Hybridization process for back-illuminated silicon Geigermode avalanche photodiode arrays
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Schuette, Daniel R. et al. “Hybridization process for backilluminated silicon Geiger-mode avalanche photodiode arrays.”
Advanced Photon Counting Techniques IV. Ed. Mark A. Itzler &
Joe C. Campbell. Orlando, Florida, USA: SPIE, 2010. 76810P-7.
© 2010 COPYRIGHT SPIE
As Published
http://dx.doi.org/10.1117/12.849356
Publisher
SPIE
Version
Final published version
Accessed
Thu May 26 18:55:49 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/61643
Terms of Use
Article is made available in accordance with the publisher's policy
and may be subject to US copyright law. Please refer to the
publisher's site for terms of use.
Detailed Terms
Hybridization process for back-illuminated silicon Geiger-mode
avalanche photodiode arrays
Daniel R. Schuette, Richard C. Westhoff, Andrew H. Loomis, Douglas J. Young, Joseph S. Ciampi,
Brian F. Aull, and Robert K. Reich
MIT Lincoln Laboratory, 244 Wood Street, Lexington MA 02420-9108
ABSTRACT
We present a unique hybridization process that permits high-performance back-illuminated silicon Geiger-mode
avalanche photodiodes (GM-APDs) to be bonded to custom CMOS readout integrated circuits (ROICs) - a hybridization
approach that enables independent optimization of the GM-APD arrays and the ROICs. The process includes oxide
bonding of silicon GM-APD arrays to a transparent support substrate followed by indium bump bonding of this layer to
a signal-processing ROIC. This hybrid detector approach can be used to fabricate imagers with high-fill-factor pixels and
enhanced quantum efficiency in the near infrared as well as large-pixel-count, small-pixel-pitch arrays with pixel-level
signal processing. In addition, the oxide bonding is compatible with high-temperature processing steps that can be used
to lower dark current and improve optical response in the ultraviolet.
Keywords: Geiger-mode avalanche photodiode, back-illumination, oxide-bonding, photon-counting, imager
1. INTRODUCTION
Bandwidth requirements associated with processing a signal at the single-photon level are a challenge for single-photon
detectors. In fast time-resolved applications traditional single-photon detectors, such as photomultiplier tubes and microchannel plates have a relatively modest number of channels to process signals at the single-photon level. This introduces
a limit on the instantaneous dynamic range [photons/sec] offered by these detectors. Increasingly, however, there is a
need for fast time-correlated optical sensing, driven by applications in areas including communications1, active
imaging2,3, and biology4. This need has pushed the development of multi-channel arrays of single-photon sensitive
detectors. While multi-channel versions of traditional technologies have been demonstrated, the most aggressive
examples of scaling to higher array sizes leverages integrated circuit techniques [references] to enable signal processing
electronics required at the pixel level.
Recently, there have been a number of successful efforts to integrate single-photon-sensitive Geiger-mode avalanche
photodiodes (GM-APDs) into commercial CMOS processes5,6,7. These efforts have yielded devices with thousands to
tens-of-thousands of channels, enabling a high degree of parallel signal processing that overcomes the instantaneous
dynamic range limits of more conventional detectors. These gains, however, come with a substantial sacrifice in
performance both because the detectors arrays share area with the pixel processing circuitry leading to poor fill factors
and the detector architecture and fabrication processes are not optimized for optical performance leading to modest
quantum efficiency (QE, < 50%) over a limited waveband (400 nm – 600 nm).
An alternative approach to integrating the GM-APD with the signal processing circuitry at the pixel level is fabrication
of a hybrid detector where the signal processing electronics and the detector arrays are fabricated onto separate layers.
An integration technique such as bump-bonding is used to connect detector elements with signal-processing pixels.
