CDS&E: AI-RHEO: Learning coarse-graining of complex fluids

斯托克斯型复杂流体流动描述了小尺度上的传输现象。例如生物细胞内的流动、毛细血管和微流控装置中的血液流动、DNA流体力学、芯片实验室工业和医疗设备、细菌流动和聚合物流动。通过数值模拟了解和预测斯托克斯型复杂流体的行为，对于理解机械-生物机理、微流控装置的设计、医疗机器人和许多其他应用是基本的。复杂流体流动的模拟具有挑战性，因为它们涉及固液相互作用、运动界面和复杂几何形状、空间和时间上的非局部和多尺度耦合以及高度非线性和通常是混沌的动力学。该项目的重点是开发能够显著提高复杂流体模拟技术的最先进水平的方法。该项目的具体目标包括以下几个方面。(A)设计高性能计算(HPC)算法，将降维和深度学习方法与积分方程式方法相结合，使斯托克斯式复杂流体流动的预测模拟速度加快数量级。(B)设计和部署高性能计算软件基础设施，以自动化配置采样、操作员拆分、深度网络培训和对一大类复杂流体的推理。(C)对所提出的方法在三个问题上的评价：微流控器件的有效性质的计算、参数估计和形状优化。在该项目中开发的复杂流体解算器将影响科学和工程中涉及移动界面和微观结构演化问题的广泛学科。此外，该项目旨在吸引一些研究生和本科生在高性能计算机、机器和深度学习、积分方程式和复杂流体的界面上学习。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Stokesian complex fluid flows describe transport phenomena at small scales. Examples include flows inside biological cells, blood flow in capillaries and microfluidic devices, DNA hydrodynamics, lab-on-a-chip industrial and medical devices, bacterial flows, and polymer flows. Understanding and predicting the behavior of Stokesian complex fluids using numerical simulations is fundamental in understanding mechano-biological mechanisms, design of microfluidic devices, medical robotics, and many other applications. Complex fluid flows are challenging to simulate because they involve solid-fluid interaction, moving interfaces and complex geometries, non-local and multiscale couplings in both space and time, and highly nonlinear and often chaotic dynamics. This project focuses on the development of methods that will significantly advance the state-of-the-art of complex fluid simulation technologies. The specific goals of this project include the following. (A) The design of high-performance computing (HPC) algorithms that integrate dimension reduction and deep learning methods with integral equation methods and result in orders-of-magnitude speedups of predictive simulations of Stokesian complex fluid flows. (B) The design and deployment of HPC software infrastructure that automates configuration sampling, operator splitting, deep network training, and inference for a large class of complex fluids. (C) Evaluation of the proposed methodology on three problems: calculation of effective properties, parameter estimation, and shape optimization of microfluidic devices. The complex fluid solvers developed in this project, will impact a broad spectrum of disciplines in sciences and engineering that involve problems with moving interfaces and microstructure evolution. Furthermore, the project aims to engage a number of graduate and undergraduate students at the interface of HPC, machine and deep learning, integral equations, and complex fluids.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.