This document describes the first part of the technical aspects of the design study for CIRPASS (Cambridge IR PAnoramic Survey Spectrograph). The main objectives of this study are:
Throughout this document it is assumed that CIRPASS will be used on an 8m telescope.
CIRPASS is primarily a multi-object spectrograph and the scientific specification calls for a large multiplex gain and the ability to do very deep spectroscopy. Another requirement of CIRPASS is that it must record the spectra at a spectral resolution of R~3000 to avoid the strong OH lines. All of these requirement demand a very fast camera (especially on an 8m telescope) so that as much spectral information as possible can be crammed on to the detector and also so that the noise characteristics of the detector are reduced by minimising the number of pixels used for each measurement. This latter point is particularly important for the NIR because of the high spectral resolution requirement and because, unlike CCDs, IR arrays have relatively high readout noise and dark current and it is not possible to on-chip bin.
To compare a multi-slit design with a multi-object
design we have looked at the following criteria: throughput, field
of view, survey speed, sky-subtraction, versatility and cost.
A summary of our conclusions is presented in Table 1.
Throughput. For
a fibre feed in the NIR it should be possible to achieve an efficiency
of 80% or better compared to a slit of the same aperture. We are
assuming here that the loss due to light that misses the fibre
or slit is the same for both.
Field of View.
For a fibre fed system this is limited by the telescope and not
the spectrograph. For a multislit version of CIRPASS one array
is 6.8 arcmin in the spatial direction if we assume there is no
anamorphism. [In practise CIRPASS has to be a reflection grating
system and not a grism system such as LDSS-2 because of the high
spectral resolution requirement. There is therefore likely to
be some anamorphism which will reduce the field of view in the
spatial direction.] To avoid serious loss of spectral coverage
we have assumed that ~100 pixels in the dispersion direction can
be deployed for positioning the slits so the total field of view
for 4 arrays is ~18 arcmin2.
Survey Speed. This
depends on the multiplex gain, the wavelength coverage and the
time it takes to reach the required spectral resolution. For a
fibre system each 1.5 arcsec fibre is 4 pixels in the dispersion
direction and each spectrum uses 10 rows of pixels. This gives
0.12m of wavelength coverage for one array at R=3000 and we can
have 100 fibres per array. The multiplex-gainwavelength-coverage
product is therefore 48microns. For a multislit system with 1.2
arcsec wide slits (3 pixels) and ~900 pixels for the wavelength
coverage we have a total wavelength coverage of 0.14microns. A
slit length of 20 arcsec is 50 pixels so we have a total multiplex
gain of 80 (for a limiting magnitude of J=23 there are ~160 galaxies
in a field of 18 arcmin2) for 4 arrays giving a value
of 11.2microns for the multiplex-gainwavelength-coverage product.
Even allowing for the better throughput of a multislit device
and for the fact that not all fibres are placed on objects, a
fibre system is likely to complete a particular survey faster
than a multislit one.
Sky-subtraction.
This is the major problem with fibres and the comparison of the
sky-subtraction capability of fibres versus slits is a complex
issue. A detailed discussion of this is given in the next section.
The basic conclusion presented is that it is possible to do very
good sky subtraction with fibres and there is no fundamental reason
for believing that fibres will not give good sky-subtraction.
Imaging. Clearly,
a fibre spectrograph that sits on the dome floor and can only
be fed with fibres cannot be used as a general-purpose imager.
However, for CIRPASS this is not an important science driver.
For a multislit system it is possible to replace the grating with
a mirror to use it as an imager. However, given the fast f/1.2
camera and the 0.4 arcsec/pixel scale the utility of doing this
is not ideal - an image scale of 0.1-0.2 arcsec/pixel is more
appropriate. Furthermore, at f/1.2 a broadband image will saturate
the array faster than it can be read out.
Transferability.
For a fibre system matching the f-ratio of the telescope is achieved
by the choice of microlens that feeds the fibre at the telescope
focal plane. Transfer to another telescope therefore is simply
a case of making a new fibre feed and making a new plate-holder
unit to suit the new telescope. For a multislit system transfer
to another telescope requires making a new collimator. The mechanical
interface requirements are also more complicated because the whole
spectrograph has to go on the telescope.
