This page offers guidelines on how to configure the wavelength grids for a panchromatic simulation, and by extension, the wavelength bias distributions for primary and secondary sources.

It is wise to review the following topics before reading the guidelines provided below:

The discussion on this page is organized as follows:

Note: Before adopting the wavelength grids and especially the spectral resolution values recommended in the text below, it is important to conduct appropriate convergence tests. In other words, try the same simulation with "better" grids and more photon packets and compare the results.

Primary sources

The primary source system requires the user to configure a wavelength range that limits the wavelengths of emitted primary photon packets. Within this range, the wavelengths are directly sampled from the relevant source spectrum without the need for (additional) discretization.

As a general rule, limit the primary source wavelength range to the range that is relevant for the study at hand. In other words, avoid emitting photon packets that will never be detected by any instrument or probe and that will not contribute to the relevant part of the radiation field. On the other hand, if the simulation includes kinematic effects, the photon packet wavelengths will shift before being detected. It might then be useful to have a somewhat wider source range, compared to the instrument wavelength grids, to ensure a proper contribution level in the outer wavelength bins. If a model is placed at nonzero redshift (see Nonzero redshift & CMB dust heating), one should take the corresponding wavelength shift into account as well.

The default wavelength bias distribution and wavelength bias factor are appropriate for most simulations, and especially if the wavelength range spans multiple decades. The default composite biasing scheme then achieves a constant spectral resolution over the entire range, modulated with the source luminosity.

For narrow wavelength ranges, perhaps corresponding to a particular spectrograph or spanning a given emission line, a linear distribution of the photon packet wavelengths might be more appropriate. To this end, SKIRT offers a built-in linearly uniform wavelength bias distribution (LinWavelengthDistribution) in addition to the default logarithmic distribution (LogWavelengthDistribution). Users can also define a custom distribution for maximum flexibility (FileWavelengthDistribution and ListWavelengthDistribution). For example, one might want to strongly favor one or more wavelength intervals of special interest.

To evaluate whether the photon packet distribution as a function of wavelength is appropriate for the source spectrum and for the study at hand, proceed as follows:

add a LuminosityProbe and a LaunchedPacketsProbe instance to the list of probes in the configuration.
run the simulation with a fair number of photon packets.
use the PTS visual/plot_sources command (pts.visual.do.plot_sources) to plot the luminosity of and the number of photon packets launched by each source as a function of wavelength, e.g.

pts plot_sources . --wmin 1 --wmax 10 --dex 3

Radiation field

A panchromatic simulation with secondary emission always tracks and stores the radiation field in every spatial cell. If the simulation does not include secondary emission, the user can still opt to record the radiation field for probing at the end of the simulation. In both cases, the user must configure a wavelength grid that will be used to discretize the spectral variation of the radiation field (in every spatial cell).

The radiation field wavelength grid must have disjoint bins and often the bins will be consecutive, although that is not a requirement.

We first consider a panchromatic dust continuum simulation that produces synthetic observations from UV to submm wavelengths for a typical spiral galaxy at redshift zero. We expect the radiation field at shorter wavelengths to dominate the dust heating process, with the longer wavelengths having a minimal effect. We can also presume that the precise wavelength of an incoming photon packet might not be so important, as long as its energy is properly categorized, possibly allowing fairly wide wavelength bins. Tests indeed indicate that, for such a simulation, it is sufficient to configure a radiation field wavelength grid with just 40 points from 0.02 to 10 micron, distributed evenly in log space using a LogWavelengthGrid.

For other dust models, such as compact galaxies at higher redshift or dust disks surrounding young stars or galactic nuclei, the dust opacity may be substantially higher, necessitating iteration over secondary emission to self-consistently determine dust self-absorption. In this case, the radiation field will need to be tracked for longer wavelengths as well, and the radiation field wavelength grid must be correspondingly extended.

Models that self-consistently calculate secondary line emission by gas will need additional radiation field wavelength bins near each of the relevant lines to properly resolve the line profiles. The CompositeWavelengthGrid class should come in handy to configure such a grid. If the model does not contain any dust, the radiation field may not need to be tracked away from the lines, in which case the radiation field wavelength grid may contain (large) gaps.

Secondary emission

Thermal emission from dust

The dust emission wavelength grid controls the resolution of the dust emission spectrum calculated for each spatial cell. Especially when taking into account the stochastic heating of small dust grains, this spectrum contains many narrow infrared features. It is thus desirable to configure a grid that can properly resolve these features. Memory usage is not an issue because the emission spectrum is stored just once per execution thread. The performance impact is very limited as well because sampling these emission spectra is not the bottleneck of the calculation.

The dust emission wavelength grid must have disjoint and preferable consecutive bins. For a typical dust continuum simulation, one can employ a NestedLogWavelengthGrid with a resolution of 100 bins per dex in the overall range from 0.2 to 2000 micron, and 200 narrower bins in the range from 3 to 25 micron, for a total of just over 500 bins.

Note: Even if the wavelength range of all instruments and probes is smaller, it is still important to configure a dust emission wavelength grid with the full spectral range described above for the emission calculations to be correct.

For thermal dust emission, the wavelength bias distribution and the wavelength bias factor can be left to their default values; there seems to be no specific benefit in tweaking this aspect of the configuration.

Line emission from gas

In this case, there is no need to discretize the spectrum. For each line, photon packet wavelengths are sampled from a Gaussian profile that depends on the local thermal velocity of the originating species.

By default, the emitted photon packets are distributed between the lines and over each line profile according to a composite biasing scheme balancing the relative luminosities on the one hand (for 50 per cent) and the logarithmic wavelength bias distribution on the other hand (for the other 50 per cent). This might be improved by configuring a linear bias distribution and a wavelength bias factor closer to one. Alternatively, one could provide a more complex bias distribution that favors the most important lines for the study at hand.

