基于扩散模型的流场超分辨率重建方法

A FLOW FIELD SUPER-RESOLUTION RECONSTRUCTION METHOD BASED ON DIFFUSION MODEL

低分辨率的流场数据具有较少的信息量,不能充分捕捉流场中的细节演化过程.尤其对于湍流的随机脉动特征和小尺度涡旋细节特征更加难以获取,这限制了对流场演化机理进行深入研究.为了解决这一局限性,并从低分辨率流场中重建高分辨率数据,文章提出一种流场超分辨率重建的生成扩散模型FlowDiffusionNet.该模型以低分辨率流场数据输入作为约束条件,采用去噪分数匹配方法,来实现高分辨率流场数据的复现.FlowDiffusionNet在结构设计上充分考虑了流场数据的低频信息与高频空间特征,采用基于扩散过程的建模方法,用于重建高分辨率流场数据的残差.该模型结构便于实现迁移学习,可在不同程度的退化流场上应用.将该方法在多种经典流场数据集上进行测试,并与双三次插值(bicubic)、超分辨率生成对抗网络(SRGAN)、超分辨率卷积神经网络(SRCNN)等方法进行比较.结果表明,该方法在各种流场上的重建性能达到最佳水平,特别是对于包含小尺度涡结构的4倍下采样流场数据,客观评价指标SSIM达到0.999.

Low-resolution flow field data contains limited information,which fails to fully capture the detailed evolutionary processes of the flow field.Especially for the random turbulent features and small-scale vortex details in turbulence,they are even more challenging to obtain,thereby restricting the in-depth investigation of flow field evolution mechanisms.In order to address this limitation and reconstruct high-resolution data from low-resolution flow fields,this paper proposes a generative diffusion model called FlowDiffusionNet for flow field super-resolution reconstruction.The model takes the low-resolution flow field data input as the constraint condition,and utilizes a denoising fraction matching method to reproduce high-resolution flow field data.FlowDiffusionNet's structural design takes into consideration both the low-frequency information and high-frequency spatial features of flow field data,employing a diffusion-based modeling technique to reconstruct the residuals for high-resolution data.The proposed model's architecture is amenable to transfer learning,allowing its application to degraded flow fields at different levels.The performance of FlowDiffusionNet is evaluated on various classical flow field datasets and compared against other methods such as bicubic interpolation,super-resolution generative adversarial network(SRGAN),and super-resolution convolutional neural network(SRCNN).The results demonstrate that the proposed method achieves the best reconstruction performance on various flow fields,especially for flow field data with small-scale vortex structures down sampled by a factor of 4,where the objective evaluation index structural similarity index measure(SSIM)reaches 0.999.