A&A 560, A35 (2013) Astronomy

DOI: 10.1051/0004-6361/201322281 &

c ESO 2013 Astrophysics

Using 3D Voronoi grids in radiative transfer simulations

P. Camps, M. Baes, and W. Saftly

Sterrenkundig Observatorium, Universiteit Gent, Krijgslaan 281, 9000 Gent, Belgium

e-mail: [email protected]

Received 15 July 2013 / Accepted 5 October 2013

ABSTRACT

Context. Probing the structure of complex astrophysical objects requires effective three-dimensional (3D) numerical simulation of the
relevant radiative transfer (RT) processes. As with any numerical simulation code, the choice of an appropriate discretization is crucial.
Adaptive grids with cuboidal cells such as octrees have proven very popular; however, several recently introduced hydrodynamical
and RT codes are based on a Voronoi tessellation of the spatial domain. An unstructured grid of this nature poses new challenges in
laying down the rays (straight paths) needed in RT codes.
Aims. We show that it is straightforward to implement accurate and efficient RT on 3D Voronoi grids.
Methods. We present a method for computing straight paths between two arbitrary points through a 3D Voronoi grid in the context
of a RT code. We implement this grid in our RT code SKIRT, using the open source library Voro++ to obtain the relevant properties
of the Voronoi grid cells based solely on the generating points. We compare the results obtained through the Voronoi grid with those
generated by an octree grid for two synthetic models, and we perform the well-known Pascucci RT benchmark using the Voronoi grid.
Results. The presented algorithm produces correct results for our test models. Shooting photon packages through the geometrically
much more complex 3D Voronoi grid is only about three times slower than the equivalent process in an octree grid with the same
number of cells, while in fact the total number of Voronoi grid cells may be lower for an equally good representation of the density
field.
Conclusions. The benefits of using a Voronoi grid in RT simulation codes will often outweigh the somewhat slower performance.
Key words. hydrodynamics – radiative transfer – methods: numerical

1. Introduction to their intrinsic 3D nature, these techniques allow the inclu-

sion of, for example, a clumpy medium (Witt & Gordon 1996,
The radiation observed from most astrophysical systems has 2000; Bianchi et al. 2000; Doty et al. 2005; Stalevski et al. 2012;
been substantially affected by the gas and/or dust residing in Schechtman-Rook et al. 2012), polarization (Code & Whitney
or in front of the system under study. Many of these systems 1995; Goosmann & Gaskell 2007), or kinematical informa-
have a complicated three-dimensional (3D) geometry, for ex- tion (Matthews & Wood 2001; Baes & Dejonghe 2002; Baes
ample arm structures in spiral galaxies (Misiriotis et al. 2000; et al. 2003). Recent RT applications explicitly using 3D include
Fritz et al. 2012), filaments and clumps in star-forming regions models of young stellar objects (Wolf et al. 1998), protostellar
(Goldsmith et al. 2008; Paron et al. 2013; Fallscheer et al. 2013), to protoplanetary disks (Indebetouw et al. 2006; Niccolini &
and bow-shocks around evolved stars (Decin et al. 2012; Wang Alcolea 2006), reflection nebulae (Witt & Gordon 1996), molec-
et al. 2013). Properly probing the structure of these systems re- ular clouds (Pelkonen et al. 2009; Steinacker et al. 2005), spiral
quires a 3D numerical treatment of the relevant radiative trans- galaxies (Bianchi 2008; Schechtman-Rook et al. 2012; De Looze
fer (RT) processes (Disney et al. 1989; Witt et al. 1992; Baes et al. 2012a), interacting and starburst galaxies (Chakrabarti
& Dejonghe 2001), which may include dust, line, ionizing UV, et al. 2007; Hayward et al. 2011), and active galactic nuclei
Lyα, neutron, and neutrino RT. (Schartmann et al. 2008; Stalevski et al. 2012).
Because of the complexities involved in multiple anisotropic For the purpose of numerical computation, the domain un-
scattering, absorption, and reemission, the full RT problem der study must be discretized. Since memory requirements and
is highly nonlocal and nonlinear. It is not feasible to di- computation time rapidly increase with the number of grid cells,
rectly integrate the equations in three dimensions except modern 3D RT codes employ an adaptive grid, placing more and
for the simplest problems. Consequently virtually all mod- smaller cells in areas that require a higher resolution. Starting
ern RT codes use ray tracing or Monte Carlo techniques from a cuboidal root cell that spans the complete spatial do-
(Gordon et al. 2001; Ciardi et al. 2001; Kurosawa & Hillier main, an adaptive mesh refinement (AMR) scheme recursively
2001; Juvela & Padoan 2003; Wolf 2003; Stamatellos & subdivides each cell into k × l × m cuboidal subcells until the
Whitworth 2003; Bethell et al. 2004; Harries et al. 2004; required resolution is reached. In the special case of an octree,
Wood et al. 2004; Doty et al. 2005; Ercolano et al. 2005; k = l = m = 2 so that each cell is subdivided into eight sub-
Niccolini & Alcolea 2006; Jonsson 2006; Pinte et al. 2006; cells (hence the name of the data structure). Adaptive-mesh grids
Verhamme et al. 2006; Tasitsiomi 2006; Steinacker et al. 2006; and especially octree grids are well established (Kurosawa &
Bianchi 2008; Laursen et al. 2009; Baes et al. 2011; Robitaille Hillier 2001; Steinacker et al. 2002; Wolf 2003; Harries et al.
2011; Heymann & Siebenmorgen 2012; Abdikamalov et al. 2004; Niccolini & Alcolea 2006; Jonsson 2006; Bianchi 2008;
2012; Lunttila & Juvela 2012; Steinacker et al. 2013). In addition Laursen et al. 2009; Robitaille 2011; Lunttila & Juvela 2012;

