Geiger-mode avalanche photodiodes for three-dimensional imaging … · 2020. 1. 14. · • AULL,...

• AULL, LOOMIS, YOUNG, HEINRICHS, FELTON, DANIELS, AND LANDERSGeiger-Mode Avalanche Photodiodes for Three-Dimensional Imaging

VOLUME 13, NUMBER 2, 2002 LINCOLN LABORATORY JOURNAL 335

Geiger-Mode AvalanchePhotodiodes for Three-Dimensional ImagingBrian F. Aull, Andrew H. Loomis, Douglas J. Young, Richard M. Heinrichs,Bradley J. Felton, Peter J. Daniels, and Deborah J. Landers

■ We discuss the properties of Geiger-mode avalanche photodiodes (APDs) andtheir use in developing an imaging laser radar (ladar). This type ofphotodetector gives a fast electrical pulse in response to the detection of even asingle photon, allowing for sub-nsec-precision photon-flight-time measurement.We present ongoing work at Lincoln Laboratory on three-dimensional (3D)imaging with arrays of these diodes, and the integration of the arrays with fastcomplementary metal-oxide semiconductor (CMOS) digital timing circuits.

L been involvedwith the development of laser radar (ladar)systems, which, like their microwave counter-

parts, measure the flight time of transmitted pulses ofradiation to determine the distance to an object of in-terest. Real-time acquisition of such range informa-tion is central to a wide variety of remote-sensing ap-plications. The optical radar, because of its shorterwavelength, can resolve objects subtending a smallerangular field of view. This finer cross-range resolutionenables us to measure the range to multiple points ina scene, thus acquiring a ladar image, or three-dimen-sional (3D) image. Clearly, 3D images complementconventional intensity images in terms of informa-tion content.

One approach to building an imaging ladar systemis to add scanning optics to a single-point ladar sys-tem. The laser sends out multiple light pulses, eachdirected to a different point in the scene by the scan-ning mechanism, and each resulting in a range mea-surement obtained by using a single detector. Thescanner, however, limits the speed of image acquisi-tion and adds significantly to the weight, volume,power, and cost of the system.

For real-time imaging of moving targets, it is desir-

able to capture the 3D image with a single laser pulse.For small, lightweight, fieldable systems, the trans-mitter and receiver components must be compact andmust consume modest power. Single-pulse ladar im-aging is achieved by flood-illuminating the scene ofinterest and imaging the returning light onto an arrayof detectors, each with its own timing circuit to mea-sure the range to the corresponding point in thescene. This process presents the additional challengeof achieving both detector sensitivity and speed. Forresolved targets, the fraction of transmitted opticalenergy returning to the receiver falls off as the inversesquare of the range. In a single-pulse ladar imager thisenergy is further divided among multiple detectors.Therefore, for many scenarios with interesting imagesizes and target ranges, a particular detector circuitmust time the arrival of a weak optical pulse. If arange precision of even a few centimeters is needed,the timing precision must be less than a nanosecondand the detection circuit bandwidth correspondinglyhigh. High bandwidth, however, means high noisecompeting with the weak signal.

Lincoln Laboratory is developing 3D imaging sys-tems based on two enabling technologies. The first isthe diode-pumped solid-state microchip laser, in-


336 LINCOLN LABORATORY JOURNAL VOLUME 13, NUMBER 2, 2002

vented at the Laboratory. Passively Q-switched fre-quency-doubled Nd:YAG (neodymium-doped yt-trium-aluminum-garnet) microchip lasers have beendeveloped that produce very short (250 picosec) opti-cal pulses at 532 nm, with pulse energies of 30 µJ.The microchip laser systems, including power supply,are very compact. This microchip laser fulfills the re-quirements for an imaging ladar transmitter: a smallpackage that delivers many photons in a very shortpulse. This technology has been presented in detail ina previous issue of the Lincoln Laboratory Journal [1].

The second enabling technology, which is the sub-ject of the present article, is a ladar receiver based onarrays of Geiger-mode avalanche photodiodes (APDs)integrated with fast complementary metal-oxidesemiconductor (CMOS) time-to-digital convertercircuits [2–5]. Geiger mode, discussed in detail later,is a way of operating an APD so that it produces a fastelectrical pulse of several volts amplitude in responseto the detection of even a single photon. With simplelevel shifting, this pulse can trigger a digital CMOScircuit incorporated into the pixel. Single-photonsensitivity is achieved along with sub-nanosecondtiming precision. Because the timing information isdigitized in the pixel circuit, it is read out noiselessly.An integrated APD/CMOS array would also be acompact, low-power, all-solid-state sensor.

An Avalanche Photodiode Primer

An APD is a variation of a p-n junction photodiode.When a p-n junction photodiode is reversed biased,an electric field exists in the vicinity of the junctionthat keeps electrons confined to the n side and holesconfined to the p side of the junction. When an inci-dent photon of sufficient energy (>1.1 eV in the caseof silicon [Si]) is absorbed in the region where thefield exists, an electron-hole pair is generated. Underthe influence of the field, the electron drifts to the nside and the hole drifts to the p side, resulting in theflow of photocurrent in the external circuit. The timeintegral of the current can be shown to be one elec-tron charge. An electron-hole pair can also be ther-mally generated, resulting in leakage current, which isalso called dark current because it is present even inthe absence of incident light.

The physical process of electron-hole generation,

drift, and collection for a p-n junction photodiode isshown in Figure 1. The vertical axis represents thespatial dimension along the direction of the electricfield (perpendicular to the junction plane); the hori-zontal axis is time. The electron trajectory is repre-sented by a solid arrow, and the hole trajectory by adashed arrow. The slope of a trajectory is the drift ve-locity (which, for high electric fields in Si, is about107 cm/sec for both electrons and holes). When aphotodiode is used to detect light, the number ofelectron-hole pairs generated per incident photon, ametric known as the quantum efficiency, is at bestunity. Losses due to reflection or absorption in zero-field regions usually lower the quantum efficiency.

