CDS&E: Collaborative Research: Deep learning enhanced parallel computations of fluid flow around moving boundaries on binarized octrees

CDS

• 批准号: 1953222
• 负责人: Amir Barati Farimani
• 金额: $ 24.5万
• 依托单位: Carnegie-Mellon University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-08-15 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1953222&HistoricalAwards=false
• 关键词: CDS& Collaborative; Research; Deep; learning

Computer simulations of heat and fluid flow find applications in many aspects of science and engineering. Notable examples are aerodynamic design of aircrafts and automobiles, and weather forecasting. These simulations are often computationally expensive, and they are performed on supercomputers. Special methods are used to implement the equations of heat and fluid flow as a simulation software. The end goal is to create an accurate computer code that can make optimal use of available computing power. However, this end goal is becoming challenging on modern extreme-scale supercomputers that deploy a large of number of computing processors to work in parallel. Existing algorithms face performance bottlenecks and do not realize the full potential of a modern supercomputer. The project team will develop new algorithms to overcome this performance bottleneck. The successful completion of this award is expected to result in an open-source heat and fluid flow simulation software. The project team will develop educational tutorials to pique the interest of high-school students in new capabilities of computer simulation and machine learning techniques in science and engineering. The technical objective is to enhance parallel performance of simulations of incompressible fluid flow around moving boundaries. A recently developed binarized octree generation technique will be further developed as an open-source parallel adaptive mesh refinement software infrastructure to solve the fluid flow equations on Cartesian domains with deep levels of mesh adaptations. Machine learning techniques and deep neural nets will be adopted in ways to ease potential bottlenecks that are expected to degrade scalability of parallel computations when large number of processors are deployed in simulations. The project team will develop multiple deep learning algorithms such as convolutional neural networks and generative adversarial networks to learn the fluid flow around complex geometries and apply the learning for rapid and accurate field estimation at arbitrary points. To successfully incorporate the effect of boundary conditions at the interface, conditional generative adversarial networks will be trained on different coarse and fine grids to learn the communication pattern among the blocks. This award by the Division of Chemical, Bioengineering, Environmental and Transport Systems within the NSF Directorate of Engineering is jointly supported by the NSF Office of Advanced Cyberinfrastructure.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

热和流体流的计算机模拟在科学和工程的许多方面都可以找到应用。著名的例子是飞机和汽车的空气动力学设计以及天气预报。这些模拟通常在计算上很昂贵，并且在超级计算机上进行。特殊方法用于实现热和流体流程作为仿真软件的方程。最终目标是创建一个准确的计算机代码，以最佳使用可用的计算能力。但是，这个最终目标是在现代极限规模的超级计算机上变得具有挑战性，这些超级超级计算机部署了大量的计算处理器以并行工作。现有的算法面临性能瓶颈，并且没有意识到现代超级计算机的全部潜力。项目团队将开发新的算法来克服这种性能瓶颈。预计该奖项的成功完成将导致开源热和流体流量模拟软件。项目团队将开发教育教程，以激发高中生在科学和工程学中的计算机模拟和机器学习技术的新功能中的兴趣。技术目标是增强动态界限周围不可压缩流体流的模拟平行性能。最近开发的二进制OCTREE生成技术将进一步开发，作为一种开源平行的自适应网格改进软件基础架构，以解决具有深层网格适应性的笛卡尔域上的流体流动方程。机器学习技术和深神网将以缓解潜在瓶颈的方式采用，这些瓶颈有望在模拟中部署大量处理器时降低并行计算的可扩展性。项目团队将开发多种深度学习算法，例如卷积神经网络和生成的对抗网络，以了解复杂几何形状周围的流体流，并将学习应用于任意点的快速，准确的现场估计。为了成功地纳入界面处的边界条件的效果，有条件的生成对抗网络将在不同的粗网格上训练，以了解块之间的通信模式。 NSF工程局内化学，生物工程，环境和运输系统的该奖项由NSF高级网络基础设施办公室共同支持。该奖项反映了NSF的法定任务，并被认为是值得通过基金会的知识分子和更广泛影响的评估来通过评估来支持的，这是值得的。

项目成果

Deep learning for reduced order modelling and efficient temporal evolution of fluid simulations

• DOI: 10.1063/5.0062546
• 发表时间: 2021-10-01
• 期刊: PHYSICS OF FLUIDS
• 影响因子: 4.6
• 作者: Pant, Pranshu;Doshi, Ruchit;Barati Farimani, Amir
• 通讯作者: Barati Farimani, Amir

Adaptive grey wolf optimizer

• DOI: 10.1007/s00521-021-06885-9
• 发表时间: 2022-01-11
• 期刊: NEURAL COMPUTING & APPLICATIONS
• 影响因子: 6
• 作者: Meidani, Kazem;Hemmasian, AmirPouya;Farimani, Amir Barati
• 通讯作者: Farimani, Amir Barati

Mesh deep Q network: A deep reinforcement learning framework for improving meshes in computational fluid dynamics

网格深度 Q 网络：用于改进计算流体动力学中的网格的深度强化学习框架

• DOI: 10.1063/5.0138039
• 发表时间: 2023
• 期刊: AIP Advances
• 影响因子: 1.6
• 作者: Lorsung, Cooper;Barati Farimani, Amir
• 通讯作者: Barati Farimani, Amir

A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction

• DOI: 10.1016/j.jcp.2023.111972
• 发表时间: 2022-11
• 期刊: J. Comput. Phys.
• 作者: Dule Shu;Zijie Li;A. Farimani

Reduced-order modeling of fluid flows with transformers

• DOI: 10.1063/5.0151515
• 发表时间: 2023-05
• 期刊: Physics of Fluids
• 影响因子: 4.6
• 作者: AmirPouya Hemmasian;Amir Barati Farimani