Article published by EDP Sciences A35, page 1 of 9

 A&A 560, A35 (2013)

this case, both memory usage and run time scale roughly linearly
with the number of cells.
Furthermore, RT simulations frequently serve to pre-
dict the observable properties of artificial systems resulting
from (magneto-)hydrodynamical (MHD) simulations (Juvela &
Padoan 2003; Bethell et al. 2004; Stamatellos & Whitworth
2005; Jonsson et al. 2010; Acreman et al. 2010; Hayward et al.
2011; Robitaille 2011; Lunttila & Juvela 2012; Juvela et al.
2012). Hydrodynamical simulation codes historically employ
one of two schemes: a Lagrangian formulation based on moving
particles (smoothed particle hydrodynamics or SPH), for exam-
ple Gadget (Springel 2005; Dolag & Stasyszyn 2009; Pakmor
et al. 2012) and SEREN (Hubber et al. 2011); or a Eulerian ap-
proach based on a non-moving spatial grid, often an AMR grid,
for example RAMSES (Fromang et al. 2006), Enzo (Collins
et al. 2010; Bryan et al. 2013), and AMR-VAC (Keppens et al.
2012).
Recent codes including TESS (Duffell & MacFadyen 2011)
and AREPO (Springel 2010, 2011) introduce a new scheme
that employs a moving mesh based on a Voronoi tessellation of
the spatial domain (see Sect. 2.1 and Fig. 1). This new scheme
is claimed to combine the best features of SPH and the tradi-
Fig. 1. A Voronoi tessellation for 400 random sites (in gray), bounded tional Eulerian approach, and it is becoming increasingly pop-
by a cube. Voronoi cell edges are shown in red, Delaunay edges in blue. ular. It has already been applied to various problems includ-
ing the formation of stars, galaxies, and cosmological structures
(Greif et al. 2011; Bauer & Springel 2012; Sijacki et al. 2012;
Heymann & Siebenmorgen 2012; Saftly et al. 2013) and several Kereš et al. 2012; Vogelsberger et al. 2012; Torrey et al. 2012;
methods have been investigated to make them as efficient as pos- Nelson et al. 2013; Marinacci et al. 2013). While the output from
sible (Saftly et al. 2013). a moving mesh code can be re-gridded to an AMR grid to per-
Adaptive-mesh grids seem to be an obvious choice. It is form RT, the resampling process unavoidably introduces inaccu-
straightforward to construct an appropriate grid for any density racies and represents additional overhead; it seems preferable to
field, whether defined by an analytical model or by a collec- perform both aspects of the simulation (MHD and RT) on the
tion of smoothed particles; and it is easy to calculate a straight same grid.
path through the grid, since the boundaries of the cuboidal cells These considerations lead to the question of whether it is
are lined up with the coordinate axes and each cell has a lim- possible to perform accurate and efficient RT on unstructured
ited number of neighbors (Saftly et al. 2013). This second point Voronoi grids.
is very important in the context of RT because ray tracing and
One approach is to approximate a straight path through the
Monte Carlo RT codes determine the radiation field in each grid
grid by a sequence of non-collinear segments connecting neigh-
cell by laying down random rays (i.e., straight paths) through the
boring sites. For example in the SimpleX code (Paardekooper
domain. The simulation run time is often dominated by the por-
et al. 2010) and in the LIME code (Brinch & Hogerheijde 2010)
tion of the code that identifies the grid cells crossed by each path
radiation travels along the edges of the Delaunay triangulation
and calculates the lengths of the corresponding path segments.
corresponding to the Voronoi grid (see Sect. 2.1; the Delaunay
Adaptive-mesh grids also have drawbacks. First of all, for a edges are shown in blue in Fig. 1). While it facilitates calculating
given density field and required resolution, an AMR grid may the paths, this approximation requires additional mechanisms to
not be the kind of grid with the least number of cells. To illus- compensate for errors in path length (see Fig. 5 in Paardekooper
trate this, consider a density field defined by a set of smoothed et al. 2010) and direction (see Fig. 4 in Brinch & Hogerheijde
particles. An octree grid constructed such that each cell encloses 2010). The distance covered by the path inside a particular grid
at most one particle usually has over three times more cells than cell becomes a fuzzy concept, while this is an important quan-
there are particles; i.e., two out of three cells are empty1 . In con- tity in many RT codes, e.g., for tracking the amount of energy
trast, an unstructured grid based on a Voronoi tessellation of the absorbed in the cell. And finally the spread on direction makes it
spatial domain (see Sect. 2.1 and Fig. 1), using the given particles hard to produce high-resolution images of the simulated object.
as generating sites, has exactly the same number of cells as there
are particles. While not an issue in many situations, minimiz- In Sect. 2 we present instead an efficient method of calculat-
ing the number of cells is sometimes crucial. For example, con- ing a straight path between two arbitrary points through a 3D
sider a panchromatic RT simulation involving small dust grains Voronoi grid, applicable in any RT code based on ray tracing
not in local thermodynamic equilibrium (non-LTE) conditions. or Monte Carlo techniques. The path segments inside each grid
Because each cell stores radiation field data per wavelength bin, cell are calculated to high precision using a straightforward al-
memory requirements are substantial. Moreover, the simulation gorithm that relies on the mathematical properties of Voronoi
run time is most likely dominated by the calculation of the non- tessellations. In Sect. 3 we introduce an implementation of the
LTE heating and re-emission of the dust grains in each cell. In method in our dust RT code SKIRT (Baes et al. 2011). We
demonstrate the method’s reliability, accuracy, and efficiency by
1
To verify this claim, we ran a few tests with particles distributed uni- comparing results obtained through the Voronoi grid with those
formly over the spatial domain, and with particles representing a galaxy generated by existing well-tested grids. In Sect. 4 we summarize
generated by a hydrodynamical simulation. our conclusions.

A35, page 2 of 9
 P. Camps et al.: 3D Voronoi grids in radiative transfer

leaves the domain. This is illustrated in Fig. 2 for a single cell.

