling.

Simple Full-Spectrum Correlated k-Distribution Model based on Multilayer Perceptron