气固流态化系统广泛存在于化工、材料、能源、资源和食品医药等过程工业中。在此类系统中颗粒和流体之间、颗粒和颗粒之间有着复杂的相互作用，是典型的非线性非平衡系统，有着明显的多尺度结构。随着计算能力的提高和模型、算法的发展，数值模拟已成为研究该系统的有力工具。双流体模型将颗粒处理成流体，在网格尺度描述颗粒的运动，无法提供颗粒尺度的信息。同时颗粒相的粘度和压力等本构关系也难以准确建立。离散颗粒方法则直接在颗粒尺度上计算颗粒间的作用力，跟踪颗粒的运动，可以方便地研究多粒径体系、非规则颗粒、颗粒间复杂作用力、颗粒停留时间分布等。但对于实际体系，由于颗粒数量巨大，并且颗粒间作用力的计算需要较小的时间步长，离散颗粒方法还鲜有应用。目前即使采用超级计算，离散模拟也多用于研究小规模实验装置。而随着过程工业的发展，对模拟精度与规模的要求不断提高，对离散颗粒方法进行深入的改进与扩展显得尤为重要。本论文试图通过对模型和计算方法的系统改进，从时间和空间尺度上扩展离散模拟的适用范围，为工业过程的模拟提供可能。论文的主要创新点如下：从源头上减少跟踪颗粒数目、加大时间步长是提高离散模拟速度、扩大其规模的根本途径，为此论文提出了以颗粒团代替单颗粒的粗粒化离散模型EMMS-DPM。该模型根据EMMS模型描述的团聚物尺寸和空隙率分布限定了所定义的颗粒团，即粗颗粒(coarse-grained particle, CGP)的大小，流体与粗颗粒之间的作用力采用EMMS曳力封闭。为了保证粗颗粒的碰撞能量耗散和原系统相当，我们根据颗粒动理论对恢复系数进行了修正。通过对固定床压降、循环流化床空隙率分布的模拟验证了模型的正确性。内容详见第三章。鉴于颗粒流体系统的多尺度特征和颗粒间以近程力为主的相互作用方式，发展了耦合利用多核与众核处理器(如Graphics Processing Unit, GPU)的离散模拟大规模多尺度并行计算方法，达到了计算与通信操作的高度重叠，显著提高了计算速度、效率与可扩展性。采用系统共享内存，实现了CPU端计算流体的同时在GPU端进行颗粒的计算，同时避免了通过文件交换数据带来的IO开销。计算程序通过了固定床压降和提升管空隙率分布等计算的验证。内容详见第二章。在上述改进的基础上，论文对气固流态化系统完整实现了“先分布、后演化”的EMMS计算范式。首先根据宏尺度EMMS模型预测的空隙率分布确定模拟的初始条件，然后据此非均匀地划分并行模拟区域以平衡计算负载，最后采用上述粗粒化方法模拟系统的动态行为。在典型的循环床提升管模拟中，采用该范式进一步提高了计算速度20%左右。集成上述工作并与可视化以及控制程序在线耦合运行，建立了循环流化床虚拟过程演示系统，实现了准实时的交互模拟。应用该系统，在对中试规模提升管的精度、速度和效率优先模拟中分别达到了与单颗粒离散模拟相当的准确性，二维实时模拟和目前所知最高的单GPU颗粒更新速率，表明了在现有工作基础上最终实现虚拟过程的可能性。内容详见第四章。基于上述工作，论文通过模拟实例表明了离散模拟已经能够提供实验与连续介质模拟难以获得的全面而详细的流场信息。一方面，对工业规模的流态床模拟时间可达小时量级，每天演化进度超过10分钟，为了解的很多操作和反应过程的颗粒与产物停留时间分布等信息提供了有力手段。比如对甲醇制烯烃工艺中的流态化反应器的模拟时间已超过6800s，对了解其中的催化剂结焦过程有很大帮助。另一方面，对接近中试规模的循环流化床实验装置可实现三维全回路离散模拟。比如对一个包含30公斤80微米直径的A类颗粒的系统采用302万30倍粒径的粗颗粒达到了每天8秒的计算速度。模拟获得的循环时间分布、停留时间分布、空隙率分布等重要参数均与实验数据及理论分析吻合。上述研究有力证明了本论文改进的离散模拟的实际应用能力与优势。内容详见第五章。综上所述，本论文提出的EMMS-DPM通过基于介尺度结构分析的粗粒化方法在基本保持准确性的前提下显著降低了离散模拟的计算量；提出的CPU-GPU耦合并行计算方法实现了CPU和GPU计算以及计算与通信操作的重叠，显著提高了计算速度。通过上述工作与宏尺度模型耦合减少了负载不均衡对并行计算速度的影响，通过与可视化程序、控制程序的耦合实现了在线交互模拟；通过模型、算法和软件、硬件的统一，在时空尺度上扩展了离散模拟的应用范围，初步建立了气固流态化系统虚拟过程的计算模式。

Gas-solid fluidization is widely found in process engineering, such as in chemical, material, energy, resource, food and pharmaceutical industries. Due to the complex interactions between the fluid and particles, it is a typical non-linear and non-equilibrium system featuring multiscale spatial-temporal heterogeneity. With the increase of computation capacity and development of physical models and numerical algorithms, simulation becomes a powerful tool in the study of fluidization. In the two fluid model (TFM), the particles are treated as a continuum phase and solved on Eulerian grids. The particle-scale behaviors are therefore averaged and no reliable constitutive laws for the solid phase properties such as viscosity and pressure have been established. On the other hand, the discrete particle method (DPM) tracks particles directly while the fluid is solved on Eulerian grids. It can be used to study particles with different diameters, irregular shapes, complex forces and particle residence time distribution. In real systems, since the particle number is very large and a very small time step is needed to resolve the collision process, practical application of this method is rarely seen. Even accelerated with supercomputers, only small fluidized with large particles can be simulated so far. With the development of process engineering and the increasing demand on simulation accuracy and scales, improvement of the simulation method in both models and algorithms is highly desirable.This dissertation tried to improve discrete method from physical models to computation methods and expanded its application scope both spatially and temporally. The main innovations are as follows:Reducing the number of tracked particles and increasing time step is the fundamental approach to expanding the application of DPM. We proposed a coarse grained method, EMMS-DPM, which is based on the analysis of meso-scale structures. In EMMS-DPM, the number of tracked particles is reduced significantly and a larger time step can be used. The diameter and inner voidage of coarse grained particles are defined by the cluster properties solved from the EMMS model. The restitution coefficient of the coarse grained