颗粒流体两相流是一类典型的复杂多相系统，其中气固系统因其流体力学不稳定性呈现显著的多尺度动态非均匀结构，已引起广泛关注。而液固系统通常呈现较为均匀的结构，其复杂性更多体现在颗粒间存在较强的相互作用时形成的聚团，但相关研究还相当薄弱。本论文以聚乙烯（polyethylene，PE）颗粒—烷烃液固系统为例，力图分析液固系统在不同尺度的复杂性：微尺度上颗粒表面的复杂性；介尺度上聚团的生长、聚并、破碎等过程；以及宏尺度上颗粒流体非均匀结构对整体流动的影响。为此，需要在不同尺度上发展不同的基于颗粒解析（particle-resolved）的模拟方法进行更准确快速的数值模拟，从而形成全面的液固两相流模拟能力。本论文分别采用三种不同尺度的基于颗粒解析的模拟方法——颗粒离散元方法（discrete element method，DEM）、直接数值模拟（direct numerical simulation，DNS）和离散颗粒模型（discrete particle model，DPM）研究了PE颗粒—烷烃液固系统，并使用基于三种高性能计算设备——中央处理器（central processing unit，CPU）、众核（many integrated core，MIC）架构处理器和图形处理器（graphic processing unit，GPU）的并行计算技术优化了这些方法。论文中各种模拟方法的研究内容包括：（1）颗粒尺度的DEM方法。选用非均匀有理B样条（non-uniform rational basis-splines，NURBS）描述颗粒的复杂表面形状，实现了收敛速度更快的NURBS曲线/曲面接触判断算法。该方法基于经典的颗粒接触力模型计算颗粒间相互作用力，建立了具有较高精度和计算效率的基于NURBS的DEM，并通过一对颗粒间碰撞的轨迹定量分析验证并展示该方法的精度与效率，同时说明了其处理更复杂问题的可行性。采用多线程并行优化了前述DEM，并在CPU和MIC上测试了优化的效果。（2）团聚物尺度的DNS方法。采用并行性良好的格子玻尔兹曼法（lattice Boltzmann method，LBM）与球形颗粒DEM耦合，模拟了环流反应器提升管内PE颗粒—烷烃液固系统的流动过程。针对环流反应器提升管的驱动方式，提出了两段法周期边界条件（two-region periodic boundary condition，TPBC），并首先在颗粒团聚物明显的气固两相流模拟中验证了其可行性。然后，模拟了环流反应器提升管中的液固两相流动，获得了流速分布、颗粒浓度等流动结构。最后优化了前述DNS的CPU+MIC+GPU多尺度异构并行计算，并通过测试说明了多尺度并行计算的良好性能。（3）反应器尺度DPM方法。采用包含颗粒间粘附力模型的多尺度并行DPM，模拟了实验室规模的环流反应器。对于PE颗粒—烷烃液固系统，建立了反映浸润在烷烃中的溶胀PE颗粒易接触固结的粘附力模型，并应用于LBM-DEM耦合的DNS及CFD-DEM耦合的DPM中。应用两种方法分别模拟了环流反应器的提升管中溶胀PE颗粒—烷烃的流动过程，验证了该粘附力模型的合理性。最后，应用含粘附力DPM模拟了溶胀PE颗粒—烷烃的环流反应器实验装置，并分析了其中的流动状态及提升管内溶胀颗粒的聚团特性。本论文根据PE颗粒—烷烃系统的特点在三个尺度上提出了精度更高的物理模型，并基于CPU+MIC+GPU异构并行计算优化了对应的三种模拟方法，，从而构建基于颗粒解析的液固两相流计算方法。;Particle-fluid two-phase flows are typical complex multi-phase systems, among which gas-solid systems present remarkable multi-scale dynamic heterogeneous structures mainly due to hydrodynamic instability, and has already attracted great attention. However, liquid-solid systems generally present more homogeneous structures, and their complexity is more manifested as the formation of agglomerates when strong interaction exists between particles, but the relevant research is still lacking. As in the polyethylene (PE)-alkane liquid-solid system, the complexities of different scales are manifested as: the complex surface shapes of the swollen PE particles at micro scale; the process of growth, agglomeration and fragmentation for particle agglomerates at meso scale; and the heterogeneous effects on the overall flow at macro scales. Therefore, it is necessary to develop different particle-resolved simulation methods at different scales for more accurate and rapid numerical simulation, establishing a comprehensive simulation platform of liquid-solid systems.There particle-resolved simulation methods at three different scales - discrete element method (DEM), direct numerical simulation (DNS) and discrete particle model (DPM) are picked to analyze the PE-alkane liquid-solid system, and the parallel techniques based on three high performance hardware – central processing unit (CPU), many integrated core (MIC) and graphic processing unit (GPU) are used to optimize above methods. The works on these methods are including:1. DEM at particle scale. The Non-Uniform Rational Basis-Splines (NURBS) is chosen to describe the complex shapes of the PE particles, and then fast contact detection and interaction computation algorithms are developed for two NURBS curves/surfaces based on classical contact force models. The algorithms have been implemented with multi-thread parallelization on CPUs and MICs. The established NURBS-based DEM thus has higher accuracy and efficiency, which is validated and demonstrated by a quantitative analysis of the trajectories after collision between a pair of particles, and its feasibility is also illustrated for more complex questions.2. DNS at cluster scale. The DNS of the PE-alkane system in the riser of a loop reactor is carried out by coupling the lattice Boltzmann method (LBM) with traditional DEM. A novel two-region periodic boundary condition (TPBC) is proposed resampling the natural flow driven mode in the riser, and its feasibility is validated in the simulation of a gas-solid system with remarkable particle clustering then. DNS of the riser of the loop reactor is carried out, revealing spatio-temporal distribution of its flow velocity and solids concentration. This DNS is implemented in the CPU+MIC+GPU multi-scale heterogeneous parallel computing mode with good performance.3. DPM at reactor scale. The multi-scale parallel DPM with the proposed cohesive force model is implemented to simulate a lab loop reactor. A cohesive force model reflects the sintering of swollen PE particles contacting in alkane is established, and implemented in LBM-DEM coupled DNS and CFD-DEM coupled DPM. The DNS and DPM are both used to simulate the swollen PE particle-alkane flow in the riser of a loop reactor, which validates the rationality of the cohesive force model. Then, the labs of the swollen PE particles in the loop reactor are simulated in the established DPM, and the flow structure and the agglomeration of the swollen PE particles are investigated.For these methods, the models with higher accuracy are proposed based on the PE-alkane system, and optimized by CPU+MIC+GPU heterogeneous parallel computing， thus building the particle-resolved simulation of complex liquid-solid flows.