Cost. Comparing
a fibre sytem with a multislit sytem many of the component are
the same (gratings, cameras, detectors etc.). For a fibre system
there is the extra cost of the fibres themselves. For a multislit
system there is the increased cost of the mechanical structure
which has to be light-weight, compact and flexure-free because
it has to be mounted on the telescope. In the case of an IR instrument
such as CIRPASS this structure also has to be cryogenic. Furthermore,
to carry out imaging at ~f/3 more optics are required. The cost
of meeting these requirements more than offsets the cost of the
fibre feed and a multislit system will be more expensive than
a multi-fibre system. So far in the design study we concentrated
on fibre systems and our current cost estimates are ~£500-600K
(excluding detectors). We do not have a detailed cost estimate
of a multislit system.
Capability | ||
Throughput (above atmosphere to data file) | ||
Field of view | ||
Multiplex-gainwavelength-coverage product | ||
Sky-subtraction | ||
Imaging | ||
Transferability between telescopes | ||
Cost |
It is generally believed that sky-subtraction with
a slit-based system is more accurate than with a fibre-based system
and to do spectroscopy at the faintest possible levels one should
avoid using fibres because they simply cannot be made to work.
This view is however incorrect and the limiting signal-to-noise
ratio that can be achieved for both fibres and slits is fundamentally
set by Poisson statistics. Several workers have made very deep
observations with fibres to demonstrate this. Elston and Barden
(NOAO newsletter #19, 1 Sept 1989) reached a magnitude of R=22.3
(approximately B=24) by using a beam-switching technique. This
magnitude limit compares well with that of deep multislit surveys
(e.g. LDSS-2). They used a plug-plate fibre system on the 4m telescope
at Kitt Peak and their on-source exposure time was 3 hours. More
recently Cuby and Mignoli (SPIE vol 2198, 99, 1994) showed that
the theoretically faintest limits could be achieved without beam-switching
thus getting the largest possible multiplex gain. They used MEFOS
on the ESO 3.6m telescope and successfully observed objects as
faint as bJ=22.6 in an exposure time of only 90 minutes.
So why are most people sceptical about using fibres
for ultra-faint spectroscopy? Basically it is because with fibres
one has to take much greater care when making the observations
and doing the data-reduction in order to eliminate systematic
errors. In order to achieve good sky-subtraction one has to derive
an object-plus-sky spectrum and a sky spectrum that very accurately
estimates the sky component of the object-plus-sky spectrum. For
a multislit system the sky spectrum and the sky-plus-object spectrum
are adjacent both on the sky and on the detector and the systematic
errors in both are practically the same and therefore cancel out
in the sky-subtraction process. Interpolating along the slit using
blank sky regions either side of the object helps enormously here.
However, for a fibre system the two spectra to be subtracted from
one another are typically neither adjacent on the detector or
on the sky and so the systematic errors associated with each often
do not cancel completely if a simplistic data reduction algorithm
is used. The most dominant systematic errors are due to a lack
of adjacency at the detector and include wavelength sampling,
scattered light and spectrograph vignetting. What the astronomers
from NOAO and ESO mentioned above have demonstrated is that with
a little care and attention these systematic effects can be successfully
dealt with to give results as good as those of a multislit system.
So one can perhaps justify the poor reputation that
fibres have because they are indeed a little harder to use. However
there are further reasons why fibres have a much poorer reputation
than they deserve. A common misconception about fibres systems
is that their transmission varies with time (for a given field
configuration) because the fibres twist and bend as the telescope
tracks, making flat-fielding inherently inaccurate. This is simply
not true for modern robot based systems and is probably a throw-back
to the very early days when loose-fitting fibres where sometimes
used in poorly manufactured plug-plates. Well engineered plug-plates
and fibre ferrules can certainly have the stability required.
