球床式高温气冷堆内的球流运动是一种缓慢稠密循环颗粒流，其研究对于高温气冷堆技术的发展具有重要意义。单区堆与双区堆是球床式高温气冷堆的两种技术路线，均是由大量离散球组成的宏观复杂系统其堆积与流动特性直接影响堆芯经济性与安全性。本文以单/双区球流为对象，结合数值模拟与深度学习方法，基于离散单元法（DEM）研究不同颗粒性质与球床结构下的球流性能，重点分析堆积结构与流动特性，并通过卷积神经网络（CNN）对球流图像的时间特性与未来状态分布进行智能预测。首先，对于单区球流运动，建立三维循环球流数值模型，提出堆积结构与流动特性定量评价指标，研究颗粒性质的影响与球床结构优化问题。关于颗粒性质的影响，研究表明颗粒密度基本不影响堆积结构，但颗粒直径或大颗粒数量比增加时，将导致空隙率上下波动幅度增大，且波动传递的范围变广。另外，在两种密度混合流动下， 轻球将在壁面出现聚集现象，该现象随着密度差异增加而越为明显。几何优化方面，分析了不同结构（包括不同的底部结构设计与出口直径）下的流型、均匀性与滞留性，并对流动临界转变进行了定量分析，发现外凸弧形结构以及采用固定锥高增加开口大小的方式可有效改善堆芯流动。其次，针对双区球流运动，基于数量概率分析，提出了适用于三维双区球流的界面分析法，给出中心区边界和交混区宽度的定量确定方法，并从中心区边界、交混程度、卸球比三方面定量验证了三维双区堆芯的稳定性。此外，通过参数效应研究，重点探讨加载比和颗粒密度对双区球流的影响。研究表明，加载比对滞留性基本没有影响，但会影响双区形状与交混程度；增加环形区域燃料球密度对于减少滞留区和总滞留时间有积极作用。最后，设计并提出针对球流图像的智能预测算法，实现深度学习在球流研究中的结合应用。为预测真实场景下的球流剩余时间特性，自主设计RT-Net网络，将传统卷积核设计为可合并的多分支卷积组件，有效提高了准确度并降低计算时长，预测误差在0.05 s内的测试集比例达到96.7%；针对球流运动的时间演化过程，利用提示学习的思想，设计并提出图像生成网络 Pre-Net，在真实球流实验图像数据上实现了球流未来状态的智能预测，探索并开拓了深度学习技术在反应堆球流研究中的新研究路径。

The pebble flow of pebble bed High-Temperature Gas-cooled Reactor (HTGR) is a slow and dense circulating pebble flow, which is of great significance for the development of HTGR technology. There are two types of HTGR pebble beds, single-region and double-region designed pebble bed, both of which are macroscopic complex systems composed of a large number of discrete pebbles. Inside the pebble bed, the packing structure and flow characteristics directly affect the core economy and safety. In this paper, the single and double region pebble flows are both studied by combining numerical simulation and deep learning methods. The flow characteristics of pebble flows under different pebble properties and bed structures are investigated based on the Discrete Element Method (DEM), with the focus on the packing structure and flow characteristics. In addition, the intelligent prediction of pebble flow images is performed with the aid of Convolutional Neural Network (CNN).Firstly, for the single-region pebble flow, a three-dimensional numerical model for circulating pebble flow is established to study the influence of pebble properties and the optimization of bed structure, and quantitative evaluation indexes for the packing structure and flow characteristics are proposed. Regarding the influence of pebble properties, it is shown that the pebble density basically does not affect the packing, while with the increase of pebble diameter or the number ratio of large pebbles, the porosity fluctuates more drastically and the fluctuations can be transmitted more widely. Besides, the aggregation phenomenon of light pebbles near the wall region is observed for mixing pebble flow, and this phenomenon becomes more obvious with the increase of density difference. In terms of structure optimization, the flow pattern, uniformity, stagnation and quantitative analysis of flow transition under different structures (including different bottom structures and outlet diameters) are given. Results show that the convex arc-shaped structure and the increase of the opening size with the fixed cone height can significantly improve the flow performance.Secondly, regarding the double-region pebble flow, an interface analysis method applicable to the three-dimensional double-region pebble bed is proposed based on the quantitative probability analysis, and the quantitative determination methods of the central region boundary and the mixing region width are given. The stability of the three-dimensional double-zone pebble flow is quantitatively verified from the perspectives of the central region boundary, mixing degree, and unloading ratio. Besides, a parametric effect study is conducted, focusing on the effects of loading ratio and pebble density on the flow characteristics. It is found that the loading ratio has basically no effect on stagnation, but affects the shape of central region boundary and the degree of mixing; increasing the fuel pebble density in the annular region has a positive effect on reducing the stagnation region and the total retention time.Finally, intelligent prediction algorithms for pebble flow images are proposed to realize the combined application of deep learning in pebble flow research. To predict the remaining time characteristics from pebble flow images in real scenes, RT-Net is designed based on the multi-branch convolutional component, which effectively improves the accuracy and reduces the calculation time. The proportion of test sets with prediction errors within 0.05 s reaches 96.7%. For the time evolution process of pebble flow, using the idea of guided learning, an image generation network Pre-Net is proposed, which can predict the future distribution of pebble flow on the real experimental image data, providing a new approach for the deep-learning-based research of pebble flow.