鼓泡塔反应器在过程工业中已经得到广泛的应用，其内部气液两相流动的快速和准确模拟对这类反应器的设计和放大具有十分重要的意义。然而，气液两相的多尺度流动结构、复杂内构件以及气液两相湍流等使得鼓泡塔反应器的快速和准确模拟非常困难。针对这些问题，本文以格子波尔兹曼方法（Lattice Boltzmann Method，LBM）为基础，研究了鼓泡塔反应器内的微尺度、介尺度和宏尺度气液两相流动问题；实现了LBM方法对气液两相流中微、介尺度物理现象的直接数值模拟以及基于LBM的混合物模型的气液两相流中的宏尺度现象的模拟，并将LBM与浸入边界法（Immersed Boundary Method，IBM）耦合，用以模拟带有复杂内构件和复杂边界条件的反应器内的流动；进一步研究了湍流模拟中LBM-RANS耦合模型显式求解的数值收敛性问题和LBM-RANS耦合模型隐式求解的计算加速问题，具体内容如下：1、气液两相流动中微尺度和介尺度问题的直接数值模拟（LBM-DNS）：应用直接数值模拟方法，一方面可以强化对气液两相流动微介尺度物理现象的认识，另一方面也可以为上层连续介质模型本构关系的建立和验证提供依据。当前，在高运动粘度比、高雷诺数流动条件下，基于LBM的气液直接数值模拟存在模拟计算稳定性差的问题。本文发展了基于LBM的气液两相流直接数值模拟方法，实现了气液运动粘度比1:103、高雷诺数、低莫顿数的气液两相流动模拟。以此为基础研究了单气泡、双气泡以及气泡群的运动。模拟能捕捉到空气-水体系2mm气泡Z字型上升过程以及特定条件下水平并列的两个气泡的波浪形上升过程，结果表明气泡水平方向上的运动主要是由气泡尾涡引起的。气泡群运动模拟中气泡之间聚并现象明显，未发现气泡破碎，这可能是由于气液直接数值模拟中液相湍动微弱，无法达到气泡破碎条件引起的，需要进一步提高网格数量才可能模拟出聚并与破碎共存的现象。2、基于LBM的气液两相混合物模型（LBM-Mixture）：由于计算量的限制，直接数值模拟目前尚无法直接应用于工业规模的反应器的模拟，一个较为实际的方法是应用连续介质层次的混合物模型，并用采用LBM方法求解和GPU加速计算，这可以为反应器的初步概念设计提供一种快速的计算方法。基于此设想，本文发展了基于LBM的混合物模型，该模型将压力项的作用以外力源项的方式添加到LBM方程中，克服了文献报道的模型中无量纲松弛时间τ必须为1才能恢复到宏观NS（Navier-Stokes，NS）方程的限制。在LBM模拟中，流体的运动粘度v与τ以及时间步长以及空间步长相关，即。如取τ=1，模拟中必须选取较小的才能模拟低运动粘度的流动，因而难以模拟大尺度、低粘度流动问题。本文发展的LBM混合物模型被成功应用于中心或偏心进气的二维鼓泡塔的模拟（空气-水体系）。模拟结果表明该方法能准确捕捉到二维鼓泡塔中的气相水平摆动过程，模拟预测的摆动频率和时均液速与文献报道的实验结果吻合。进一步比较了GPU并行加速的LBM-Mixture模型求解速度与Fluent中双流体求解速度。相同网格和时间步长情况下，4块GPU卡加速的LBM-Mixture模型的计算速度约为Fluent求解的双流体模型（4个CPU核）的250倍。LBM-Mixture模型的GPU并行加速方法为气液体系实时模拟和VPE（Virtual Process Engineering）的实现打下了基础。3、LBM耦合浸入边界法（LBM-IBM）模拟复杂构体和复杂边界的流动问题：工业规模的鼓泡塔或者气液搅拌釜内往往有十分复杂的内构件或不规则的几何边界，如换热管、搅拌桨、折流板等等。对于LBM模拟而言，一方面在经典的LBM框架下，LBM处理复杂边界时具有一定的局限性；另一方面，使用基于有限体积的LBM处理复杂边界，需要生成贴体网格，增加了网格生成的时间成本。本文采用IBM来处理复杂边界，缩短了前处理时间。建立了GPU并行加速的LBM耦合IBM模型，并通过模拟Rushton桨搅拌釜内的流动来验证模型的准确性。单GPU卡的计算速度为CPU核的50倍以上。相比于16核CPU并行计算的基于Fluent的滑移网格模拟，4块GPU卡鼓泡塔反应器在过程工业中已经得到广泛的应用，其内部气液两相流动的快速和准确模拟对这类反应器的设计和放大具有十分重要的意义。然而，气液两相的多尺度流动结构、复杂内构件以及气液两相湍流等使得鼓泡塔反应器的快速和准确模拟非常困难。针对这些问题，本文以格子波尔兹曼方法（Lattice Boltzmann Method，LBM）为基础，研究了鼓泡塔反应器内的微尺度、介尺度和宏尺度气液两相流动问题；实现了LBM方法对气液两相流中微、介尺度物理现象的直接数值模拟以及基于LBM的混合物模型的气液两相流中的宏尺度现象的模拟，并将LBM与浸入边界法（Immersed Boundary Method，IBM）耦合，用以模拟带有复杂内构件和复杂边界条件的反应器内的流动；进一步研究了湍流模拟中LBM-RANS耦合模型显式求解的数值收敛性问题和LBM-RANS耦合模型隐式求解的计算加速问题，具体内容如下：1、气液两相流动中微尺度和介尺度问题的直接数值模拟（LBM-DNS）：应用直接数值模拟方法，一方面可以强化对气液两相流动微介尺度物理现象的认识，另一方面也可以为上层连续介质模型本构关系的建立和验证提供依据。当前，在高运动粘度比、高雷诺数流动条件下，基于LBM的气液直接数值模拟存在模拟计算稳定性差的问题。本文发展了基于LBM的气液两相流直接数值模拟方法，实现了气液运动粘度比1:103、高雷诺数、低莫顿数的气液两相流动模拟。以此为基础研究了单气泡、双气泡以及气泡群的运动。模拟能捕捉到空气-水体系2mm气泡Z字型上升过程以及特定条件下水平并列的两个气泡的波浪形上升过程，结果表明气泡水平方向上的运动主要是由气泡尾涡引起的。气泡群运动模拟中气泡之间聚并现象明显，未发现气泡破碎，这可能是由于气液直接数值模拟中液相湍动微弱，无法达到气泡破碎条件引起的，需要进一步提高网格数量才可能模拟出聚并与破碎共存的现象。2、基于LBM的气液两相混合物模型（LBM-Mixture）：由于计算量的限制，直接数值模拟目前尚无法直接应用于工业规模的反应器的模拟，一个较为实际的方法是应用连续介质层次的混合物模型，并用采用LBM方法求解和GPU加速计算，这可以为反应器的初步概念设计提供一种快速的计算方法。基于此设想，本文发展了基于LBM的混合物模型，该模型将压力项的作用以外力源项的方式添加到LBM方程中，克服了文献报道的模型中无量纲松弛时间τ必须为1才能恢复到宏观NS（Navier-Stokes，NS）方程的限制。在LBM模拟中，流体的运动粘度v与τ以及时间步长以及空间步长相关，即。如取τ=1，模拟中必须选取较小的才能模拟低运动粘度的流动，因而难以模拟大尺度、低粘度流动问题。本文发展的LBM混合物模型被成功应用于中心或偏心进气的二维鼓泡塔的模拟（空气-水体系）。模拟结果表明该方法能准确捕捉到二维鼓泡塔中的气相水平摆动过程，模拟预测的摆动频率和时均液速与文献报道的实验结果吻合。