Predicting 3D distribution of soot particle from luminosity of turbulent flame based on conditional-generative adversarial networks

Abstract

Optical detection has been fully applied and developed in combustion research, but its results are severely limited by detection methods and detection equipment with no expanded space. In this paper, a new method was developed to research soot formation and distribution in turbulent jet flame with Reynolds number range of 2000–8000, SSIM (structure similarity index) was selected as the metric. The main work of this paper included three parts: MART (Multiplicative Algebraic Reconstruction Technique) was used to implement the three-dimensional reconstruction of the turbulent jet flame by luminosity signal, the iteration of reconstruction algorithm would be stopped when the accuracy is greater than 0.9. C-GAN (Conditional-Generative Adversarial Network) was used to implement the prediction of two-dimensional soot signal, the results of soot prediction model showed that the prediction accuracies in different conditions are in the range of 0.83–0.91. Combining the previous two methods, the three-dimensional distribution of soot particle can be reconstructed by inputting the luminosity projections of jet flame, which provided new insight into the soot formation and distribution. The accuracy of this method can be verified by a unique validation experiment through oblique-sections of the flame, the validation accuracy is up to 0.92 in all conditions. The method can be applied to predict 3D soot field by using a well-trained deep learning model and several luminosity images to decrease the cost and limitation of experiment. It has the potential to be used for predicting other optical signals such as laser induced fluorescence and chemiluminescence if these datasets can be observed and trained. Therefore, this method is expected to be a solution for the visualization research of actual combustion devices.