The presented method can easily be adjusted for other domain
geometries, or for rays that originate outside the domain.
During a setup phase, before any straight paths are calcu-
lated, the following data are prepared:
1. The domain boundaries ( Dmin , Dmax ).
2. The positions of the sites { pm , m = 1 . . . M }.
3. For each site pn , the indices { mn,i , i = 1 . . . In } of all sites
neighboring that site or, equivalently, of all cells neighboring
the cell corresponding to that site. Domain walls are repre-
sented by special (negative) index values.
Data items (1) and (2) are externally specified as part of the prob-
lem definition. The neighbor lists (3) can easily be derived from
Fig. 2. A single Voronoi cell (in red) with its neighboring sites (in gray) a Voronoi tessellation or Delaunay triangulation for the specified
and corresponding Delaunay edges (in blue). A straight path through set of sites, since nearest neighbor information is the defining
the cell is shown (in green) with its intersection points with the cell characteristic of these concepts. No information is needed on the
boundary at entry and exit. vertices, edges or faces of the polyhedra delimiting the Voronoi
cells.
To begin calculating a straight path, the current point r is set
to the starting point, and the current cell index mr is set to the
2. Method index of the cell containing the starting point. By definition of a
Voronoi tessellation, finding the cell containing a given point r
2.1. Voronoi tesselations of 3D space is equivalent to locating the site pi nearest to r. This is a straight-
Given a set of points { p1 , p2 , . . . pn } in 3D space, called sites, forward operation that can easily be optimized as described later
the corresponding Voronoi tessellation (Dirichlet 1850; Voronoi in Sect. 2.3. For the time being we assume that there is a function
1908) is a set of cells {Ci } where each cell Ci consists of all C(r) that returns the index m of the cell containing a given point.
the points p at least as close to pi as to any other site. The cor- Once initialized, the method loops over the algorithm that
responding Delaunay triangulation (Delaunay 1934) is a graph computes the exit point from the current cell, i.e., the intersec-
created by placing a straight edge between any two sites that tion of the ray formed by the current point r and the direction k
share a cell boundary in the Voronoi tessellation. Thus every site with the current cell’s boundary. The algorithm also produces the
is connected to its nearest neighbors. index of the neighboring cell without extra cost. If an exit point
An example Voronoi tessellation is shown in Fig. 1, and a is found, the loop adds a path segment to the output, updates the
single Voronoi cell is illustrated in Fig. 2. A Voronoi cell is de- current point and the current cell index, and continues to the next
limited by a convex polyhedron. A Delaunay edge, i.e., a line iteration. If the exit is towards a domain wall, the loop is termi-
segment that connects two sites sharing a polygonal face, is per- nated. Because of computational inaccuracies it may occur that
pendicularly bisected by the plane containing the face, although no exit point is found. In that case, no path segment is added to
the bisection point may lie outside the face. For a set of sites the output, the current point is advanced in the direction k over
chosen randomly from a uniform distribution, the number of an infinitesimal distance   ||Dmax − Dmin ||, and the new current
nearest neighbors (or equivalently the number of cells sharing cell index is determined by calling the function C(r).
a face with any given cell) has an expectation value of 15.54 The algorithm computing the exit point from the current cell
(van de Weygaert 1994). requires the following input data: the current point r; the ray’s
To obtain an optimal grid in the context of a RT simulation, direction k as a unit vector; the index mr of the current cell;
the Voronoi sites should be more densely packed (generating the indices { mi , i = 1 . . . I } of all cells neighboring the current
smaller cells) in regions where a higher resolution is desired. For cell, with domain walls represented by special (negative) values;
example one could select the positions randomly from a proba- the positions of the sites { pm , m = 1 . . . M }; and the domain
bility distribution that follows the density of the RT medium, boundaries (Dmin , Dmax ).
and perhaps place extra sites near sharp edges or large gradients. The line containing the ray under consideration can be writ-
If the density field is defined by a set of smoothed particles, the ten as L(s) = r + s k with s ∈ R. The exit point can similarly
particle locations form natural Voronoi sites. And of course if the be written as q = r + sq k with sq > 0, and the distance covered
density field is already defined on a Voronoi mesh, the original within the cell is given by sq . The index of the cell next to the
site locations can be used. exit point is denoted mq and is easily determined as follows.
1. Calculate the set of values {si } for the intersection points be-
2.2. A straight path through a Voronoi grid tween the line L(s) and the planes defined by the neighbors
mi (see below for details on this calculation).
We consider a cuboidal spatial domain D defined by its corner 2. Select the smallest nonnegative value sq = min{si |si > 0}
points ( Dmin , Dmax ), and a set of sites { pm ∈ D, m = 1 . . . M }. in the set to determine the exit point and the corresponding
All sites are inside the domain, and the corresponding Voronoi neighbor mq .
cells are clipped by the domain walls, as illustrated in Fig. 1. 3. If there is no nonnegative value in the set, no exit point has
Given a ray describing the path of a photon package, defined by a been found.
starting point r0 ∈ D and a direction k, our aim is to calculate the
ray’s consecutive intersection points with the Voronoi cell walls To calculate si in step (1) for a regular neighbor mi > 0, intersect
– or equivalently, the distance travelled in each cell – until the ray the line L(s) = r + s k with the plane bisecting the sites p(mi )

