AF: Small: Fully Implicit Methods for Partial Differential Equations and Software for Hybrid Architecture

该项目的第一个目标是开发新的区域分解算法和软件，用于多物理应用中产生的一些高度非线性、耦合的偏微分方程组的数值解。对于简单的问题，隐式方法相对容易从给定的显式或半隐式方法发展而来，但对于一些多物理问题，要开发出一种允许高性能实现的完全隐式方法是相当困难的。现有的大多数技术都是非光滑的，因此很难用牛顿型求解器来求解。在该项目中，将开发一些离散化技术，包括用于捕捉区域信息的高阶非光滑离散化和用于建立牛顿迭代的雅可比系统的低阶局部光滑离散化。低阶离散化作为一种非线性预条件，加快了收敛速度，但不改变解的精度。该项目的第二个目标是在具有大量处理器的高性能计算机上开发所提出的线性和非线性预处理方法的有效实施。包括算法和软件在内的相对成熟的技术可以用来解决许多类型的单一物理问题，但对于耦合的多物理问题，迫切需要健壮性和可扩展性的技术，特别是对于具有加速器的大规模并行计算机。所提出的算法和软件将对全球大气流动和生物流体模拟等几个重要的应用领域产生重大影响，也将对计算科学中需要求解大型线性和非线性方程的其他领域产生重大影响。为了扩大研究的影响，该软件将与广泛使用的PETSc包完全兼容。对于对高性能计算和通用计算科学与工程感兴趣的研究生和本科生来说，这项研究是一个充满机会的领域。

The first goal of the project is to develop new domain decomposition algorithms and software for the numerical solution of some highly nonlinear, coupled systems of partial differential equations arising from multi-physics applications. For simple problems, implicit methods are relatively easy to develop from a given explicit or semi-implicit method, but for some multi-physics problems it is quite difficult to develop a fully implicit method that allows a high performance implementation. Most of the existing techniques are non-smooth and therefore difficult to solve with Newton type solver. In the project, some discretization techniques will be developed that involve high order non-smooth discretization for capturing the domain information, and low order locally smooth discretization for building the Jacobian system of the Newton iterations. The low order discretization serves as a nonlinear preconditioner that speeds up the convergence, but doesn't change the accuracy of the solution. The second goal of the project is to develop an efficient implementation of the proposed linear and nonlinear preconditioning approaches on high performance computers with a large number of processors. Relatively mature technologies including algorithms and software are available for solving many types of single physics problems, but for coupled multi-physics problems, robust and scalable techniques are badly needed, especially for large scale parallel computers with accelerators. The proposed algorithms and software will have a great impact on several important application areas, such as the simulation of global atmospheric flows and the bio-fluids, and will also have substantial influence on other areas of computational sciences where large linear and nonlinear equations need to be solved. To broaden the impact of the research, the software will be made fully compatible with the widely used PETSc package. The research is a rich area in opportunities for both graduate and undergraduate students interested in high performance computing and general Computational Science and Engineering.