Evolution and prospects for single-photon avalanche diodes and quenching circuits

Transcription

1 journal of modern optics, 15 june 10 july 2004 vol. 51, no. 9 10, Evolution and prospects for single-photon avalanche diodes and quenching circuits S. COVA, M. GHIONI, A. LOTITO, I. RECH and F. ZAPPA Politecnico di Milano, Dipartimento di Elettronica e Informazione, and IFN-CNR, Piazza Leonardo da Vinci, Milano, Italy; cova@elet.polimi.it (Received 17 July 2003) Abstract. The evolution of solid-state avalanche detectors of single optical photons is outlined and the issues for further progress are discussed. Physical phenomena that underlay the operation of the single-photon avalanche diodes (SPAD) and determine the performance are considered and their role is assessed (detection efficiency; dark-counting rate; afterpulsing; photon timing resolution; etc.). The main technological issues that hamper the development of detectors with wide sensitive area and of array detectors with high filling factor are illustrated. Silicon SPADs are the main focus of attention; infrared-sensitive SPADs in germanium and in compound semiconductors are also dealt with. The role of the active-quenching circuits (AQC) is assessed and the evolution is outlined up to integrated AQCs, which offer the prospect of monolithic integration of complete photon counter instruments. 1. Introduction Photon counting and time-correlated photon counting (TCPC) techniques were developed relying on photomultiplier tubes (PMT), that is, vacuum tube detectors with high internal gain. High performance PMTs were produced industrially with sophisticated technologies from the 1960s; solid-state avalanche detectors of single photons became available to experimenters much later, in the 1990s. However, the path leading to the development of these solid-state detectors started much earlier, during studies on the avalanche phenomenon in p-n junctions at the Shockley laboratory in the 1960s [1, 2]. In p-n junctions specially devised to obtain a uniform behaviour over their small area (diameter of a few microns) and reverse biased above the breakdown level, Shockley scholars observed macroscopic voltage pulses triggered by the absorption of single optical photons. Thanks to the work of this group and in particular of R. H. Haitz, a basic insight into the phenomenon was gained. The physical phenomena involved in the generation of the pulses (photon pulses; dark-current pulses; afterpulses) were identified and a correct model of the behaviour of the device biased above the breakdown was devised [2]. The application of these devices as photodetectors, however, was strongly limited by the features of the available junctions and avalanche-quenching circuits (see section 5). In the 1970s, during the early development of silicon avalanche photodiodes (APD), fundamental contributions to the understanding of the avalanche phenomenon and of its statistical properties were given by R. J. McIntyre and P. P. Webb [3]. They also observed the generation of single-photon Journal of Modern Optics ISSN print/issn online # 2004 Taylor & Francis Ltd DOI: /

2 1268 S. Cova et al. pulses in their avalanche photodiodes biased above breakdown [4], but the application was still limited by the features of the device and quenching circuit. The p-n junction devices that can operate biased above the breakdown voltage and generate macroscopic current pulses in response to single-photons are called single-photon avalanche diodes (SPAD). As pointed out in figure 1, their operation is radically different from that of ordinary avalanche photodiodes (APDs), which are biased near to the breakdown voltage V BD, but below it. In APDs, avalanche multiplication is exploited to produce linear amplification of the primary photogenerated electrical signal. APDs are somewhat similar to PMTs, but with a basic difference: there is an intrinsic positive feedback in the amplification due to the avalanche process. An electron impact generates a secondary electron hole pair; the hole is then accelerated by the field in the opposite direction to the electron and it can, in turn, generate by impact an electron hole pair. This positive feedback loop strongly enhances the statistical fluctuations of the multiplication process, which increases much more than proportionally to the gain. Hence, whereas PMTs provide a gain with a mean value around 10 6 and moderate fluctuations, the mean value of the gain obtainable in APDs with acceptable fluctuations is lower than 1000 in the best cases. Consequently, APDs can detect single photons only in the most favourable cases and their performance as single photon detectors is not very satisfactory. The intrinsic positive feedback that represents a significant drawback for APDs, however, is deliberately exploited in SPADs to obtain a different operating mode, similar not to an amplifier, but to a bistable circuit. This operation in principle is similar also to that of the Geiger Muller counters of ionizing radiations; hence, SPADs are also termed Geiger-mode (c) Figure 1. Regions of APD operation and SPAD operation in the reverse I V characteristics of a p-n junction; Schematic cross-section of planar p-n diode developed by Haitz et al. to investigate avalanche behaviour above breakdown [1, 2]; (c) Schematic cross-section of avalanche photodiodes with reach-through structure developed by McIntyre and Webb [3].

