LINEAR-MODE SINGLE-PHOTON-SENSITIVE AVALANCHE PHOTODIODES FOR GHZ-RATE NEAR-INFRARED QUANTUM COMMUNICATIONS

Transcription

1 LINEAR-MODE SINGLE-PHOTON-SENSITIVE AVALANCHE PHOTODIODES FOR GHZ-RATE NEAR-INFRARED QUANTUM COMMUNICATIONS Andrew Huntington, Madison Compton, Sam Coykendall, George Soli, and George M. Williams Voxtel, Inc. Beaverton, Oregon ABSTRACT We present design and performance data for a highspeed telecom-band (1.3 μm) single-photon-sensitive receiver based on a new class of multi-stage InGaAs avalanche photodiode (APD) operated in the proportional mode (linear mode). Peak photon detection efficiency of 7% was measured at 1.64 μm. Unlike Geiger-mode single-photon-sensitive APDs (SPADs), a multi-stage linear SPAD is operated below its avalanche breakdown voltage and need not be gated to suppress afterpulsing, permitting operation at a much higher maximum count rate (MCR) without an associated increase in dark count rate (DCR). Further, the linear-mode APD preserves signal amplitude information, whereas the response of a Geiger APD is binary. The multi-stage APDs demonstrated can be operated with a linear gain in excess of M = 8, and have silicon-like multiplication noise characterized by an effective ionization coefficient ratio of k =.3 out to M = 1. INTRODUCTION Detection of a single photon is required in many quantum communication applications. Much research activity (both theoretical and experimental) on singlephoton detectors has been reported. 1,2 Though excellent low-noise single-photon-detection performance has been reported for superconducting devices, low operating temperature requirements (normally <4 K) limit their practical use. In real applications, APDs are used more often due to their lower cost, smaller package, and ease of employment. GHz-rate APD-based photon counting systems are attractive for quantum communication applications, but currently available high-speed telecom-band APDs suffer from high multiplication noise (k.4) and low gain (M < 3), preventing their application to photon counting. These limitations are a function of the InGaAs-on-InP material system required to efficiently detect telecomband photons in the μm spectral range. The noisy multiplication process is a function of the bulk properties of the alloys (typically InP or InAlAs) that are compatible with InGaAs, and the noisy multiplication process makes stable operation at high gain infeasible. Photon counting in Geiger mode is possible with an InGaAs APD, but Geiger-mode operation is undesirable for high-count-rate applications because of a fundamental tradeoff between DCR and MCR. If a Geiger APD is cooled to reduce its DCR, then it must be quenched below its breakdown voltage in between detection events to avoid afterpulsing. This substantially reduces its MCR. Development of a telecom-band linear APD with very high gain and very low noise eliminates the trade between DCR and MCR because the receiver need not be gated off between detection events if the APD is operated in linear mode. A basic linear-mode APD receiver consists of the APD detector element and a transimpedance amplifier (TIA). The APD converts incident photons to primary photocarriers, and amplifies the resulting primary photocurrent through internal avalanche gain. The TIA converts the APD s current signal into a voltage signal; if a capacitive feedback TIA (CTIA) is used, then the voltage is proportional to the total multiplied charge delivered by the APD. When the receiver is configured as a thresholded detector i.e., where a count is registered if the voltage exceeds a programmed threshold then a direct comparison to Geiger SPADs can be made. Like a Geiger-mode SPAD, a thresholded linear-mode photoncounter can be configured to discard signal amplitude information. Such a thresholded receiver can be considered single-photon-sensitive if it has a reasonably high probability of registering a count in response to a single-photon signal. Alternatively, the signal amplitude data may be recorded when a count is registered, preserving that information. However, even in the case of a very low-noise APD, the statistical distribution of the APD s gain is wide enough to limit the accuracy of the amplitude measurement for weak signals. LINEAR APDS, GEIGER APDS, AND AFTERPULSING In linear mode, the reverse bias applied to the APD is held constant, and the primary photocurrent generated in the APD s absorber is amplified by a proportional multiplication factor that is independent of signal strength (below saturation). The output of a linear APD is /8/$ IEEE 1 of 6