An APD detects light by using the same principle.The difference between an APD and an ordinary p-njunction photodiode is that an APD is designed tosupport high electric fields. When an electron-holepair is generated by photon absorption, the electron(or the hole) can accelerate and gain sufficient energyfrom the field to collide with the crystal lattice andgenerate another electron-hole pair, losing some of itskinetic energy in the process. This process is known asimpact ionization. The electron can accelerate again,as can the secondary electron or hole, and create moreelectron-hole pairs, hence the term “avalanche.” Aftera few transit times, a competition develops betweenthe rate at which electron-hole pairs are being gener-ated by impact ionization (analogous to a birth rate)and the rate at which they exit the high-field region

FIGURE 1. Photon detection in a photodiode represented ina simple space-time diagram. The absorption of the photoncreates an electron-hole pair, and the two oppositelycharged particles drift in opposite directions under the influ-ence of the electric field in the vicinity of the reverse-biasedp-n junction.

n contact

p contact

Electron

HolePhoton

Time

Spa

ce



and are collected (analogous to a death rate). If themagnitude of the reverse-bias voltage is below a valueknown as the breakdown voltage, collection wins thecompetition, causing the population of electrons andholes to decline. Figure 2 illustrates the avalanche-multiplication process in a space-time diagram.

This situation represents the most commonlyknown mode of operation of APDs: measuring theintensity of an optical signal and taking advantage ofthe internal gain provided by impact ionization. Eachabsorbed photon creates on average a finite numberM of electron-hole pairs. The internal gain M is typi-cally tens or hundreds. Because the average photocur-rent is strictly proportional to the incident opticalflux, this mode of operation is known as linear mode.

The amplification process adds noise to the signalabove and beyond the amplified shot noise thatwould be projected from scaling up by a factor of Mthe photocurrent that would flow in a regular photo-diode. Impact ionization is a statistical process. WhileM might be the average number of electron-hole pairsper absorbed photon, the actual number varies. Thisgain fluctuation produces excess noise, or multiplica-tion noise, which gets progressively worse as the aver-age gain M of a particular diode is increased by raisingthe reverse bias. Once the point is reached where themultiplication noise dominates over the noise intro-duced by downstream circuitry, further increases ingain deteriorate the system signal-to-noise ratio.

The severity of multiplication noise also dependson material properties. In general, electrons and holesare not equally likely to initiate impact ionization. InSi, for example, electrons are much more likely to im-pact ionize than are holes. For linear-mode operation,this property leads to low multiplication noise whenthe avalanche is electron-initiated. Figure 3 showsspace-time diagrams for two hypothetical semiconduc-tor materials. The one on the top represents a materialin which holes are incapable of initiating impact ion-ization, and the one on the bottom represents a mate-rial in which electrons and holes are equally likely toimpact ionize. Both diagrams indicate the same aver-age value of gain M. In the top diagram, the multipli-cation process occurs in a single pass, with one elec-tron entering the high-field region and M electronsexiting it. In this case, as the avalanche progresses,there are many electrons in the high-field region andtherefore many impact-ionization events occurring inparallel. The law of large numbers works to reducethe variance of the gain. If one electron falls short ofthe average number of ionizations, another electron is

FIGURE 2. Avalanche multiplication illustrated in a space-time diagram. The primary electron (the companion hole isnot shown), on the left, starts a chain of impact-ionizationevents. The solid arrows depict electron trajectories, and thedashed arrows depict hole trajectories. Points a, b, and drepresent electron-initiated impact ionizations; points c ande represent hole-initiated impact ionizations.

FIGURE 3. Two hypothetical materials. On the top, only elec-trons can initiate impact ionization. On the bottom, elec-trons and holes are equally likely to initiate impactionization.

a

b

c

d

e

Time

Spa

ce

Time

Spa

ce

Time

Spa

ce



likely to exceed it. In the bottom diagram, on theother hand, a prolonged volley of alternating electronimpacts and hole impacts achieves the gain. While theaverage duration of this process is assumed to lead tothe same average gain as in the top diagram, at anyone instant in time there may be only one or two car-riers in the high-field region. The ionizations tend tooccur sequentially rather than in parallel; therefore,the failure of an electron or hole to ionize can lead tothe termination of the avalanche. Because termina-tion can occur at any time during the volley, it leadsto large statistical variations in gain. The avalancheprolongation also slows the response and limits thegain-bandwidth product of the detector.

What Is Geiger Mode?

Consider what occurs when the APD is reverse biasedabove the breakdown voltage by using a power supplythat can source unlimited current. When the reversebias exceeds the breakdown voltage, the electrons andholes multiply by impact ionization faster, on aver-age, than they can be extracted. (This criterion, infact, is the best definition of the avalanche breakdownvoltage.) The space-time diagram in Figure 4 illus-trates the concept. The population of electrons andholes in the high-field region and the associated pho-tocurrent grow exponentially in time. The moreabove breakdown the APD is biased, the faster thegrowth time constant. This growth of current contin-ues for as long as the electric fields in the device arenegligibly altered by the presence of the growing cur-rent and the growing population of electrons andholes. If there is series resistance in the diode, how-ever, more and more voltage is dropped across thatresistance as the current grows. This effect reduces thevoltage dropped across the high-field region, andtherefore slows down the rate of growth of the ava-lanche. Ultimately, a steady-state condition is reachedin which the voltage across the high-field region is re-duced to the breakdown voltage, where the genera-tion and extraction rates balance. At this point thecurrent neither grows nor decays, and the series resis-tance provides negative feedback that tends to stabi-lize the current level against fluctuations. A down-ward fluctuation in current, for example, causes adecrease in the voltage drop across the series resis-

tance, and an equal increase in the drop across theAPD high-field region, which in turn increases theimpact-ionization rates and causes the current to goback up [6]. If the level of steady-state current is nottoo small (less than a few tens of microamps), it con-tinues to flow indefinitely [7]. Therefore, an ava-lanche initiated by the absorption of a single photoncauses the diode current to grow to some resistance-limited value. The turn-on transient of this current isfast, typically lasting tens of picoseconds.