3 Evolution and prospects for SPADs and quenching circuits 1269 avalanche diodes. In the quiescent state the SPAD device is biased at voltage V a above the breakdown voltage V BD and no current flows (OFF-state): in the junction depletion layer the electric field is very high, but no free carriers are present. When even a single charge carrier is injected into the high-field region, it is strongly accelerated and can impact and generate a secondary electron hole pair, starting a self-sustaining avalanche multiplication process. The current then swiftly grows until the space charge effect limits its value to a constant level. This level is proportional to the excess bias voltage V E ¼ V a V BD, hence an avalanche resistance can be defined. The SPAD device is thus switched to the ONstate, where a constant macroscopic avalanche current flows (in the milliampere range). The fast onset of the current marks the time of arrival of the photon that generated the initial charge carrier. The device remains in this ON-state until the avalanche is quenched by an external circuit, which drives the applied voltage down to V BD (or lower). The circuit then concludes the operating cycle by resetting the voltage to the quiescent level. An avalanche current pulse with standard amplitude is thus generated. The detector is insensitive to any subsequent photon arriving in the time interval from the avalanche onset to the voltage reset, which is the detector dead time. It is evident that speaking of gain in the case of SPADs does not make sense, just as in the case of a bistable circuit. In a junction that has to operate as a SPAD, it is mandatory to avoid all factors that produce a local concentration of the electric field, such as edge effects or local defects of the material (metal precipitates etc.) called microplasmas [1, 2]. The higher electric field lowers the breakdown voltage level in this restricted area. The device may work as a SPAD only within the restricted sensitive area and with correspondingly low current, or it may not work as a SPAD at all, going into steady breakdown as soon as biased. The absence of local field concentration is the necessary condition to attain Geiger mode operation, but further stringent conditions must be fulfilled to attain acceptable SPAD performance, as discussed in section 2. Figures 1 and 1(c) outline the structure of the silicon devices where Geiger mode operation was first observed. Haitz et al. [1, 2] developed these devices in planar technology (figure 1), with a thin junction depletion layer (about 1 mm) and a correspondingly low V BD value (from 15 to 50 V). A deep diffused guard ring avoided the edge effects and defined a small sensitive area (diameter less than 10 mm); low-doped p-type silicon wafers currently exploited in industrial fabrication were employed. McIntyre and Webb [3] employed a dedicated technology and special ultralow-doped p-silicon wafers to develop the reach-through avalanche photodiodes (figure 1(c)). The devices had a thick depletion layer (from 20 to 100 mm) and a correspondingly high V BD value (from 100 to 500 V). The active area (diameter from 50 to 500 mm) was defined and edge effects were avoided by a p þ implanted enrichment and by a reduced thickness of silicon over it, obtained by accurately etching the wafer. 2. Device physics and detector performance The first feature in the performance of a photodetector is the quantum detection efficiency (QE). Besides the physical phenomena that determine the performance of semiconductor photodiodes in general (reflection, absorption, etc.), other physical effects have key importance in SPADs. For a photon to be detected, not only must it be absorbed in the detector active volume and generate a primary electron hole pair.

4 1270 S. Cova et al. It is also necessary that the primary carrier succeeds in triggering an avalanche. The avalanche initiation probability increases with the excess bias voltage V E, because it is enhanced by a higher electric field. Theoretical and experimental studies [5, 6] have shown how this probability first increases linearly with low V E and then tends to saturate at high V E. It is worth stressing that it is necessary to consider the V E value not alone, but together with that of V BD. For instance, V E ¼ 5 V is a high excess bias for a SPAD with V BD ¼ 15 V, but a low excess bias for a SPAD with V BD ¼ 300 V. Data of QE experimentally obtained are in good agreement with the computed values, as illustrated in figure 2. The second basic feature is the detector noise. Free carriers are thermally generated also in absence of illumination and produce dark current pulses. The dark-counting pulse rate has the same role as the dark current in ordinary photodiodes, that is, its Poissonian fluctuations are the internal noise source of the detector. Low-noise SPADs, competitive with good PMTs, should have a dark-counting rate of a few thousand counts per second (kcps) or lower. The darkcounting rate of SPADs, as illustrated in figure 2, increases with the excess bias voltage V E. The rise is due not only to the avalanche initiation probability, which increases also the QE, but also to effects due to the electric field that enhance Figure 2. Quantum detection efficiency at 850 nm wavelength versus excess bias voltage for a planar epitaxial SPAD with V BD ¼ 17 V; Dark-counting rate versus excess bias voltage for the same device. The active area diameter is 10 mm.

