标题： 基于GPU的非均匀网格可压缩格子玻尔兹曼模拟，使用标准C++并行性：从最佳实践到空气动力学、气动声学和超音速流模拟 显示英文标题

标题： GPU-based compressible lattice Boltzmann simulations on non-uniform grids using standard C++ parallelism: From best practices to aerodynamics, aeroacoustics and supersonic flow simulations

摘要： 尽管经过数十年的研究，在非均匀网格上创建准确、鲁棒且高效的格子玻尔兹曼方法（LBM）并实现无缝GPU加速仍然是一个挑战。本工作引入了一种新策略来解决这一挑战，通过整合简单但有效的组件：(1) 现代C++中的并行算法，(2) 保守的单元中心网格细化，(3) 局部边界条件，以及(4) 鲁棒的碰撞模型。我们的框架支持多种晶格（D2Q9、D2Q13、D2Q21、D2Q37），针对不同的流动条件进行了优化。它包括具有多项式和数值平衡的碰撞模型，一种用于多原子行为的第二分布函数，一种类似Jameson的激波传感器，并推广了Rohde的细化策略。该框架的准确性与鲁棒性在多种基准测试中得到了验证，包括驱动腔流、风成噪声、30P30N机翼空气动力学、无粘性黎曼问题以及跨音速和超音速状态下NACA机翼的粘性流动。现代C++进一步使我们的框架达到GPU原生性能，同时确保高可移植性、模块化和易于实现。值得注意的是，弱可压缩LBMs在非均匀网格上实现了最先进的GPU效率，而完全可压缩LBMs在最计算密集的情况下受益于相当于数千个CPU核心的加速效果。我们的先进性能模型包含了邻居列表和异步时间步进效应，为非均匀网格上LB模拟的性能分解提供了新的见解。总体而言，这项研究为可移植的、基于树的LBMs设定了新标准，证明了精心选择的组件组合可以在各种流动条件下实现高性能、高精度和鲁棒性。作为最终的概念验证，提出了适用于亚音速和超音速应用的自适应网格细化。 显示英文摘要

摘要： Despite decades of research, creating accurate, robust, and efficient lattice Boltzmann methods (LBM) on non-uniform grids with seamless GPU acceleration remains challenging. This work introduces a novel strategy to address this challenge by integrating simple yet effective components: (1) parallel algorithms in modern C++, (2) conservative cell-centered grid refinement, (3) local boundary conditions, and (4) robust collision models. Our framework supports multiple lattices (D2Q9, D2Q13, D2Q21, D2Q37) tailored to various flow conditions. It includes collision models with polynomial and numerical equilibria, a second distribution for polyatomic behavior, a Jameson-like shock sensor, and generalizes Rohde's refinement strategy. The framework's accuracy and robustness is validated across diverse benchmarks, including lid-driven cavity flows, Aeolian noise, 30P30N airfoil aerodynamics, inviscid Riemann problems, and viscous flows past a NACA airfoil in transonic and supersonic regimes. Modern C++ further enables our framework to reach GPU-native performance, while ensuring high portability, modularity, and ease of implementation. Notably, weakly compressible LBMs achieve state-of-the-art GPU efficiency on non-uniform grids, while fully compressible LBMs benefit from acceleration equivalent to thousands of CPU cores in the most compute-intensive cases. Our advanced performance models incorporate neighbor-list and asynchronous time-stepping effects, providing new insights into the performance decomposition of LB simulations on non-uniform grids. Overall, this study sets a new standard for portable, tree-based LBMs, demonstrating that a combination of well-chosen components can achieve high performance, accuracy, and robustness across various flow conditions. As a final proof-of-concept, adaptive mesh refinement is proposed for subsonic and supersonic applications.