When we measure the current-voltage characteris-tic of an APD on a curve tracer, the point of ava-lanche breakdown is evident as a sharp rise in currentwith a slope equal to the inverse of the device resis-tance. While the term “breakdown” is applied to thisphenomenon, it is not a destructive phenomenonsuch as the dielectric breakdown that occurs when thefield is strong enough to dislocate atoms in the mate-rial. Avalanche breakdown is stable against thermalrunaway. As the device temperature rises, other ki-netic-energy-loss mechanisms, such as lattice vibra-tions, increasingly compete with impact ionization,causing the breakdown voltage to rise and the steady-state avalanche current at a particular bias voltage todecrease.

Simply connecting an APD to a low-impedancepower supply, however, gives no way to either detectthe turn-on or shut off the avalanche so that the APDis ready to detect another photon. Shutting off theavalanche current is called quenching, and is accom-

FIGURE 4. Concept of avalanche breakdown voltage. In Gei-ger mode, in which the avalanche photodiode (APD) is bi-ased above the avalanche breakdown voltage, the growth inthe population of electrons and holes due to impact ioniza-tion outpaces the rate at which they can be extracted, lead-ing to exponential growth of current.

Time

Spa

ce



plished by two types of circuit: passive quenching andactive quenching. In a passive-quenching circuit, theAPD is charged up to some bias above breakdownand then left open circuited. Once the APD hasturned on, it discharges its own capacitance until it isno longer above the breakdown voltage, at whichpoint the avalanche dies out. An active-quenchingcircuit senses when the APD starts to self-discharge,and then quickly discharges it to below breakdownwith a shunting switch. After sufficient time toquench the avalanche, it then recharges the APDquickly by using a switch.

Figure 5(a) shows the simple passive-quenchingcircuit and Figure 5(b) shows the same circuit with afirst-order circuit model inserted to describe the APDbehavior during discharge. The model assumes thatonce the APD has turned on and reached its resis-tance-limited current, the ensuing self-discharge isslow enough that the APD will behave quasi-stati-cally, following its DC current-voltage characteristicas it discharges down to breakdown. The correspond-ing model is a voltage source equal to the breakdownvoltage in series with the internal resistance R of theAPD. The model predicts exponential decay of the

current to zero and voltage to the breakdown with atime constant RC [8].

Once the avalanche has been quenched, the APDcan be recharged through a switch transistor. Anotherscheme is to connect the APD to a power supplythrough a large series resistor Rs that functions as avirtual open circuit (Rs >> R) on the time scale of thedischarge, and then recharges the APD with a slowtime constant RsC. This circuit has the benefit of sim-plicity, and the APD fires and recharges with nosupervision.

In ladar applications, where the APD detects onlyonce per frame, the slow recharge time, typically mi-croseconds, imposes no penalty. There is also interest,however, in using the Geiger-mode APD to countphotons to measure optical flux at low light levels.With passive quenching, the count rate will saturateat low optical fluxes because many photons will arrivewhen the APD is partially or fully discharged, andtherefore unresponsive. With a fast active-quenchingcircuit, the APD can be reset after each detection on atime scale as short as nanoseconds, enabling it tofunction as a photon-counting device at much higheroptical intensities.

Geiger-Mode APD Performance Parameters

In linear mode the multiplication gain of the APDhas statistical variation that leads to excess noise. InGeiger mode the concept of multiplication noise doesnot apply. A Geiger-mode avalanche can, by chance,die out in its earliest stages. If it does, no detectableelectrical pulse is observed and the photon that initi-ated the avalanche goes undetected. If the avalancheprogresses to completion, however, the total numberof electron-hole pairs produced is fixed by the exter-nal circuit, not by the statistics of the impact-ioniza-tion process. In the simple passive-quenching case,for example, the avalanche has no further opportu-nity to die out until the APD has discharged from itsinitial bias down to the breakdown voltage. This dis-charge fixes the amplitude of the voltage pulse and,therefore, the total amount of charge collected in theprocess, typically >107 electron-hole pairs per detec-tion event.

The user of a Geiger-mode APD is concerned notwith multiplication noise, but with detection probabil-

FIGURE 5. Passive-quenching circuits. (a) In Geiger mode,the APD is charged up to some bias above the breakdownvoltage V and then left open circuited. (b) Subsequently,once an avalanche has been initiated, the APD behaves ac-cording to a simple circuit model.

Bias > Vbreakdown

Vbreakdown

+–

Bias

C

R

(b)(a)



ity, the probability that an incident photon will pro-duce a detection event. This probability is the prod-uct of the quantum efficiency, which is the probabil-ity that the photon will be absorbed in the activeregion of the device, and the avalanche probability,which is the probability that the photoelectron (orhole) will initiate an avalanche that does not termi-nate prematurely. Figure 6 is a qualitative plot of thephoton detection probability versus bias in Si forboth electron-initiated and hole-initiated avalanches.The more the APD is biased above breakdown, thehigher the avalanche probability and, therefore, thedetection probability.

Another important difference between linear andGeiger modes is that a particular Geiger-mode detec-tion event does not give intensity information. Theelectrical pulse produced by the detection of a photonis indistinguishable from that produced by the detec-tion of many simultaneously absorbed ones. (In pas-sive imaging applications, intensity is measured bycounting multiple events, or by measuring the timethat elapses before the first detection event aftercharging the APD.)