5 Evolution and prospects for SPADs and quenching circuits 1271 the rate of generation of carriers. It is well known that in silicon and in other semiconductors the transition of carriers from a band to the other occurs through the generation recombination (GR) centres, that is, local levels with mid-gap energy. The Poole Frenkel effect caused by a high electric field can enhance the emission from the local level to the band at higher energy, giving rise to a fieldassisted generation process. At very high field intensity a tunnel-assisted direct band-to-band transition may take place, that is, a generation of free carriers in the junction without the assistance of GR centres. Even a very small probability of band-to-band tunneling, corresponding to an extremely small tunnel current, can cause a significant increase of the dark-counting rate: a tunnel current of A corresponds to the generation of more than ten thousand carriers per second that can trigger the avalanche. Tunnel-assisted generation is not reduced by lowering the temperature and sets a limit to the reduction of the dark-counting rate obtained by cooling the detector. Some important conclusions can be drawn for the design of SPADs. In order to avoid enhancing the detector noise, the electric field intensity must not be much higher than the level necessary to attain avalanche multiplication. With such a constraint, high quantum detection efficiency can be obtained by means of a field distribution that ensures a sufficiently extended region of avalanche multiplication within the junction. The noise in SPADs is further enhanced by an effect that does not play any role in ordinary APDs. Impurities and crystal defects cause not only GR centres at mid-gap, but also local levels at intermediate energy between mid-gap and band edge, called deep levels. Some of these deep levels act as minority carrier traps. During each avalanche pulse, a few avalanche carriers can be trapped in these levels and are subsequently released, as outlined in figure 3. The released carriers can re-trigger the avalanche, thereby generating correlated afterpulses [2, 7, 8]. The release is statistical; the emission probability per unit time has a characteristic value for each type of level involved, whose reciprocal is the time constant (trap lifetime) of the exponential depopulation transient for that level type. As illustrated in figure 3, the overall transient of afterpulse generation after an avalanche can be measured with a technique developed in our laboratory [7] and the exponential components can be identified. The afterpulsing effect introduces a positive feedback loop that very significantly increases the dark-counting rate, unless the overall probability of generating an afterpulse is fairly small [9]. Since the population of the trap levels is far from saturation [8, 9], it linearly increases with the charge that flows during the avalanche pulse. Therefore, in order to reduce the afterpulsing effect, the avalanche pulse charge should be reduced as far as possible. The features of the avalanche quenching circuit can play an important role to this purpose [8]. The circuit can further contribute to reduce afterpulsing by enforcing a longer hold-off time after each avalanche pulse [8], of course at the expense of a correspondingly increased counting dead time. For instance, for the silicon SPAD of figure 3 it may be checked that a hold-off time of 500ns reduces the probability to less than 1% and that the count losses due to the increased dead time are acceptable in many applications. However, the lifetimes of the traps exponentially increase if the device is cooled for reducing the noise caused by thermal generation. As shown by the Arrhenius plot in figure 3(c), a temperature reduction of a few tens of kelvin increases the lifetime by an order of magnitude; in fact, the activation energy of each component corresponds to the energy difference of a few 0.1 ev from the trap level to the band. Hence, to reduce

6 1272 S. Cova et al. (c) Figure 3. Trapping and release of an electron by a deep level; Probability density of afterpulse generation after an avalanche in a silicon SPAD operating at room temperature. (c) Arrhenius plot of the lifetimes of components of the afterpulsing transient at different temperatures. the afterpulsing probability in a SPAD device operating at cryogenic temperature a very long dead time is necessary: tens of microseconds and more. Such a long dead time causes very significant count losses and a strong reduction of the data collection rate. This is a main drawback of SPAD devices intended to operate in the infrared spectral range, which are necessarily fabricated in semiconductor materials with narrower bandgap than silicon and must operate well below ambient temperature to reduce the primary dark-counting rate (see section 4). The device design is important for obtaining SPADs with high sensitivity and low noise, but the main challenge is the development of suitable technologies for device fabrication. It is necessary to reduce as far as possible the defects that cause GR centres and trap levels; in particular, metal impurities must be reduced to vanishingly small concentration. The key problem is to develop gettering process steps that are very efficient to that purpose and at the same time are compatible with the design of the SPAD structure. 3. Photon timing The onset of the avalanche pulse is correlated with the arrival time of the photon that generates the initial carrier. Due to various physical effects, however,

7 Evolution and prospects for SPADs and quenching circuits 1273 the delay of the instant at which the onset is sensed with respect to the true arrival time of the photon is not constant, but subject to statistical fluctuations. A typical delay distribution is reported in figure 4 for the planar device structure developed by Haitz (figure 1). Two main components are evident. The narrow peak has a full width at half-maximum (FWHM) of about 60 ps and is due to carriers photogenerated in the junction depletion layer, which are immediately accelerated by the electric field (figure 4). The slow tail is due to carriers photogenerated in neutral regions nearby, that walk around by diffusion and eventually reach the edge of the depletion layer, where they are finally accelerated by the electric field. This time-dependent diffusion effect was studied in various device structures with a simple simulation approach [10]. The tail limits the resolution attained in photon timing; furthermore, its amplitude and duration are wavelength-dependent, due to the dependence of the optical absorption coefficient (i.e. of the penetration length). This is a further remarkable drawback in applications where the light source is not monochromatic. It is clear that SPAD devices Figure 4. Photon arrival time distribution measured in a standard time-correlated photon counting set-up with laser diode pulses of about 20 ps duration at 830 nm wavelength and a planar Si-SPAD device of the type in figure 1; Sketch of the avalanche generation by a photon absorbed in the SPAD depletion layer and by a photon absorbed in a neutral region nearby.