2 proportional to the level of illumination it receives, and so a linear APD can read signal amplitude. Linear APDs are typically used in optical receivers to boost weak signals above the noise floor of the receiver s amplifier. In Geiger mode, the reverse bias is modulated, and the APD s response is binary. Geiger-mode operation of an APD involves momentarily biasing the diode above its avalanche breakdown voltage V br. The excess voltage applied is called the overbias. In this active state, avalanche breakdown of the diode junction can be triggered by as little as a single primary carrier, which is why the technology has been used for photon counting. However, the current that flows during avalanche breakdown is determined by the characteristics of the external circuit rather than the number of primary carriers that initiated the breakdown, so the breakdown current of a Geiger APD is essentially the same for all signal strengths and it is very large. Geiger APDs suffer long reset times following each detection event, due to the afterpulse phenomenon. That is, the very large breakdown current that flows in a Geiger APD during a detection event populates traps in the detector that release their trapped carriers over time (Figure 1). The sooner a Geiger APD is returned to service after it fires, the more likely it is to trigger off of a carrier that was trapped during the previous detection event, registering a spurious count. Quench times >1 µs are generally necessary, and the lower the temperature the Geiger APD is operated at to suppress its dark count rate, the longer the required quench duration. This is the physical origin of the tradeoff between DCR and MCR encountered with Geiger APDs. For quantum communication systems (e.g. quantum cryptography), today s low-speed single-photon Geigermode APDs are a bottleneck. Dark Carrier Genera on Rate APD Current Bias V br Geiger event (fills traps) Overbias Extra Dark Carriers Released by Traps A erpulses (Dark counts triggered by carriers released from traps) Quench (when pulse detected) Hold OFF Gate ON τ (Maximum count rate = 1/τ) Time Figure 1: Illustration of the afterpulse phenomenon that limits the count rate of a Geiger APD. In contrast, the current that flows in a linear-mode APD is too small to fill any appreciable population of traps, so linear APDs recover from detection events as soon as the current pulse clears the diode junction typically in 1 ns. In principle, a linear-mode APD can be cooled to reduce its DCR without lowering its MCR. Thus, linearmode photon-counting APD technology is one way to address the twin requirements of low DCR and high MCR. In the absence of afterpulsing, the MCR of a linearmode APD is determined by its impulse response. This is the time it takes for the avalanche of carriers in its multiplication region to complete (called the avalanche buildup time ) and for the last of the secondaries to clear the junction (related to the saturation drift velocity of the slowest carrier, which for holes in InGaAs is about cm s 1 ). The fastest InGaAs telecommunication APDs can operate at an avalanche gain of M = 1 with a 1-GHz bandwidth. This is achieved by minimizing the thickness of the junction, and operation at lower gain (to minimize the avalanche buildup time). Photon-counting APDs that must be sensitive to 1.55 μm light when operated cold require thicker absorption layers (perhaps 2 μm of InGaAs or more), and must be operated at considerably higher gain (M > 4). Nonetheless, a bandwidth of hundreds of MHz is achievable. The possibility of circumventing the fundamental tradeoff between DCR and MCR using linear-mode APD technology is exciting, but requires sensitive amplifiers. InGaAs APDS WITH BULK MULTIPLIERS Manufacturing low-noise APDs with high responsivity in the telecom wavelengths is a challenge because the III-V compound semiconductor alloys that are compatible with efficient InGaAs absorbers have fairly high bulk values of k, which is the ratio of the material s ionization coefficient for holes to that for electrons. For instance, InP the most common alloy used in telecommunications APDs has a k value of 2.5. The inverse ratio (k =.4) is often quoted to enable side-by-side comparison with electron-avalanche materials like Si or InAlAs, for purposes of calculating the excess noise factor, F(M). Lower k is better: 2 M 1 F ( M ) = M 1 (1 k). 3 (1) M On that basis, conventional silicon APDs (with k.2) greatly outperform SWIR APDs made from bulk InGaAs/InP (k.4) or InGaAs/InAlAs (k.3). DESIGN OF MULTI-STAGE APD INGAAS APDS APDs and other optoelectronic devices such as lasers are commercially manufactured in both the InGaAs/InP 2 of 6