particles are calculated from the kinetic theory of granular flow (KTGF), ensuring a same level of energy dissipation due to collisions. This model was verified by the simulations on the pressure drop in a fixed bed and the flow fields in two different risers.Considering the multiscale structure of particle fluid systems and the short range collision forces, a multiscale parallel computation method was developed combining the use of many core processors and multi core processors (such as GPUs). The computation is overlapped with communication and the computation speed, efficiency and scalability are improved significantly. Using system shared memory, the CPU and GPU in each process can run concurrently and eliminate the IO cost of data exchange between solid phase and gas phase. This method was verified with the simulations on the pressure drop of a fixed bed and the voidage distribution in a CFB riser. The details can be found in Chapter 2.Based on these innovations, the EMMS paradigm was implemented completely for gas-solid fluidization. The initial state of the particles is generated according the voidage distribution predicted by the macro-scale EMMS model. The simulation domain is then divided non-uniformly according to this distribution, resulting in a balanced load among different processes. Finally, the dynamic behavior of this system is simulated using the coarse grained method we developed. For the simulation of the riser in a typical CFB, the computation speed is improved about 20%. By integrating these works and coupling the computational program with visualization and control programs, a demo system for virtual process engineering (VPE) of CFB was developed. A typical CFB riser was simulated with different parameters focusing on accuracy (A), capability (C) and efficiency (E), respectively. The simulations achieved comparable accuracy of traditional DPM (A), real time speed (C) and highest particle update rate (E), respectively. This indicates that VPE of gas-solid fluidization is on the horizon already. With these developments we went on to demonstrate that EMMS-DPM is capabable of practical applications already. On one hand, simulation of industrial scale fluidized beds in EMMS-DPM can afford to track the process for hours, at a speed of about 10min, providing a powerful tool for the understanding of long-term behavior of particle and reactant in the fluidization process, such as resistance time distribution. For example, our simulation on a fluidized bed in the methanol to orifin (MTO) process lasted more than 6800s, which is very helpful to the study of carbon depositon of the catalyst particles. On the other hand, 3D full loop simulation of quasi-pilot-scale CFBs becomes possible. For example, we have simulated a whole CFB with 30kg 82mm particles of the Gedart A type, with 302 million coarse-grained particles representing 4226 billion real particles running at a speed of 8s/day. The particle cycling time, particle residence time, voidage distributions compare well with experimental data and/or theoretical analysis. These applications well demonstrated the capability and potential of EMMS-DPM and the related works in this dissertation.In summary, a coarse grained discrete particle method, EMMS-DPM, was proposed in this dissertation. It is based on the analysis of the meso-scale structures in fluidization and can reduce the computation cost of DPM significantly. A new parallel algorithm for the simulation of particle fluid systems which overlaps CPU and GPU computation and communication was also proposed, and the EMM paradigm was fully implemented in the simulations with the initial conditions and load balance basis provided by the macro-scale EMMS model. The computing software is further coupled with visualization and control programs to achieve quasi-realtime interactive simulation, a prototype of VPE. These works expanded the scope of DPM simulation both spatially and temporally, setting up a computational mode for VPE of gas-solid fluidization.