8 1274 S. Cova et al. intended for photon-timing applications should have a structure designed for minimizing the diffusion tail effects. The FWHM of the main peak exceeds the contribution due to the noise in the pulse-timing circuits. It was therefore ascribed to statistical fluctuations in the build-up of the avalanche, from the first seed (the photogenerated electron hole pair) up to the macroscopic level set by the sensing threshold of the timing circuit. Various research groups have endeavoured to explain and compute it in terms of a local statistical build-up, with essentially one-dimensional models of the current rise. This approach, however, led to FWHM values shorter than 10 ps. As possible source of further fluctuations, the lateral propagation of the avalanche from the first seed to the whole detector area was then investigated in our laboratory. A simple lateral diffusion of carriers within the junction was too slow for playing a role in such a fast propagation. However, Lacaita et al. [11] finally found that a multiplication-assisted lateral diffusion of carriers takes place and produces a lateral propagation with high velocity. The sketch in figure 5 outlines this process, which essentially proceeds as follows. As the first filament of avalanche current is started, the strong gradient of carrier concentration at its edge drives the lateral diffusion. A few carriers diffuse and in turn start avalanche multiplication in the sheath surrounding the first filament, thereby pushing swiftly the high gradient region outwards. Lateral diffusion then occurs at the new outer edge of the current; and so on. The rise time of the avalanche current corresponds to the time taken by the current to propagate and fill the whole detector area. This time systematically depends on the position of the initial seed; in fact, with SPAD devices of suitably stretched geometry the effect can even be exploited to obtain a Figure 5. Sketch of phenomena that support lateral propagation of the avalanche current started by a photon: lateral diffusion of carriers assisted by avalanche multiplication; absorption within the detector area of photons emitted by hot carriers in the avalanche current.

9 Evolution and prospects for SPADs and quenching circuits 1275 position-sensitive single-photon detector [13]. On the other hand, if the position of the initial seed within the area is random, also the delay from start to threshold crossing is randomly distributed. Another contribution to the lateral propagation is brought by the photon emission from hot carriers in the avalanche, as pointed out by McIntyre [14]. The emission probability is very low: on average, about one photon is emitted every carriers crossing the junction [15]. As outlined in figure 5, the process is inherently random: photons emitted by the first avalanche current filament are randomly absorbed at different locations within the depletion layer and start other current filaments; these filaments in turn emit photons, and so on. In planar Si-SPADs devices, due to the thin junction depth the photon-assisted propagation is not very efficient and the multiplicationassisted diffusion is dominant. On the other hand, in devices with thick depletion layer, such as the reach-through avalanche diodes, the propagation by photons is much more efficient and is actually dominant. Anyway, the efficiency of both processes is enhanced by a higher electric field; hence, by increasing the excess bias voltage V E the photon-timing resolution is improved in all cases [8]. An accurate model of the SPAD device, developed by Spinelli and Lacaita [16], is nowadays available for evaluating and analysing by numerical simulation the timing performance of any Si-SPAD device. 4. SPAD devices 4.1. Silicon SPAD detectors Devices specifically devised for operating as SPADs have been developed in silicon. A good basis was provided by the deep technological know-how produced over decades of activity in industrial and research laboratories on the development of silicon microelectronics, but specific efforts were anyway necessary. Good SPAD performance is obtained with two different device structure, outlined in figure 6. The planar epitaxial devices outlined in figure 6 were devised at our laboratory [17] for overcoming the drawbacks of the simple planar structure of figure 1. The epistrate substrate p-n junction limits the neutral region from which carriers are collected, thereby reducing the diffusion tail effect in the photon-timing distribution. The high-field area (that is, the detector active area) is defined by the p þ implant in the upper central region. The outer lightly doped p region is the guard ring that avoids edge breakdown. The buried p þ layer provides a low resistivity path to the avalanche current; the internal device resistance is typically about 1 k. The breakdown voltage V BD is typically 20 V; by tailoring the epistrate thickness and the doping profile, it can be adjusted in the range 15 to 50 V. The depletion layer is typically 1 mm thick. The corresponding photon detection efficiency is high in the visible and declines in the red and near infrared (NIR) spectral range: typical values are over 40% at 500 nm, 30% at 630 nm, 15% at 730 nm, 10% at 830 nm and a few 0.1% at 1064 nm. Higher quantum efficiency can be obtained with a thicker depletion layer, which, however implies also higher V BD. A ring of heavy phosphorous predeposition and deep diffusion surrounds the device, providing electrical isolation and effective gettering action on the active area. Low dark-counting rate at room temperature (a few kcps or lower) and low afterpulsing probability (a few per cent or lower) are obtained in devices with diameter of the active area up to 30 mm. The photon-timing resolution is

10 1276 S. Cova et al. Figure 6. Schematic cross-section of silicon device structures designed for operating as SPADs: planar epitaxial device developed at authors laboratory [17]; Slik TM device developed at the former RCA ElectroOptics, now PerkinElmer Optoelectronics [20]. remarkable, as shown in figure 7. The peak has better than 30 ps FWHM; the diffusion tail is one decade lower and has simple exponential shape with about 270 ps lifetime [17]. To attain even better photon-timing resolution, completely free from a diffusion tail, the special device structure shown in figure 8, called a dual-junction SPAD, has been developed [18]. In correspondence to the detector active area the neutral p-layer is eliminated by patterning the p þþ buried layer and so the depletion layers of the active junction and of the epistrate substrate junction meet each other. The carriers diffuse only in the thin n þ upper layer and therefore the diffusion tail is practically eliminated (figure 8). Planar epitaxial Si-SPADs, thanks to the low voltage, have low power dissipation during the avalanche, at most a few hundred milliwatts, and it is not necessary to cool the detector under operating conditions. The devices are fabricated in ordinary silicon wafers with a processing technology compatible with the fabrication of integrated circuits and/or arrays of detectors on a chip. The fabrication yield is high; the devices are rugged and reliable and can be produced at low cost in high volume. In recent years, devices of this kind have been developed in various laboratories [19]. The Slik TM device sketched in figure 6 was devised at the former RCA Optoelectronics, now PerkinElmer Optoelectronics, and there employed to produce the highly successful single-photon counting modules SPCM [20]. Slik TM stands for Super low k, where k denotes the ratio of the ionization coefficient of holes to that of electrons. The device represents a remarkable evolution of the

