High-performance Ge-on-Si photodetectors

High-performance Ge-on-Si photodetectors

(Parte 2 de 4)

Th e dark current density of Ge-on-Si photodiodes with thin

Ge or SiGe buff er layers is typically of the order of 10 mA cm–2. A notable feature of these photodiodes is that the dark current keeps increasing with applied electric fi eld under reverse bias, and it does not saturate. Temperature-dependent current–voltage data shows that the activation energy for the dark current decreases with

(113) facet x ( μm) y (μ m)

(1) facet z ( μ m)

Ge-0.7% SiOxide -siliconα

100 nm

600 nm 200 nm a b

Figure 3 | Selective growth of germanium. a, Atomic force microscope image of a selectively grown Ge mesa in a SiO window. Faceting can be clearly seen at the mesa sidewalls. b, Selective, trench-fi lled Ge. Processintegrated structure of a p-i-n diode using Ge-0.7% Si on a silicon-oninsulator (SOI) substrate. α-silicon, amorphous silicon; BOX, buried oxide. Figure b reproduced with permission from ref. 46, © 2008 SPIE.

530 NATURE PHOTONICS | VOL 4 | AUGUST 2010 | w.nature.com/naturephotonics applied electric fi eld42, whereas it increases with temperature43,4. Th e authors of these studies conclude that generation and tunnelling processes through defect states in intrinsic Ge under an electric fi eld are the major sources of dark current. Minimizing the operating voltage is therefore desirable for achieving both low dark currents and integration with CMOS circuits.

Several approaches are currently being developed to reduce the dark current density further. Low-temperature Ge buff er layers have been heavily doped with boron to prevent the depletion region from extending into the defective layer, and have also been annealed at high temperature to improve the material quality before intrinsic Ge growth45. Th is approach has been shown to decrease the dark current density to 1 mA cm–2. Ge photodetectors selectively grown to fi l submicrometre narrow trenches have demonstrated a dark current density of the order of 0.1 mA cm–2 (ref. 46). Th is technique is highly relevant for waveguide-integrated photodetectors, as will be discussed in the next section. It was recently demonstrated that the dark current density of Ge/Si photodiodes grown by low-energy plasma-enhanced CVD is more than two orders of magnitude lower than devices grown by ultrahigh vacuum or low-pressure CVD with the same threading dislocation density15. A dark current density of 4 × 10–5 A cm–2 at a bias voltage of −1 V has been achieved in this case, which is approaching that of bulk single-crystal Ge detectors. Th is result may indicate an eff ective defect passivation by plasma-introduced atomic hydrogen.

One important performance parameter for photodetectors is their 3 dB bandwidth. Owing to their relatively high carrier drift velocity, Ge photodetectors exhibit fast response times while also showing high responsivity. Table 1 shows the development of detector performance over time for both normal-incidence and waveguide-coupled Ge detectors. Th e bandwidth of Ge-on-Si photodetectors has been improved from several gigahertz31,3 to greater than 30 GHz (refs 34,35,37,40,41).

An ideal normal-incidence photodetector would have a high responsivity, low dark current density and high bandwidth. Unfortunately, there are several trade-off s that make such devices diffi cult to achieve. For example, normal-incidence detectors must be of a reasonable size for fi bre coupling; a larger size increases the capacitance and dark current, and therefore also limits the bandwidth and sensitivity. In addition, a large spacing between the anode and cathode reduces the drift velocity of the carriers at a given reverse bias, and therefore limits the bandwidth. Table 1 shows that because of these trade-off s it has so far not been possible to fabricate a normal-incidence Ge detector that simultaneously exhibits high bandwidth, high responsivity and low dark current. Most notably, Ge detectors with metal–semiconductor–metal designs have very high dark currents, most likely due to the poor Schottky contact with Ge.

Ge detectors fabricated on ultrathin silicon-on-insulator substrates have reached bandwidths of up to 29 GHz for operation at 850 nm (refs 9,10). Th e thin silicon bottom layer made a bottom contact impractical; these detectors therefore featured interdigitated top contacts. For high-speed performance the spacing of the top contacts was optimized for effi cient carrier collection, and because of this constraint the Ge thickness was limited by the ability to extend the electric fi eld through the Ge layer and reach saturation velocity for the carriers. Th is limited Ge thickness and increase in bandgap induced by interdiff usion between the Si and Ge layers resulted in low internal quantum effi ciencies at longer wavelengths. Th ese restrictions highlight the limitations common to all free-space detectors in terms of the bandwidth–effi ciency product. To increase the speed of these detectors, the Ge layer must be thin to decrease the carrier transit time. On the other hand, a thinner Ge layer absorbs less and so results in a lower photocurrent, particularly in the telecommunications wavelength range of 1,300–1,550 nm.