Examples of this approach applied to GM-APD arrays are InP-based detectors produced by the Optoelectronic Materials
group at MIT Lincoln Laboratory8. These devices utilize flip-chip bump bonding to produce a hybrid detector where
each detector element is directly connected to CMOS signal processing electronics, providing a degree of parallel
processing that scales directly with the array size and enabling array sizes of thousands to millions of detector elements.
These compound semiconductor single-photon detector arrays benefit from a transparent InP substrate that provides
mechanical support during the flip-chip bonding process, making detector integration ultimately possible. While thick
silicon photodiodes have been bump bonded, the relatively thin silicon GM-APDs needed for high-speed applications
have not been amenable to this integration approach.
Advanced Photon Counting Techniques IV, edited by Mark A. Itzler, Joe C. Campbell,
Proc. of SPIE Vol. 7681, 76810P · © 2010 SPIE · CCC code: 0277-786X/10/$18
1
doi: 10.1117/12.849356
Proc. of SPIE Vol. 7681 76810P-1
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms
Figure 1: Cross section of the hybrid GM-APD detector assembly illustrating the relative locations of the
quartz substrate, the thin GM-APD device layer, the CMOS ROIC and the wirebonds for the control and
readout of the device.
In the remainder of this paper, we will outline a silicon GM-APD hybrid detector fabrication method that combines a
silicon GM-APD back-illumination process, wherein the thick silicon substrate of a GM-APD device wafer is replaced
by a quartz substrate and GM-APD is integrated with the CMOS readout integrated circuits (ROIC) by flip-chip bump
bonding. The final hybrid detector structure is illustrated in cross section in Figure 1 and will be described in detail in the
following sections. The key advantage of this approach is the independent optimization of the GM-APD array and the
ROIC, enabling detector performance that is optimized (in terms of fill factor, dark count rate, quantum efficiency, and
detection timing) for a given application while making it possible to leverage commercial integrated circuit foundries to
achieve the necessary miniaturization for very-high-element-count arrays. To-date, this work has focused on the active
imaging applications of ladar and adaptive optics wavefront sensing.
2. SILICON GM-APD HYBRID DETECTOR ARCHITECTURE
Before presenting an overview of the GM-APD hybridization process a hybrid-detector architecture is described using,
as an example, a GM-APD detector architecture that has been developed for ladar. This example architecture is similar
to the design, described by Aull2, in which the GM-APD utilizes a separate absorber-multiplier structure to collect
photo-electrons from deep in the silicon bulk. The structure is a reach-through APD built in an epitaxial layer of silicon
grown on a heavily doped silicon substrate. The GM-APD is formed by a p-type buried anode implant into the silicon
with the dopant located at a small (sub-micron) separation from a surface cathode implant. A depletion region forms
when a voltage is applied to the n+ cathode relative to the silicon bulk. As the voltage is increased this depletion region
extends first to the anode implant then punches through extending the depletion region into the absorber. Because of the
small separation between the anode and the cathode implants, the electric field in the multiplier is significantly higher
than the absorber. By proper design of the buried implant dose and depth the multiplier electric field strength is
optimized for high probability of avalanche breakdown and for depletion regions that extend through most of the silicon
absorber. This results in high photon-detection-efficiencies (PDE = quantum efficiency × avalanche breakdown
probability, > 50%) while collecting photoelectrons from a significant volume at modest breakdown overbias levels (< 6
V). If the light enters from the cathode-side (also referred to as the front side) of the GM-APD photons are blocked by
the metal interconnects used to electronically access the detector. As a result, the scalability of this architecture to largeelement-count arrays is limited and also the optical path into the photosensitive regions has a number of interfaces that
cause reflections and absorbing layers that lead to losses, particularly at short wavelengths.
