Noiseless imaging detector for adaptive optics with kHz frame rates
John Vallerga
*a
, Jason McPhate
a
, Bettina Mikulec
b
, Anton Tremsin
a
, Allan Clark
b
and Oswald
Siegmund
a
a
Space Sciences Laboratory, 7 Gauss Way, Univ. of California, Berkeley, CA USA 94720-7450;
b
University of Geneva, 24, quai Ernest-Ansermet, 1211 Geneva 4, Switzerland
ABSTRACT
A new hybrid optical detector is described that has many of the attributes desired for the next generation AO
wavefront sensors. The detector consists of a proximity focused MCP read out by four multi-pixel application specific
integrated circuit (ASIC) chips developed at CERN (“Medipix2”) with individual pixels that amplify, discriminate and
count input events. The detector has 512 x 512 pixels, zero readout noise (photon counting) and can be read out at 1kHz
frame rates. The Medipix2 readout chips can be electronically shuttered down to a temporal window of a few
microseconds with an accuracy of 10 nanoseconds. When used in a Shack-Hartman style wavefront sensor, it should be
able to centroid approximately 5000 spots using 7 x7 pixel sub-apertures resulting in very linear, off-null error correction
terms. The quantum efficiency depends on the optical photocathode chosen for the bandpass of interest. A three year
development effort for this detector technology has just been funded as part of the first Adaptive Optics Development
Program managed by the National Optical Astronomy Observatory.
Keywords: adaptive optics, wavefront sensors, MCP, Medipix, GaAs photocathodes
1. INTRODUCTION
The white paper entitled “A Roadmap for the Development of Astronomical Adaptive Optics”
1
spells out the desired
properties for the next generation of detectors for wavefront sensing: very fast, very low noise, and many pixel elements.
Specific goals cited were 512 x 512 arrays with 1 kHz frame rates, 1-3 electrons noise per pixel and optical/infrared
quantum efficiencies (QE) > 80%. Currently, most array-based wavefront sensor (WFS) detectors are silicon charge
coupled devices (CCDs) which can achieve the QE goal but currently have readout schemes that are incapable of
simultaneously achieving fast frame rates and low readout noise. For example, the Palomar AO wavefront sensor
detector (EEV39 CCD 64 x 64 pixels) can run at frame rates of 1100 Hz at 7.5 e
-
rms noise or 50 Hz at 3e
-
rms noise
2
.
In a CCD, the charge generated by the photon flux at each pixel is transferred to a high gain, low noise amplifier to
create an analog voltage for digitization. The noise associated with this amplification increases at faster clocking rates.
A typical read noise for 1 MHz clocking rates is ~5 e- noise rms. Scaling from these numbers, a 32 x 32 CCD could be
read out in 1 millisecond at a 1MHz pixel rate. So to read out an array of 512 x 512 would require effectively 256 CCD
readouts in parallel. And to reduce the noise by a factor of two, the bandwidth would have to be reduced by a factor of 4,
resulting in 1024 CCD type readouts to achieve the frame rate goals of the AO roadmap, given current amplifier
technology.
A recent new CCD technology is the “Low Light Level” CCD (L3CCD) by E2V Technologies. It uses a special
serial shift register that amplifies the pixel charge at every pixel transfer before the output preamp, resulting in output
gains of up to 1000 e
-
per input photon, easily overwhelming the few electron readout noise of the amplifier.
Unfortunately, this amplification process in the shift register is not a noiseless process, and at the higher gains, the
variance of the signal charge doubles
3
. So to achieve the same shot noise limited signal to noise ratio, you would have to
collect twice the number of photons. Also, for the larger CCDs (E2V 97 512x512) the maximum readout rate is 10MHz
corresponding to 40 frames/sec.