Waveguide integration Th e trade-off between quantum effi ciency, bandwidth and a relatively high dark current for free-space detectors can be overcome by using the Ge detector as part of a waveguide. Figure 4 illustrates how the effi ciency–bandwidth product depends on the detector size, thus providing a fi gure-of-merit for detector design. Th e blue and pink curves show the performance of normal-incidence Ge detectors at 1,550 nm. Two diff erent Ge thicknesses are depicted to show the trade-off between higher quantum effi ciency (d = 2 μm) and high speed (d = 0.5 μm). Th e orange curve shows the theoretical performance of a waveguide-integrated Ge detector at 1,550 nm. Due to the waveguiding nature of the Ge detector, the absorption length is increased and therefore decoupled from the carrier collection path. Th is confi guration allows waveguide-integrated detectors to exhibit high speed while reaching almost 100% quantum effi ciency. Furthermore, the device area of a waveguide-integrated photodetector can be almost ten times smaller than that of a freespace detector, so the absolute dark current is signifi cantly lower for the same dark current density. Because noise is determined by the absolute dark current (not the dark current density), waveguide integration also enhances the sensitivity of Ge photodetectors. Dark current density therefore stops being a limiting factor for device performance in such small devices. Table 1 shows that the dark current for waveguide-integrated detectors is generally below 1 μA, with a responsivity of the order of 1 A W–1 at 1,550 nm. Th e concept of waveguide integration was fi rst successfully dem- onstrated by Ahn et al.47 using a top-coupled Si3N4 channel waveguide. Th e reported bandwidth–effi ciency product of 6.5 GHz at a bias of

0.1 V was already signifi cantly higher than what could be achieved by free-space detection. Th ese results were confi rmed by Vivien et al.48, who used a Ge detector coupled to a silicon ridge waveguide.

Th ree waveguide–detector confi gurations are possible. Th e fi rst is to have a waveguide on the top of the Ge detector, with the light coupling evanescently to the Ge detector47. Th e second is to have a waveguide underneath the detector, again with the light evanescently coupling to the detector. However, in this bottom coupled design, the waveguide is made from silicon because the Ge requires a crystalline

Transit time limitRC time limit

Bandwidth × quantum efficiency (


Waveguide-integrated photodetector

Discrete free-space photodetectors d = 2 µm d = 0.5 µm

Figure 4 | Product of bandwidth and quantum effi ciency as a function of detector area for diff erent Ge detector designs at an operating wavelength of 1,550 nm. Curves for two diff erent Ge thicknesses are shown, highlighting the trade-off between higher quantum effi ciency (blue, d = 2 μm) and high speed (pink, d = 0.5 μm), respectively. The green square shows the performance of a 5 μm × 20 μm waveguide-integrated detector with an internal quantum effi ciency of 90%.

NATURE PHOTONICS | VOL 4 | AUGUST 2010 | w.nature.com/naturephotonics 531 template for epitaxial growth49–51. Th e third confi guration is to buttcouple the waveguide to the detector, making the detector an extended part of the waveguide48,49,52. Both the top- and bottom-coupled detectors exploit the fact that light can be easily coupled evanescently from a lower-index material to a higher-index material as long as the index diff erence is small. In evanescent coupling, the electromagnetic mode, and hence the optical power, is slowly transferred to the high-index material. Effi cient butt-coupling requires mode-matching conditions for waveguide and detector modes. If these are met, the optical power in the waveguide is directly transferred to the detector.

Although butt-coupling is the most effi cient coupling mechanism that allows for a very short Ge detector, bottom-coupling is most oft en used because of the ease of integration into a CMOS process. For applications such as transceivers, limitations on detector size are not severe, which allows for larger devices. Eventually, as more photonic devices are integrated into a single CMOS-based chip, smaller detectors will be needed to increase the density of the photonic devices.

Th e highest reported effi ciency–bandwidth product for a waveguide-coupled Ge detector is 30 GHz at a bias of 1 V (ref. 53). Several Ge detectors with effi ciency–bandwidth products of around 20 GHz have been reported, and although these detectors all vary in design, their performances are similar and are generally much higher than normal-incidence devices. Th ere is still a performance gap between the most effi cient fabricated diodes and the theoretical limit (Fig. 4). Capacitance and series resistance associated with the metal contacts could be an extrinsic factor that explains the discrepancy in bandwidth between theoretical limit and experimental data.