11 Evolution and prospects for SPADs and quenching circuits 1277 Figure 7. Photon-timing resolution tests: planar epitaxial device of figure 6 with an active area diameter of 10 mm; Slik TM device of figure 6 in a PerkinElmer SPCM-AQR module [18]. Tests carried out in a time-correlated photon counting setup with laser diode pulses of 20 ps duration at 830 nm wavelength. reach-through avalanche diode structure of figure 1(c). It is built in special ultrapure high-resistivity silicon wafers with a dedicated technological process; various device features and technological process steps are proprietary and covered by patents [21]. The active area of the detector is fairly wide (diameter about 200 mm). It is defined by a p þ implant and deep diffusion in the central region and by a reduction of the wafer thickness to about 30 mm in correspondence to it, obtained by accurately etching the back of the wafer. A light diffused n guard-ring around the shallow n þ zone contributes to avoid edge breakdown. The internal device resistance is typically about 500. The device is illuminated from the back (figure 6). Diffusion of photogenerated carrier takes place only in the surface p þ layer (a few micrometres thick), since the lightly doped p region (from 20 to 30 mm thick) is fully depleted. The decrease of the electric field from its maximum at the n-p junction is gradual, hence the avalanche region is fairly extended. The

12 1278 S. Cova et al. Figure 8. Schematic cross-section and photon timing resolution of the dualjunction SPAD developed at the authors laboratory [18]. breakdown voltage V BD is high and varies from sample to sample in a wide range, from 250 to 500 V. Thanks to the thick depletion layer, the photon detection efficiency is very high in the visible region and fairly good in the NIR up to about 1 mm wavelength. The typical value is remarkably higher than 50% over all the range from 540 to 850 nm wavelength and is still about 3% at 1064 nm [20]. A proprietary recipe produces excellent gettering action: notwithstanding the remarkable volume of the depletion layer, the dark-counting rate at room temperature is very low, typically a few kcps. The afterpulsing probability is also strongly reduced, typically well below 1%. The photon timing resolution is fairly good, as shown in figure 7: the peak has about 350 ps FWHM and the exponential diffusion tail is one decade lower and has about 160 ps lifetime. The Slik TM devices have very good performance, but also a number of practical drawbacks. Due to the high breakdown voltage V BD, the power dissipation during the avalanche is high, from 5 to 10 W: very effective cooling of the detector in operating conditions is mandatory [20] (with Peltier cells or other means). The special fabrication technology is inherently complex, the production yield of good devices is low and the cost is high. The devices are delicate and degradable and there is no perspective of producing chips that monolithically integrate various detectors (i.e. quad-cells, arrays, etc.) or a detector and the associated electronic circuitry. Table 1 summarizes the main features of the two types of silicon SPAD devices reviewed here.

13 Evolution and prospects for SPADs and quenching circuits 1279 Table 1. Comparative summary of the main features of Si-SPAD devices of figure 6. For simplicity, planar epitaxial devices are denoted as Thin Si-SPAD s and Slik TM devices as Thick Si-SPAD s. Thin Si-SPADs Thick Si-SPADs. Good photon detection efficiency. Very good photon detection efficiency. Low dark-counting rate (low noise). Low dark-counting rate (low noise). Photon-timing at few 10 ps. Photon-timing at few 100 ps resolution resolution. Sensitive area diameter up to 30 mm. Sensitive area diameter 200 mm. Low voltage: 15 to 50 V. High voltage: 250 to 500 V. Low power dissipation: cooling not necessary. High power dissipation: Peltier cooler mandatory. Standard Si substrate. Ultrapure high-resistivity Si substrate. Planar fabrication process compatible with array detectors and integrated circuits (ICs). Dedicated fabrication process not compatible with array detectors and integrated circuits (ICs). Robust and rugged. Delicate and degradable. Low cost. High cost 4.2. SPAD detectors for the infrared spectral range In order to detect photons with wavelength longer than 1.1 mm, devices in semiconductor materials with bandgaps smaller than the E G ¼ 1.1 ev of silicon (germanium and compound semiconductors) must be employed and the detector must be cooled below room temperature. Germanium devices have structures similar to silicon devices, but the material has less favourable properties and the technology is much less developed. Ge-diodes specifically designed for operating as SPADs have not been reported. Commercially available devices are designed for operating as APDs and selected samples with sufficiently uniform breakdown over the active area can work also in Geiger mode. In the early 1970s, investigations of the physics of germanium avalanche diodes operating in Geiger mode were reported [22], but not until the early 1990s was an extensive study carried out on the performance of germanium diodes in photon counting and timing [23]. A summary of the situation is given here and a full report and discussion can be found in [23]. Ge-SPAD devices must be cooled to cryogenic temperatures to reduce the dark-counting rate to acceptable levels. The available devices are designed to have high electric field and so tunnel effects (section 2) set a limit to the reduction of carrier generation rate. The dominant role is played by the enhancement of the dark-counting rate caused by the very strong afterpulsing effects. The trap concentration is high and the trap lifetime at low temperature is quite long, about 10 ms or more. The excess bias voltage must be carefully selected due to a trade-off with the dark-counting rate on one side, and the photon detection efficiency and timing resolution on the other side (see sections 2 and 3 and [23]). Optimum sensitivity at 1.3 mm wavelength is typically achieved at an excess bias of a few 0.1 V, with detection efficiencies of a few per cent and dark-counting rate of a few 10 kcps. Good performance can be achieved in free running operation with a fairly long hold-off time, at least a few microseconds. On the other hand, gated detector operation [8] is easy and very effective in the reduction of the dark-current enhancement caused by trapping

