Fast method for predicting spectral radiative properties of alumina particles based on the neural network

Efficient evaluation of the solid rocket motor (SRM) plume radiation signature relies on the accurate and fast prediction of spectral radiative properties of gases and particles in the plume, while traditional spectral radiative properties prediction methods suffer from high computation cost. Although many methods based on machine learning have been developed to predict the spectral radiative properties of common gas species, similar method for alumina particles in the plume is lacking. Thus, the stochastic noise-added ensemble averaging technique and neural network are combined to establish a novel machine learning-based method for fast and accurate prediction of spectral radiative properties of alumina particles when evaluating the spectral radiative intensity of SRM plume. In terms of prediction accuracy, the spectral radiative properties of alumina particles predicted by the machine learning-based method agree well with the results from the Mie theory method, and the average relative errors for predicting effective gray scattering efficiency, absorption efficiency and asymmetry factor are 1.51%, 1.43% and 0.90%, respectively. For the prediction efficiency, the machine learning-based method is 2–4 orders of magnitude faster than Mie theory calculation. Additionally, its model memory size is only 3 MB, which is significantly smaller compared with the look-up table method based on the database of 0.3 GB in memory size, offering greater convenience in practical applications. • The stochastic noise-added ensemble averaging technique and neural network are combined to establish a novel machine learning-based method for fast and accurate prediction of spectral radiative properties of alumina particles. • In terms of prediction accuracy, the average relative errors of proposed method for predicting effective gray scattering efficiency, absorption efficiency and asymmetry factor are 1.51%, 1.43% and 0.90%, respectively. • For the prediction efficiency, the proposed method is 2–4 orders of magnitude faster than Mie theory method. • The memory size of machine learning model is 3 MB, which is significantly smaller and more convenient compared with the look-up table method based on the database of 0.3 GB in memory size.