As depicted in Figure 2, the detector can be back-illuminated by removing the heavily doped silicon substrate. In this
structure, a photon enters the detector from the back side (anode side) of the devices, opposite the metal interconnects,
and is then converted into a photo-electron in the absorber region. As with the front-side GM-APDs, reach through fields
in the absorber draw the photo-electron into the multiplier region where it initiates an avalanche breakdown. Because the
absorber region can be tens of microns thick this structure can have good quantum efficiency into the near infrared. In
addition, using backside passivation techniques developed for scientific imagers (e.g. molecular-beam epitaxy-MBE), it
is possible to extend the performance of the detector deep into the ultraviolet. In both front- and back-side devices, the
electric field within the multiplier region primarily determines the avalanche breakdown probability. Since the electric
field extends through most of the absorber region, high photon-detection efficiencies can be obtained over a broad band
of wavelengths at modest overbias levels.
Proc. of SPIE Vol. 7681 76810P-2
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms
Figure 2: Example cross section of a back-illuminated GM-APD device. For a measure of scale, the GMAPD device layer is typically between 5 and 20 µm thick (depending on device architecture) while the
diameter of the anode diameter ranges from 20 µm to 50 µm. Light enters through the top of the quartz
substrate and is converted to photoelectrons in the absorber. The anti-reflection coating and backside
passivation are used to optimize the optical interface between the quartz and silicon.
Electrical connection to the GM-APD is made with a bump bond directly between the GM-APD cathode and a signalprocessing pixel on the CMOS ROIC. The electrical connection to the substrate is made through a ring of bump bonds
around the edge of the GM-APD array which biases the silicon bulk and the GM-APD anode. Bump bonding enables a
(high) density of interconnects that distinguishes this detector from other GM-APD arrays that use optimized processes
to improve detector performance (e.g. commercial devices offered by Hamamatsu, PerkinElmer, and others). These
commercial detectors are small, typically linear, arrays with less than one-hundred elements. In contrast, present bump
bonding technology supports hybridization of very-large-element-count detector arrays, with routine fabrication of
bump-bonded hybrid imagers with tens-of-thousands of pixels and commercial examples of multi-mega-pixel devices
(e.g. the 4k × 4k hybrid imagers produced by Teledyne Scientific Imaging). With this hybridization process, it is
possible to fabricate single-photon-sensitive arrays with element-count, pixel-pitch and pixel-size comparable to
scientific imagers (the example you give in Figure 1 make it seem like the pixels are big).
While the example GM-APD architecture discussed above is for photon-timing applications, such as 3-D imaging ladar2,
the back-illumination process presented in the following section is versatile and can be applied to other GM-APD
structures as well (e.g. photon-counting detectors such as the 100% fill-factor quad-cell array developed for adaptive
optics wavefront sensors reported in3). Using the hybridization approach, a variety of GM-APD detector architectures
can be fabricated on a wafer and integrated to existing CMOS readouts that use different foundry process technologies.
This enables efficient and flexible use of the GM-APDs and CMOS ROICs in support of many applications (e.g. ladar,
adaptive optics, communications, etc.). Basic architectures, already mentioned, are photon-timing and photon-counting.
These basic architectures can have variations that are optimized for particular applications and include features like
coincidence measurements and synchronous or asynchronous arming of the GM-APD detectors. By placing the detector
on a separate tier, modified circuit architectures may be realized without affecting the performance of the detector array.
In addition, as integrated circuit technology advances to higher-density, higher-performance fabrication nodes, the two
tier approach allows the CMOS ROIC to advance without necessitating redesign of the detector array. Finally, the pixel
pitch achievable with two-tier hybridization should have an advantage over the monolithic approach, because it avoids
detector fill-factor trade-offs and can use the finer-feature-size processes offered by CMOS foundries.