One noise source in a linear-mode APD is the shotnoise due to multiplied dark current. In a Geiger-mode APD a single thermally generated electron orhole can initiate an avalanche, leading to an electricalpulse that is indistinguishable from a photon detec-tion. In ladar applications, such an event represents afalse alarm whose probability needs to be minimized.For photon counting to measure intensity, suchevents inflate the count. The average error can bemeasured and subtracted, but its statistical varianceconstitutes a noise source. The number of dark-cur-rent-induced events per second is the dark count rate.The use of high-quality material and proper devicefabrication techniques minimizes the number of im-purities and defects that contribute to dark current.Geiger-mode APDs have been made in Si with room-temperature dark count rates below 1000 counts persecond. In a ladar, this means that the APD can beactivated for 10 µsec with a false-alarm probabilityless than 1%. A 10-µsec detector active time corre-sponds to a range window of about 1.6 km (1 mile).

Since the electrical pulse from the APD is used in aladar to measure the arrival time of an optical pulse,

the user must be concerned with the statistical varia-tion of the time interval between the pulse arrival andthe resulting electrical signal from the APD. Suchtiming jitter arises from several sources. First, the pho-ton detection probability is generally less than unityand the transmitted optical pulse has finite time dura-tion. For feeble returns the APD can detect a photonat the leading edge, the middle, or the trailing edge ofthe pulse. The associated statistical variation contrib-utes to timing jitter [9].

Second, the photoelectron requires finite time todrift from where it is first generated to the high-fieldlayer where the avalanche starts. These regions are of-ten separated in the APD device structure. Depend-ing on how deeply into the detector the photonpropagates before it is absorbed, the photoelectronmay have a shorter or longer drift delay. Because thespatial probability density for photon absorption isexponential, the depth at which the photon is ab-sorbed has a standard deviation equal to the absorp-tion length. The absorption length in Si monotoni-cally increases with wavelength, so this contributionto the timing jitter also increases with wavelength.For a Si APD operating at 800 nm, for example, theabsorption length is roughly 10 µm. Because the satu-

FIGURE 6. Qualitative plot of the photon detection probabil-ity versus bias in silicon (Si) for both electron-initiated andhole-initiated avalanches. The probability of photon detec-tion is the probability that the incoming photon creates aphotoelectron or photohole that initiates an avalanche thatdoes not prematurely die out. In Si, electrons have a higherimpact-ionization probability than holes, so the probabilityof detection is higher for electron-initiated avalanches thanfor hole-initiated avalanches.

1

1 20 3 4 5 60

Electron-initiated avalanche

Hole-initiated avalanche

Bias in excess of breakdown (V)

Pho

ton

dete

ctio

n pr

obab

ility



rated carrier velocity (107 cm/sec) is the fastest pos-sible drift velocity, this value translates to a minimumtiming-jitter contribution of 100 psec.

The third contribution to the timing jitter arisesfrom statistical variation in the time it takes the ava-lanche current to grow to its resistance-limited value.The space-time diagrams used previously to describethis current growth are oversimplified, because theyneglect the fact that the avalanche starts as a localizedfilament of current that spreads laterally until it isuniformly distributed over the whole area of the junc-tion. In diodes with a small active volume this spread-ing occurs predominantly by multiplication-assisteddiffusion, and occurs over a time scale of tens of pico-seconds. If the initial filament is near the diode pe-riphery, the process takes longer than for a filament inthe center because the distance over which the ava-lanche must spread is greater. In larger-volume di-odes, a different mechanism predominates. The hotelectrons and holes in the avalanche filament give offphotons that have sufficient energy to be reabsorbedelsewhere within the diode, thereby initiating otherfilamentary avalanches. Avalanche buildup typicallyoccurs over hundreds of picoseconds.

The fourth contribution to timing jitter comesfrom thermal noise due to APD resistance. This noiseproduces jitter because the APD is followed by cir-cuitry that uses a thresholding operation to detectwhen the APD is discharging, and amplitude fluctua-tions result in fluctuations in threshold-crossing time.

In monolithic arrays of Geiger-mode APDs, theemission of photons by discharging devices intro-duces crosstalk; the photons from one APD can trig-ger avalanches in others. Measurements of the radia-tion from Si devices indicate that about 3 × 10–5

photons are emitted with energies greater than thebandgap for every carrier that crosses the junction.Above 1.7 eV the spectral distribution of this emis-sion is a Maxwellian distribution with an effectivetemperature of 4000 K [10]. Because the number ofphotons emitted is proportional to the amount ofcharge traversing the junction during the discharge,the crosstalk can be minimized by minimizing theparasitic capacitance that the APD must discharge.Active quenching circuitry is also helpful, because itshunts most of the charge through a switch.

The most important performance parameters for aGeiger-mode APD used for photon timing are pho-ton detection probability and timing jitter. For aphoton-counting passive imager, dark count rate andspatial uniformity of detection probability are impor-tant, and jitter is not particularly important. For bothapplications, we want low crosstalk.

APD Fabrication and Characterization

Lincoln Laboratory has fabricated several lots ofAPDs on 4-in and 6-in Si wafers. The diodes are fab-ricated in 4 × 4 and 32 × 32 arrays, and there are alsoa number of test structures. Figure 7 shows a cross-sectional diagram of the basic APD device structure.The Si substrate is heavily p-doped (p+) with 1018 bo-ron atoms/cm3) and overgrown by a lightly p-doped(1014 boron atoms/cm3) epitaxial layer. The diode isfabricated by ion implantation of n-type (arsenic andphosphorus) and p-type (boron) dopants. The struc-ture that results, rather than being a simple p-n diode,is p-π-p-π-n (π denotes very lightly p-doped). Thelower π layer is where the photons are absorbed whenthe device is illuminated from the substrate side,which is the intended mode of operation. Reverse-bi-asing at the proper operating voltage establishes amodest electric field (104 V/cm) in this photon ab-sorption layer that causes the photoelectron to driftup into the upper π layer. The field in the upper πlayer is much stronger (several 105 V/cm), sufficientto cause impact ionization that initiates an avalanche.

