Three-dimensional high-sampling super-resolution reconstruction of swirling flame based on physically consistent diffusion models

Abstract

Three-dimensional swirling flame flow fields are often limited by factors such as system complexity and operational difficulty, resulting in relatively low achievable spatial resolution in experimental measurements. Providing high-quality visual data is crucial for optimizing the design of combustion chambers. This paper proposes a three-dimensional high-sampling super-resolution reconstruction method based on a physically consistent diffusion model to enhance combustion diagnosis capabilities. When basic diffusion models are used for super-resolution reconstruction, they may introduce artifacts or blurring. This can disrupt the inherent physical connections among flames, adversely affecting the reconstruction of flame details. Therefore, we have introduced a physically consistent encoder designed to process flame swirling data. This encoder allows the model to delve deeper into the intrinsic flow structure of the flame data, capturing flame resolution features across various scales and levels. It improves the accuracy of texture detail reconstruction in areas of intense combustion. During the training process, we have incorporated structural similarity loss into the loss function to assist the model in generating detailed and consistent edge combustion feature within the flame flow structure. These methods ensuring high fidelity and visual quality in the reconstructed flame. With the total voxel number 8× and 64× super-resolution tasks of the three-dimensional temperature fields of the swirling flames, the experimental results have shown that the method not only yields higher peak signal-to-noise ratio (PSNR) values and lower mean absolute error (MAE) compared to the baseline methods but also results in a more realistic visual representation of flame details.