Proc. of SPIE Vol. 7681 76810P-3
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms
1. Oxide Bonding to Silicon
Handle Wafer
2. Device Wafer Thinning
3. Backside Processing of
Device Layer
Si Handle Wafer
Si Handle Wafer
Si Handle Wafer
Device Layer
Device Layer
5. Handle Wafer Removal,
Device Layer Transfer,
Contact Opening
6. Bump Bonding
Device Wafer
4. Oxide Bonding to Quartz
Handle Wafer
Si Handle Wafer
CMOS ROIC
Device Layer
Device Layer
Device Layer
Quartz Handle Wafer
Quartz Handle Wafer
Quartz Handle Wafer
Figure 3: GM-APD back-illumination process flow illustrating the major steps to process a wafer of GMAPD devices that has completed front-side processing through to hybridization with a CMOS ROIC. To
simplify the illustration, the relative positions of the components of the bonded wafer stack are maintained
in all process description panels.
3. GM-APD HYBRIDIZATION PROCESS OVERVIEW
As summarized in Figure 3, the process for fabricating hybrid detectors with back-illuminated silicon GM-APD detector
arrays has six major steps that are described below. The process sequence begins after GM-APD arrays have been
fabricated through the front-side processing, which includes device cathode and buried anode implants and metal layers.
The silicon starting material may either be an epitaxial or a high-resistivity silicon wafer and the device architecture can
be optimized for specific applications such as photon timing or photon counting, as previously discussed.
3.1. Silicon-handle wafer oxide bonding
GM-APD devices are presently processed on 150-mm wafers. The final thinned GM-APD device layer is typically less
than 20-µm thick and is too fragile to handle freestanding so the GM-APD wafer is bonded to a temporary support that
will be removed at a later stage in the process. To provide this support a silicon-handle wafer is attached to the front side
(cathode side) of the GM-APD device wafer. Net-stress in the device and handle wafers must be limited so that the bow
in each wafer is a few microns or less. Oxide layers are then deposited and planarized on both the GM-APD device
wafer and the silicon-handle wafer to produce surfaces with local average smoothness of angstroms. When brought in
contact, the activated oxide surfaces initiate a bond that is then further strengthened through a thermal anneal at 175 C
for three hours. During this anneal it is important that gasses, primarily water vapor, that evolve at the bond interface be
able to diffuse out of the structure. Due to the thickness of the handle-wafer and unthinned device wafer these gasses
cannot escape if the bonding interfaces are strictly planar; therefore, lateral venting trenches are etched into the bonding
oxide to provide the requisite diffusion path9. With the oxide-bonded silicon-handle wafer, even a GM-APD device layer
that is only a few microns thick can be processed with standard microelectronic fabrication tools.
Proc. of SPIE Vol. 7681 76810P-4
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms
Figure 4: The photograph is a 150 mm GM-APD device wafer oxide bonded to a quartz handle. There are
very few defects (voids) within the interior of the array with most bonding losses occurring along the edge
of the wafer pair.
3.2. Device wafer thinning
Because the final device will be illuminated through the back side of the wafer, it is important that photoelectrons are
collected from the full thickness of GM-APD device absorber layer for high quantum efficiency and good timing
resolution. The final detector layer thickness is a trade-off between detector response time and photon-detection
efficiency at long wavelengths. Depending on the needs of the particular application, GM-APD device wafers supported
by silicon-handle wafers are thinned to a thickness of few microns to tens of microns. Operationally, this is done by first
grinding off approximately 600 microns of silicon and then using a combination of wet etching and chemical mechanical
polishing to generate a defect-free surface with the desired thickness.
3.3. Backside processing
Backside processing of the GM-APD device layer is important for ensuring the high optical performance of backilluminated devices as compared to front-illuminated devices. The passivation and anti-reflection coating steps optimize
the optical interface through minimization of reflection and absorption losses. Reflection losses are minimized at the
quartz-to-silicon and quartz-to-air interfaces by application of an anti-reflection-coating. Absorption losses are
minimized by adapting back-side processing techniques developed for back-illuminated CCD imagers. For devices
needing high quantum efficiency in the ultraviolet, the thinned GM-APD device wafer is passivated by growing a layer
of highly boron doped silicon that is only a few nanometers thick, a delta-doping10, on the back side of the device using
MBE. This layer is used to passivate the back surface of the device, which reduces dark counts as well as produces an
electric field within the device that drives photo-electrons toward the multiplier region of the GM-APD. Ionimplantation with laser-annealing is also used, in place of MBE, where short wavelength response is not as important.