FIGURE 7. Cross-section of the basic APD device struc-ture. The APD is fabricated in a lightly p-doped epitaxiallayer grown on a heavily p-doped (p+) substrate. The heavilyn-doped (n+) implant forms the n-side contact of the diodeand the substrate form the p-side contact. The p+ implantcreates a sheet of charge that segments the structure intoregions of low electric field (below the implant), where pho-ton absorption takes place, and high electric field (above theimplant), where impact ionization occurs.

p+ substrate

Metal

p+ implant

n+ implant

10 mµ

0.5 mµ



The photoelectron and the secondary electrons arecollected at the top n layer, and the photohole andsecondary holes are collected at the substrate.

Note that the p-π-p-π-n structure exists only in thecentral portion of the APD. Because of the absence ofa p implant in the diode’s peripheral portion, this por-

tion is a simple p-π-n structure, in which the field isintermediate in strength between the photon-absorp-tion region and the avalanche region. This peripheraldiode serves as a “guard ring” that performs two func-tions. First, it tailors the electric-field profile so thatavalanche breakdown occurs in the central portion ofthe diode, not at the periphery. Second, it ensuresthat electrons generated outside the region directlyunder the p implant do not drift to the avalancheregion, but are collected in the peripheral portion ofthe n layer without being able to start an avalanche.This collection minimizes the volume from whichdark current is multiplied, and therefore minimizesthe dark count rate. The price paid is that the fractionof the chip area sensitive to light is limited. In imag-ing applications that cannot waste photons, the effi-ciency needed can be reclaimed by using a microlensarray to concentrate the incident light in each pixel’sregion so that a high percentage of it is detected.

Figure 8 shows a 6-in-diameter Si wafer on whichAPD arrays have been fabricated. There are 4 × 4 and32 × 32 arrays with pixel-to-pixel spacings of 100 µmand 150 µm. APD active area diameters also varyfrom 30 µm to 50 µm. Figure 9 shows measurementsof detection probability and timing jitter. Room tem-perature dark count rates of less than 1000 counts persecond have been achieved in recent lots.

FIGURE 9. APD performance measurements. On the left is a plot of measured de-tection probabilities on uncoated, front-illuminated APDs from several different wa-fers. On the right is a histogram of measured photon detection times obtained by re-peatedly illuminating an APD with a laser pulses of 250-psec duration, attenuated sothat the detection probability is low (10%). The width of the histogram (290 psec) isdue to a combination of the duration of the laser pulse and the timing jitter of theAPD. The latter is inferred to be less than 150 psec.

FIGURE 8. Avalanche photodiode arrays fabricated on a6-in-diameter Si wafer in the Lincoln Laboratory Microelec-tronics Laboratory. Each die includes 32 × 32 APD arrays,4 × 4 arrays, and test devices.

6 in

0 500–500

Cou

nts

Measured photon detection time (psec)

2 4

= 532 nm

Uncoated devicesilluminated throughcathode region

60

45

40

35

30

25

20

15

10

5

0

Det

ectio

n pr

obab

ility

(%)

Overbias (V)

λ



Ladar Imaging

The development of APD-based ladar focal planeshas occurred in three stages. In the first stage,4 × 4 APD arrays have been packaged as discrete de-vices and incorporated into a printed circuit boardwith circuitry to amplify the pulses from the APDs.The amplified pulses are then conveyed by coaxial ca-bling to commercial rack-mounted timing modules.A ladar system, known as the Gen I brassboard, wasdeveloped by using this scheme. In the second stage,4 × 4 APD arrays were hybrid integrated with CMOSchips that have sixteen timing circuits. In the thirdstage, 32 × 32 APD arrays were fully integrated with32 × 32 arrays of CMOS timing circuits. The com-panion article in this issue describes the brassboardand the imagery obtained from it. Subsequent sec-tions in this article describe the development of theCMOS timing circuitry and hybrid-integrated andfully integrated APD/CMOS imagers.

CMOS Timing Circuits

Using the MOSIS foundry service, we haveprototyped arrays of time-to-digital converter pixelcircuits. Figure 10 is a block diagram of the pixel cir-cuit of the timing chip. It consists of a pseudorandomcounter clocked by a 500-MHz master clock that isbroadcast to all pixels. The pseudorandom counter isa shift register with a feedback path that has an exclu-sive OR gate; if the shift register is n bits long, itcycles through a sequence of 2n–1 distinct states. Theclock is fed to the counter through a transparent latchwhose output is frozen when the photon is detected.The state of the counter thus encodes the number ofclock cycles that elapsed from the start of counting tothe photon detection time. The state of the transpar-ent latch, which is also read out, indicates whether thephoton was detected in the high-clock or low-clockportion of the clock period. Thus 1-nsec photon-flight-time resolution is achieved. An additional ver-nier bit is created by generating a secondary clock de-layed by 90° with respect to the master clock andstoring its state in a second transparent latch, and 0.5-nsec resolution is achieved. In ladar applications thisapproach yields a range resolution of 3 in (7.6 cm).

The pseudorandom count value must be decoded

by table lookup or reverse encoding after it is readout. The benefit of this implementation is that it givesa compact pixel circuit; a conventional binarycounter would require much more chip real estate.The vernier bit scheme facilitates achievement of sub-nsec precision without requiring 2-GHz clocking.Therefore, we can use an established commercialfoundry process with 0.5-µm or 0.35-µm gate geom-etries. A 16-pixel array that uses a 17-bit version ofthe counter circuit has been demonstrated by usingthe MOSIS HP 0.5-µm CMOS process, and 32 × 32arrays of 15-bit and 10-bit counter circuits have beendemonstrated by using the MOSIS TSMC (TaiwanSemiconductor Manufacturing Company) 0.35-µmprocess. Bench tests of the 32 × 32 CMOS chips con-firm functionality at clock rates over 650 MHz. Fig-ure 11 shows a portion of a 32 × 32 timing chip.

