Some applications of Deep Learning techniques to fluid dynamic solutions

Danesh K. Tafti

William S. Cross of Engineering

Dept. Mechanical Engineering

Virginia Tech, Blacksburg, USA

讲座时间：

2023年9月10日 上午10:00-11:00

讲座地点：

三号实验楼307

讲座内容简介：

        考虑到计算流体动力学 (CFD) 解决方案的复杂、昂贵及不确定性，人工智能 (AI) 和深度学习 (DL) 方法在计算成本及结果准确度上的优势让它在流体力学领域获得了越来越多的重视。该法可通过更好地参数化现有模型来实现，因而其常被应用在湍流建模、新模型开发或者通过加速传统求解算法或开发降阶代理模型来降低 CFD 成本等方面。然而，纳维-斯托克斯方程中包含的复杂非线性物理特征与高昂的生成训练数据的成本是这些方法推广过程中亟待解决的问题与挑战。该讲座将探讨深度学习在流场预测与偏向工程导向方面应用，内容包含两方面的案例研究：第一个案例研究了随机分布的柱状颗粒集合中随时间变化的混沌流场的未来状态预测；第二个案例研究了不同堆积密度和雷诺数下随机分布的长椭球颗粒集合中的稳定流场的预测。此外，研究还对通过预测流场计算出的颗粒受力等物理量的准确性进行了评估，证实了当前模型的准确性与可靠性。

Introduction:

Computational Fluid Dynamics (CFD) solutions are complex, expensive, and uncertain. Artificial Intelligence (AI) and Deep Learning (DL) methods have the potential to give accurate results at much less computational cost. This can be through better parameterization of existing models, e.g. turbulence modeling, or through the development of new models where none were possible, or through reducing the cost of CFD by accelerating conventional solution algorithms or by the development of reduced-order surrogate models. However the complex non-linear physics embedded in the Navier-Stokes equations and the cost of generating training data (data paucity) are some of the challenges that impede the generalizability of these methods. The seminar will explore the prediction of flow fields and downstream engineering tasks such as determining forces acting on embedded objects through the use of DL techniques. In the first case study, the future state prediction of a time-dependent chaotic flow field in a random array of cylinders is investigated. In the second case study the prediction of steady flow fields in different random assemblies of prolate ellipsoids under different packing densities and Reynolds numbers is investigated. In both case studies, the accuracy with which engineering quantities such as drag forces can be found using the DL predicted flow fields is also evaluated.

主讲人简介：

        Danesh. K. Tafti 教授拥有三十余年计算流体力学相关的研究与工作经验，研究方向涵盖了大涡模拟及高性能并行运算算法开发、扑翼飞行的空气动力学分析、颗粒-流体两相流及泥沙输移高精度仿真等方向，共发表论文二百七十余篇，其中有学术期刊论文一百四十余篇，总引用数达7179次（来源：谷歌学术，2023年9月）。Tafti教授曾任《ASME J. Heat Transfer》副主编，目前为《International Journal of Heat and Fluid Flow》、《Journal of Applied and Computational Mechanics》及《International Journal of Rotating Machinery》期刊编委。Tafti 教授多年来一直保持着与美国国家能源实验室在流-固两相流方向的的深度合作，到目前为止承担来自美国国家科学基金会、美国能源部、国家超级计算应用中心等政府部门以及企业的项目经费共计21,900,295美元，其中个人承担7,613,767美元，近三年个人获得研究经费共计501,212美元。Tafti 教授执教期间共指导了28名博士（四名在读）、14名博士后以及28名硕士（2名在读），拥有丰富的研究生教学、指导经验。