湍流是一种典型的非平衡非线性系统并且广泛存在于自然界和各种工程应用当中，被学术界广泛认为是难以攻克的难题之一。随着计算机技术的进步，数值模拟成为人们研究湍流问题的重要手段。目前，雷诺平均方法是解决工程湍流切实可行的方案，然而，传统的湍流模型往往假定计算网格内流体总是处于充分湍流状态，忽视了流动中的层流部分，导致模拟的准确性不足。能量最小多尺度（Energy Minimization Multi-Scale，EMMS）湍流模型是基于EMMS原理的介尺度湍流模型，该模型借鉴两相流分相思想，视单相湍流由湍流流体成分和层流流体成分组成，通过调节湍流流体体积分数f刻画计算网格内的非均匀结构，能够计算层、湍共存的流动状态，有效解决了非均匀湍流系统定量模拟。格子Boltzmann方法（Lattice Boltzmann Method，LBM）作为一种高效的流体求解器，具有显式计算、易于并行和能够处理复杂边界等特点，十分适合GPU的大规模长流线作业。因此将EMMS湍流模型与LBM相结合，并在此基础上实现算法的多GPU并行，将可以提升工业湍流模拟的速度和准确性。基于LBM和EMMS湍流模型，本论文主要章节安排如下：第一章为文献综述。首先回顾了当前主要的湍流模式理论，在此基础上概述了考虑了层流及湍流两相相互作用的EMMS湍流模型，最后介绍了用于求解流场的LBM方法；第二章实现了LBM耦合EMMS湍流模型算法的多GPU并行。提出了一种EMMS湍流模型与LBM的耦合方法，利用CUDA和MPI实现了算法的多GPU并行，并测试了算法在Tesla K80和Tesla P100上的加速性能；第三章在第二章的基础上进行了算法的数值验证。通过顶盖驱动方腔流和后台阶流，验证了LBM耦合EMMS湍流模型算法的有效性。发现在未使用壁面函数的情况下，模型可以获得一定精度的流场且弱网格相关。最后，认识到模型可以一定程度上预测流动转捩。第四章对LBM耦合EMMS湍流模型算法的大规模并行计算进行了初步探索。提出了一种三维复杂构型的网格生成技术，在此基础上对多级涡轮和F22中的绕流初步进行了大规模的模拟，展示了LBM耦合EMMS湍流模型的多GPU并行算法强大的并行能力和良好的工业应用前景。第五章总结本论文工作的主要成果与结论，并展望了LBM耦合EMMS湍流模型在理论基础与大规模算法实现方面的几个努力的方向。;Turbulence is a typical nonlinear non-equilibrium system and it is widespread in nature and various engineering applications, and turbulence is widely considered as one of the most difficult problems to be solved. With rapid development of computer technology, numerical simulation has become an important means to investigate turbulence. At present, the Reynolds-averaged Navier–Stokes model is a feasible solution for engineering turbulence. However, the traditional turbulence models always assumes that the fluids in the computational grid are fully turbulent states and neglects the laminar portion of the fluid flow, which leads to a source of inaccuracies in modeling practical engineering flows .The Energy Minimization Multi-Scale (EMMS) based turbulence model is a mesoscale turbulence model originated from the principle of compromise-in-competition between viscosity and inertia. The model draws on the idea of phase separation in two-phase flows and regards single-phase turbulence as a mixture of turbulent fluid components and laminar fluid components. The inhomogeneous structure in the computational grid is characterized by adjusting the turbulent fluid volume fraction f. The EMMS-based turbulence model can calculate the flow state of laminar and turbulent coexistence and effectively solve the quantitative simulation of non-uniform turbulence system. As an efficient fluid solver, explicit scheme, easy parallelism and capability of handling complex boundaries are some of the characteristics of lattice Boltzmann method and it is very suitable for large-scale computation on GPUs. Therefore, combining EMMS-based turbulence model with the lattice Boltzmann method and implementing the parallel algorithm on multi-GPUs will improve the speed and accuracy of industrial turbulence simulation.Based on the LBM and EMMS-based turbulence model, the following chapters are as follows:Chapter 1 is literature review. Firstly, the current main turbulence models are reviewed. Then, EMMS –based turbulence model with consideration of the interaction between laminar and turbulent phases is summarized. Finally, the LBM is used to solve the flow field is introduced.Chapter 2 implements the algorithm of multi-GPUs parallelism of the LBM coupled EMMS-based turbulence model. A method coupled EMMS-based turbulence model and LBM is presented. The multi-GPUs parallelization of the proposed algorithm is implemented using CUDA and MPI. The acceleration performance of the proposed algorithm on Tesla K80 and Tesla P100 is tested.Chapter 3 mainly numerically verified the proposed algorithm. The proposed algorithm coupled LBM and EMMS-based turbulence model is validated by the cases of the lid-driven cavity and backward-facing step flows. It is found that the proposed model can obtain a certain accuracy of flow field without using the wall functions and features weak dependence on the mesh spacing of the computational grid. Finally, the implication is found that the proposed model can predict flow transitions to some extent.Chapter 4, the large-scale parallel computing of LBM coupled EMMS-based turbulence model algorithm is preliminarily explored. A grid generation technique for three-dimensional complex configuration is proposed. Then large-scale simulations of flow through the multi-stage turbine and F22 are carried out, which demonstrates the powerful parallelism and good industrial application prospects of the multi-GPUs parallel algorithm of LBM coupled EMMS –based turbulence model.Chapter 5 summarizes the main achievements and conclusions of the work of this thesis, and looks forward to some efforts in the theoretical basis and large-scale algorithm implementation of the LBM coupled EMMS-based turbulence model.