Figure 12 shows the device that was constructed tovalidate the operation of the CMOS/APD combina-tion. A 4 × 4 APD array was bonded to a large padin the middle of the 16-pixel MOSIS chip, andwire bonds were used to make an electrical connec-tion from each APD to a corresponding timing cir-cuit. The 16 timing circuits were daisy chained andread out serially. Timing values were observed bothin response to dark counts and in response to

FIGURE 10. Block diagram of the pixel circuit of the timingchip, which functions as a “stopwatch” and times the detec-tion of the photon by the APD. A 17-bit shift register with anexclusive OR (XOR) feedback loop is a pseudorandomcounter. Photon detection causes the clocks that drive thecounter to freeze. The counter and the stored states of theclocks encode the time of detection.

ClocksHold

(from APDcircuit)

Serial data out

Serial data in

Vernierbits

Control

Latch

4-to-1multiplexor

0 1 2 3 4 5 6 7 8 9

16

18

17

15 14 13 12 11 10



optical pulses from a frequency-doubled passivelyQ-switched Nd:YAG microchip laser. The correct-ness of the timing values was validated. This chip wasclocked through an off-chip source and worked up toa clock frequency of about 350 MHz. (A design flawin the clock distribution circuitry limited the speed.)The level of pixel-to-pixel crosstalk was also checkedby illuminating one APD with the laser pulses andmonitoring the timing values from the pixel circuitattached to the neighboring APD. The timing valuesobserved were random and showed no correlationwith those from the illuminated pixel. Thus thecrosstalk probability is low (<1%).

APD/CMOS Integration

Clearly, as the array size is scaled up, the wire-bond-ing approach used for the 4 × 4 array becomes im-practical. We can envision monolithic integration ofAPDs and timing circuits. This integration, however,presents two problems. The voltage requirements andfabrication process requirements of a Geiger-modeAPD are quite different from those of a CMOS logicprocess. A standard CMOS process cannot be used,

forcing costly development of a specialized processthat might necessitate compromises between circuitperformance and detector performance. Second, amonolithic process most likely involves a partition ofthe pixel real estate into detector and circuit sections,thus limiting the fill factor of the detector array. Inthe long run, silicon-on-insulator technology andthree-dimensional integration techniques are likely tomature and yield monolithic implementations with-out these drawbacks. In the short run, our approachhas been to use inexpensive foundry services to proto-type the timing circuits, which are received as singleunpackaged chips, and to pursue in-house fabricationof full wafers of APDs. The integration of the two isthen achieved by bonding the chips face to face withthe APD arrays, and illuminating the APD array fromthe substrate side (so-called back illumination).

A common technique, used for focal planes basedon III-V or II-VI semiconductors, is to bump-bondthe detectors with readout circuits by using arrays ofindium (In) or solder bumps. For example, such “flip-chip” methodologies have been used for indium-gal-

FIGURE 11. Portion of a 32 × 32 timing chip. This chip wasfabricated through MOSIS by using a complementary metal-oxide semiconductor (CMOS) process with 0.35-µm gatelengths. The white rectangle in each pixel is the bonding padto which the corresponding APD connects.

FIGURE 12. 4 × 4 APD/CMOS array, which is a “chip within achip.” The larger chip is a MOSIS-fabricated timing chipwith circuitry for 16 pixels, fabricated in a CMOS processwith 0.5-µm gate lengths. The smaller chip epoxied in thecenter of the MOSIS chip is a 4 × 4 APD array. The connec-tions from the APDs to the timing circuits are made withwire bonds.



lium-arsenide (InGaAs) photodiodes grown epitaxi-ally on indium-phosphide (InP) substrates. In thiscase, nature is kind, providing a substrate materialthat is relatively transparent at the wavelength wherethe epitaxial detector is used. Flip-chip bonded de-vices can be constructed and used in back-illumi-nated mode without removing the substrate.

The Si APDs, however, are homoepitaxial devices.The Si substrate is optically opaque at the wave-lengths where the detector is useful. Photoelectronscreated in the substrate are almost immediately anni-hilated by holes and do not reach the high-field re-gion in the epitaxial layer. A viable flip-chip bondingtechnique must therefore include removal of the sub-strate, leaving a detector structure that may be only afew microns thick. In such a bump-bonding process,mechanical sturdiness might be ensured by filling thespaces between the bumps before detector substrateremoval, which is illustrated in Figure 13(a).

Questions and concerns about the scalability ofthis type of process to large array sizes lead us to de-velop an alternative process known as bridge bond-ing, which is illustrated in Figure 13(b) [11]. First,

the CMOS chip is epoxied face to face with the corre-sponding APD array without making any electricalconnection whatsoever. The APD wafer, with severalCMOS chips bonded to it, is epoxied, substrate up,to a handle wafer. The resulting APD/CMOS/handle“sandwich” can then undergo the same types of clean-room fabrication processes that are applied to normalfull wafers. The APD substrate is removed by using anelectrochemically selective etch developed at LincolnLaboratory for fabrication of back-illuminated CCDfocal planes. A shallow p+ implant is added to the ex-posed backside of the APD wafer, followed by a laseranneal, in order to replace the p-side electrical contactlayer formerly provided by the substrate. Vias are thenetched in between the APDs and metal “bridges” pat-terned within the vias to connect each APD with thecorresponding timing circuit. The backside of theAPDs is also metalized to provide a common electri-cal contact to the p sides. In operation, this backsidecontact is held at a negative voltage slightly smaller inmagnitude than the breakdown voltage, and the nside of each APD is charged up to a positive voltage(typically 4 or 5 V) by the pixel circuit.

