明渠湍流中存在超过3h（h为水深）甚至可超20h的大尺度结构（large-scale motions, LSMs），该大尺度结构对总紊动强度贡献显著，并在明渠物质输移方面发挥重要作用。迄今有关明渠LSMs的研究仅见于实验量测或小计算域（不超过25h）直接数值模拟（Direct numerical simulation, DNS）。实验量测多局限于单点测量，需进行时间-空间信号转换来间接推演LSMs；小计算域DNS无法捕捉LSMs全貌。有待通过长计算域DNS对明渠挟沙水流大尺度结构的主导尺度、床面附着结构、平均床面切应力等对象进行研究，并澄清明渠和其他流型LSMs间的异同。 本文基于作者自主开发的DNS求解器 CP3d，利用预乘谱、线性相干谱等手段，对清水光滑床面、含悬沙光滑床面、固定粗糙以及有推移质运动床面的明渠大尺度结构统计特征进行研究。计算域长度不低于100h，网格数最多达84亿，使用浸没边界法（Immersed Boundary Method, IBM）解析的可动颗粒数目达220万。 本文提出基于积分平均的简化润滑力模型，改进二维区域分解并行框架，使用压力泊松方程 (Pressure Poisson Equation, PPE) 最优化选取方法，使PPE求解器计算速度提升约50%. DNS-DEM (Discrete Element Method, 离散元)-IBM 的计算效率比文献中给出的速度高数十倍。求解器在直至8000个CPU下并行效率良好。 清水光滑明渠大尺度结构方面。首次通过 DNS 捕捉到流向速度 (u) 的流向预乘谱中明显的双峰分布。与封闭槽道相比，明渠中u以及展向速度 (w) 在大尺度部分的脉动更强，且床面处明渠的流速梯度谱在大波长部分已高于槽道。另外，在 y >0.7h范围内，明渠床面附着结构的临界波长和倾角均小于槽道。 含悬沙光滑明渠大尺度结构方面。当 y<0.8h，含悬沙速度场总脉动强度和清水算例接近，但大尺度脉动受到抑制，而小尺度脉动强度增强。浓度场(c)的流向大尺度峰值波长几乎不随悬沙平均浓度cv变化，而u的峰值波长随cv增大而减小。床面附着结构方面，c的附着波长大于u，附着倾角比u小4°~5°. 固定粗糙床面的明渠大尺度结构方面。和光滑床面相比，u的流向超大尺度峰值减弱，大于 30h 部分的脉动强度略有增强。另外波长小于 5h 的脉动对床面切应力的贡献增强。颗粒受力方面，大尺度流向脉动贡献了超 65% 的颗粒扭矩总脉动。颗粒释放后，床面形态不断发展最终形成沙纹形态，且 u的流向预乘谱中出现和沙纹波长接近的峰值；远离床面后，尺度介于 10-30h 之间的脉动强度进一步减弱。

Large-scale motions (LSMs), whose length scale are greater than 3h (h is water depth) and might be longer than 20h, exist in turbulent open-channel flows (OCFs). These large-scale coherent structures make great contributions to turbulent fluctuation intensities, and play an important role in the mass transfer of OCFs. Up to now, researches related to LSMs in OCFs are limited to experimental works and direct numerical simulation (DNS) in small computational domains (Streamwise domain size no longer than 25h). The related experimental works mainly use the temporal data recorded by an one-point fixed sensor. The spatial statistics of LSMs are not available, and the transformation between temporal data to spatial data is mandatory. DNS with small computational domains can not capture the full picture of LSMs. It is highly necessary to perform DNS of OCFs with large domain size to study the characteristics of LSMs (dominant scales, wall attached motions, mean shear stress), and to clarify the similarities and differences of LSMs in OCFs. In this work, the author developed a DNS solver called CP3d (Channel-Particle 3d) from scratch, and then CP3d was utilized to study the statistic behavior of LSMs in three cases: OCFs with smooth bed and particle-free flows, OCFs with smooth bed and suspended load, OCFs with fixed roughness bed and bedload transport. The main analysis tools are pre-multiplied spectra and linear coherence spectra. The streamwise computational domain is no less than $100h$. The number of computational meshes is up to 8.4 billion, and the number of moving particles whose surfaces are resolved by immersed boundary method (IBM) is up to 2.2 million. We proposed a simplified lubrication force model based on the integral average, and developed the two-dimensional pencil-like parallel framework, and then adopted an optimal pressure Poisson equation (PPE) solver. Compared to the traditional method, the resulting solver achieved a 50% speedup when computing PPE. In our test, CP3d is more than ten times faster than other similar solver by DNS-IBM-DEM coupling method reported in the literature. A nearly perfect linear strong scaling performance is achieved up to 8,000 computational cores. In terms of LSMs in OCFs with smooth bed and particle-free flows, two distinct peaks in streamwise pre-multiplied spectrum of streamwise velocity (u) are captured by DNS for the first time. (1) Compared to closed-channel flows, the spectra of u and spanwise velocity (w) in OCFs are stronger in large wavelength scale, and such differences can also be observed even in the spectra of streamwise velocity gradient at the wall. (2) In the region y>0.7h, the wavelengths and inclined angles of wall-attached LSMs in OCFs are smaller than those in closed-channel flows. As for LSMs in OCFs with smooth bed and suspended load: (1) In the region y<0.7h, the total fluctuation intensities of velocity with suspended load are close to those in particle-free OCFs, while the small-scale fluctuation are enhanced, and the fluctuation of LSMs are reduced. (2) The streamwise peak wavelengths of concentration (c) spectrum alters very little as the mean concentration cv changes, while the peak wavelengths of u decrease as cv increases. (3) As for the wall-attached motion, the wavelengths of wall-attached LSMs for c are larger than those for u, while the inclined angles for c are about 4° ~ 5° smaller than those for u. When it comes to LSMs with fixed roughness bed: (1) Compared with the smooth cases, the streamwise peak of very-large-scale motion for u is damped, while the fluctuations of wavelength larger than 30h for u increase slightly. (2) In addition, the contribution to the mean wall-shear stress from streamwise wavelengths smaller than 5h enhances. (3) The streamwise LSMs contribute more than 65% of the particle torque fluctuation intensity. After bed particles are released, the general flat rough bed develops gradually into a ripple-like bedform: (1) The pre-multiplied spectra of streamwise and wall-normal velocity show peaks near the mean wavelength of the ripple. (2) In the region far away from the bed, the fluctuation intensities with wavelengths within 10h to 30h are further decreased.