自适应快速模拟细节丰富的流体技术研究

Physically based fluid simulation has a broad application prospect in digital entertainment and virtual reality, thus are of important research significance. However, still there are many challenging problems in realistic modeling of complex fluid scenes, especially for real-time applications. The research objectives of this project mainly revolves around three aspects: Put forward a new simulation architecture to overcome the shortage of the current models; Optimize the proposed architecture and use hardware acceleration for speed up; Synthesize fluid details to enhance scene realism. The specific research plans include: To propose a hybrid simulation architecture, which as an effective combination of euler grid method and meshless SPH method can play the advantages of both; Research using adaptive technology to efficiently optimize the hybrid architecture, with low computation cost to achieve high computation speed; To propose a turbulence model that can be superimposed onto the hybrid model, to increase fluid details; Studying on porting the whole algorithm onto GPU to greatly increase speed. These technologies will be combined to quickly simulate large-scale complex fluid dynamic scenes with rich details.

基于物理的流体动画技术在数字娱乐、虚拟现实等领域有着广泛的应用前景，具有重要研究价值。对复杂流体场景进行实时的真实感模拟与绘制, 仍然存在着不少有挑战性的问题。本项目的研究目标主要围绕三个方面展开：提出新的模拟架构，克服当前方法的不足；优化算法性能和采用硬件加速，提升模拟速度；合成流体细节，增强场景真实感。具体的研究内容包括：研究提出一种混合的模拟架构，有效结合欧拉网格方法和无网格SPH方法，发挥两种方法各自的优势；研究利用自适应技术来高效优化混合模型的性能，以低的计算代价获取高的计算速度；研究提出一种湍流模型叠加到混合模型中，合成丰富的流体细节，增强流体真实感；研究采用硬件加速技术将整个算法移植到GPU上执行，大幅提升计算速度。本项目将综合利用这些技术以快速模拟规模较大的具有丰富细节的复杂流体动态场景。

实时模拟高度真实感的复杂流体现象是计算机图形学领域的难点与热点研究问题之一。本项目重点研究以较低的计算代价快速模拟规模较大的细节丰富的真实感流体效果：（一）提出一种涡粒子与网格结合的混合模拟架构，捕捉到了丰富的流体涡旋细节，并使用GPU对其最耗时的步骤加速，大大提升了模拟速度；（二）提出一种自适应光滑长度的SPH流体模拟方法，改进插值精度获得了更真实的模拟效果；（三）利用八叉树提出一种基于MultiFlip的两相流模拟算法，自适应跟踪精确的气液边界，模拟了气泡和液体的复杂交互作用；（四）提出了三种生成/捕捉高频流体细节的方法：基于提升小波亥姆霍兹-霍奇分解(LWHHD)的湍流细节合成、基于对流时角动能损失的流体细节合成、一种高精度对流插值算法，增强了流体的真实感。在该项目的资助下，共发表论文14篇、申请发明专利5项、获得CCF推荐国际会议最佳论文奖1项，另申请到项目资助2项。本项目的研究成果在广告媒体、影视动画、电脑游戏以及虚拟现实等诸多领域中具有重要的参考与应用价值。