An alternative to charge integrating arrays are photon counting detectors. Each photon is detected and registered as
one count, and hence, no “read noise”. Examples in the UV and optical include avalanche photodiodes (APDs) and
imaging MCP detectors. Silicon based APDs are fast and have the high QE of silicon in the optical and near IR but are
not yet incorporated in large arrays. Imaging MCP detectors can have large area (100 x 100 mm), high spatial resolution
(20 µm FWHM), low background dark count, and event timing resolution less than 1 nanosec
4
. Their QE is determined
by the characteristics of the photocathode material that absorbs the initial photon and releases the photoelectron. The QE
of these photocathodes is now approaching 50%, in particular GaAs and GaAsP (Fig. 11).
*jvv@ssl.berkeley.edu; phone 1 510 643 5666
Advancements in Adaptive Optics, edited by Domenico Bonaccini Calia, Brent L. Ellerbroek,
Roberto Ragazzoni, Proceedings of SPIE Vol. 5490 (SPIE, Bellingham, WA, 2004)
0277-786X/04/$15 · doi: 10.1117/12.553017
1256
CERN-OPEN-2006-062
22/06/2004
The problem with using this type of MCP detector as a WFS right now it the inherently serial nature of the readout.
Most high resolution readout anodes handle photon events one at a time. For example, delay line anodes, which use the
time difference of the anode signal traversing across the anode to determine the photon position, can only allow one
event on the anode at a time, otherwise “pileup” would lead to an incorrect position determination. Delay line lengths of
~50 ns are typical for larger detectors, so even global event rates of 2 MHz result in a 10% pileup fraction, assuming no
additional event overhead (ADC conversion, etc). The count rate expected for a strawman Shack Hartman WFS with
5000 lenslets, each focusing to a spot of 1000 events, and updating at 1 kHz rates corresponds to 5000 x 1000 x 1000 = 5
Giga events per second, 3 orders of magnitude faster than some of the fastest imaging photon counters. Clearly some
type of integrating detector is required for the WFS application, yet keeping the “noiseless” operation of a photon
counter is very desirable, as it is a crucial characteristic needed for accurate centroid positioning.
Most Shack Hartman WFS image the individual sub-pupils sampled by the lenslet array onto a 2x2 pixel subarray
(the “quad cell”), and use a normalized difference algorithm to measure the centroid of the spot which represents the
slope of the wavefront, e.g. x
cent
= (A-B)/(A+B). However, if the input spot illumination is symetrically shaped with a
width on the order of a pixel, then the measured centroid of the input vs. its true centroid is not a linear function of the
true position. Furthermore, as the illumination input changes width, the slope of the relationship between measured and
true centroid changes. This slope is related to the gain of the feedback to the actuators. So in variable seeing conditions,
where the spot sizes are changing, poor phase control can result.
If more pixels are applied to the measurement, this “sampling distortion” can be substantially reduced. In Fig. 1 are
shown simple one dimensional models of various Gaussian width input functions whose measured centroid is plotted
against their true input centroid. No read noise or input shot noise was assumed, deviations from linearity are strictly due
to the distortions inherent in the assumed sampling algorithm. In the quad cell case (2 samples in one dimension) the
non-linearity vs. input centroid position is apparent for a single input width, as is the varying slope for the different
widths of the input function. As the number of samples of the Gaussian are increased, (right side of Fig. 1) the non-
linearity is decreased as is the sensitivity to width variations. So even with infinite signal to noise ratio, the quad cell can
generate errors in the output centroid and therefore errors in the phase determination.
So why do most AO Shack Hartman systems with array detectors use quad cell algorithms? The answer is noise and
speed: less pixels mean faster readouts. The centroid variance due to fixed readout noise per pixel for an N x N sampling
box scales as N
2
(if the same spatial area is sampled, i.e. the error is in microns, not pixels). If the goal is a centroid
determination limited only by the shot noise of the input flux, then more input flux is required to achieve the same SNR.