A35, page 3 of 9
 A&A 560, A35 (2013)

and p(mr ). An unnormalized vector perpendicular to this plane bounding box can be easily calculated from the list of its ver-
is given by tices. This requires fully constructing the cell; however, this is
needed anyway to calculate the cell volume for use in other ar-
n = p(mi ) − p(mr ) (1) eas of the RT simulation (e.g., determining the specific energy
absorbed per unit mass by the medium in the cell). Regardless,
and a point on the plane is given by
the cell geometry is needed solely during setup and does not have
p(mi ) + p(mr ) to be retained thereafter.
p= · (2) The function C(r) receives the following input data: the
2
query point r; for each block Bk , the indices { mk, j , j = 1 . . . Jk }
The equation of the plane can then be written as of all Voronoi cells that possibly overlap that block; the posi-
tions of the sites { pm , m = 1 . . . M }; and the domain boundaries
n · (x − p) = 0. (3)
( Dmin , Dmax ).
Substituting x = r + si k and solving for si provides For each query, the function C(r) performs these steps:

n · ( p − r) 1. Verify that the query point is inside the domain; if not return
si = · (4) a special (negative) index value.
n· k 2. Locate the block containing the query point; this is trivial
If n · k = 0 the line and the plane are parallel so that there is no since blocks are on a regular linear grid.
intersection, and the above equation produces si = ±∞. When 3. Retrieve the list of cells possibly overlapping that block, and
using standard IEEE 754 floating point arithmetic there is no thus possibly containing the query point.
reason to test for this special case, since the infinite value will 4. Calculate the squared distance from the query point to the
never be selected as the exit point in step (2). sites for each of these cells. By definition of a Voronoi tesse-
To calculate si in step (1) for a domain wall mi < 0, substitute lation, the closest site determines the Voronoi cell containing
the appropriate normal and position vectors for the wall plane in the query point.
Eq. (4). For example, for the left wall one has n = (−1, 0, 0) and
p = (Dmin,x , 0, 0) so that 3. Tests, results, and discussion
Dmin,x − rx 3.1. Implementation
si = · (5)
kx
The SKIRT code (Baes et al. 2011) performs 3D continuum
In an actual implementation of this algorithm there is no need to RT using the Monte Carlo technique. It is used for studying
accumulate the complete set of {si } values; one can simply keep dusty astrophysical objects including spiral galaxies (De Looze
track of the smallest nonnegative value. As a further optimiza- et al. 2012a,b; De Geyter et al. 2013) and active galactic nu-
tion, part of the intersection calculation can be avoided for about clei (Stalevski et al. 2012, 2013). Input models can be defined
half of the planes by noting that the sign of n · k determines the through a range of built-in geometries, or imported from the re-
sign of si in Eq. (4). Indeed, the site p(mr ) and the current point r sults of a MHD simulation. The SKIRT code also offers vari-
are on the same side of the plane defined by Eq. (3) (unless r lies ous dust grid options, including regular grids and adaptive grids
in the plane), so that the numerator of Eq. (4) is always positive (Saftly et al. 2013).
(or zero if r lies in the plane). We implemented a Voronoi dust grid in SKIRT according to
the method presented in Sect. 2. This allowed us to use the built-
in geometries for creating synthetic test models, and to compare
2.3. Finding the cell containing a given point
the results with those produced by the existing and well-tested
We now return to the implementation of the function C(r) that grids.
identifies the Voronoi cell containing a given query point. By We employed the open source library Voro++ (Rycroft
definition of a Voronoi tessellation, this operation is equivalent 2009) to set up the input data described in Sects. 2.2 and 2.3.
to finding the site closest to the query point. There are many so- The library and its data structures are used only during setup.
phisticated ways to accelerate this nearest neighbor search, for All relevant information is extracted and stored in our own data
example by building a k-d tree (Friedman et al. 1977) or an structures for reference after setup.
R-tree (Guttman 1984) data structure. We chose to use a sim-
ple mechanism, since this function is usually invoked only once 3.2. Test models
per path and thus its performance is not overly critical.
We assume the domain D is partitioned in a set of cuboidal We tested the Voronoi dust grid with two synthetic models of our
blocks { Bk , k = 1 . . . K } according to a regular linear grid. own making, called torus and spiral, and we ran the RT bench-
During the setup phase described in the beginning of Sect. 2.2, mark described by Pascucci et al. (2004). We first present the
an additional data structure is constructed containing, for each results for our models, and in Sect. 3.5 we discuss the results for
block Bk in the partition of the domain D, the indices { mk, j , j = the Pascucci benchmark.
1 . . . Jk } of all Voronoi cells that possibly overlap that block. The torus model consists of a central light source surrounded
Determining these lists in principle requires an intersection test by an axisymmetric dusty torus, as might be present in the center
between each block and each cell. In practice it suffices to con- of active galactic nuclei. The dust geometry is described by a ra-
sider the cell’s bounding box, which can be easily intersected dial power-law density from a given inner to outer radius, with an
with the blocks. opening angle of 50 degrees. A cut through the dust distribution
One might be tempted to derive a Voronoi cell’s bound- is shown in the top row of Fig. 4. The sites for the Voronoi dust
ing box from the positions of the neighboring sites; however, grid are selected randomly from a uniform distribution over the
the convex hull of a cell’s neighboring sites does not necessar- cuboidal domain enclosing the torus. Since the model is axisym-
ily fully enclose the cell. Because a Voronoi cell is convex, its metric, we can compare the results of the Voronoi grid with those

A35, page 4 of 9
 P. Camps et al.: 3D Voronoi grids in radiative transfer

104 cells 104.5 cells 105 cells 105.5 cells 106 cells 106.5 cells

Fig. 3. A cut through the dust density distribution of the torus model, discretized on Voronoi grids with a resolution varying from 104 cells (left) to
106.5 cells (right). All grids were constructed from a set of uniformly distributed sites.

