EI、Scopus 收录
中文核心期刊
罗仁宇, 李奇志, 祖公博, 黄云进, 杨耿超, 姚清河. 基于卷积神经网络的超分辨率格子Boltzmann方法研究. 力学学报, 待出版. DOI: 10.6052/0459-1879-24-248
引用本文: 罗仁宇, 李奇志, 祖公博, 黄云进, 杨耿超, 姚清河. 基于卷积神经网络的超分辨率格子Boltzmann方法研究. 力学学报, 待出版. DOI: 10.6052/0459-1879-24-248
Luo Renyu, Li Qizhi, Zu Gongbo, Huang Yunjin, Yang Gengchao, Yao Qinghe. A super-resolution lattice boltzmann method based on convolutional neural network. Chinese Journal of Theoretical and Applied Mechanics, in press. DOI: 10.6052/0459-1879-24-248
Citation: Luo Renyu, Li Qizhi, Zu Gongbo, Huang Yunjin, Yang Gengchao, Yao Qinghe. A super-resolution lattice boltzmann method based on convolutional neural network. Chinese Journal of Theoretical and Applied Mechanics, in press. DOI: 10.6052/0459-1879-24-248

基于卷积神经网络的超分辨率格子Boltzmann方法研究

A SUPER-RESOLUTION LATTICE BOLTZMANN METHOD BASED ON CONVOLUTIONAL NEURAL NETWORK

  • 摘要: 对于隧道突涌水、飞行器和汽车外形设计等与流动相关的问题, 通常使用计算流体力学 (CFD) 来预测流场特征和分析结构性能, 为设计的快速迭代提供技术支撑. 然而, 高精度CFD仿真需要大量的计算资源. 近年来, 基于机器学习的超分辨率流场重构方法在流体力学领域取得了重大进展. 本文首次基于格子Boltzmann方法(LBM), 结合卷积神经网络, 建立新的超分辨率流场重构模型SRLBM, 将介观分布函数从低分辨率重建至高分辨率, 进而还原宏观速度场与涡量场. 首先, 使用LBM模拟了不同雷诺数下的二维圆柱绕流, 从多方面与文献进行对比, 验证了LBM的准确性. 然后, 将二维圆柱绕流数据作为SRLBM的训练集, 并对比不同缩放系数下SRLBM的重建效果. 结果表明, SRLBM在不同缩放系数下均能准确恢复高分辨率分布函数. 在8倍缩放系数下, 相比双三次插值重建方法, SRLBM重建的分布函数误差降低了近60%, 宏观场误差降低了近70%. 即使在32倍缩放系数下, SRLBM还原的宏观场与直接数值模拟结果基本保持一致. 固体体积分数作为额外输入通道可有效提高SRLBM的预测能力, 在32倍缩放系数下, 可使圆柱区域相对误差降低近40%. SRLBM具有一定的泛化能力, 当缩放系数为8时, 在一定雷诺数范围内重建的高分辨率流场误差小于3%. 因此, SRLBM在经过充分训练后, 具备成为高精度复杂流场快速重构方法的潜力.

     

    Abstract: For fluid problems, such as water inrush disaster in tunnels, shape design and optimization of aircraft and automobile, computational fluid dynamics (CFD) is commonly used to predict flow characteristics and analyze structural performance. However, high-precision CFD simulations require significant computational resources. In recent years, the super-resolution flow field reconstruction method based on machine learning has made significant progress in fluid mechanics. In this study, a novel super-resolution flow field reconstruction model SRLBM is firstly proposed, which is based on the lattice Boltzmann method (LBM) combined with convolutional neural networks, to reconstruct mesoscopic distribution functions from low-resolution to high-resolution, thereby restoring macroscopic velocity fields and vorticity fields. First, the two-dimensional flow around a cylinder at different Reynolds numbers is simulated using LBM, and compared with published data regarding various aspects to validate the accuracy of LBM. Then, the data from the two-dimensional cylinder flow are utilized as the training dataset for SRLBM, and the reconstruction performance of SRLBM under different scaling factors are compared. The results show that SRLBM can effectively restore high-resolution distribution functions for different scaling factors. At a scaling factor of 8, compared to bicubic interpolation reconstruction methods, the error of SRLBM is reduced by nearly 60% regarding distribution functions, and is reduced by nearly 70% regarding the macroscopic fields. Even at a scaling factor of 32, the macroscopic field restored by SRLBM is generally consistent with the results from the direct numerical simulation. Incorporating solid volume fraction and distribution function as input channels can effectively enhance the predictability of SRLBM, which can reduce the relative error in the cylinder region by nearly 40% when the scaling factor is 32. The SRLBM demonstrates a good generalization ability, when the scale factor is 8, the error in the reconstructed high-resolution flow field within a certain range of Reynolds numbers is below 3%. Therefore, after adequate training, SRLBM has the potential to become an effective method for reconstructing high-resolution complex flow field.

     

/

返回文章
返回