进一步比较了GPU并行加速的LBM-Mixture模型求解速度与Fluent中双流体求解速度。相同网格和时间步长情况下，4块GPU卡加速的LBM-Mixture模型的计算速度约为Fluent求解的双流体模型（4个CPU核）的250倍。LBM-Mixture模型的GPU并行加速方法为气液体系实时模拟和VPE（Virtual Process Engineering）的实现打下了基础。3、LBM耦合浸入边界法（LBM-IBM）模拟复杂构体和复杂边界的流动问题：工业规模的鼓泡塔或者气液搅拌釜内往往有十分复杂的内构件或不规则的几何边界，如换热管、搅拌桨、折流板等等。对于LBM模拟而言，一方面在经典的LBM框架下，LBM处理复杂边界时具有一定的局限性；另一方面，使用基于有限体积的LBM处理复杂边界，需要生成贴体网格，增加了网格生成的时间成本。本文采用IBM来处理复杂边界，缩短了前处理时间。建立了GPU并行加速的LBM耦合IBM模型，并通过模拟Rushton桨搅拌釜内的流动来验证模型的准确性。单GPU卡的计算速度为CPU核的50倍以上。相比于16核CPU并行计算的基于Fluent的滑移网格模拟，4块GPU卡加速的LBM-IBM模拟的计算速度是其270倍左右。GPU并行加速的LBM耦合IBM模拟具有准确和快速的优点，进一步可与本文发展的LBM-DNS/LBM-Mixture/LBM-RANS等模型结合，用于模拟带有复杂内构件的鼓泡塔气液两相流动。4、LBM-RANS耦合模型的数值收敛和计算加速问题：相比于直接数值模拟和大涡模拟，采用LBM-RANS耦合模型模拟实际工业反应器中的湍流和多相流是较为经济的方法。一方面，显式求解的RANS模型与LBM耦合的计算效率较高，但计算稳定性差；另一方面，隐式求解的RANS模型与LBM耦合的计算稳定性好，但计算效率低。本文研究了二维显式求解的标准k-ε模型与LBM耦合的数值稳定性问题，并给出了可提高数值稳定性的数值格式组合方式。然而，这些数值格式组合很难保证三维计算的稳定性。隐式求解RANS模型能使计算更加稳定，但计算效率低。本文提出了一种时空异步多尺度方法，提高了LBM耦合隐式求解RANS模型的计算效率，解决了LBM耦合隐式求解RANS模型快速计算的关键技术问题。以上四个关键问题的解决，将为LBM方法模拟实际工业过程的气液两相流动建立基础。本文的研究工作表明，LBM方法可研究从微、介尺度的气泡动力学(LBM-DNS)，到反应器尺度的多相流动。后者涵盖了从反应器多相流的模拟和初步概念设计（LBM-Mixture）到复杂内构件（LBM-IBM）和湍流的模拟（LBM-RANS）。加速的LBM-IBM模拟的计算速度是其270倍左右。GPU并行加速的LBM耦合IBM模拟具有准确和快速的优点，进一步可与本文发展的LBM-DNS/LBM-Mixture/LBM-RANS等模型结合，用于模拟带有复杂内构件的鼓泡塔气液两相流动。4、LBM-RANS耦合模型的数值收敛和计算加速问题：相比于直接数值模拟和大涡模拟，采用LBM-RANS耦合模型模拟实际工业反应器中的湍流和多相流是较为经济的方法。一方面，显式求解的RANS模型与LBM耦合的计算效率较高，但计算稳定性差；另一方面，隐式求解的RANS模型与LBM耦合的计算稳定性好，但计算效率低。本文研究了二维显式求解的标准k-ε模型与LBM耦合的数值稳定性问题，并给出了可提高数值稳定性的数值格式组合方式。然而，这些数值格式组合很难保证三维计算的稳定性。隐式求解RANS模型能使计算更加稳定，但计算效率低。本文提出了一种时空异步多尺度方法，提高了LBM耦合隐式求解RANS模型的计算效率，解决了LBM耦合隐式求解RANS模型快速计算的关键技术问题。以上四个关键问题的解决，将为LBM方法模拟实际工业过程的气液两相流动建立基础。本文的研究工作表明，LBM方法可研究从微、介尺度的气泡动力学(LBM-DNS)，到反应器尺度的多相流动。后者涵盖了从反应器多相流的模拟和初步概念设计（LBM-Mixture）到复杂内构件（LBM-IBM）和湍流的模拟（LBM-RANS）。

Bubble column reactors are widely applied in process industries. Accurate and fast simulation of gas-liquid two-phase flow in bubble column reactors is important for the design and scale-up of bubble columns. Multiscale problems in gas-liquid two-phase flow, accompanied with the highly turbulent flow and complex geometry of internals, make it difficult to achieve the fast and accurate simulation.In view of these problems, the gas-liquid flows at micro-scale, meso-scale and lab-scale have been investigated by Lattice Boltzmann Method (LBM). A multiscale modeling strategy is put forward: a LBM-based direct numerical simulation (DNS) is developed to simulate the micro and meso-scale problems, and a LBM-based mixture model is used to simulate the flow at the reactor scale. Then LBM is coupled with Immersed Boundary Method (IBM) to settle the problems of complex boundary condition in LBM simulation. Finally, the numerical and computation acceleration issues are analyzed in the coupled LBM-RANS (Reynolds-Averaged Navier-Stokes, RANS) model to deal with the highly turbulent flow. 1. Direct Numerical Simulation (DNS) for microscale and mesoscale problems in gas-liquid