Another problem is that examples of faint work with fibres that
would convince the sceptics are extremely rare. One reason for
this is that fibre systems typically can access a very large field
compared to a multislit device but until quite recently it has
been very hard to obtain the photometry and astrometry for a very
faint sample over such a large field because photography rather
than CCD images were the only data available. Conversely, when
deep CCD images are available to provide targets for a deep spectroscopic
survey these have been well matched to the field sizes of multislit
spectrographs and therefore these have been used to do the work
because of the perceived risks of fibre work. Finally, whenever
fibres are used the temptation to deploy most of the fibres on
objects and very few on blank sky has also tended to push fibre
users away from faint samples.
In the case of CIRPASS there are several features which will greatly help make the sky-subtraction process accurate. CIRPASS will have a small number of configuration choices and can therefore be operated essentially as a fixed-format device. This means that many calibration procedures to determine wavelength calibration, scattered light contributions and spectrograph vignetting can be repeated and the properties of the system can be accurately measured and understood and stored as a database. This coupled with a sophisticated, dedicated data reduction package will make sky-subtraction both accurate and painless for the observer. A white pupil Baranne design also helps because fibres at the ends of the slit have similar spectrograph vignetting to those at the centre and so the colour of the sky spectrum to be subtracted off any individual fibre is invariant . For redshift work which just requires the identification of features and the determination of their positions, the sky to be subtracted off is relatively featureless because of the digital OH suppression. Finally, the spectrographs do not move with respect to the gravity vector giving great stability.
The number of photons per pixel detected from the
sky background is very low so it is vital that CIRPASS contributes
significantly less than this in terms of the background due to
thermal emission from within the spectrograph. Since the detectors
have a dark current of about 0.1 electron/second/pixel and this
is below the background from the sky a useful goal is to ensure
that the thermal emission is less or equal to this.
To calculate the thermal background detected we have
made the following assumptions.
The detected flux for a small wavelength bin (0.01microns in this case) can be calculated from the blackbody equation allowing for the detector QE, the filter transmission and the transmission of the optics. Integrating a series of these values over the range 1.0-2.6microns then gives the total detected flux. Note that the flux emitted from a patch of black-body of size equal to one detector pixel in to a solid angle defined by f/1.2 is the same as the flux landing on a single pixel in the absence of other losses due to the conservation of surface brightness - so calculating the detected flux is equivalent to calculating the flux emitted from a pixel-sized patch of a black-body.
In doing this calculation the main uncertainties
are the values for the blocking factor of the filter, t, and the
emisivity, e. The following temperatures for the spectrograph
were required to make the thermal background equal to 0.1 electrons/sec/pixel.
Emisivity=0.1 | Emisivity=1.0 | |
t=1e-04 | 233.6K | 217.1k |
t=1e-05 | 236.6K | 221.4K |
A blocking factor of 1e-04 is quite realistic and
it may be possible for the filter manufacturers to achieve 1e-05.
It should also be possible to reduce the emisivity of the inside
of the spectrograph by making a lot of the surfaces shiny. Furthermore,
each pixel in the detector can really only see the optical components
of the spectrograph and not the mechanical structure and these
will typically have low emisivity. It can be seen from the above
table that in order to keep the thermal background to an acceptably
low level the temperature of the spectrograph has to be between
217K and 236K or -56C to -37C. Obviously if the spectrograph is
colder than 217K the background will be reduced to even lower
values. The analysis shows that despite the uncertainties in knowing
e and t for the real instrument the uncertainty in the required
temperature is only about 20 degrees.
The spectrum of the thermal background for each of
the 4 cases in the table is shown in the figures below. It can
be seen that the spectrum consists of 2 peaks - one near 1.8microns
just before the filter switches on and one near 2.5m where the
very strong thermal emission comes through despite the blocking
filter. It can be seen that even when the blocking filter has
the less demanding specification of e=1e-4 a significant fraction
of the background still comes from the 1.8microns peak.