produced by a regular two-dimensional (2D) cylindrical grid, in Taking into account our naive cell placement, the Voronoi
addition to those produced by an adaptive (3D) octree grid. In grid compares well with the highly tuned adaptive octree grid,
Fig. 3 we illustrate the effect of the number of Voronoi grid cells thus verifying this aspect of our implementation.
for the torus model.
The spiral model represents an idealized spiral galaxy with 3.4. Results
three arms, similar to the spiral model presented in Saftly et al.
(2013). The stellar distribution includes a flattened Sérsic bulge Shooting photon packages through the grid is the most important
and a double-exponential disk with a spiral arm perturbation. test in the context of this study.
The dust is distributed in a thinner, similarly perturbed double- The middle row of Fig. 4 shows the dust temperature calcu-
exponential disk. Cuts through the dust distribution are shown in lated by a panchromatic simulation for the torus model, using
the top row of Figs. 5 and 6. In this case, the sites for the Voronoi the three grids describe above. All quantities, including the radi-
dust grid are selected randomly from the dust distribution, as op- ation field and the amount of dust absorption, are discretized on
posed to a uniform distribution. Areas with a higher dust density the same grid as the dust density. In each simulation, the central
are thus, on average, covered with smaller cells. light source emits 105 photon packages for each of 100 wave-
length bins on a logarithmic grid. Scattering events cause addi-
tional photon packages to be created, which is particularly rel-
3.3. Test grids evant for this model because of the high optical depth of the
torus. In the end, each simulation traces about 700 million pho-
For the torus model we ran simulations with three different dust ton packages through the dust grid.
grids: a regular 2D cylindrical grid with 2502 = 62 500 cells; The bottom rows of Figs. 4–6 show the flux density calcu-
an adaptive octree grid with ≈950 000 cells; and a Voronoi grid lated by a monochromatic simulation for each model and grid
with about the same number of uniformly distributed cells. The combination. The Poisson noise is caused by the statistical na-
top row of Fig. 4 shows a cut through the gridded dust density ture of the Monte Carlo technique. In each simulation, the light
distribution for each of these grids. The cylindrical grid captures sources emit 10 million photon packages at a fixed wavelength,
the sharp edges of the model perfectly, because the cylindrical and scattering events again cause additional photon packages to
coordinate axes are lined up with the edges. The octree grid does be created.
a fine job as well because of its adaptive nature: smaller cells Other than the effects of grid resolution and the unavoidable
are automatically created along the sharp edges. The Voronoi noise, the calculated temperature and flux density maps are the
grid does not do particularly well at the edges because of the same for the various grids. In particular, as noted in Sect. 3.3,
random placement of its cells. This would not be an issue when the Voronoi grid does not resolve the central area of the dust
importing a grid from a moving mesh code, because the cell sizes distribution as well as the other grids, causing some deviation in
would already be properly adjusted to the underlying gradients. the central area of the calculated flux density field. This effect is
ultimately due to the naive placement of the Voronoi cells in our
For the spiral model we ran simulations with two different tests, and would not be present for a properly adjusted grid.
dust grids: an adaptive octree grid with ≈1 350 000 cells; and a These results validate the accuracy of our straight path cal-
Voronoi grid with about the same number of cells, placed using a culation method for Voronoi grids.
weighed distribution according to the dust density (smaller cells Table 2 provides an indication of the processing time spent
in higher density areas). The top rows of Figs. 5 and 6 show a cut per cell crossing for each simulation. To obtain these numbers,
through the gridded dust density distribution for each of these the elapsed time for the photon shooting phase of a simulation
grids. The differences between the grids are most easily seen in is divided by the number of grid cells crossed during that phase.
the lower density areas. The result thus includes some overhead for generating the ran-
Although this study does not focus on grid quality, we still dom paths and for storing results, in addition to the grid traversal
need to ensure that our Voronoi grid implementation properly calculation itself. The tests were performed on a typical desktop
represents the theoretical dust densities defined by the synthetic computer using a single core. The last column lists the ratio be-
models. To obtain an objective quality measure, we sample the tween the cell crossing times for the Voronoi and octree grids.
theoretical dust density ρt and the gridded dust density ρg at a The Voronoi grid performs roughly three times slower than our
large number of random points uniformly distributed over the highly optimized octree implementation (which maintains, for
domain. We use the standard deviation of the difference ρt − ρg example, a neighbor list for each cell to accelerate the process
as a measure for how well the grid reflects the theoretical density of finding the next cell on a path). This seems surprisingly fast
distribution. Table 1 lists the resulting numbers for the various in view of the high geometric complexity of a Voronoi grid (il-
grids and models. For each model the value for the octree grid is lustrated in Figs. 1 and 2) compared to the cuboidal cells in an
normalized to unity. octree. Moreover, as noted in the introduction, an octree grid

A35, page 5 of 9
 A&A 560, A35 (2013)

ρdust – cylindrical grid ρdust – octree grid ρdust – Voronoi grid

T dust – cylindrical grid T dust – octree grid T dust – Voronoi grid

f – cylindrical grid f – octree grid f – Voronoi grid

Fig. 4. Illustration of the results for the torus model with three different dust grids. Rows – top: the dust density distribution (cut through the central
edge-on plane); middle: the calculated dust temperature (cut through the central edge-on plane); bottom: the calculated flux density escaping from
the model (edge-on view). Columns – left: regular 2D cylindrical grid with 2502 = 62 500 cells; middle: adaptive octree grid with ≈950 000 cells;
right: Voronoi grid with ≈950 000 uniformly distributed cells.

ρdust – octree grid ρdust – Voronoi grid

f – octree grid f – Voronoi grid

Fig. 5. Illustration of the results for the spiral model, edge-on view. Rows – top: the dust density distribution (cut through the central edge-on
plane); bottom: the calculated flux density escaping from the model (edge-on view). Columns – left: adaptive octree grid with ≈1 350 000 cells;
right: Voronoi grid with ≈1 350 000 cells with a non-uniform, weighed distribution.

A35, page 6 of 9
 P. Camps et al.: 3D Voronoi grids in radiative transfer

ρdust – octree grid ρdust – Voronoi grid

f – octree grid f – Voronoi grid

Fig. 6. Illustration of the results for the

spiral model, face-on view. Rows –
top: the dust density distribution (cut
through the central face-on plane); bot-
tom: the calculated flux density escap-
ing from the model (face-on view).
Columns – left: adaptive octree grid
with ≈1 350 000 cells; right: Voronoi
grid with ≈1 350 000 cells with a non-
uniform, weighed distribution.