14 1280 S. Cova et al. effects. The DC bias is set just below the breakdown level and a bias pulse with amplitude corresponding to the excess bias voltage is employed to gate-on the detector during the short intervals when the optical signal arrives. During the long quiescent intervals between the gate-on times the trap levels release the carriers without triggering the avalanche [8, 23]. Regarding photon-timing, good resolution with better than 85 ps FWHM has been demonstrated. Other drawbacks are met in the operation of Ge-SPADs, such as transient shifts of the breakdown voltage due to self-heating and to majority carrier trapping. The latter effect emphasizes how far the germanium technology lags behind silicon: majority carrier trapping effects were observed in silicon devices during the 1960s [24], but they were completely eliminated in the following decade. The driving force for the development of APDs in III V compound semiconductors (in particular InGaAs with E G ¼ 0.73 ev) is the development of optical fibre communication systems operating at 1.55 mm wavelength. From the standpoint of SPADs, the perspective of attaining with these devices higher detection efficiency at longer wavelengths and lower noise is particularly interesting, because at 77 K the germanium cut-off wavelength is shifted down to 1.45 mm. APD devices with SAM structure (separate-absorption-and-multiplication) have been introduced and developed for communication systems. IR photons are absorbed in a fully depleted lightly doped InGaAs layer with low energy gap (E G ¼ 0.73 ev), whereas the high field is established and multiplication occurs in a InP layer, which has more favourable properties for the avalanche process (E G ¼ 1.35 ev). The concept evolved to the SAGM structure (separate-absorption-and-gradedmultiplication), where a graded quaternary InGaAsP layer is interposed to make a more gradual the transition of the band structure, as outlined in figure 9. In 1985 at the ATT Bell Labs it was demonstrated that such devices can detect single photons, although with severely limited performance [25]. In 1995, a thorough investigation of the performance of SAGM devices was carried out [26] and opened the way to their practical application. A full report can be found in [26] and a condensed summary will be given here. Device samples must be selected that have not only sufficiently uniform breakdown over the active area, but also a sufficiently high value of the breakdown voltage V BD at room temperature. More precisely, V BD must be sufficiently higher than V RT (reach-through voltage), at which the depletion layer is extended enough to include the InGaAs layer, and IR photons can consequently be detected. The latter condition ensues from the fact that V BD markedly decreases with the temperature, whereas V RT is fairly constant. If V BD at the operating temperature becomes lower than V RT the detector is of no use for the IR range, since only photons absorbed by InP can be detected. Even a value of V BD just slightly higher than V RT is not suitable, since in such conditions the field at the hetero barrier will be low and it will be more difficult for a carrier photogenerated in the InGaAs to cross the barrier and reach the avalanche region, as sketched in figure 9. The device must be cooled, but the thermal generation of carriers can be reduced to acceptable levels even at moderately low temperature, attainable with Peltier cooling. However, the carrier trapping effects are extremely intense, due to a concentration of trapping levels much stronger than in Ge-devices: the consequent afterpulsing effects are by far the dominant limitation to the performance and application of InGaAs/InP devices. Free running SPAD operation is, in practice,

15 Evolution and prospects for SPADs and quenching circuits 1281 Figure 9. Cross-section of InGaAs/InP avalanche photodiode with SAGM structure; Schematic band diagram and sketch of a free carrier crossing the hetero-interface. out of the question, but interesting performance can be attained in gated detector operation, in particular with very short gate-on intervals in the nanosecond range. Regarding photon-timing with IR photons, the performance of available devices is limited mainly by effects at the hetero interface (see figure 9), which cause random delays in the transit of carriers photogenerated in the InGaAs layer [26]. Resolution of a few hundred picoseconds has been achieved and there is no reason why a suitably designed InGaAs/InP device should not attain timing resolution well below 100 ps, as with silicon and germanium devices, since the avalanche physics is the same in all types of semiconductor. The very high density of minority carrier trapping levels found in III V devices [26] is a clear signature that the processing technology is not only far from reaching the silicon standard, but is still not developed and clean as the germanium one. There is plenty of room for improvement, as pointed out also by recently