As mentioned earlier, minimizing the operation voltage of Ge photodetectors is desirable for both dark current reduction and CMOS integration. Optimizing the built-in electric fi eld in the intrinsic Ge layer by adequately controlling the doping profi le in a p-i-n diode structure enables full responsivity for photovoltaic operation at zero bias33,47,51. In recent years, bandwidths at zero bias have also been improved signifi cantly, particularly for waveguideintegrated devices, because the thickness or width of the intrinsic Ge region can be reduced to obtain a stronger built-in electric fi eld at zero bias for vertical or lateral p-i-n structures, respectively, without aff ecting the optical absorption path length in the longitudinal direction. For example, Ahn et al.47 reported a bandwidth of 6.6 GHz at zero bias, which is nearly 90% of the full bandwidth at a reverse bias of −3 V. Yin et al.50 and Feng et al.53 reported bandwidths of 16 and 17.5 GHz at zero bias, respectively. Photovoltaic operation at zero bias is especially benefi cial for achieving high energy effi ciency in large-scale electronic–photonic integration because the photodetectors consume almost no power at all.

CMOS integration Th ere are several approaches to integrating Ge detectors into a CMOS process. Because Ge is already in use at the ‘front end of the line’ (the fi rst stage of integrated circuit fabrication), a straightforward approach is to insert the Ge epitaxy step in an established CMOS process fl ow aft er poly-gate formation and before contact module or metallization. Th is approach was successfully demonstrated by Luxtera in a 130 nm CMOS technology node39 and by the Massachusetts Institute of Technology in collaboration with BAE Systems in a 180 nm node46. Although the performance of these detectors was suffi cient for the applications for which they were developed, optimization of the device for more advanced CMOS technology nodes is limited by the fi xed layer thicknesses and process fl ow. It would be desirable to have access to a CMOS process that could be adapted to optimize electronic as well as photonics devices. Unfortunately, such a process is not currently available, and the main limitation lies in the access to pure Ge growth. Foundries usually have access to SiGe growth with up to 50% Ge. In many cases, wafers are shuttled from a foundry, where the transistor level is fabricated, to a fabrication plant with Ge growth capability, where the metallization is performed. Th ere are several ongoing projects to achieve the integration of Ge detectors in a single fabrication laboratory.

It would be ideal to integrate Ge photodetectors with ‘back end of the line’ processes (in which active components are interconnected with wiring on the wafer) in the upper interconnect levels, such that these photonic devices do not compete with CMOS transistors at the single-crystalline-Si level. However, growing high-quality Ge on amorphous layers poses a signifi cant challenge in such cases, and research in this area is still at an early stage. Development of Ge

Table 1 | Performance comparison for diff erent Ge photodetector designs. Responsivity (A W) @ 1,550 nm3 dB bandwidth (GHz)Dark current density (mA cm)Dark current (μA)Diode designYearReference

The table is divided into normal-incidence devices and waveguide-coupled devices. The references have been listed in chronological order. All performance data are for 1 V reverse bias, unless otherwise stated. The data have been extracted from the references as indicated. MSM, metal–semiconductor–metal. *Data calculated using the referenced material.

532 NATURE PHOTONICS | VOL 4 | AUGUST 2010 | w.nature.com/naturephotonics photodetectors on dielectric or amorphous-Si layers for low fabrication temperatures has been reported54–57. Th ere are also eff orts to develop submicrometre Ge photodetectors through integration with near-infrared dipole antennas, potentially achieving bandwidths of >100 GHz at very low power consumption58. Incorporating Ge epitaxy aft er poly-gate formation and before metallization will remain a dominant solution for CMOS integration in the near future.

Avalanche photodetectors Avalanche photodetectors (APDs) are widely used in high-bit-rate, long-haul optical communication systems. Compared with their p-i-n counterparts, APDs off er ~5–10 dB better sensitivity due to their internal multiplication gain59. Ge/Si APDs combine the excellent optical absorption of Ge at telecommunications wavelengths with the outstanding carrier multiplication properties of Si. In the high-electric-fi eld gain region of Si, photogenerated electrons from the Ge absorption layer undergo a series of impact ionization processes, which consequently amplifi es the photocurrent and improves the sensitivity. Th e key fi gure of merit of a multiplication material is the ionization ratio k, which is defi ned as the ratio of the multiplication rate of one carrier type to the other, such that the value is less than 1. Th is parameter aff ects the excess noise, the gain–bandwidth product, and thus the sensitivity of an APD59. Ideally k should be minimized such that the multiplication process is dominated by either electrons or holes to reduce excess noise and increase the bandwidth. Si has a much smaller k (<0.1) than typical multiplication materials used in group i–v APDs such as InP (k ~ 0.5), making it more advantageous for use as a multiplication layer. By combining the success of epitaxial Ge diodes with a highly desirable Si gain region, Ge/Si APDs promise to be a competitive candidate in high-speed optical communication systems with higher gain–bandwidth products and better sensitivity over traditional group i–v APDs. Th e high performance and monolithic CMOS integration capability can also expand the application of Ge/Si APDs to other fi elds, such as 3D imaging and single-photon detection60.

(Parte 2 de 4)