Instruments and probes

All SKIRT instruments require a wavelength grid for binning the observed fluxes contributed by detected photon packets. Some probes similarly require a wavelength grid for binning the probed spectral quantity. Each instrument and eligible probe can be configured with its own distinct wavelength grid. To avoid repetition in configurations with several instruments and probes, a default wavelength grid can be configured for the instrument system. The default grid is adopted by instruments and probes for which no individual grid has been specified. In the paragraphs below, we focus the discussion on instruments.

Instruments can use disjoint or band wavelength grids. For a disjoint grid the binning process is straightforward. For a band wavelength grid, an arriving photon packet is registered for each band in the wavelength grid after multiplying its luminosity contribution by the band’s transmission factor corresponding to the packet’s wavelength. This amounts to "on-the-fly" convolution of the detected flux with the transmission curve of each band.

If the study under consideration does not require a both spectrally and spatially resolved IFU data cube, one useful technique is to configure two instruments for the same line of sight: an SEDInstrument for recording a high-resolution spectrum, and a FrameInstrument for recording a set of broadband images. Each instrument performs different binning on the same set of arriving photon packets: the SED instrument spatially integrates fluxes into narrow spectral bins, while the frame instrument does a spectral convolution for the fluxes in every spatial pixel. This leads to a very efficient use of the photon packets being traced through the system. The alternative is to run the simulation using a regular wavelength grid with narrow bins and perform the convolution after the fact. For proper results, the instrument wavelength grid must resolve all spectral features of the sources, including emission or absorption lines which may be Doppler shifted because of kinematic effects. This requires a large number of spectral bins with correspondingly large memory requirements.

Note: Always place instruments with the same line of sight consecutively in the ski file. SKIRT then recognizes the similarity and applies a significant performance optimization.

Memory usage

There are obviously countless areas in the SKIRT code that consume memory. In many cases, however, the overall memory requirements are dominated by just a few components, namely the radiation field storage and the instrument data cubes.

Consider a typical panchromatic simulation. We assume that the radiation field wavelength grid has $N_{λ, r f}$ points and that the spatial grid has $N_{c e l l}$ points. The data structure for storing the radiation field then has a size in bytes of

$S_{r f} = 8 N_{λ, r f} N_{c e l l} .$

If the simulation iterates over secondary emission to self-consistently determine dust self-absorption, it tracks the radiation field resulting from primary and secondary emission separately, so that the data structure has twice this size.

We further assume a frame instrument recording in $N_{λ, i n s}$ wavelength bins and $N_{x}$ by $N_{y}$ image pixels. The data structure for storing the detected fluxes has a size in bytes of

$S_{i n s} = 8 N_{λ, i n s} N_{x} N_{y} .$

If the instrument is requested to keep track of individual flux components (such as direct and scattered light), or if there are similar instruments at other viewing angles, the required memory becomes a multiple of this size.

Memory usage can often be limited by properly configuring the radiation field and instrument wavelength grids according to the guidelines provided in the previous sections, for example by using broadband-based wavelength grids for frame instruments. On the other hand, some studies will need a spatially and spectrally resolved instrument data cube in a given wavelength range to simulate integral-field spectroscopy observations and evaluate the effects of kinematics. This affects $N_{λ, i n s}$ without changing $N_{λ, r f}$ in the equations above. Also, when including line interactions with gas media, the number of wavelengths $N_{λ, r f}$ of the radiation field storage will need to increase as well.

Discrete panchromatic simulations

In panchromatic simulation mode, SKIRT samples all wavelengths in a given wavelength range. In contract, some other codes only sample a set of discrete wavelengths in that range. Depending on the form of the involved spectra, these two methods may produce different results. While the SKIRT method is designed to more closely reflect the physical reality, these differences can be very annoying when comparing SKIRT results with those of other radiative transfer simulation codes as part of, for example, a benchmark effort.

SKIRT therefore includes the DiscreteWavelengthDistribution class, which creates the ability to mimick codes that emit photon packets at discrete wavelengths rather than across a continuous range, while still using panchromatic simulation mode. Configuring such a simulation, however, involves some advanced options and requires great care, as described in more detail below.

When a DiscreteWavelengthDistribution instance is used as the wavelength bias distribution for a source with a composite bias factor of one, the source will emit photon packets only at the characteristic wavelengths of the associated grid, and the photon packets will be distributed with equal probability among those wavelengths. In this situation, it makes little sense to configure wavelength grids with different bins for detecting photon packets in other areas of the simulation. Doing so might cause unexpected results. For example, some bins might receive no photon packets at all and thus incorrectly report zero influx. It is therefore best to configure the same wavelength grid throughout the simulation.

To configure a simulation that uses discrete wavelengths in panchromatic mode, follow these guidelines:

Choose a wavelength grid with consecutive bins (i.e. representing a continuous range without gaps), such as a LogWavelengthGrid or a FileWavelengthGrid with relativeHalfWidth set to zero.
Configure all primary and secondary sources (including, e.g. dust emission) with a wavelength bias factor of one, and a wavelength bias distribution of type DiscreteWavelengthDistribution using the chosen wavelength grid.
Explicitly configure all wavelength grids in the simulation to be the same chosen wavelength grid. This includes the radiation field wavelength grid, the dust emission wavelength grid, and the wavelength grids for all instruments and probes.
Avoid using kinematics in the simulation. Small Doppler shifts from the emitted discrete wavelengths will stay within the same wavelength bin anyway. Larger shifts will cause photon packets to move between wavelength bins in a discontinuous way.