16 1282 S. Cova et al. reported results. New InGaAs/InP SAGM-APDs (developed in recent years for optical fibre communication systems) tested as SPADs have shown improved performance in photon counting and timing, though still limited to gated detector operation [27]. A SAM avalanche photodiode specifically designed for Geiger mode operation has been reported, although intended for detecting 1.06 mm photons and therefore with absorbing layer not in InGaAs, but in quaternary InGaAsP [28]. The device is still limited to gated detector operation, but provides interesting performance at moderately low temperature and can operate up to room temperature, though with very high dark-counting rate. Alternatives to the improvement of the InGaAs/InP device technology may be given by the development of new technologies. APD s have been announced that exploit absorption in InGaAs and avalanche multiplication in Silicon, built with sophisticated wafer bonding techniques [29]. Operation of these devices in Geiger mode is expected to be possible, but has not yet been demonstrated. Also modern Si Ge device technologies may be employed to build IR-sensitive avalanche photodiodes. Operation in Geiger mode of devices of this kind has been reported, but so far the SPAD performance obtained is marginal, with very low photon detection efficiency and very high dark counting rate [30]. A general conclusion can be drawn, valid for any given technology capable of producing avalanching junctions suitable for Geiger mode operation: in order to obtain SPAD s with good performance the ultimate challenge is to develop technological fabrication processes that reduce to negligible level the density of states that cause minority carrier trapping and afterpulsing, besides that of centres giving rise to carrier generation. 5. Role and evolution of the quenching circuit The circuit that quenches the avalanche and resets the bias voltage plays a key role in the SPAD detector performance. A thorough analysis and discussion of the matter can be found in [8] and the main issues will be summarized here. The early studies on avalanche diodes in Geiger mode employed the simple passive quenching circuit (PQC), outlined in figure 10. The bias voltage is applied through a ballast resistor R L with high value; a small resistor R S is connected to the other terminal to observe the current pulse. The avalanche current discharges the total capacitance C T at the diode terminal, the sum of the junction capacitance C D and of the stray capacitance C S.The voltage on the diode decreases towards V BD and the avalanche current correspondingly decreases. When the voltage approaches V BD the rate of decrease slows down; practically all the avalanche current flows through R L and is reduced to the value (V a V BD )/R L.IfR L is high enough to reduce the current to about 20 ma, the number of avalanche carrier is small, the probability of interruption of the multiplication chain is high and the avalanche is finally quenched. The voltage then starts to recover slowly towards the bias voltage V a, as the small current in R L recharges C T with a long time constant R L C T. During the recovery the diode voltage is higher than V BD and the avalanche can be triggered, but the diode fires at a voltage lower than V a (figure 10). It then operates with lower photon detection efficiency and impaired photon-timing resolution. After each avalanche pulse, the triggering probability has a continuous evolution, starting from practically nil and finally reaching the steady value. The behaviour of the detector is thus

17 Evolution and prospects for SPADs and quenching circuits 1283 (c) (d) Figure 10. Schematic circuit diagram of a SPAD in the passive quenching arrangement (typical values R L ¼ 500 k and R S ¼ 50 ); Waveform of the avalanche current (upper) and of the voltage applied to the SPAD (lower); (c) Schematic diagram of a SPAD in the active quenching circuit (AQC); (d) Output pulses from the AQC at about 2 Mc/s pulse counting rate (horizontal scale 10 ns/div; vertical scale 0.5 V/div). The dead time of 36 ns is clearly observable. peculiar: it does not have a well-defined dead time, but rather a time-dependent sensitivity to triggering events. The question is further complicated by the fact that the voltage and current pulses produced during the recovery have smaller amplitudes, as shown in figure 10. Since a comparator is employed for sensing them, pulses smaller than the threshold level are discarded and a dead time is introduced, which is neither well known, nor very stable. The result is a loss of linearity at high counting rates, which may be measured empirically, but for which equations for accurate correction of the count losses are not available. The drawbacks of the passive quenching can be reduced, though not eliminated, with modern circuit technologies that reduce significantly the stray capacitance C S and thus make faster the recovery transient. With surface mounting techniques and miniature components C S can be reduced to a few picofarads. Monolithic integration of detector and ballast resistor, nowadays possible at least for silicon SPADs, can reduce C S well below 1 pf and the recovery transient well below 1 ms. A solution that completely avoided the drawbacks of passive quenching was the active quenching circuit (AQC), first devised in 1975 [31]. The principle is simple: to sense the rise of the avalanche pulse and react back on the SPAD, forcing with a controlled bias voltage source the quenching and reset transitions in short times. As outlined in figure 10(c), the sensing comparator sends a command to a voltage driver, which switches the bias voltage down to the breakdown voltage V BD or below it. After an accurately controlled hold-off time, the bias voltage is switched