flow. DNS for gas-liquid flow might be an effective approach to understand the physical background of gas-liquid flow and establish or validate constitutive closure models for the models at upper scales. Currently, there are a number of numerical issues in LBM-based DNS model in simulating gas-liquid flow, such as numerical instability for large kinematic viscosity ratios or high Reynolds numbers. In this work, a LBM-based DNS model for gas-liquid flow is developed. It can simulate the gas-liquid flow with large kinematic viscosity ratio (1:103), high Reynolds number and low Morton number. The developed model has been used to study the bubble dynamics from a systematic and multiscale perspective, that is, progressively probing the behaviors of a single bubble, a bubble pair and a bubble swarm. The Z-type rising process of a single 2mm bubble in water and the wavy rise process of two bubbles in given conditions are captured, and these lateral movements of bubbles are essentially induced by bubble wakes. Coalescence or breakage phenomenon is not captured due to the insufficient turbulence intensity in simulation, which could be further settled by increasing the turbulence intensity and using much more fine grids and GPU-Accelerated simulation. 2. A LBM-Mixture model: due to the limitation in computational resources, DNS cannot be used in the simulation of industrial-scale bubble column reactor directly at present, and the continuum-based mixture model solved by LBM and accelerated by multiple GPUs may supply an efficient and fast approach for preliminary concept design of bubble column reactors. The LBM-Mixture model is developed and can overcome the weakness of the original LBM model, in which the no-dimensional relaxation time τ should be set as 1 to recover LBM to the macro-scale equations of fluid flow. In LBM, the kinematic viscosity v is a function of τ, the time step and the spatial resolution through the equation . If τ is set as 1, has to be smaller to simulate the low kinematic fluid flow, leading to much higher computational cost. The developed model has been applied in the gas-liquid flow simulation of a central or partial aerated flat rectangle bubble column (air-water system). The lateral oscillations of air phase are captured by simulations. The predicted oscillation frequency and time-averaged liquid velocity are in accordance with the reported experimental data. Then we compared the computational speed of multiple GPU-accelerated LBM-Mixture model with Two-Fluid model solved by Fluent. The computation speed of 4 GPU cards accelerated LBM-Mixture model is about 250 fold faster than Two-Fluid model solved by Fluent using 4 CPU cores, when the numerical setups are the same. LBM-Mixture model accelerated by multiple GPUs has established a solid foundation for Real-time simulation or VPE（Virtual Process Engineering）in gas-liquid system.3. Integrating the LBM and IBM for