FIGURE 13. APD/CMOS integration. (a) In bump bonding (left), indium or solder bumpsare formed on one or both chips and they are then pressed together. For Si detectors the airspaces between the bumps must be filled, so that when the substrate is removed to allowlight to reach the detectors, the device is mechanically solid and sturdy. (b) In bridge bond-ing (right), the two chips are epoxied together, and the APD substrate removal is carriedout. Electrical connections are made last by etching vias between the APDs and patterningmetal connections in the vias.

Detector array

Bump bonding Bridge bonding

CMOS readoutcircuit

(a) (b)



Successful development of the bridge-bondingprocess required overcoming a number of technologi-cal hurdles. Thinning must be uniform. Curing of theepoxies used must not lead to destructive mechanicalstresses due to the large thermal-expansion coefficientmismatch between semiconductors and epoxies. Thevias through the epoxy must have sloped sidewalls toallow good step coverage of the bridge metal. Becauseof the vias, most of the required photolithographicsteps are done on a nonplanar surface, which presents

challenges of nonuniform photoresist thicknesses andexposure depth of focus issues. Finally, the handlingof the APD must not result in excess increases in leak-age current or dark count rate.

After several iterations of the process, which ini-tially used dummy CMOS chips and then functionalones, the first two bridge-bonded 32 × 32 APD/CMOS arrays were fabricated, and initial functional-ity tests performed on one of these arrays. Figure 14 isa photomicrograph of a 32 × 32 array and Figure 15is a magnified view showing the features of individualpixels.

Initially, the device was tested by using spatiallyvarying patterns of steady light, either room light orcoronal light from a flashlight, projected onto the de-vice. Illuminated APDs detected a photon very earlyduring the timing interval, and the correspondingpixels reported out very low detection-time values.APDs that were masked from the light, in most cases,did not fire at all during the timing interval, and thesepixels reported out timing values corresponding tothe duration of the timing interval (250 nsec). Thetiming values were displayed as gray scale (black cor-responding to immediate detection, light gray corre-sponding to 250 nsec), and therefore the displayedimages are “negative tone” replicas of the illuminationpatterns. Figure 16 shows the images obtained in the

FIGURE 14. Photomicrograph of a bridge-bonded APD/CMOS device showing the 32 × 32 array.

FIGURE 15. Photomicrograph of the same bridge-bonded APD/CMOS device shown in Figure 14,at much higher magnification, showing the features of individual pixels.

APDactivearea

APDp contact

APDn contact

CMOScircuit

CMOScontact

pad



dark, in room light, and with coronal flashlight illu-mination through a pinhole, through a washer hole,and around the edges of a square piece of black plas-tic. The white pixels represent values in excess of 250nsec, which indicates that a bit error was made.

In this experiment the p side of the APDs was bi-ased at about –23.5 V, a few tenths of a volt in magni-

FIGURE 16. “Negative tone” intensity images from bridge-bonded APD/CMOS array obtained by illumi-nation with different patterns of steady light. The black pixels represent small time values resulting fromearly detection events in the illuminated portions of the array.

tude below the avalanche breakdown voltage, and theCMOS circuit turned the APDs on by charging the nside to +4 V. The magnitude of the p-side bias on theAPDs was decreased to more than a volt below theavalanche breakdown to confirm that the response ofthe APDs to the light was indeed Geiger-modepulses. This decrease resulted in the disappearance of

FIGURE 17. The range image on the right was obtained by illuminating the central portion of anAPD/CMOS array with a sub-nsec pulse from an 800-nm-wavelength diode laser. The histo-gram on the left shows that 164 pixels reported detection events during the first 50 clock peri-ods, and that nearly all of these were during the 21st period, which corresponds to the arrivaltime of the laser pulse.

20 30 40 5010164

00

200150100

50

In the dark In room light

800- m pinholeµ Square piece of plastic7-mm washer hole



the black pixels, indicating that the images are notdue to linear-mode photocurrents.

A fast near-infrared diode laser (λ = 800 nm, pulseduration <1 nsec) was used to illuminate the centralportion of the device to verify its timing functionality.The pulse was timed to occur 60 nsec after the start ofa 500-nsec timing interval. Figure 17 shows the re-sulting image, and a timing-value histogram. Thedark gray area corresponding to the laser illuminationis clearly visible in the center of the image. The hori-zontal axis of the histogram is elapsed clock periods,and the clock period used was about 3 nsec. The peakin the histogram corresponds to the time of the laserpulse. Work is in progress to characterize and opti-mize the detection probability and timing precisionof these devices.

Conclusion

The work reported here represents only the initialstages in the development of a new and very powerfulfocal-plane-array technology. The images presentedin the companion article illustrate the efficacy of 3Dimaging for target recognition and image segmenta-tion. APD/CMOS technology is clearly an enablingtechnology for 3D imaging. The next phase of thetechnology development effort will be to refine theperformance of the imagers and achieve high detec-tion probability, low dark count rate, and high spatialuniformity. If this phase is successful, we hope to scaleup the number and density of the pixels and the time-measurement precision of the CMOS timing circuits.

Acknowledgments

This work was sponsored by the Department of theAir Force and the Defense Advanced ResearchProjects Agency.

R E F E R E N C E S1. J.J. Zayhowski, “Microchip Lasers,” Linc. Lab. J. 3 (3), 1990,

pp. 427–46.2. B.F. Aull, A.H. Loomis, J.A. Gregory, and D.J. Young, “Gei-

ger-Mode Avalanche Photodiode Arrays Integrated withCMOS Timing Circuits,” Mtg. of the Boston Chapter of theIEEE Electron Devices Society, Lexington, Mass., Nov. 1998.

3. B.F. Aull, “Geiger-Mode Avalanche Photodiode Arrays Inte-grated with CMOS Timing Circuits,” 56th Annual Device Re-search Conf. Dig., Charlottesville, Va., 22–24 June 1998, pp.58–59.