Which of these reasons is dominant must depend on the particular instrument and application (brightness and color of
guide star, size and speed of CCD, quality and stability of seeing), but if an array detector with zero readout noise existed
and was fast, there would be no hesitation to use more pixels in the centroid determination.
Fig. 1. Model of measured centroid vs. true centroid using pixellated sampling of a Gaussian input function in one dimension.
The plotted curves represent different widths of the Gaussian input as a fraction of pixel size and the centroid algorithm is
a
“center of gravity” type where x
cent
is the sum of the used pixel values weighted by the pixel number and divided by the sum o
f
the pixel values. Note the input centroid scales are one pixel wide while the output centroid determinations are different for th
e
two plots. The slope of the quad cell algorithm curves change dramatically with input width while the 5x5 algorithm slopes are
similar for different in
p
ut widths exce
p
t for the narrowest in
p
ut function.
5x5 algorithm error for Gaussian input
-0.5
0
0.5
-1 -0.5 0
Gaussian Centroid true position (center pixel)
Calculated position
Sigma = 0.2
Sigma = 0.4
Sigma = 0.6
Sigma = 0.8
Sigma = 1.0
Quad cell (2x2) algorithm error for Gaussian input
-1
0
1
-0.5 0 0.5
Gaussian Centroid true position
Calculated position
Sigma = 0.2
Sigma = 0.4
Sigma = 0.6
Sigma = 0.8
Sigma = 1.0
Proc. of SPIE Vol. 5490 1257
Another issue to point out is that for future large area telescopes using laser guide stars, the location of the scattering
layer in the atmosphere, whether it be the ~ 100 km of the sodium layer or closer for Rayleigh scattering, will not be in
focus, i.e. the wavefront will be curved when a distant astrophysical source wavefront is parallel. This will put most of
the Shack Hartman spot images “off null”, away from the center of the standard quad cell configuration where the non-
linearities are greater. More pixels used in the spot sampling results in more accurate and repeatable centroids off null.
2. A NOISELESS, IMAGING MCP DETECTOR AT KHZ FRAME RATES
To take advantage of the noiseless operation of a photon counting detector using MCPs requires a specialized
readout that can localize and count the events in a massively parallel way, i.e. pixellated counters. We soon came up with
the idea of an idealized application specific integrated circuit (ASIC) pixellated readout behind an MCP in an optical
tube with a high QE photocathode. Each 50 x 50 micron pixel of the ASIC would amplify the ~10
4
electrons per photon
event into a pulse that would trigger a discriminator which could then be integrated by a counter until read out (and reset)
digitally. We envisioned a 512 x 512 array to support 70 x 70 Shack Hartman spots each focused on a set of 5x5 pixels.
The count rate per spot from the strawman requirement was 1000 events at 1 kHz or a megahertz, so the maximum rate
per pixel would be at most 200 kHz. The array would have to read out in approximately 1 ms. We have been using
various ASICs recently to read out cross strips anodes in our high resolution MCP detectors (< 5µm FWHM
4
) that
include linear arrays of 128 amplifiers, shapers and discriminators, and realized that such an idealized ASIC would be
difficult and expensive to develop on our own. Fortunately, we soon discovered it already existed, being designed and
constructed by the Microelectronics Group at CERN for the multi-national MEDIPIX collaboration
(http://www.cern.ch/medipix).
Our novel detector scheme should achieve the first
three of the specific goals listed above with QEs
approaching 40%. The detector (Fig. 2) is a
microchannel plate (MCP) image tube with a gallium
arsenide (GaAs) photocathode and a new pixellated
CMOS readout chip called the “Medipix2”
5
. Photons
interacting with the photocathode release a photoelectron
that is proximity focused to the MCP input face. The
MCP amplifies this single photoelectron with a gain on
the order of 10
4
. The resultant charge cloud exits the
MCP and lands on the input pad of a Medipix2 pixel
where it is counted as one event. The counter will
integrate until it is read out in a digital, noiseless
process. Also, because the data is digital, it can be read
out at ~246 MHz pixel rates, which corresponds to a
frame readout time of 266 µs for the current Medipix2
chip.