18 1284 S. Cova et al. back to the operating level V a. A standard pulse synchronous to the avalanche rise is derived from the comparator output, to be employed for photon counting and timing. The approach is fairly simple and bears some similarity to an approach employed in past times for working with true Geiger Mueller gas detectors of ionizing radiations. However, completely new problems arise with SPADs, due to the much faster time scale and to the role played by capacitances, as discussed in [8]. The basic advantages offered by the AQC approach are the fast transitions from quenched state to operating level and conversely, the short and well defined duration of the avalanche current and of the deadtime. These features have made possible to carry out accurate photon counting measurements up to high counting rates of million of counts per second (Mc/s) (see figure 10(d)) and also to maintain high resolution in photon timing at such high rates, overcoming the strong limitations previously met with PQCs [8]. The introduction of the AQC principle in 1975 opened the way to practical applications of SPADs, but it was only in 1981 that extensive work on the subject was started. It was demonstrated with AQCs high resolution photon timing [32] and fast gating of the detector [33] and, looking at the essential features of the circuits, we introduced the terminology of active quenching circuits and passive quenching circuits now universally adopted [32]. A steady evolution of the AQC design followed with contributions from various laboratories (see bibliography in [8]). The technological progress made possible progressive reduction of size and power consumption. Compact modules were produced with surface mounting techniques [20, 34]. Work was undertaken in our laboratory to design and produce integrated AQCs, in order to attain true miniaturization of photon-counting modules. We first integrated the low-voltage circuitry, that is, all the AQC except the quenching voltage pulse driver [35]. By exploiting new fabrication technologies, recently made available at silicon foundries, we finally achieved full monolithic integration of the AQC [36]. Figure 11 shows a photograph of an integrated active quenching circuit (i-aqc) chip developed in high voltage CMOS technology and figure 11 shows the corresponding quenching pulse waveform. As shown in section 3, in order to attain high resolution in photon timing the circuit should extract the time information from the very first part of the avalanche rise. This requirement may be conflicting with other constraints set on the design of an AQC. In fact various AQCs are unable to exploit the timing performance demonstrated for the SPAD detector at low counting rate with a simple PQC [37]. A simple approach was devised to overcome this limitation and patented [38]. The approach was to have a suitable pick-up of the fast avalanche current signal from the SPAD terminal biased at high voltage, leaving the SPAD terminal at ground potential connected to the AQC. As outlined in figure 12, the solution is a simple AC (alternating current coupling) pickup. An essential feature is a short differentiating time constant, in contrast with the usual art in AC coupled amplifiers and other circuits, which employs long time constants to transmit faithfully the pulse shape. In our case a long time constant causes degradation of the timing performance at high counting rate. This ensues because the AC coupling inherently generates after each pulse a small, but very long tail with opposite polarity. The random superposition of such tails causes a random shift of the baseline, hence a random walk of the time at which a pulse crosses the threshold of the fast comparator that senses its arrival. Figure 12 compared

19 Evolution and prospects for SPADs and quenching circuits 1285 Figure 11. Photograph of the silicon chip of an integrated active quenching circuit (i-aqc) developed at authors laboratory; Quenching pulse waveform (horizontal scale 20 ns/div, vertical scale 10 V/div). with figure 7 gives an example of the improvement obtained with the timing pick-up. 6. Conclusions Silicon SPAD technology is fairly advanced and can be further improved. Developing highly efficient and low-cost Si-SPAD detectors appears nowadays feasible. The availability of monolithic integrated AQCs makes it possible to develop really miniaturized detector modules, possibly even monolithic integrated modules. The i-aqcs are important also from the standpoint of the development of SPAD arrays. Si-SPAD array detectors are a realistic prospect, at least for applications where a certain degree of crosstalk between pixels can be tolerated. In fact, the emission of photons by the hot carriers in the avalanche current (see section 3 and figure 5) is an intrinsic source of optical crosstalk between pixels [39]. It is also possible to pursue the elimination (or at least a strong reduction) of such a crosstalk in SPAD arrays by interposing some absorbing or reflecting structure around each pixel, at the expense of a significant complication of the fabrication technology. All the technologies presently available for developing infrared sensitive SPADs in materials different from silicon (Ge, III V and II VI compound semiconductors, Si Ge, wafer-bonded Si and III V) lag far behind silicon technology and require further progress. The key challenge is to develop processing steps that provide a strong reduction of the density of the trap levels that give rise

20 1286 S. Cova et al. Figure 12. Pulse pick-up circuit to obtain improved photon timing resolution with existing active quenching circuits; Photon-timing resolution obtained by introducing the pulse pick-up circuit in a PerkinElmer SPCM-AQR module; to be compared with the standard one in figure 7. to afterpulsing effects. On the other hand, the development of a suitable technology to produce IR-sensitive SPADs may open remarkable new perspectives: devoting remarkable efforts to such an objective appears to be worthwhile. Acknowledgements This work was supported by Ministero dell Istruzione, Universita e Ricerca (MIUR) under FIRB contract No. RBNE01SLRJ. References [1] GOETZBERGER, A., MCDONALD, B., HAITZ, R. H., and SCARLETT, R. M., 1963, J. Appl. Phys., 34, [2] HAITZ, R. H., 1965, J. Appl. Phys., 35, 1370; HAITZ, R. H., J. Appl. Phys., 36, [3] MCINTYRE, R. J., 1972, IEEE Trans. Electron Devices, Ed-19, 703; WEBB, P. P., MCINTYRE, R. J., and CONRADI, J., 1974, RCA REVIEW, 35, 234. [4] WEBB, P. P., and MCINTYRE, R. J., 1970, Bull. Am. Phys. Soc., Ser. Ii 15, 813. [5] MCINTYRE, R. J., 1973, IEEE Trans. Electron Devices, Ed-20, 637. [6] OLDHAM, W. O., SAMUELSON, R. R., and ANTOGNETTI, P., 1972, IEEE Trans. Electron Devices, Ed-19, 1056.