After backside passivation, metal layers that can act as light shields or return current paths and optical coatings that
prevent reflection losses at the quartz-to-silicon interface may be added.
Proc. of SPIE Vol. 7681 76810P-5
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms
3.4. Quartz wafer oxide bonding
Quartz (fused silica) is chosen as a final handle wafer because of its compatibility with the oxide-bonding process and
high transparency from ultraviolet to near-infrared wavelengths. In the fourth major step of the process, the quartzhandle wafer is attached to the backside of the thinned GM-APD device wafer by depositing an oxide layer on both the
GM-APD device wafer and the quartz-handle wafer, then planarizing these layers until local roughness is on the order of
angstroms. As with the silicon-handle wafer, when brought into contact the activated surfaces of these oxide layers
initiate a bond. Strengthening this bond is more complicated than in conventional silicon-to-silicon wafer bonding due to
the thermal mismatch of silicon and quartz. Specifically, it has been found that annealing at 125 C for approximately
three hours is sufficient to remove residual water from the bond interface, critical in preventing the development of postanneal voids, but low enough to prevent slipping of the bond surface during the anneal, which would generate
unacceptable post-anneal wafer stress. As with the bonding of the silicon handle wafer (section 3.1) venting trenches are
etched into the bonding oxide to prevent the formation of voids due to outgassing.
3.5. Silicon-handle wafer removal and final front-side processing
Following oxide-bonding of the quartz handle wafer, the thin GM-APD device layer is held between the quartz and
silicon wafers. The silicon handle wafer is removed with grinding and selective chemical etching, a process similar to
that used in wafer thinning. The bonding oxide provides a convenient etch stop so the process is easier. After the handle
wafer is removed, the only silicon remaining is the thin active GM-APD device layer, now supported by a quartz
substrate. The final front-side processing is completed by etching the oxide from the APD bump-bond pads in
preparation for flip-chip hybridization. A photograph of a completed back-illuminated GM-APD wafer is shown in
Figure 4.
1111
__________
III
IOhiIiI liii liii 1101
uasta1an1J14 U3a1
14IH*Il1UI1Ii4
a,s
(a)
(b)
Figure 5: (a) shows an optical image GM-APD array, viewed through the quartz substrate, bump bonded to
a CMOS ROIC. The dark ring around the array edge is an epoxy used to underfill the device and (b) shows
an infrared image of the same device where the array of small circles are the bump bonds.
3.6. Bump-bonding
To fabricate the final hybrid detector an electroless plating process, called zincation electroless nickel gold (ZENIG)11, is
used to grow an under-bump metal on the CMOS readout integrate circuit. The ZENIG process avoids plasma deposited
metal processes that can damage the CMOS. Also the ZENIG process enables under bump metallization of small chips
that otherwise would require lithographic processing, which is difficult on the single chip scale. The GM-APDs have
Ti/Ni/Au under bump metals deposited on the APD pad opening followed by the evaporation of 7-μm high indium
bumps. A photoresist lift-off process is used to pattern the bumps. Detector arrays, diced out of the quartz-oxide-bonded
Proc. of SPIE Vol. 7681 76810P-6
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms
GM-APD wafer, are then flip-chip bonded to the CMOS circuitry. Finally, an epoxy is used to underfill the GM-APD
array and thereby ensure mechanical reliability. Figure 5 shows photographs of a completed hybrid detector.
CONCLUSIONS
Large-element-count GM-APD arrays have been fabricated using a back-illumination process that enables independent
optimization of the detector and CMOS ROIC. In the case of the detector, optimization occurs on two fronts: first, the
optical interface may be optimized to minimize light lost in the path to the detectors active region, through highperformance back side processing techniques (developed for solid-state imaging system such as CCDs and active pixel
sensors) and application of anti-reflective coating; and second, the size and depth of the photo-collecting region may be
structured to optimize for trade-offs between timing accuracy, dark count rate, and long-wavelength quantum efficiency.