2.1 The Medipix2 ASIC
The Medipix2 chip is a pixel detector readout chip consisting of 256 x 256 identical elements, each working in a
single photon counting mode. Each pixel consists of a preamplifier, a discriminator, and a 13 bit pseudo-random counter.
The counter logic, based on a shift register, also behaves as the input/output register for the pixel. Each cell also has an 8
bit configuration register which allows masking, testing and 3-bit individual threshold adjust for the discriminator (Fig.
3). It was designed and manufactured in a six-metal 0.25µm CMOS technology. Each 55 x 55 µm pixel contains 508
transistors (Fig. 4). The total active area of the chip is 1.98 cm
2
and is 3-side abuttable to support larger arrays (Fig. 6).
Fig. 5 shows the electrical schematic of the Medipix2 layout and how the shift registers relate to the parallel readout.
The input referred noise of the Medipix2 cell has been measured to be ~100 e
-
rms and the thresholds can be
adjusted to a consistency of ~120 e
-
rms.
5
That is why we have baselined an MCP gain of 10
4
e
-
per detected photon or a
SNR of 100. Note that this noise is in the detection circuit and all we require of the circuit is a minimum of false events.
For the minimum threshold setting of 1000 e-, a false trigger would be a 10 sigma noise fluctuation. There is a different
style of noise in the amplification process in the MCP as it is statistical in nature, starting with a single photoelectron
creating an avalanche of secondaries. At gains of 10
4
, the distribution of pulse heights is slightly peaked, but with a finite
fraction of events with lower pulse height, so the setting of the thresholds for each pixel will have an effect on the
P
hoton
e
-
Q = 10
4
e
-
Pij = Pij + 1
Window
Medipix2
MCP
Photocathode
Fig. 2. Schematic of a sealed microchannel plate image sensor.
A
single photon results in a single count in the pixel Pij. Ou
r
scheme uses the Medipix2 ASIC as the readout.
1258 Proc. of SPIE Vol. 5490
variation of the sensitivity of the device. We expect this variation to be small (~ 1%) and easily calibrated with high
signal to noise flat fields, since the input event rate can be so high.
Fig 3. Schematic of major functional blocks in a single Medipix2 pixel.
A fast charge event on the input is amplified and shaped by the preamp,
discriminated and counted at the shift register (if Shutter is disabled). The
digital number count is clocked out at high rate through the shift register
s
of the pixel column when the Shutter signal is enabled. Digita
l
configuration bits are input through this same shift register to contro
l
thresholds, masking and electrical testing
Fig 4. CAD Layout of a single pixel of the
Medipix2 ASIC. Amplifier and discriminator are
on the left side and the 13 bit shift register is on the
right.
Fig. 6. Image of input face of Medipix2 chip. The
dimensions are 1.6 cm high and 1.4 cm wide. The active
area has 256x256 pixels and is abuttable on 3 sides.
Fig. 5. Schematic of readout architecture of Medipix2 chip
organized into 3328 bit columns read out via a 256 bit fas
t
shift register with a 32 bit parallel readout
Input
Preamp
Disc.
Disc.
logic
Mux
.
13 bit
counter –
Shift
Register
Clock out
Shutter
Lower Thresh.
Disc.
Mux.
Previous Pixel
Mask bit
Analog
Digital
Upper Thresh
.
Next Pixel
Mask bit
Polarity
Proc. of SPIE Vol. 5490 1259
The Medipix2 has two working modes depending on the Shutter input. When the Shutter signal is low, the pixel is in
acquisition mode and the discriminator clocks the counter. When the Shutter signal is high, acquisition stops and an
external clock is used to shift the data from pixel to pixel out along the columns (Fig. 5). Again note that this is a
noiseless process, digital bits are being transferred, not charge like in the case of a CCD. For the Medipix2, this has been
measured at 80 MHz
6
and modeled up to 100MHz so to read out the whole array will take 266µs. Note that the current
design does not allow data collection while reading out, so the 266µs is a fixed deadtime per frame readout. At a kHz
rate, the sensor is active 73.4% of time which translates into less sensitivity. This is also true for frame transfer CCDs.