4. B.F. Aull, “Geiger-Mode Avalanche Photodiode Arrays for Im-aging Laser Radar,” Solid State Research Report, LincolnLaboratory (Mar. 1997), pp. 31–33.

5. B.F. Aull, “Geiger-Mode Avalanche Photodiode Arrays for Im-aging Laser Radar,” Solid State Research Report, LincolnLaboratory (Mar. 1996), pp. 39–41.

6. There is another mechanism that acts like an additional seriesresistance. The electrons and holes generated by impact ioniza-tion require finite time to transit the device to their respectiveextraction electrodes. As the avalanche grows, electrons tend to“pile up” at the n side of the high-field region and holes at thep side. The electrostatic screening due to the associated spacecharge reduces the field in the high-field region. Since this re-duction in field grows in proportion to the electron and holepopulation (and therefore in proportion to the current), theeffect can be modeled as a series resistance.

7. At lower levels of steady-state current, the population of elec-trons and holes is sufficiently small that there is a significantprobability that they will all transit out of the high-field regionat some point in time without ionizing. The steady-state cur-rent level below which this phenomenon is observed is knownas the latching current.

8. Strictly speaking, the voltage source needs to be slightly belowbreakdown so that the avalanche current decays at a rate thatmatches the RC time constant of the circuit. In practice, how-ever, such circuit time constants are sufficiently long that thisis a very small correction.

9. In addition to this random timing variation, there is a system-atic dependence of the average detection time on pulse inten-sity. The more intense the returning pulse, the more likely de-tection will occur at its leading edge. This effect, known asrange walk, or range bias, is not unique to Geiger-mode ava-lanche photodiodes, but occurs with any detection scheme inwhich thresholding is used.

10. A.L. Lacaita, F. Zappa, S. Bigliardi, and M. Manfredi, “On theBremsstrahlung Origin of Hot-Carrier-Induced Photons inSilicon Devices,” IEEE Trans. Electron Devices 40 (3), 1993, pp.577–582.

11. A.H. Loomis and B.F. Aull, “Bridge Bonding of Geiger-ModeAvalanche Photodiode Arrays to CMOS Timing Circuits,”Solid State Research Report, Lincoln Laboratory (Mar. 1999),pp. 33–37.



. develops 3D imaging andphoton-counting focal planesby using Geiger-mode ava-lanche photodiodes as a staffmember in the AdvancedImaging Technology group.He joined the Laboratory in1985, after earning a Ph.D.degree in electrical engineeringfrom MIT. During his firstdecade at the Laboratory heled teams that demonstratedmultiple-quantum-well spatiallight modulators with gallium-arsenide/aluminum-gallium-arsenide (GaAs/AlGaAs)charge-coupled device (CCD)arrays for electrical addressingand developed optoelectronicswitching and neural process-ing devices by using quantum-well modulators and double-barrier resonant tunnelingdiodes. During the past twoyears Brian has developedarrays of Geiger-mode ava-lanche photodiodes integratedwith high-speed CMOS tim-ing circuits that are beingintegrated into a 3D ladarsystem. He also holds a B.S.degree in electrical engineeringfrom Purdue University.

. is an associate staff memberwith the Advanced ImagingTechnology group. He super-vises the production of special-ized back-illuminated CCDsand Si sensors and developsadvanced wafer-processingtechniques for various Siimagers. His current researchinterests are in wafer bondingand specialized back-end CCDfabrication techniques. Beforejoining Lincoln Laboratory in1986, he worked two years forat GCA corporation inBedford, Massachusetts, devel-oping laser scanning systemsfor defect identification. In1984, Andy received a B.S.degree in electronic engineer-ing from Wentworth Instituteof Technology, Boston, Massa-chusetts.

. is an associate staff member ofthe Advanced Imaging Tech-nology group and responsiblefor the fabrication of CCDimagers and photodiodearrays. He is also area engineerfor i-line photolithographyand etch areas of the Micro-electronics Laboratory. Beforejoining the Laboratory in1989, he was a technicalinstructor for MIT’s Micro-electronics Educational Facil-ity. Doug holds a B.S. degreein electrical engineering fromthe University of Vermont.

. leads the Laser and SensorApplications group, whichstudies direct detection andcoherent laser remote sensing.Rick joined Lincoln Labora-tory in 1986, after completinga postdoctoral position innonlinear fluid dynamics atthe University of California,Santa Barbara. He holds anS.B. degree in electrical engi-neering from MIT and aPh.D. degree in physics fromthe University of Massachu-setts.



. provides computer program-ming and CAD support forprojects in the AdvancedImaging Technology group.Currently, Brad is workingwith the APD ladar develop-ment team to provide dataacquisition and display rou-tines. Brad joined the Labora-tory in 1977. He has a B.S.degree in mathematics fromthe University of Lowell (for-merly Lowell TechnologicalInstitute). He has also pursuedgraduate-level courses inmathematics and computerscience at University of Lowell.Brad’s other interests includecircuit design verification andimage processing.

. supervises a microelectronicpackaging facility in the Ad-vanced Imaging Technologygroup. Pete holds a B.E.Tdegree in electrical engineeringfrom Northeastern University.He is a member of the SigmaEpsilon Rho Honor Society.He joined the Laboratory in1964.

. is a project technician for theAdvanced Imaging Technologygroup. She packages Labora-tory-processed semiconductordevices used in ladar rangingsystems and in CCDs used inground and space based as-tronomy. Before joiningLincoln Laboratory in 1984,Debbie packaged semiconduc-tors for missile systems atMicrowave Associates inBurlington, Massachusetts.

Date post:	21-Jan-2021
Category:	Documents
Upload:	others
View:	8 times
Download:	0 times

Geiger-mode avalanche photodiodes for three-dimensional imaging … · 2020. 1. 14. · • AULL,...

Documents