A simple embedding method for solving partial differential equations on surfaces
Introduction
Many applications in the natural and applied sciences require the solutions of partial differential equations (PDEs) on surfaces or more general manifolds. Examples of such application areas arise in biological systems, image processing, medical imaging, mathematical physics, fluid dynamics and computer graphics. For example, applications in image processing include the generation of textures [27], [28], the visualization of vector fields [7] and weathering [8], while applications in fluid dynamics include flows and solidification on surfaces [17], [18], and the problem of evolving surfactants on interfaces [29].
A popular approach to solving PDEs on surfaces is to impose a smooth coordinate system or parameterization on the surface, express differential operators within these coordinates, and then discretize the resulting equations. See, for example [10] for a tutorial and survey of methods for parameterizing surfaces. However, the required coordinates can be complicated or impractical to construct, and the coordinates may introduce artificial singularities, such as at the poles in spherical coordinates. Indeed, as pointed out in [10], “parameterizations almost always introduce distortion in either angles or some region”. Also, equations that are simple when written using intrinsic derivatives, such as surface diffusion, become substantially more complex when written in a coordinate system, involving nonconstant coefficients and more derivative terms.
Another common approach to solving partial differential equations on surfaces is to solve the PDE directly on a triangulation of the surface. This approach can be effective for certain classes of equations, however, as a general technique it leads to a number of difficulties. These are discussed in [3], [4]. In particular, this approach leads to nontrivial discretization procedures for the differential operators, as well as difficulties in accurately computing geometric primitives, such as surface normals and curvature [3]. In addition, convergence of numerical schemes on triangulated surfaces remains less well understood than methods on Cartesian grids [11].
An alternative approach to treating PDEs on surfaces is to embed the surface differential equations of interest within differential equations posed on all of , so that the solution of the embedding equations, when restricted to the surface, provides the solution to the original problem. With such an approach, the ultimate goal is to develop a method that allows the treatment of general, complex surface geometry, while still retaining the simplicity that comes from working in standard Cartesian coordinates. With this in mind, Bertalmío et al. [3] introduced an embedding method for solving variational problems and the resulting Euler–Lagrange evolution PDEs on surfaces. In their approach, the underlying surface is represented as the level-set of a higher dimensional function and the evolution corresponding to the surface PDEs is carried out via PDEs that are posed on all of . This leads to equations that can be discretized and solved using Cartesian grid methods. A further improvement to this method was proposed by Greer [11]. In his approach, the evolution equation is modified away from the surface to maintain greater regularity of the solution near the surface of interest. See also [1] for related work on the finite element approximation of elliptic partial differential equations on implicit surfaces via level-set methods.
Embedding methods based on level-set methods have a number of limitations, however. Most obviously, these methods do not immediately allow for open surfaces with boundaries, or filamentary objects of codimension-two or higher, although it is in principle possible to represent such objects by introducing additional level-set functions. Another limitation is that these methods result in an embedding PDE posed on all of space, and complications arise when they are solved in a restricted computational band around the surface. Such “narrow banding” requires the imposition of appropriate boundary conditions, and how to best impose these conditions is not generally understood. For example, even using the regularity improvements proposed by Greer [11], a degradation of the order of convergence is observed when banding is used in diffusive problems. At a more technical level, level-set based methods also either lead to degenerate diffusion equations or require the use of an additional diffusion step when applied to parabolic equations [11].
Similar to other embedding methods, the approach we present here discretizes the partial differential equations using a fixed Cartesian grid in the embedding space. However, our method is based on the use of a closest point representation of the surface (cf. [16]) rather than a level-set representation. In conjunction with this change in representation, we abandon the concept of solving an embedding PDE for all time, and instead use the embedding PDE to advance the solution near the surface for one time step (or one stage of a Runge–Kutta method). This leads to a new method for solving PDEs on surfaces which has great simplicity, as well as additional desirable features. Most importantly, the embedding PDE is the obvious analog of the surface PDE, and involves only the standard Cartesian differential operators. In addition, the method can treat open surfaces and is not limited to objects of codimension-one. It also naturally allows the computation to be done on a grid defined in a narrow band near the surface without any degradation in the order of accuracy and without imposing artificial boundary conditions. To distinguish our approach from other embedding procedures, and to emphasize its essential reliance on the closest point representation, we refer to the method as the closest point method. As an aside, note that a special case of this procedure was introduced in our recent paper on the diffusion-generated motion of curves on surfaces [16]. It was in this context that the general potential of the approach was first realized, although it occurs only as a special, simple procedure for approximating in-surface curvature motion according to the diffusion-generated motion algorithm.
The paper unfolds as follows. Section 2 gives the method, describes its implementation and provides an analysis. In Section 3, we give a number of two-dimensional convergence studies to validate the method. Section 4 considers a variety of three-dimensional convergence tests and provides some examples that are relevant to biology: a Fitzhugh–Nagumo and a morphogenesis simulation. In Section 5, we give a summary and discuss some of our ongoing projects in the subject. Finally, Appendix A concludes with a discussion on the computation of geometric quantities defined on the surface.
Section snippets
The closest point method
In this section, we describe the method as well as its motivation, analysis and implementation. We begin by discussing the motivation for the method and its surface representation.
Numerical experiments in 2D
We now provide some studies of numerical convergence in two dimensions. In all of our examples analytical solutions are determined from the corresponding one-dimensional periodic systems.
For simplicity, second-order centered differences are used to carry out spatial discretizations and all time-stepping is carried out explicitly (using forward Euler or the third-order TVD Runge–Kutta method). All computations are carried out on a uniform grid defined on the relevant computational band. See
Numerical experiments in 3D
We now examine the numerical behavior of the method in three dimensions. Where analytical solutions exist, numerical convergence studies are carried out.
Similar to the previous section, second-order centered differences are used to carry out spatial discretizations and all time-stepping is carried out explicitly (using forward Euler or the third-order TVD Runge–Kutta method). All computations are carried out on a band around the surface.
Summary and future work
In this work, we present the closest point method, which is a new embedding method for solving partial differential equations on surfaces. The method is designed to make solving PDEs on surfaces as close as possible to the familiar process of solving PDEs in , in effect hiding all the geometric complexities. Central to the method is the choice of the closest point representation of the surface. This representation naturally gives an extension step which leads to embedding PDEs that are simply
References (29)
- et al.
Variational problems and partial differential equations on implicit surfaces
Journal of Computational Physics
(2001) Fitzhugh–Nagumo simplified cardiac action potential model
Biophys. J.
(1961)- et al.
Fourth order partial differential equations on general geometries
J. Comput. Phys.
(2006) - et al.
Efficient implementation of weighted ENO schemes
J. Comput. Phys.
(1996) - et al.
Diffusion generated motion of curves on surfaces
J. Comput. Phys.
(2007) - et al.
A mathematical model for atmospheric ice accretion and water flow on a cold surface
Int. J. Heat Mass Transf.
(2004) - et al.
A fixed grid method for capturing the motion of self-intersecting interfaces and related PDEs
J. Comput. Phys.
(2000) Simple chemical-reaction systems with limit-cycle behavior
J. Theor. Biol.
(1979)- et al.
Efficient implementation of essentially nonoscillatory shock-capturing schemes
J. Comput. Phys.
(1988) - et al.
A new Eulerian method for the computation of propagating short acoustic and electromagnetic pulses
J. Comput. Phys.
(2000)
Barycentric Lagrange interpolation
SIAM Review
Variational problems and partial differential equations on implicit surfaces: bye bye triangulated surfaces?
Numerical Analysis
Cited by (0)
- 1
The work of this author was partially supported by a grant from NSERC Canada.