Table 1. Grid quality. Table 2. Run time.

Model Cylindrical Octree Voronoi Time per cell crossing (ns)

Model Simulation type Octree Voronoi Vor./Oct.
Torus 0.82 1 1.75
Spiral – 1 1.68 Torus monochromatic 219 693 3.2
Torus panchromatic 400 1006 2.5
Notes. The difference between the theoretical and gridded dust density Spiral monochromatic 309 903 2.9
is sampled at a large number of random points, uniformly distributed Spiral panchromatic 442 1095 2.5
over the domain. The standard deviation on this difference is used as a
quality measure for the grid. In the table, the value for the octree grid Notes. The elapsed time for the photon shooting phase of a simulation is
is normalized to unity for each model. Smaller numbers indicate better divided by the number of grid cells crossed during that phase. The result
quality. is an indication of the time spent per cell crossing, including grid traver-
sal calculations and some overhead for generating the random paths and
for storing results. The tests were performed on a typical desktop com-
puter using a single core. The last column lists the ratio between the run
may need many more cells than the Voronoi grid to represent a times for the Voronoi and octree grids. Larger numbers indicate slower
particular density field, further balancing performance in favor performance.
of the Voronoi grid.
As discussed in Sect. 2.2, the cell crossing algorithm may
occasionally fail to find an exit point because of computational plane of the dust density distribution is shown in the left panel
inaccuracies. In our tests this occurred at most once per 50 mil- of Fig. 7.
lion cell crossings, so this issue does not affect the algorithm’s We ran panchromatic simulations for this model with opti-
performance. cal depths τ = 0.1, 1, and 10 using a 3D Voronoi grid consisting
of one million cells randomly placed according to a 1/r distri-
3.5. The Pascucci benchmark bution. This distribution serves to properly resolve the intense
radiation field in the center of the model. In each simulation, the
The Pascucci benchmark (Pascucci et al. 2004) models a star central light source emits 105 photon packages for each of 150
embedded in a circumstellar disk with an inner cavity free of wavelength bins on a logarithmic grid.
dust, prescribing an analytical 2D distribution and a set of optical The two rightmost panels of Fig. 7 compare the spectral en-
depths and viewing angles. A cut through the central edge-on ergy distribution (SED) produced by our SKIRT simulations

A35, page 7 of 9
 A&A 560, A35 (2013)

SEDs for i =12.5 ◦ SEDs for i =77.5 ◦

ρdust – Voronoi grid τ =10 τ =10
τ =1 τ =1
τ =0.1 τ =0.1
107 107

λ Fλ (W m−2 )

λ Fλ (W m−2 )
106 106

105 105

104 -1 104 -1
10 100 101 102 10 100 101 102
λ (µm) λ (µm)

Fig. 7. Illustration of the results for the Pascucci benchmark (Pascucci et al. 2004). The left panel shows a cut through the central edge-on plane
of the dust density distribution discretized on a 3D Voronoi grid with one million cells randomly placed according to a 1/r distribution. The other
panels show the simulated spectral energy distribution (SED) for disk inclinations equal to 12.5◦ (center) and 77.5◦ (right), for optical depths
τ = 0.1, 1, and 10. Dots indicate benchmark reference points; solid lines represent our simulation results using the 3D Voronoi grid shown in the
left panel.

with the corresponding benchmark results published in Pascucci of the interface, and allows cooperation even when source code
et al. (2004). The center panel shows the SEDs for the various is not publicly available. In contrast to this approach, for exam-
optical depths at a nearly face-on disk inclination of 12.5◦ , the ple, the Sunrise RT code (Jonsson 2006) directly invokes parts
right panel at a nearly edge-on inclination of 77.5◦ . Dots indicate of the non-public Arepo moving mesh code (Springel 2011) to
benchmark reference points; solid lines represent our simulation implement the interface2 .
results.
For higher optical depths our simulation results deviate
slightly because the SKIRT code is not optimized for opera- 4. Conclusions
tion in this regime; running the benchmark with a 2D axially
The choice of an appropriate discretization is crucial in any nu-
symmetric logarithmic grid results in the same deviation (not
merical simulation code. Because of the large dynamic range of
shown). These results further validate our method for calculat-
the physical quantities, in most problems the resolution of the
ing straight paths through a Voronoi grid.
grid must scale with the field densities or gradients. Adaptive
grids with cuboidal cells, such as octrees or more generally
3.6. Applicability AMR grids, have proven very popular in part because of their
relative ease of implementation. However, several recent codes
From Sect. 2 it follows that the presented method requires as have adopted unstructured grids based on Voronoi tessellations,
input data solely the coordinates of the Voronoi sites, plus any or equivalently, Delaunay triangulations. These grids tend to
relevant physical properties (such as mass densities) for the cell more closely reflect dynamic ranges in the model with fewer
surrounding each site. In other words, the interface between the cells, presenting cell boundaries that are more adjusted to the
input model and the RT code is very thin, opening up a wide underlying gradients. Since a Voronoi grid is defined solely by
range of possibilities. An input model can be defined by SPH its generating points, the cell size and distribution can be easily
particles, serving as Voronoi sites; or by a Voronoi mesh pro- fine-tuned by placing these sites in the appropriate locations.
duced by an MHD code; or by appropriately distributed ran- In a RT simulation the Voronoi grid can be a very flex-
dom points generated from (semi)-analytical density or opacity ible tool. Appropriate sites can be generated randomly, dis-
fields, similar to the approach in Paardekooper et al. (2010), for tributed according to the input model’s density or opacity fields;
example. if needed extra sites can be added in high-gradient areas. In the
We also note in Sects. 2.2 and 2.3 that the path calculation case of a particle-based input model, the particle locations them-
algorithm itself requires no information on a Voronoi cell other selves can serve as sites; and for an input model already based
than its bounding box and the locations of its own site and all on a Voronoi mesh no re-gridding is required at all.
neighboring sites. The required data structures can be easily built In this work we have shown that it is straightforward to im-
from the input data using a publicly available Voronoi library. plement accurate and efficient RT on Voronoi grids. In spite
The library code is invoked only during the initialization phase, of the geometric complexity of the cell boundaries, calculating
minimizing its impact on performance and robustness, and al- straight paths between two arbitrary points through a 3D Voronoi
lowing it to be easily replaced by another code if the need arises. grid is only about three times slower than a highly optimized
For example, while we are happy with the Voro++ library’s ease octree implementation with the same number of cells, while in
of use and with its performance during the tests, we may in the practice the total number of Voronoi grid cells may be lower
future consider using a parallelized method (Lo 2012; Springel for an equally good representation of the density field. The pre-
2011). sented method automatically yields the precise distance covered
As a consequence, the presented method allows the RT code
to support a Voronoi grid while remaining uncoupled from the 2
http://code.google.com/p/sunrise/wiki/
code producing the input model. This decreases the complexity RunningWithArepo