3 NEP [fw/hz 1/2 ] dp = 5 na R =.875 A/W k =.4 k =.2 k =.2 Field (arbitrary units) / Number of Ioniza ons Electron-ini ated ioniza on Hole-ini ated ioniza on Electric field profile InAlAs Al.24 Ga.24 In.52 As Avalanche Gain (M) Figure 2: Effect of k on NEP calculated for APDs. and InGaAs/InAlAs alloy systems. Although both alloy systems span a similar band gap range (InP has a roomtemperature band gap of 1.34 ev; for InAlAs it is 1.46 ev), the band offset ratio is higher in AlGaInAs than in InGaAsP (.7 versus.4) and the impact ionization rate for electrons in InAlAs is higher than that for holes, but vice versa for InP. When grown as bulk alloys, InP and InAlAs have high values of k, and consequently poor excess noise performance in accordance with Equation (1) above. InP has a natural k value of.4; InAlAs is only slightly better, with k.3. The problem, then, is that the InGaAs alloy which absorbs SWIR efficiently is not compatible with an avalanche material that naturally has low excess noise. Our multi-stage APD design is an engineered InAlAs/InGaAs structure in which the k value is closer to that of silicon (k.2.4), so that a near infrared (NIR)-sensitive APD can be made with low excess multiplication noise. The impact of a reduction in k from.4 to.2 on the noise-equivalent power (NEP) of an APD receiver is illustrated in Figure 2. The desirability of lower k should be apparent. As noted earlier, conventional InGaAs APDs typically have bulk InP multiplication layers, characterized by an ionization coefficient ratio of k.4. We manufacture advanced low-noise APDs that have thinner InAlAs multiplication layers, characterized by k.2. We applied the same engineering principles to achieve k =.3 at a gain of M 1. We accomplished this through an advanced application of impact-ionization engineering (I 2 E). 4 The goal of I 2 E is to reduce excess multiplication noise by designing semiconductor structures in which the impact ionization events will naturally be correlated. In general, two tools are used: (1) the so-called dead-space effect and (2) localized enhancement of the ionization rate. Both reduce the number of possible ionization chains and, hence, narrow the distribution of the multiplication Loca on (arbitrary units) Figure 3: Monte Carlo simulation of a multiplication layer in which a change in alloy composition has been used to localize impact ionization. gain through spatial localization of the ionization events. Ionization events tend to be localized inside thin multiplication regions because, following each ionizing collision, carriers must pick up energy across a certain distance the dead space before they are capable of causing subsequent ionizations. By effectively eliminating some of the places where impact ionization can occur in the structure, the dead-space effect reduces the number of possible ionization chains; this is the source of the noise reduction observed for such APDs. A thin multiplication layer introduces little noise because the number of possible ionization chains is small, but by the same token, the long ionization chains necessary to produce high gain do not fit inside a thin multiplication layer. Higher gain can be obtained from a thin multiplication layer by increasing the field strength, but in doing so, feedback from hole-initiated ionizations is intensified and noise suppression is lost, precisely because a larger number of ionization chains can now fit into the same space. And not only do stronger fields degrade excess noise performance, they enhance dark current leakage through mechanisms such as band-to-band tunneling and thermionic field emission, so increasing field strength is counterproductive. Any technique that acts to localize impact ionization and eliminate some of the structure s possible ionization chains will also reduce its excess noise. The practice of building heterostructure multiplication regions from materials with dissimilar ionization thresholds falls into this category. A typical structure of this variety has a wide-band gap region on the p side of the multiplication layer and a narrower-band gap region on the n side; electrons that pick up energy in the wide-gap material do not trigger ionizations until they hit the lower-threshold narrow-gap material, whereas the holes generated by those collisions in the narrow-gap material fail to gain 3 of 6

4 Field (arbitrary units) / Number of Ioniza ons InAlAs Electron-ini ated ioniza on Hole-ini ated ioniza on Electric field profile Al.24 Ga.24 In.52 As Field (arbitrary units) / Number of Ioniza ons abc Electron-ini ated ioniza on Hole-ini ated ioniza on Electric field profile Loca on (arbitrary units) Figure 4: Monte Carlo simulation of a multiplication layer in which the electric field profile has been used to localize impact ionization. sufficient energy to ionize when they drift through the wide-gap region (Figure 3). Unfortunately, use of narrowgap alloys for part of the multiplication region is limited by the commensurate increase in dark current from tunneling. This consideration limits the contrast in ionization coefficients that can be practically achieved through material selection alone. A new concept we developed for the multi-stage structure is I 2 E through electric field control (Figure 4). The impact ionization rate in a semiconductor typically has an exponential dependence upon the local electric field. Just as alloys of different ionization threshold can be ordered so as to enhance the ionization rate of