Algorithmic Approach

Our Algorithmic Approach is based on finite difference discretizations of PDEs on locally structured grids, combined with techniques for introducing adaptivity and irregular geometry, plus methods for representing kinetic problems using particles. Our approach corresponds to a line of investigation that has been the subject of research on the part of the PIs in this proposal for over a decade. It has been highly successful in a variety of applications areas, and represents a good match to the requirements of the applications described above.

High-resolution finite difference methods on structured grids

In grid-based methods for numerical PDEs, it is necessary to make a fundamental choice of discretization technologies. We have chosen to use locally structured grids, i.e. ones that are based on defining discrete unknowns on a rectangular discretization of the spatial independent variables. This is in contrast to unstructured grids, in which the discretized dependent variables are defined at nodes in space corresponding to a graph with more general connectivity. We will focus on finite difference methods on such grids, although much of the lower-level infrastructure could be used equally well for finite element discretizations. For the time-dependent problems described here, we believe that there are significant advantages to the structured grid approach (although we also note that unstructured finite element methods remain the methods of choice for problems such as computing eigenmodes in complex geometries). In the applications described here, structured grid methods have been used extensively, so there is a body of experience in those communities with numerical methods based on that approach. It is well-understood how to construct stable and arbitrarily high order accurate discretizations to PDEs on such grids. Regularity in space leads to regular data layouts, making it easier to optimize for various types of data locality. The regular geometric structure of the spatial distribution of unknowns leads to fast efficient iterative solvers for elliptic and parabolic problems based on multigrid iteration. Adaptive gridding based on local refinement of structured grids, discussed below, is actually a more mature technology for time-dependent problems in which one must regrid often: comparably efficient grid generation for unstructured grid methods is still an open research question. Coupling to particle methods is inexpensive and well-understood. Locating the particles on the grid is trivial, and interpolation between particles and fields has an extensive body of practical experience and mathematical analysis.

Volume-of-Fluid Methods for Irregular Boundaries

There are two closely-related requirements for representing irregular boundaries. To represent engineering geometries, we need to compute solutions in domains whose whose shape (and trajectory, if they are moving) are known. In addition, in the gas jet problem for plasma-laser accelerators, and in detailed modeling of the initial breakup of a fuel jet for combustion, we need to be able to represent free boundaries, for which PDE's are being solved on both sides of a boundary whose location and shape is changing as a function of time and is computed as part of of the solution. In both cases, we will use a volume-of-fluid representation of the discretized solution near the boundary. In this approach, the surface is represented by its intersection with an underlying rectangular grid. This leads to a natural, finite-volume discretization of the PDE on irregular control volumes adjacent to the boundary. In the case where the surface is a domain boundary, this approach is variously referred to as a Cartesian grid, embedded boundary, or cut cell method (in this proposal, we will use the term embedded boundary). One of the principal advantages of the embedded boundary method is that the problem of generating the description of the geometry on the grid starting from surface tesselations produced, for example, by a CAD system has been completely solved. In addition, there has been considerable progress on the development of discretizations for this type of geometry representation. In this approach, the primary dependent variables are centered at the centers of the rectangular grid cells, even if those cells correspond to irregular control volumes. Because the values are centered at regular points on the grid, it is routine to construct finite difference approximations to various partial differential operators on the irregular control volumes that are consistent, conservative, and stable/well-conditioned.

In the case in which the PDE is being solved on both sides of a free boundary, the method is referred to as a volume-of-fluid method. We use a similar approach for the case of free boundaries. Solutions are double-valued in each cell, with both values centered at the Cartesian cell centers. The construction of difference approximations to various operators follows along the lines of that in the embedded boundary case, except that the appropriate well-posed boundary conditions are applied along the free boundary. In using these ideas to represent two-phase flow at low Mach numbers, we have a variety of options, from tracking only the density jump and representing all of the other dependent variables as single-valued, to making all quantities double-valued, as was done for tracking hyperbolic fronts.

Adaptive Mesh Refinement

Our adaptive mesh methods are based on the block-structured adaptive mesh refinement (AMR) algorithms of Berger and Oliger. In this approach, the regions to be refined are organized into rectangular patches of several hundred to several thousand grid cells per patch. Thus one is able to use the high-resolution rectangular grid methods described above to advance the solution in time. Furthermore, the overhead in managing the irregular data is amortized over a relatively large amounts or floating point work on regular arrays. For time-dependent problems, refinement is performed in time as well as in space. Each level of spatial refinement uses its own stable time step, with the time steps at a level constrained to be integer multiples of the time steps at all finer levels. AMR is a mature technology for problems without geometry, with a variety of implementations and applications for various nonlinear combinations of elliptic, parabolic and hyperbolic PDE's. In particular, the collection of applications here address most of the major algorithmic issues in developing adaptive algorithms for the applications described above, such as adaptive multigrid solvers for Poisson's equation, the representation of non-ideal effects in MHD, and the coupling of particle methods to AMR field solvers. AMR has also been successfully coupled to the volume-of-fluid algorithms described above.

Particle Methods for Kinetic Problems

To solve the kinetic equations as they arise in plasma physics, particle-beam modeling, and spray modeling, particle-in-cell methods are most often used. In this approach, the distribution function of the particles is represented by a collection of macro-particles, while the spatial field variables are represented as functions on a discrete grid in physical space, with the equations for those fields discretized using finite differences or finite elements. The coupling between the particles and the fields is implemented using interpolation. Quantities such as charges and currents that are required as sources in the field equations are interpolated onto the grid from the particle representation, and field variables required for the particle dynamics are interpolated to the particle locations from the grid. Within this approach, there are a range of design variations, based on tradeoffs between accuracy and computational cost. In an electrostatic PIC code, one can interpolate charge onto the grid, solve Poisson's equation for the electric field potential, and interpolate a difference approximation to the electric field onto the particles. One could also interpolate the dipole field onto the grid, solve three Poisson's equations for the electric field components, and interpolate the field directly from the grid. The latter is more accurate, at the cost of increasing the number of Poisson solutions.

In addition to pursuing the classical PIC approaches to particle methods for kinetics problems, we will also provide support for some of the methods that have been developed in the mathematics community. There are a variety of fast and accurate algorithms for evaluating electrostatic potentials that may be of use in this setting, such as the fast multipole method (FMM) and the method of local corrections (MLC). For both of these methods, the nonlocal component of the potential induced by a collection of N particles can be computed in no more than O(N log N) operations in a way that essentially eliminates the self-force problem and decouples the effect of the short-range smoothing of the potential from the grid spacing (in fact, FMM is a grid-free method). In addition, the MLC has been applied to the problem of interpolating forces onto the grid for general particle / mesh methods, and in particular can be applied to the case of Vlasov-Maxwell, as well as to the particle methods arising in the simulation of sprays in combustion.