A35, page 8 of 9
 P. Camps et al.: 3D Voronoi grids in radiative transfer

by the path inside each grid cell, and eliminates the need for Hayward, C. C., Kereš, D., Jonsson, P., et al. 2011, ApJ, 743, 159
approximate corrections or work-arounds required by alternate Heymann, F., & Siebenmorgen, R. 2012, ApJ, 751, 27
Hubber, D. A., Batty, C. P., McLeod, A., & Whitworth, A. P. 2011, A&A, 529,
approaches where the radiation travels only along the Delaunay A27
edges. The method requires only a thin interface with the in- Indebetouw, R., Whitney, B. A., Johnson, K. E., & Wood, K. 2006, ApJ, 636,
put model and with the actual construction of the grid, allowing 362
codes to remain largely uncoupled and enabling the use of a pub- Jonsson, P. 2006, MNRAS, 372, 2
licly available Voronoi library. Jonsson, P., Groves, B. A., & Cox, T. J. 2010, MNRAS, 403, 17
Juvela, M., & Padoan, P. 2003, A&A, 397, 201
While we implemented and tested the method in our con- Juvela, M., Malinen, J., & Lunttila, T. 2012, A&A, 544, A141
tinuum RT code SKIRT, focusing on the effects of dust, it is Keppens, R., Meliani, Z., van Marle, A., et al. 2012, J. Comput. Phys., 231, 718
widely applicable to all RT codes using ray tracing or Monte Kereš, D., Vogelsberger, M., Sijacki, D., Springel, V., & Hernquist, L. 2012,
Carlo techniques. MNRAS, 425, 2027
We conclude that the benefits of using a Voronoi grid in Kurosawa, R., & Hillier, D. J. 2001, A&A, 379, 336
Laursen, P., Razoumov, A. O., & Sommer-Larsen, J. 2009, ApJ, 696, 853
RT simulation codes will often outweigh the somewhat slower Lo, S. 2012, Comput. Meth. Appl. Mech. Eng., 237 88
performance. Lunttila, T., & Juvela, M. 2012, A&A, 544, A52
Marinacci, F., Pakmor, R., & Springel, V. 2013, MNRAS, accepted
[arXiv:1305.5360]
Acknowledgements. This work fits in the CHARM framework (Contemporary
Matthews, L. D., & Wood, K. 2001, ApJ, 548, 150
physical challenges in Heliospheric and AstRophysical Models), a phase VII
Misiriotis, A., Kylafis, N. D., Papamastorakis, J., & Xilouris, E. M. 2000, A&A,
Interuniversity Attraction Pole (IAP) programme organised by BELSPO, the
353, 117
BELgian federal Science Policy Office. W.S. acknowledges the support of Al-
Nelson, D., Vogelsberger, M., Genel, S., et al. 2013, MNRAS, 429, 3353
Baath University and The Ministry of High Education in Syria in the form of a
Niccolini, G., & Alcolea, J. 2006, A&A, 456, 1
research grant.
Paardekooper, J.-P., Kruip, C. J. H., & Icke, V. 2010, A&A, 515, A79
Pakmor, R., Edelmann, P., Röpke, F. K., & Hillebrandt, W. 2012, MNRAS, 424,
2222
References Paron, S., Weidmann, W., Ortega, M. E., Albacete Colombo, J. F., & Pichel, A.
2013, MNRAS, 433, 1619
Abdikamalov, E., Burrows, A., Ott, C. D., et al. 2012, ApJ, 755, 111
Pascucci, I., Wolf, S., Steinacker, J., et al. 2004, A&A, 417, 793
Acreman, D. M., Harries, T. J., & Rundle, D. A. 2010, MNRAS, 403, 1143
Pelkonen, V.-M., Juvela, M., & Padoan, P. 2009, A&A, 502, 833
Baes, M., & Dejonghe, H. 2001, MNRAS, 326, 733
Pinte, C., Ménard, F., Duchêne, G., & Bastien, P. 2006, A&A, 459, 797
Baes, M., & Dejonghe, H. 2002, MNRAS, 335, 441
Robitaille, T. P. 2011, A&A, 536, A79
Baes, M., Davies, J. I., Dejonghe, H., et al. 2003, MNRAS, 343, 1081
Rycroft, C. H. 2009, Chaos, 19, 041111
Baes, M., Verstappen, J., De Looze, I., et al. 2011, ApJS, 196, 22
Saftly, W., Camps, P., Baes, M., et al. 2013, A&A, 554, A10
Bauer, A., & Springel, V. 2012, MNRAS, 423, 2558
Schartmann, M., Meisenheimer, K., Camenzind, M., et al. 2008, A&A, 482, 67
Bethell, T. J., Zweibel, E. G., Heitsch, F., & Mathis, J. S. 2004, ApJ, 610, 801
