Control and Data System

The Amber Engineering (AE 159) focal plane array has 128 128 detector elements, each 50 m square on a Si-CMOS multiplexer. The detector material is Si:Ga, providing sensitivity from around 5 m to 18 m in wavelength. Other detector characteristics are provided in Table 2.

 
Table 2: Detector Characteristics.

Some of the best features of this detector are its very deep wells, which coupled with fast readout electronics, make broadband imaging at the AAT possible, its very good uniformity, and operability (only four dead pixels in 16384). The sensitivity achieved to date, at the AAT is 100 mJy/arcsec/1 /1 min, = 12.5m, =1m. This however, is an on source sensitivity, and when chopping off chip is used this is a factor of 2.1 longer for actual elapsed time. Sensitivities obtained in the various filters at the AAT are included in Table 1. The quoted sensitivities were obtained at the AAT, with a sky temperature of about 5 C and 60 - 70 % relative humidity, in straight imaging mode. In polarimetry mode the sensitivity is reduced by a factor of about by the inclusion of the wire grid analyser (< 50% transmission), and the warm waveplate which is a few per cent emissive. Sensitivities better by a factor of two or more would be expected when observing from a high altitude, cold, dry site like Mauna Kea observatory in Hawaii.

The detector has two analog outputs, which are fed through trans-impedance amplifiers and some other signal conditioning to two Analogic ADAM-846-1, 400 kHz, 16 bit, sample-and-hold, A/D converters. Sixteen bit A/D converters are required so that the noise can be adequately sampled by the A/D without the requirement for different gain or offset readings. The shot noise from a full well of 2 photo-electrons is 4472 electrons, and a 16 bit A/D which covers the full dynamic range of the detector has 305 photo-electrons per A/D unit. A/D converters with fewer bits would cause excessive digitization of the noise, particularly in low background situations where the detector well may be less than half full.

The nominal Sample-and-Hold plus A/D conversion time for the ADAM-846-1 is 2.4 s per conversion, but we are running them nearer to 2 s per conversion with no discernible loss in performance. Even so, the readout rate for the data system is still A/D conversion time limited. The two channels are readout simultaneously, so the minimum readout time for the array is 1.92 s per pixel and 17.7 ms per array, which corresponds to a frame rate of 56 frames/s. This read rate, coupled with the deep wells (2 10 electrons) of the AE159 means that the readout electronics can handle detected photon rates of up to 1.1 10 ph/s.

The digitized data is then co-added on two 50 MHz Digital Signal Processors (Ariel Corp PC-25+ boards using a Texas Instruments TMS320C25/50 DSP chip). The analogue and digital electronics are separated by opto-isolators. A third DSP board generates the necessary clocking signals to operate the FPA and initiate A/D data conversion, as well as controlling the chop frequency of the oscillating secondary.

As well as cyclic clocking, the Amber AE159 chip has the ability to change its readout frequency from DC to 200 frames/s in real time, which means that it is not necessary to continue clocking and reading out the array when no useful function is performed. We operate the detector in this way to minimize non-productive time spent by the detector. The chip is always read out at the maximum possible rate for the data system (17.6 ms full frame), and longer integrations are obtained by suspending the clock for the required integration period, and then resuming clocking to readout the array. Integration times are increased from the minimum of 17.6 ms in one milli-second increments. This clocking method has no particular benefit in a STARE mode but when chopping is required we lose the minimum possible time waiting for the chop to settle. In this case the array is read out after an arbitrary length integration, clocking is then suspended, and the chop signal is generated. After the required amount of time for the chop to settle (5 - 10 ms) at the AAT, the array is clocked out, again at maximum speed, to flush unwanted integration while the secondary was moving, and then the real data integration is re-started. This means we only need to lose around 25 ms at each chop, even though the integration time may be much longer. This readout scheme is not required for broad band imaging at the fastest frame rates, but when narrow band filters are used, particularly the SIV and NeII filters, integrations become a significant fraction of a second to ensure the wells are reasonably filled and noise is dominated by photon noise rather than readout noise. Using this readout scheme does not seem to incur any penalty in extra readout noise from the array.

The DSPs are housed in a computer mounted at the telescope, and as the DSPs perform all of the time critical functions, we are able to use a relatively modest 486DX2-66MHz PC as the data acquisition control computer mounted at the telescope. The digital signal processor boards are the heart of the data acquisition system, and as throughput is generally a premium in mid-infrared instrumentation all of the time critical acquisition software was written in assembler and is uploaded to the DSP's from the control computer at the telescope. The DSP software also provides co-adding and co-subtracting of data and chop frames. The control software is designed to be as flexible as possible to allow for many different observing modes. The DSPs can change integration times, and also the number of coadded integrations per chop cycle and and number of coadded chop cycles before downloading to the control computer. In this way we optimize observing efficiency depending on conditions. In broad-band imaging we use short integrations, but with a number of coadds per chop cycle to keep the chop at what seems to be an optimal rate of a few Hz. Chopping faster than this reduces observing/integrating efficiency, but chopping slower provides increased noise.

The data is periodically downloaded to the host PC through direct memory access and then sent via an ethernet link to the main control computer where the data is processed, displayed and stored to disk. This communication and processing all occurs while the DSPs and the detector are coadding and integrating again, and so causes no efficiency loss.

Although there has been considerable discussion as to the necessity of chopping with array detectors in the thermal infrared, most evidence to date indicates that chopping at frequencies less than 1 Hz does incur an increased noise penalty. We have maintained an observing strategy of slow chopping and beamswitching to minimize the effects of sky background variations. However, we are still experimenting with other STARE type observing strategies. Given the necessity to chop, observing can only ever be 50% efficient on source, although on compact objects (i.e. a few arcseconds extension) and small chop/beamswitch throws (< 15 arcseconds) it is possible to chop and beamswitch ``on chip'' to avoid the 50% loss in efficiency. Chop and beamswitch on chip (which produces four images, two positive and two negative) also reduces the chance for misregistration of the beam switch images. Otherwise great care is required when setting up the beam switch to ensure both positive and negative images fall on the same place on the detector. Whatever the case, our readout system is 97% efficient in spending time actually integrating (whether on sky or source) in imaging mode, and about 85% efficient in polarimetric mode where there are other overheads involved with waveplate rotation etc.

A third PC, also networked to the others by ethernet, provides off-line, but very nearly real time, reduction of the incoming data. By the end of the observation on an object, data coadding is complete, including shift and add if required. Flat fielding and flux calibration are applied in post-reduction. Both of these reductions could be included in real time reduction, but to date, we have found that calibrations have tended not to be firm until the end of the run and flat fielding is often unnecessary. The uniformity of the Amber array is such that images which are produced from chopped data are usually already flat and further flat-fielding only adds more noise. An exception to this was the 20 m images where significant pixel-to-pixel non-uniformities were found and flat-fielding was required. In polarimetry mode though the external waveplate produces some vignetting of the field and more care with flat fielding is required here.

All of the control and reduction software was written in ``C'' . For image plotting routines it was found that normal ``C'' graphics routines were far too slow, so routines which directly address the PC graphics adapter memory were developed.

© Copyright Astronomical Society of Australia 1997