CIRPASS's instrument specifications are defined in the scientific case. This section is concerned with the issues of technical feasibility, practical design options, and compromises inherent to different design forms. From the optical design point of view the most important specifications are:
The three major technical issues that we have addressed so far are schemes that give a simultaneous spectrum coverage of 1-1.8m, schemes that give the required spectral resolution with a practical grating/beamsize combination and the design of fast IR cameras.
We have assumed CIRPASS will be a plug-plate fibre
system with a single fibre per galaxy. To feed the light into
the fibre efficiently a micro lens is used to image the telescope
pupil on to the fibre at f/4. This reduces focal ratio degradation
(FRD) in the fibre. A small field stop ahead of the microlens
and in the telescope's focal plane defines the 1.5arcsec aperture.
For an 8m telescope the pupil image on the fibre has a diameter
of 233m.
Given 4, 10242 detectors and a wavelength
coverage of 1-1.8m, our starting point was to attempt to produce
a design in which the whole J band uses 2048 pixels (2 detector
widths) and likewise the whole H band uses the other 2048 pixels.
This satisfies the resolution requirement as long as 2 pixels
corresponds to a spectral resolution element. To obtain sufficient
dispersion (assuming an f/1.2 camera) a 300g/mm grating working
in second order with a 120mm beam diameter is required. These
parameters have driven the optical designs we have looked at so
far.
However, with an f/1.2 camera (which is about as fast as we can achieve) the fibre size for 2 pixel matching is only 0.79" so our 1.5" fibres are 2 times too big. To fix this problem we developed a "switchyard" scheme in which a single fibre efficiently feeds its light into 7, 0.8" fibres and this is described in section 5.2. If the switchyard (which acts like an image slicer) is not used then we have to increase one or more of the beamsize, the grating ruling frequency or the spectral order to give an overall factor of 2. Pushing the design in this direction is difficult and expensive.
The exact manner in which information content is
distributed between wavelength coverage and multiplex gain is
a matter of observing strategy, technical feasibility, and the
observer's personal preference. Is simultaneous wavelength coverage
from 1-1.8m necessary or is it only a desirable feature? Our initial
design efforts were aimed at producing a large wavelength coverage
with a correspondingly reduced multiplex gain.
There are 3 ways that we can think of in which a
spectrum can be broken down into sections and captured by our
four non-buttable detectors. These are:
So far in this study we have mainly considered the
first 2 of these.
The illustration below shows how dichroic beam splitters
could be used to feed four spectrographs simultaneously. In this
manner the light from each galaxy is fed into four different spectrographs
and simultaneous wavelength coverage is achieved. In practice,
however, dichroic beam splitters are limited by the sharpness
of their spectral profiles. In mixed polarization light, the switch
from transparency to opacity spans about 0.1m. There is no trouble
using beam splitters to separate J from H as there is a very useful
water absorption band stretching from 1.35microns to 1.45microns,
and that is how the COHSI dichroic is used. But intraband splitting
is more messy and problematic.
This leads us to the second method of spectrum redistribution
which is to use tilted mirrors at an intermediate spectrum to
redirect different portions of the spectrum onto different detectors.
This is the principle used by WF/PC on the HST (albeit for an
imaging system) and is also used by the proposed AUSTRALIS spectrograph
for the VLT. This technique gives a much cleaner cut between wavelength
regions than a dichroic can.
In view of the above our initial design form uses
a dichroic beam splitter to separate J & H. Each band is then
fed into its own spectrograph and the intraband splitting is done
with redirecting mirrors at an intermediate focus.
The purpose of this module is twofold: to convert
a large fibre diameter in to several smaller ones (to get increased
spectral resolution) and to partition the spectrum in to two separate
fibre feeds (one each for J and H).