For example, a 256x256 frame transfer CCD would take 256µs for the parallel transfer assuming a 1 MHz parallel
transfer clock and a mechanical shutter would have to be closed during this time to prevent the input photons from
smearing over the columns and increasing the background levels.
The Medipix2 Shutter signal can be used as a real electronic shutter with very little deadtime, by not clocking when
it is activated. The chip will then ignore input counts by not updating the counter. This ability to gate the input data can
prove very useful for AO wavefront sensing of laser guide stars. The sodium scattering region has a ~10 km (varying)
depth which results in an extended spot when observed farther away
from the laser axis (Fig 7). Chopping or pulsing the laser into 3.3 µs
pulses would reduce the projected length of the laser guide star by a
factor of ~10 and the Shutter signal could be phased to the laser to
choose exactly what level in the scattering layer is used. The
requirement for a shutter is even more important for a Rayleigh scatter
laser guide star as the scattering layer is continuous with height and a
shutter is required to choose the acceptance height of the Rayleigh guide
star and exclude scattered light from the lower atmosphere. It also would
alleviate the need for fast shutters in the optical path such as Pokel’s
cells. It should be pointed out that there is an additional way to gate our
proposed detector by changing the bias voltage of the photocathode with
respect to the MCP. This is a standard way of achieving gate timing
accuracies at the ~nanosecond level, but the AO requirements are much
more easily achieved using this simple digital signal level provided by
the Medipix2.
The Medipix chips (Medipix1 and Medipix2) were designed at
CERN to the specification of the Medipix collaboration, a group of 15
European academic and government labs to improve x-ray and gamma
ray imaging for medical, biological and high dynamic range imaging.
(For a complete list of the consortium, see: http://web-
micfe.web.cern.ch/web-micfe/mic-fe/index.html). We have officially
joined this collaboration and were well received as the consortium
immediately saw the advantages of adapting the Medipix2 technology
to UV and optical imaging, especially in astrophysics. Future
development costs for newer versions of the Medipix2 design (mask
sets, foundry runs, and readout electronics) are shared among the collaboration. This cost sharing of chip development
and production costs is similar to that for scientific grade CCDs for large telescopes.
2.2 MCPs
Microchannel plates consist of an array of holes in a specialized glass substrate whose surface has a high secondary
electron coefficient. When biased with a high voltage across the plate, electron(s) entering a pore are accelerated and
eventually impact the channel wall, releasing more electrons which continue the process resulting in an avalanche of
electrons exiting the rear surface. Typical gains (electrons out/ electrons in) for a single MCP range up to 5 x 10
4
,
depending on the voltage applied and the length to diameter (L/d) ratio. A “standard” MCP might consist of 12µm
diameter pores spaced on 15µm centers with hexagonal packing and L/d ratios of 80:1 (1 mm thick). For high resolution
imaging using centroiding anodes, gains of 10
6
to 10
8
are usually needed to achieve the high S/N ratios for the analog
output signals. This requires either two MCPs in a “chevron” configuration where the pore bias angle (angle of pore axis
with respect to plate normal) is reversed for the second plate with respect to the first, or a “Z-stack” with 3 MCPs whose
bias angles change at the interfaces.
Fig. 7. Images of the Na laser guide star in the
Keck SH WFS focal plane showing the parallax
across the pupil resulting from observing a
n
illuminated column of emission from the side
due to the large size of the telescope. (From
Claire Max’s AO class 289C,
http://cfao.ucolick.org/~max/289C/
1260 Proc. of SPIE Vol. 5490