one carrier type over the other, the electric field profile inside the junction can be shaped to achieve the same effect. Used in conjunction with material selection, this technique is very powerful. Due to the physics of the gain process described above, I 2 E APDs have been reported on in the literature with k at very low gain, though they have not been able to sustain this level of performance much above a gain of Current [A] Light #1 Dark #1 Light #2 Dark #2 Light #3 Dark #3 Gain #1 Gain #2 Gain # Reverse Bias [V] Figure 6: I-V characteristics (left axis) and avalanche gain (right axis) measured for 1-stage Voxtel APDs. Gain Loca on (arbitrary units) Figure 5: Monte Carlo simulation of a 1-stage APD in which the multiplication layer design from Figure 4 is cascaded. Hole relaxation layers between each stage prevent hole feedback. M 4. 5 This limitation is a consequence of the fact that they derive their low-noise properties by eliminating the longer impact ionization chains; short chains produce low gain. If individual I 2 E multiplication layers cannot be operated at high gain and still preserve their low noise character, one possible solution is to operate I 2 E multiplication layers at low gain and cascade them in stages. To our knowledge, William Clark was the first to propose a multi-stage design. 6 Figure 5 depicts a Monte Carlo simulation of our multi-stage I 2 E APD. The purpose of the first repeating layer, marked a, is to have a field that (1) heats electrons traveling to the right, and (2) is insufficient to cause holeinitiated ionizations from holes traveling to the left. The second repeating layer, marked b, having both a higher field and a smaller gap, is designed to cause the hot electrons to ionize; it is too thin for cold holes to ionize in it. Finally, the hole relaxation layer, marked c, has an electric field so low that carriers can lose any extra kinetic energy gained in neighboring high-field regions. Note that c is much physically thicker than either a or b in the Excess Noise Factor Stage APD #1 1-Stage APD #2 1-Stage APD #3 k =.4 (bulk InP) k =.3 (bulk InAlAs) k =.4 k =.3 k =.2 (Si) k = (HgCdTe) Avalanche Gain Figure 7: Excess noise measurements for 1-stage APDs compared to relations characteristic of other technologies. 4 of 6

5 Figure 9: Raw count data collected from a 1-stage APD receiver operated at M = 18, both in the dark and under illumination by a sparse stream of single photons. layer structure, but does not appear this way in the Monte Carlo model due to the use of an uneven grid spacing. The purpose of the hole relaxation layer is to prevent feedback between stages by making sure that hot holes from the ionization events in subsequent gain stages lose their energy through collisions, thus preventing them from ionizing as they stream back through the gain stages. The range over which a multi-stage APD consisting of cascaded I 2 E multiplication stages can maintain low-noise operation is limited by the gain-per-stage at which the individual multiplication stages lose their low-noise character. MEASUREMENTS ON 1-STAGE APDS AND RECEIVERS We grew the 1-stage InGaAs-based APD design presented above in Figure 5 using solid-source molecular beam epitaxy (MBE). MBE techniques were required to achieve the doping precision necessary to implement the structure, but the low-noise/high-gain results reported here have proven repeatable across four different process lots. Our initial results for 5-stage APDs with a maximum gain of M = 2 and silicon-like noise (k <.2) up to a gain of M = 2 have been reported elsewhere. 7 Since then, we have manufactured a second lot of 5-stage APDs with identical characteristics, one lot of 7-stage APDs with silicon-like noise beyond M = 1, and the 1-stage APDs described here. The 2" wafers were processed into back-illuminated etched-mesa APDs encapsulated with benzocyclobutene (BCB) resin using contact photolithography. Standard detector sizes produced include APDs 3-, 5-, 75-, and 2 μm in diameter. The gain and noise data presented here were measured from 75-μm diodes, whereas the Figure 8: Photon detection efficiency measured from the count data of Figure 9. photon-counting measurements were taken on 5-μm devices. The 1-stage I 2 E APDs were shown to operate at gains as high as M = 8 (Figure 6), and maintain silicon-like shot noise levels up to at least M = 1 (Figure 7). Photon-counting thresholded linear-mode receivers were assembled from APD die hybridized to commercial 2-GHz TIA chips. Raw count rate data measured under continuous-wave illumination from a stabilized 1.64-μm diode laser source, attenuated down to a sparse stream of single photons, is shown in Figure 8; the error bars represent the standard deviation across multiple samples. Photon detection efficiency (PDE) was calculated as the ratio of the excess counts above the dark count tally to the total number of photons delivered during the sample period (Figure 9). A maximum PDE of 7% was measured, but the DCR was quite high, at nearly 1 MHz. Comparison of the DCR of a receiver both with and without the APD powered has shown that the great majority of these dark counts originate in the TIA. We estimate that band-limiting the TIA to 1 GHz will reduce these dark counts by between 2 and 4 orders of magnitude, depending upon the chosen threshold; work is underway to test this. RADIATION EFFECTS We made a preliminary assessment of the radiation hardness of the multi-stage APDs to test their suitability for a NASA optical communications application. 