In the case of the signal processing circuit, the 3-D hybrid integration removes trade-offs between detector fill-factor and
pixel circuit area found in monolithic commercial CMOS foundry-based approaches.
To-date, this hybridization process has been targeted at the active imaging applications of 3-D ladar imaging and
adaptive optics wavefront sensing. Given, however, the versatility of the approach and the potential for extended
detector response at wavelengths from into the deep ultraviolet to the near-infrared we anticipate that this technology
offers opportunities for applications in other fields as well.
ACKNOWLEDGEMENTS
This work was sponsored by the Air Force under the Department of the Air Force contract number FA8721-05-C-0002.
Opinions, interpretations, conclusions, and recommendations are those of the author and are not necessarily endorsed by
the United States Government.
REFERENCES
[1] Farr, W., Gin, J., and Nguyen, D., “GigaHertz Bandwidth Photon Counting”, Proc. SPIE 7320, 732006 (2009).
[2] Aull, B., Loomis, A., and Young, D., et.al., “Geiger-Mode Avalanche Photodiodes for Three-Dimensional
Imaging”, Lincoln Laboratory Journal 13(2), 335-350 (2002).
[3] Aull, B., Renzi, M., Loomis, A., Young, D., Felton, B., Lind, T., Craig, D., and Johnson, R., “Geiger-Mode QuadCell Array for Adaptive Optics”, Conference on Lasers and Electro-Optics/Quantum Electronics and Laser Science
and Photonic Applications System Technologies (2008).
[4] Becker, W., Bergmann, A., Biscotti, G., and Rück, A., “Advanced time-correlated single photon counting technique
for spectroscopy and imaging in biomedical systems”, Proc. SPIE 5340, 104-112 (2004).
[5] Charbon, E., “Towards large scale CMOS single-photon detector arrays for lab-on-chip application”, J. Phys. D:
Appl. Phys. 41 (2008).
[6] Rech, I., Gulinatti, A., Zappa, F., Ghioni, M., and Cova, S., “High performance silicon single-photon avalanche
diode array”, Proc. SPIE 7320, 73200H (2009).
[7] Marwick, M., and Andreou, A., “Single photon avalanche photodetector with integrated quenching fabricated in
TSMC 0.18 μm 1.8 V CMOS process”, IEEE Electronics Letters, 44, 10 (2008).
[8] Verghese, S., Cohen, D., Dauler, E., Donnelly, J., Duerr, E., Groves, S., Hopman, P., Jensen, K., Liau, Z.-L.,
McIntosh, K., Oakley, D., and Smith, G., “Geiger-Mode Avalanche Photodiodes for Photon-Counting
Communications”, Digest of the LEOS Summer Topical Meetings, 15-16 (2005).
[9] Westhoff, R., Reich, R., Loomis, A., and Gregory, J., “Integrated Process for Detector Back Illumination”, Proc.
International Image Sensor Workshop (2007).
[10] Nikzad, S., Hoenk, M., Grunthaner, P., Terhune, R., Grunthaner, F., Winzenread, R., Fattahi, M., and Tseng, H.-F.,
“Delta-Doped CCDs for Enhanced UV Performance”, Proc. SPIE 2278, 138-146 (1994).
[11] Liu, C., Hutt, A., Whalley, D., Conway, P., and Mannan, S., “Under Bump Metallisation of Fine Pitch Flip-Chip
Using Electroless Nickel Deposition”, J. of Electronics Manufacturing, 10, 3, 161-170 (2000).
Proc. of SPIE Vol. 7681 76810P-7
Downloaded from SPIE Digital Library on 11 Feb 2011 to 18.51.1.125. Terms of Use: http://spiedl.org/terms