Light from the 250micron core diameter fibre coming from the telescope is fed into 7 more manageable 130micron core diameter fibres which then deliver the light to the spectrograph. This is done by projecting the image of each 250micron fibre onto a cluster of 7 hexagonal arrays, which in turn each feed a 130micron fiber. The exact size of the output fibre and the f-ratio it is fed at depend on the input f-ratio of the spectrograph and it is simple to design the hexagonal lenslets to give the required f-ratio. The layout of the MOS switchyard is shown in Figure 1, and the details of the optical elements are shown in Figures 1a and 1b. The lens complex L1 collimates the f/4 beam from the telescope, while the doublet L2 provides the magnification necessary to project the 250micron fibre onto a 6 mm spot. A small dichroic beam splitter (not shown) is positioned at P to split J & H and feed two spectrographs simultaneously. The field lens, in conjunction with the hexagonal lenslets, is used to properly re-image the pupil onto the fibre face. Figure 1c shows the distribution of the lenslet clusters as well as how a 250micron fibre is imaged onto a lenslet cluster. Each lenslet is 3 mm from corner to opposite corner. The principle of the lenslet construction is borrowed directly from the SPIRAL concept for which the phase-A prototype has been successfully manufactured at the IoA.
This is the first spectrograph design form we explored
as it avoids the anamorphic distortion inherent in a classical
spectrograph. Anamorphic distortion prevents us from matching
the fiber diameter size to two pixels along the dispersion and
spatial direction simultaneously, thus resulting in loss of either
multiplex gain, spectral coverage, or spectral resolution. So
from an information content viewpoint, this is the preferred design
form. It also in principle allows wavelength splitting because
it has an intermediate spectrum position.
The layout of the Baranne spectrograph is shown in
Figure 2. It is essentially a modified version of the COHSI design.
The fibre slit and the spectrum are both situated on the curved
focal surface of a Schmidt camera. Unlike COHSI, where we are
making use of the water absorption band in between the J &
H spectrum to position the fibre slit at the very centre of the
Schmidt camera, we are forced to offset the slit to the end of
the spectrum, as there is no available gap in the spectrum. (Remember
that the J & H split has already occurred in the switchyard).
This, in turn, impacts adversely on the optical performance of
the system.
The primary mirror, working at f/5, generates a 120
mm diameter beam which is then projected onto the grating (standard
Milton Roy 300 g/mm grating used in 2nd order). The spectrum is
then imaged at the center of the Schmidt by the primary mirror.
The question raised at this point is how to extract
the spectrum from the center of the spectrograph, split it into
two portions to be directed onto two independent detectors. Our
attempts to do so using relay mirrors, while coming tantalisingly
close to fruition, proved unsuccessful. Using flat mirrors to
redirect the spectral portions onto different cameras suffers
from the problem that the imaging quality inherent to the Schmidt
is not adequate. Using curved mirrors and a second pass through
the Schmidt to correct the aberrations only led to the introduction
of new ones as the symmetry of the system was broken. This symmetry
can easily be recovered but is totally unaffordable (essentially
consists of cutting the primary mirror into three pieces). We
could decide not to split the spectra at the center of the Schmidt,
but generate another intermediate image and split the spectra
at this additional image. But this entails more optics, lower
throughput, and higher cost.
In short, we do not have a viable design for a system
with intraband splitting at an intermediate spectrum. Simultaneous
spectral coverage over the entire J & H band looks technologically
very difficult. We decided at this point to consider having just
one detector per spectrograph instead of two.
By dropping the requirement that the intermediate
spectrum is big enough to feed two detectors we were able to achieve
a working design of a Baranne spectrograph. Because the spectrograph
now has only half of the original wavelength coverage, this allows
us to position the entrance fiber slit much nearer to the center
of the Schmidt camera and obtain adequate image quality. This
is the system presented in Figure 2. This system could feed two
spectrographs simultaneously by simply putting a second fiber
slit, grating, camera and detector rotated 90o about
the Z-axis (the axis passing through the center of the aspheric
plate and the primary mirror). This helps to significantly reduce
the cost as only two Schmidt cameras would be required instead
of four.
A spreadsheet showing the estimated system throughput is attached. The peak efficiency is 23%. This spreadsheet has two dichroics in the light path to deploy 4 such spectrographs with maximum wavelength coverage. These dichroics are not needed if we have 4 times as many objects and 4 times less wavelength coverage. The multiplex-gainwavelength-coverage product for this design is ~38m.