8 Displacement damage by protons, quantified by nonionizing energy loss (NIEL), can be modeled by the Stopping and Range of Ions in Matter (SRIM) program based on the work of Ziegler et al. 9 The SRIM program was used to calculate NIEL for 1- and 2-MeV protons. 5 of 6

6 1 5 2 MeV, 1 11 cm 1 predic on MeV, 1 11 cm 1 measurement Dark Current [A] MeV, 1 1 cm 1 predic on 2 MeV, 1 1 cm 1 measurement Dark Current [A] Post-Irradia on, Calculated Post-Irradia on, Measured Reverse Bias [V] Figure 1: Calculated and measured dark current curves for 5-stage APDs exposed to 2-MeV protons. Note that the data obtained for 1-MeV protons, which was used to predict 2-MeV damage, was truncated above 35 V for the 1 1 cm 2 fluence; this made prediction of the 2-MeV dark current for this fluence impossible above 35 V. The calculation of NIEL is simplest and probably most accurate for purely elastic collisions mediated by the electromagnetic force, so these low proton energies were chosen to exclude the possibility of inelastic scattering involving nuclear reactions. By comparing the increase in dark current measured after irradiation to the difference in NIEL that was calculated by SRIM, it was confirmed that displacement does account for nearly all of the radiation damage observed for fluences in the cm 2 range (Figure 1). Displacement damage from elastic scattering also appears to dominate for higher proton energies. A 5-device sample of 5-stage APDs was irradiated under bias by a cm 2 fluence of 63.5-MeV protons, resulting in a 23% increase in gain-normalized dark current, measured at room temperature at a gain of M = 2. SRIM predicts that NIEL of a 63.5-MeV proton should be 4.94% that of a 2-MeV proton. The NIEL-based calculation underpredicts damage by 2% (Figure 11). LIFETIME TESTING An 11-device sample of 5-stage APDs was sealed hermetically and aged under bias at 5 C for a cumulative duration of 717 hours. Because dark current leakage is greatly increased at elevated temperature, each device was wired in series with a 5-kΩ resistor to protect it from burning out. The average current through the devices during aging was approximately 1 µa, which is well above the level during normal illuminated operation. After an initial burn-in period during which device-to-device variation decreased significantly, no overall rise in dark current was observed during this time period. A longer-term study at higher temperatures is Reverse Bias [V] Figure 11: Average measured dark current for a 5-APD sample after irradiation by a cm 2 fluence of 63.5-MeV protons, compared to a calculation for this dose extrapolated from data on a similar sample that received a 1 1 cm 2 dose of 2-MeV protons. planned to determine the activation energy of the aging process. CONCLUSION We have presented design and performance data for a new class of multi-stage InGaAs with high gain and low excess noise. When operated with a TIA as a thresholded linear-mode photon counter, high PDE and very fast MCR are obtained. DCR is presently limited by amplifier noise, but substantial improvement is possible by band-limiting the TIA. 1 S. Cova, M. Ghioni, A. Lotito, I. Rech, and F. Zappa, Evolution and prospects for single-photon avalanche diodes and quenching circuits, Journal of Modern Optics 51, pp , G. Ribordy, J. D. Gautier, H. Zbinden, and N. Gisin, Performance of InGaAs/InP avalanche photodiodes as gated-mode photon counters, Applied Optics 37, pp , R. J. McIntyre, Multiplication noise in uniform avalanche diodes, IEEE Transactions on Electron Devices ED-13, pp , J. C. Campbell, Recent Advances in Telecommunications Avalanche Photodiodes, Journal of Lightwave Technology, vol. 25, no. 1, pp (27). 5 S. Wang, J. B. Hurst, F. Ma, R. Sidhu, X. Sun, X. G. Zheng, A. L. Holmes, Jr., A. Huntington, L. A. Coldren, and J. C. Campbell, Low- Noise Impact-Ionization-Engineered Avalanche Photodiodes Grown on InP Substrates, IEEE Photonics Technology Letters, vol. 14, no. 12, pp (22). 6 W. Clark, U.S. Patent No. 6,747,296 B1 (24). 7 A. S. Huntington, M. A. Compton, and G. M. Williams, Linearmode single-photon APD detectors, Proceedings of SPIE, vol. 6771, 6771Q (27). 8 NASA SBIR contract NNG5CA28C 9 J. F. Ziegler, J. P. Biersack, and U. Littmark, The Stopping and Range of Ions in Solids (New York: Pergamon), of 6