Schechtman-Rook, A., Bershady, M. A., & Wood, K. 2012, ApJ, 746, 70
Bianchi, S. 2008, A&A, 490, 461
Sijacki, D., Vogelsberger, M., Kereš, D., Springel, V., & Hernquist, L. 2012,
Bianchi, S., Ferrara, A., Davies, J. I., & Alton, P. B. 2000, MNRAS, 311, 601
MNRAS, 424, 2999
Brinch, C., & Hogerheijde, M. R. 2010, A&A, 523, A25
Springel, V. 2005, MNRAS, 364, 1105
Bryan, G. L., Norman, M. L., et al. (The Enzo Collaboration) 2013, ApJS,
Springel, V. 2010, MNRAS, 401, 791
submitted [arXiv:1307.2265]
Springel, V. 2011 [arXiv:1109.2218]
Chakrabarti, S., Cox, T. J., Hernquist, L., et al. 2007, ApJ, 658, 840
Stalevski, M., Fritz, J., Baes, M., Nakos, T., & Popović, L. Č. 2012, MNRAS,
Ciardi, B., Ferrara, A., Marri, S., & Raimondo, G. 2001, MNRAS, 324, 381
420, 2756
Code, A. D., & Whitney, B. A. 1995, ApJ, 441, 400
Stalevski, M., Fritz, J., Baes, M., & Popovic, L. C. 2013 [arXiv:1301.4244]
Collins, D. C., Xu, H., Norman, M. L., Li, H., & Li, S. 2010, ApJS, 186, 308
Stamatellos, D., & Whitworth, A. P. 2003, A&A, 407, 941
De Geyter, G., Baes, M., Fritz, J., & Camps, P. 2013, A&A, 550, A74
Stamatellos, D., & Whitworth, A. P. 2005, A&A, 439, 153
De Looze, I., Baes, M., Bendo, G. J., et al. 2012a, MNRAS, 427, 2797
Steinacker, J., Bacmann, A., & Henning, T. 2002, J. Quant. Spectr. Rad. Transf.,
De Looze, I., Baes, M., Fritz, J., & Verstappen, J. 2012b, MNRAS, 419, 895
75, 765
Decin, L., Cox, N. L. J., Royer, P., et al. 2012, A&A, 548, A113
Steinacker, J., Bacmann, A., Henning, T., Klessen, R., & Stickel, M. 2005, A&A,
Delaunay, B. 1934, Classe des Sciences Mathématiques et Naturelles, 7, 793
434, 167
Dirichlet, L. 1850, Journal für die reine und angewandte Mathematik, 40, 209
Steinacker, J., Bacmann, A., & Henning, T. 2006, ApJ, 645, 920
Disney, M., Davies, J., & Phillipps, S. 1989, MNRAS, 239, 939
Steinacker, J., Baes, M., & Gordon, K. D. 2013, ARA&A, 51, 63
Dolag, K., & Stasyszyn, F. 2009, MNRAS, 398, 1678
Tasitsiomi, A. 2006, ApJ, 645, 792
Doty, S. D., Metzler, R. A., & Palotti, M. L. 2005, MNRAS, 362, 737
Torrey, P., Vogelsberger, M., Sijacki, D., Springel, V., & Hernquist, L. 2012,
Duffell, P. C., & MacFadyen, A. I. 2011, ApJS, 197, 15
MNRAS, 427, 2224
Ercolano, B., Barlow, M. J., & Storey, P. J. 2005, MNRAS, 362, 1038
van de Weygaert, R. 1994, A&A, 283, 361
Fallscheer, C., Reid, M. A., Di Francesco, J., et al. 2013, ApJ, 773, 102
Verhamme, A., Schaerer, D., & Maselli, A. 2006, A&A, 460, 397
Friedman, J. H., Bentley, J. L., & Finkel, R. A. 1977, ACM Trans. Math. Softw.,
Vogelsberger, M., Sijacki, D., Kereš, D., Springel, V., & Hernquist, L. 2012,
3, 209
MNRAS, 425, 3024
Fritz, J., Gentile, G., Smith, M. W. L., et al. 2012, A&A, 546, A34
Voronoi, G. 1908, Journal für die reine und angewandte Mathematik, 134, 198
Fromang, S., Hennebelle, P., & Teyssier, R. 2006, A&A, 457, 371
Wang, Z., Kaplan, D. L., Slane, P., Morrell, N., & Kaspi, V. M. 2013, ApJ, 769,
Goldsmith, P. F., Heyer, M., Narayanan, G., et al. 2008, ApJ, 680, 428
122
Goosmann, R. W., & Gaskell, C. M. 2007, A&A, 465, 129
Witt, A. N., & Gordon, K. D. 1996, ApJ, 463, 681
Gordon, K. D., Misselt, K. A., Witt, A. N., & Clayton, G. C. 2001, ApJ, 551,
Witt, A. N., & Gordon, K. D. 2000, ApJ, 528, 799
269
Witt, A. N., Thronson, Jr., H. A., & Capuano, Jr., J. M. 1992, ApJ, 393, 611
Greif, T. H., Springel, V., White, S. D. M., et al. 2011, ApJ, 737, 75
Wolf, S. 2003, Comput. Phys. Commun., 150, 99
Guttman, A. 1984, SIGMOD Rec., 14, 47
Wolf, S., Fischer, O., & Pfau, W. 1998, A&A, 340, 103
Harries, T. J., Monnier, J. D., Symington, N. H., & Kurosawa, R. 2004, MNRAS,
Wood, K., Mathis, J. S., & Ercolano, B. 2004, MNRAS, 348, 1337
350, 565

A35, page 9 of 9