complex geometry and complex boundary flow problems. The industrial-scale bubble columns or gas-liquid stirred tanks are always accompanied with complex internals or boundaries, such as heat exchange tubes, impellers, baffles and so on. This has introduced many technical and scientific problems in LBM simulations. On one hand, it is difficult for LBM to treat the complex boundaries in the classical cubic discretization framework. On the other hand, the generation of body-fitted grid for the Finite Volume based LBM is also difficult and time-consuming. In this work, we have used IBM to deal with complex boundaries and save the grid generation time. The GPU parallel computation has been used for the acceleration of the coupled model of LBM and IBM. The turbulent flow in a Rushton turbine stirred tank has been simulated for model validation. The computation speed of single GPU card accelerated LBM-IBM simulation is about 50 times faster than the computational speed of single CPU core. The computation speed of 4 GPU cards accelerated LBM-IBM simulation is about 270 times faster than 16 CPU cores accelerated Sliding Mesh simulation based on Fluent software. The integration of LBM and IBM accelerated by multiple GPUs is accurate and fast. This laid some foundation for numerical simulation of gas-liquid flow in bubble columns with complex internals, and it is also expected to be coupled with LBM-DNS, LBM-Mixture and LBM-RANS models developed in this work.4. The convergence or computational efficiency problems in the coupling of LBM and RANS models: Compared with DNS or LES for turbulent flow, LBM-RANS is more economical and practical to deal with the turbulent flow in bubble columns. However, while the explicitly-solved LBM-RANS model is efficient, it is hard to converge. On the other hand, the implicitly-solved LBM-RANS model is more stable, but time-consuming in computation. Then we investigated the numerical convergence problems for the coupling of LBM and the explicitly-solved standard k-ε model in 2D cases and found that some numerical treatment combinations could improve or ensure the numerical stability. However, those numerical treatments cannot ensure the convergence for 3D simulation. We then implemented the implicit methods for RANS model to improve the numerical stability, and proposed a new method, i.e., the Spatial-temporal Multi-Scale Asynchronous Method, to accelerate the computation.By resolving the above four issues, this work paved the way for LBM simulation to simulate the gas-liquid flows in industrial processes. We demonstrated that LBM could be an efficient approach to investigate the bubble dynamics at micro- or meso-scales with LBM-DNS and the gas-liquid flow in reactor-scale. The latter covers the preliminary concept design of bubble column reactors with LBM-Mixture model, the treatment of complex internals with LBM-IBM and the simulation of turbulence with LBM-RANS.