The difficulty inherent in the classical spectrograph
stems from the contradictory constraints of keeping anamorphism
small while keeping the camera design workable. Reducing anamorphism
requires that the camera-collimator angle be kept to a minimum.
But a smaller angle translates into a greater separation between
the grating (pupil) and the first camera lens, therefore requiring
larger lenses (which is a major problem with infrared materials).
It also does not have an intermediate spectrum which could be
used for wavelength splitting.
Our study of fast IR cameras showed that a separation of 300mm between the pupil and the first lens of the system is about as far as we can go. We therefore selected a camera-collimator angle of 30 so as to satisfy this practical constraint and live with the anamorphism introduced.
The layout of the classical spectrograph is presented in Figure 3. The beam size and grating characteristics are the same as those used in the Baranne system. Because there is no relay mirror present in the center of the Schmidt collimator, we are able to position the slit at the center of the Schmidt collimator and work at f/4 (which is more fiber friendly that f/5) while still obtaining optical performances far superior to those achieved by the Baranne Schmidt.
Note also the much smaller size of the primary mirror
and aspheric plate of this design as compared to the Baranne system.
Since we are using only a portion of the aspheric plate, cutting
a single aspheric plate into quarters will provide all four spectrograph
with their aspheric corrector.
Figure 3a presents a cut of the system showing how
the spectrum is spread onto the detector for one of the grating
settings (short-J; 1.025m - 1.20m). The fiber slit spatial direction
is into the page. This diagram also shows the details of the f/1.2
camera design. The spot diagrams of this partially optimised design,
presented in Figures 3b-d, already shows adequate imaging quality.
The optical performance for the other grating settings is of similar
quality. The exact same camera is used at all grating settings
(i.e. in practice we would purchase four identical lenses from
the manufacturer). Table 3 provides detailed information on the
characteristics of this design.
This design is of the "normal to collimator"
form with the beam diameter reduced in the spectral direction.
The fibre slit is matched to two pixels in the spatial direction.
There is no loss of multiplex gain, but the higher dispersion
leads to a smaller wavelength coverage. If the preference is to
sacrifice multiplex gain instead of wavelength coverage, then
an Ebert configuration should be selected. Which of these two
alternatives is optimal is not yet clear. Oversampling the fibre
images in the wavelength direction will help sky subtraction.
Note that it is not simply a case of choosing the grating orientation
at the telescope - the two alternatives require different beam
sizes and therefore different collimators.
A spreadsheet showing the estimated system throughput is attached. The peak efficiency is 24%. This spreadsheet has two dichroics in the light path to deploy 4 such spectrographs with maximum wavelength coverage. These dichroics are not needed if we have 4 times as many objects and 4 times less wavelength coverage. The multiplex-gainwavelength-coverage product for this design is ~34m.
A cross-dispersed echelle spectrograph with a low-order echelle grating could give the full wavelength coverage of 1.0-1.8m on one detector with a multiplex gain of ~25. In this case it not necessary to put the full wavelength coverage on to 4 chip-widths and so it is possible to have more than 2 pixels per resolution element and consequently the switchyard is not needed for either wavelength or spatial partitioning. For example an existing 75g/mm grating with a blaze angle of 26.7 could be used to give 0.926-1.813m in 7th to 12th order with a 3 pixel spectral resolution of R~2560. This would have a fibre diameter of 1.2 arcsec on an 8m and a multiplex-gainwavelength-coverage product of 88.7m. The high value here is mainly due to the fibre diameter and spectral resolution being below specification but this is clearly a powerful option. Drawbacks of the echelle approach are scattered light and the lower grating efficiency off blaze. Clearly this approach has to be investigated further.
Most applications for integral field spectroscopy require field for spatially mapping more wavelength coverage. The fibres used are also small in terms of their aperture on the sky compared to the 1.5" we require for MOS. These two points mean that the integral field mode is not really driving the spectrograph design and can therefore be fitted to the spectrographs in a straightforward manner.
Assuming that CIRPASS will be used on a telescope
with an aperture of 8m, the main conclusions of our study so far
and recommendations for further work are: