Developing high-speed communication networks is crucial for addressing the increasing demand for data transmission [1,2]. Wavelength division multiplexing [3–5] and mode division multiplexing [6–8] have enabled parallel transmission of multiple signals over a single optical fiber, substantially improving bandwidth. Communication evolution focuses on higher bandwidth density, lower cost per byte, and higher integration [9]. Photodetectors are essential in many fields, including optical switching [10,11], optical computing [12,13], light detection and ranging [14,15], and sensing [16,17].

In long-distance optical communication systems, increasing bandwidth often leads to higher link and power consumption costs [18]. Avalanche photodetectors (APDs) with higher sensitivity have been applied to improve bandwidth without compromising transmission distance [18]. For 50G passive optical network applications, APDs and a combination of a semiconductor optical amplifier and positive-intrinsic-negative (PIN) detectors are viable options. However, the latter approach is costly [19,20]. The avalanche effect in APD increases sensitivity but also introduces excess noise [21], which is closely related to the ratio of impact ionization coefficients [22].

Si has important applications in detectors because of its mature processing and excellent properties. However, Si-based photodetectors exhibit a low absorption coefficient beyond 1100 nm owing to their large bandgap. Advancements in heterogeneous bonding and heteroepitaxy techniques have led to the development of Ge on Si [23,24] and III-V on Si [25,26], extending the photodetection range into the O-band and C-band. However, the 4% lattice mismatch between Si and Ge results in a high dark current. Unlike Si-Ge APDs, III-V materials such as InP, InAlAs, and InGaAs can be lattice-matched via compositional adjustment. InGaAs offers higher electron mobility and saturation velocity, which are beneficial for increasing device bandwidth [27]. Moreover, III-V APDs have an adaptable bandgap and low noise [28].

This article presents recent research advancements in high-speed, low-dark-current APDs for optical communications.

APDs can be categorized into several basic structures based on their layer design and electric field distribution. The simplest structure is a PIN-type APD, which consists of an intrinsic (I) layer sandwiched between p-type and n-type layers. While PIN-type APDs are easy to fabricate and suitable for low-noise applications, their performance is limited by the lack of separate absorption and multiplication layers. To address this limitation, the separate absorption and multiplication (SAM) structure was developed. In SAM-type APDs, the absorption layer and multiplication layer are separated, allowing for optimized carrier transport and higher gain. However, the electric field distribution in SAM-type APDs must be carefully designed to avoid premature breakdown in the absorption layer. Further improvements were achieved with the separate absorption, charge, and multiplication (SACM) structure. SACM-type APDs introduce an additional charge layer to control the electric field distribution, reducing noise and increasing the gain-bandwidth product. This structure is widely used in high-performance APDs for high-speed optical communication. Figure 1 illustrates the principles of PIN detectors and SAM-type APDs.

The internal gain of APDs is achieved through carrier impact ionization. While higher gain improves sensitivity, it also leads to increased shot noise [29,30]: ishot2=2q(iphoto+idark)M2F(M)Δf,(1)where iphoto and idark represent the photocurrent and dark current, respectively, q represents the elementary charge (q=1.602×10−19C), M is the multiplication gain, F(M) is the excess noise factor, and Δf is the bandwidth. The shot noise current is considerably influenced by the excess noise factor F(M) [30]: F(M)=kM+(1−k)(2−1/M),(2)where k is the impact ionization coefficient ratio, defined as k=β/α, with α and β representing the ionization coefficients for electrons and holes, respectively. 1/α and 1/β represent the average distance electrons and holes travel under acceleration before impact ionization, respectively. Based on Eq. (2), a smaller k results in a lower excess noise factor F(M) and a lower shot noise current ishot. Figure 2 illustrates the avalanche processes corresponding to k=0(β=0) and k=1(α=β). When k=1, the avalanche process is chain-like, and hole and electron impact ionizations occur simultaneously, leading to a higher gain and longer avalanche build-up time at k=0 [21], limiting the bandwidth of the APD. Therefore, materials with low k values must be selected to increase device bandwidth. The gain-bandwidth product (GBP) of the APD is defined as the product of gain and bandwidth. However, GBP is not merely a mathematical product but represents the intrinsic performance limit imposed by the carrier transit time and avalanche multiplication process.

Equation (2) assumes that the impact ionization coefficient is in equilibrium, which accurately describes the excess noise factor when the multiplication layer is relatively thick [30]. Figure 3(a) illustrates that the probability distribution function (PDF) of impact ionization for newly injected carriers varies with drift distance.

Figure 4(a) illustrates that k approaches one as the electric field increases [31]. Here, a thinner multiplication layer with a higher electric field generates greater noise. However, the experimental results contradict this prediction. Figure 4(b) reveals that when the multiplication layer is thinner, the excess noise factor is lower. This discrepancy led to the proposal of a new model for impact ionization. The “dead space” is the distance a carrier travels before gaining enough energy for impact ionization [32]. This theory revises the PDF of impact ionization [Fig. 3(b)]. Considering dead space, phonon scattering becomes less substantial. Carriers achieve sufficient drift velocity and rapidly ionize upon reaching the ionization threshold energy Eth. The ionization energy is approximately 1.5 times the bandgap energy, and the length of dead space can be estimated as approximately d=Eth/qE, where E is the electric field [21]. As the electric field increases, the length of the dead space decreases considerably, and the width of the PDF contracts more rapidly, making each impact ionization distance closer to the mean value, resulting in a more stable process and reduced excess noise. Consequently, excess noise can be reduced using a thinner multiplication layer [22].

In addition to noise, a 3 dB bandwidth is a crucial parameter for high-speed APDs: 1f3dB2=1fRC2+1fT&M2,(3)where fRC is limited by the RC time constant, and fT&M is limited by the carrier crossing time and avalanche build-up time. fRC is quantitatively expressed as fRC=12πRC,(4)where R is the total circuit resistance, and C is the detector capacitance. A long transit time can reduce bandwidth. Minimizing the absorption and multiplication layer thickness shortens the carrier transit distance and time [33]. Carriers entering the multiplication layer are accelerated by a strong electric field, undergoing impact ionization to generate more free carriers. A high gain leads to more impact ionization, resulting in a longer avalanche build-up time, which limits the bandwidth of the device. Therefore, a trade-off between gain and bandwidth is necessary. The relationship between k and multiplication factor M is described by Emmons [34]: when M>1/k, the relationship between the gain and frequency at high frequencies is given by M(ω)=M0(1+(ωN(k)τM0)2)−1/2, where M(ω) is the high-frequency gain, M0 is the low-frequency gain, ω is the angular frequency, τ is the carrier transit time, and N(k) is a function of k.

The relationship between multiplication and bandwidth varies with k (Fig. 5). At low gains, the bandwidth is primarily limited by the RC constant and carrier transit time. As gain increases, the extended avalanche build-up time results in a lower bandwidth.

InP and Si are widely used in APDs. The lattice matching between InP and InAlAs suppresses dark current. Si-APDs offer higher bandwidths because of their mature processes. Figure 6 illustrates the general trend in APD bandwidth in recent years [35–54].

In addition to high-speed performance, linearity and detectivity/noise equivalent power (NEP) are also important metrics for APDs. Linearity, which reflects the proportionality between input optical power and output photocurrent, is critical for high-dynamic-range applications but it is not the focus of this work. Detectivity and NEP are key to low-light detection, representing sensitivity and minimum detectable power, respectively. However, these metrics are more relevant to applications like single-photon detection or imaging, rather than high-speed communication, where bandwidth and noise performance are prioritized.

For III-V-material-based APDs, structures are primarily categorized into vertically illuminated and waveguide types. Vertically illuminated APDs are conventional and easily coupled. Waveguide APDs can reduce absorption layer thickness, increasing the upper limit of the device bandwidth; however, coupling is challenging.

APD structures have evolved from simple to complex. Anne et al. proposed a back-illuminated APD based on an SACM structure [Fig. 7(a)]. Selective Zn doping in the P-contact layer was used to confine the active area to a small central region [55]. The thicknesses of the InGaAs absorption and InAlAs multiplication layers were set to 1.2 μm and 200 nm. The signal light was incident from the back, and the device was passivated with SiNx to prevent surface leakage current. The measured excess noise factor was 3.5, with a dark current of 17 nA and a responsivity of 0.95 A/W. The device achieved a bandwidth of 11.5 GHz at a gain of 4.5, resulting in a GBP of 140 GHz. The thickness of absorption and multiplication layers can be thinned to reduce the carrier transit time, thereby increasing the bandwidth; Lahrichi et al. reported a similar planar back-illuminated APD [56] [Fig. 7(b)]. In this device, the thickness and doping concentration of the multiplication layer are optimized, resulting in a device with a GBP of 240 GHz.

Typically, illuminated and back-illuminated configurations are used for vertically illuminated APDs. The Institute of Semiconductors, Chinese Academy of Sciences recently reported a normally illuminated planar SACM-structured APD [57]. Selective Zn diffusion in the P-contact layer confines the active area of the device [Fig. 7(c)]. This device has a 600 nm InGaAs absorption layer and a 100 nm InAlAs multiplication layer. For a 15 μm diameter device, it achieved a dark current of 3 nA at 0.9Vbr, a bandwidth of 24 GHz at a gain of 1.3, and a GBP of 360 GHz. Compared with the above structure, this design introduces an InGaAsP transition layer between the P-contact and absorption layers, reducing carrier accumulation caused by the bandgap difference between materials, thus achieving a lower dark current. This design choice favors shorter carrier transit time, leading to a higher bandwidth. However, a thinner absorption layer results in a lower responsivity. Campbell et al. incorporated a distributed Bragg reflector with 95% reflectivity at both ends of the device, creating a resonant cavity within the APD to maintain high responsivity with a thin absorption layer [Fig. 7(d)] [58]. Polyimide was used to planarize the remaining part of the device, maintaining a low capacitance of 30 fF. At 0.9Vbr, this device achieved a dark current of 10 nA, a quantum efficiency of 70%, a gain of 30 with a 200 nm multiplication layer, and a GBP of 290 GHz. Having multiple absorption layers of signal light by reflection is an effective way to improve the quantum efficiency of devices.

In planar APDs, precise control over selective diffusion depth is critical [59]. Etching techniques have been employed to confine the electric field to eliminate uncertainties arising from selective doping.

For undoped absorption layers, carrier transit time is dominated by the hole drift. Ishibashi et al. proposed the uni-traveling carrier (UTC) structure addressing slow hole transport [60]. This p-doping allows holes to relax quickly without limiting bandwidth. Electrons diffuse towards the multiplication layer under the influence of a concentration gradient. Consequently, only electrons are used as traveling carriers.

Nippon Telegraph and Telephone Corporation (NTT) reported an APD with a hybrid absorption layer structure [61] that can reduce carrier transit time while maintaining a constant thickness of the layer. Figure 8(a) illustrates the band structure of this design. The carriers in the intrinsic absorption layer move owing to the electric field. The transit time is primarily determined by hole drift because of the electron drift velocity in the InGaAs absorption layer. In the p-doped absorption layer, holes relax quickly, and free electrons move toward the intrinsic absorption layer under the concentration gradient. Thus, the total carrier transit time is determined by the hole drift and electron diffusion times.

Nada et al. calculated the carrier transit time behavior against the ratio of the thickness of the p-type absorption layer to that of the entire hybrid absorption layer [Fig. 8(b)]. A proportion of zero corresponds to a traditional absorption layer, whereas a proportion of one corresponds to a UTC absorption layer. Calculations reveal that the hybrid absorber can reduce carrier transit time compared with conventional absorption or UTC absorption layers.

NTT reported a dual-mesa APD based on an SACM structure [39] [Fig. 8(c)]. Light is partially absorbed by the absorption layer, reflected by the top electrode, and re-enters the absorption layer, improving responsivity. An edge field buffer layer is placed between the N-contact layer and the n-type field control layer to prevent edge breakdown of the device. The N-contact layer of the first mesa confines the active area of the device. Figure 8(d) depicts the electric field distribution. The electric field in the multiplication and intrinsic absorption layers is controlled at 850 kV/cm and 100 kV/cm, respectively. A dark current of 230 nA and a responsivity of 0.42 A/W were achieved at 0.9Vbr using a 400 nm hybrid absorption layer and a 100 nm multiplication layer. For a 10 μm diameter active area, the 3 dB bandwidth is 27 GHz at a gain of 5.3, with a GBP of 220 GHz. Employing a thinner hybrid absorption and multiplication layer substantially reduced the carrier transit time, resulting in a higher device bandwidth. However, the 1.1 eV bandgap difference between the InGaAs absorption layer and InAlAs p-type field control layer causes carrier accumulation at the heterojunction interface, forming an internal electric field that hindered carrier transport, limiting bandwidth.

Nada et al. addressed this issue by introducing a transition layer between the absorption and field control layers, smoothening the band and facilitating improved carrier transport [41]. Figure 9(a) illustrates that the absorption and multiplication layers were further reduced to 300 nm and 90 nm, respectively, considerably increasing the bandwidth to 42 GHz at a gain of 1.5.

A barrier exists at the InAlGaAs/InGaAs interface, reducing the device bandwidth. Nada et al. added a second p-type charge layer between the transition and absorption layers to lower the interfacial barrier, reduce carrier accumulation, and weaken space-charge effects [62]. Figures 9(b) and 9(c) illustrate the structure and band diagrams, respectively. The 3 dB bandwidth is 22 GHz at a gain of 2.8, whereas the device without the second charge layer requires a bandwidth of 20 GHz at a gain of 4.8 [Fig. 9(d)].

In traditional APD structures, photogenerated carriers in the absorption layer move toward opposite directions, creating a built-in electric field that can slow carrier movement and reduce linearity. Nada et al. proposed a dual-carrier injection structure to improve linearity [63] [Fig. 9(e)]. Traditional APDs inject only one type of carrier into the multiplication layer (electron injection for InAlAs and hole injection for InP). However, this structure injects electrons and holes into the InAlAs multiplication layer of the dual-absorption-layer APD. Electrons generated in the p-doped absorption layer diffuse toward the multiplication layer owing to the concentration gradient, and the holes relax quickly. In the undoped absorption layer, the generated electrons drift toward the N contact layer under an electric field, and the holes drift into the multiplication layer. Thus, the electrons in the multiplication layer do not accumulate near the n-type charge layer, improving the linearity of the APD [Fig. 9(f)].

A trade-off exists between responsivity and bandwidth in vertically illuminated APDs. A thicker absorption layer results in a higher responsivity; however, it requires a lower frequency response. Conversely, a thinner absorption layer results in a lower responsivity but a higher bandwidth [64]. Researchers have explored changing the incident angle to adjust the effective absorption distance for the same absorption layer thickness to address this problem [65]. The angle between the incident light and absorption layer is related to the effective absorption path length, which is maximized when the incident light is parallel to the absorption layer.

Campbell et al. introduced 100 nm waveguide layers on both sides of the absorption layer [35] [Fig. 10(a)]. Using a 200 nm absorption layer and a 150 nm multiplication layer, the device achieved a quantum efficiency of 16% and a maximum bandwidth of 28 GHz at a gain of six, with a GBP of 320 GHz. The weak coupling between the InAlGaAs waveguide and absorption layers limited the responsivity. This approach offers a promising solution to the inherent trade-off between bandwidth and responsivity.

Unlike Campbell’s approach, which places only the absorption layer between two waveguide layers, Shimizu et al. positioned the absorption, charge, and multiplication layers between the two waveguide layers [38]. Figure 10(b) illustrates this structure, with a total hybrid absorption layer thickness of 370 nm. The waveguide layer length substantially affects the device bandwidth [Fig. 10(c)]. For a length of 10 μm, the maximum bandwidth can reach 36.5 GHz at a gain of 1.7. For a 20 μm device, the bandwidth is reduced to 22 GHz. The RC constant is limited by a longer waveguide, reducing the bandwidth. The waveguide length also influences responsivity. At 1550 nm, a 10 μm waveguide results in a responsivity of 0.75 A/W, whereas a 20 μm waveguide boosts responsivity to 1 A/W. The incident light is focused around the incident end and is absorbed exponentially, causing unexpected breakdowns and localized thermal failure of the photocurrent [59].

Campbell et al. reported an evanescent coupling structure that effectively avoided these issues [37] [Fig. 10(d)] by placing the optical waveguide and the absorption layer separately. The difference in their refractive indices causes the light signal to propagate in the waveguide, gradually coupling into the absorption layer. This structure ensures a more uniform distribution of photocarriers. With an absorption layer thickness of 0.19 μm, this APD achieved a quantum efficiency of 50%.

Okimoto et al. proposed a butt-joint coupled APD structure [44] [Fig. 11(a)]. An intrinsic InGaAsP signal light transmission medium directly coupled to the APD was employed, resulting in a high coupling efficiency. This design improves the absorption efficiency of the absorption layer, reduces the carrier transit time, and balances responsivity and bandwidth. The InAlAs multiplication layer was designed to be 100 nm thick, minimizing the avalanche build-up time while ensuring adequate gain. Figure 11(b) depicts the overall device structure. An integrated metal-insulator-metal capacitor was used externally to eliminate external capacitance in the receiver. The device achieved a bandwidth of 38 GHz with a gain of 4 and a GBP of 265 GHz [Fig. 11(c)]. A similar design was recently reported [66] [Fig. 11(d)]. In this design, the multiplication-layer material was changed from InAlAs to InP, and an n-type charge layer was added to adjust the electric field within the multiplication layer. The multiplication layer thickness was set below 100 nm to obtain a higher bandwidth. The breakdown voltage is 21.6 V, with an ultralow dark current of 1.9 nA at 0.9Vbr because of Fe-doped InP passivation. The device bandwidth reached a maximum of 21 GHz at a gain of 3.3, with a GBP of 173 GHz.

III-V materials (e.g., InGaAs/InP) exhibit a high impact ionization coefficient ratio (k>0.3), which results in higher excess noise and lower gain uniformity compared to Si APDs. This is primarily due to the similar ionization rates of electrons and holes in III-V materials, leading to increased stochastic noise during the avalanche multiplication process. In contrast, Si typically has a lower ionization coefficient ratio (k≈0.02–0.1), enabling lower excess noise and more stable gain performance, making them more suitable for high-sensitivity applications such as long-distance optical communication and single-photon detection.

Due to the large bandgap structure, it is difficult to use silicon for photodetectors in optical communication. The bandgap of Ge is suitable for optical communication wavelengths. However, the use of Ge for impact ionization introduces higher noise due to its larger ionization coefficient ratio compared to Si. This limitation has motivated the development of hybrid structures, such as Si-Ge APDs with Si multiplication layers, to combine the advantages of both materials.

Both Si-Ge and InAlAs APDs share a similar layered structure with separate absorption and multiplication layers. Campbell et al. proposed a p+-i-p-i structure [67] that illustrates the specific structure and doping concentration [Fig. 12(a)]. The device features a 1 μm intrinsic Ge absorption layer, a 0.1 μm p-type Si charge layer, and a 0.5 μm intrinsic Si multiplication layer. The device was operated at 1310 nm with a responsivity of 0.54 A/W. For a 30 μm diameter device, the 3 dB bandwidth is 10 GHz at a gain of 10, and the GBP is 153 GHz. The device is punched-through at −21V, with a parasitic capacitance of 70 fF. At 0.9Vbr, the dark current density is 237mA/cm2, and k is approximately 0.1. The device bandwidth is primarily limited by the RC time constant and carrier transit time. To fabricate a high-bandwidth, low-dark-current APD, the electric field in the absorption layer must be sufficiently strong for carriers to move at saturation velocity without causing breakdown.

Campbell et al. optimized a planar device into a mesa structure and adjusted the doping profile [68]. By fine-tuning the doping concentration in the charge layer, the electric field distribution in the absorption layer was optimized, preventing the impact ionization within the Ge material. Figure 12(b) illustrates the structure and doping profile of the device. Annealing temperature adjustments during processing reduced Si-Ge interdiffusion, minimizing the impact of Ge on the multiplication layer and ensuring a lower effective k value. A floating guard ring [69] was introduced to reduce the electric field intensity at the edge of the multiplication layer, effectively preventing edge breakdown. An anti-reflective coating and passivation increased responsivity to 5.88 A/W. The effective k decreased slightly to 0.09, bandwidth reached 11.5 GHz at a gain of 20, and GBP was 340 GHz.

Huang et al. proposed a resonance-cavity-enhanced structure to improve the responsivity [70] [Fig. 12(c)]. In this device, the metal layer reflects the incident light, causing it to pass through the absorption layer twice. This double absorption effectively increases the responsivity of the detector. An n-type Si layer was incorporated to optimize the electric field distribution. For a 20 μm device, the depletion capacitance was 55 fF, and the responsivity was 0.7 A/W; the bandwidth was 34.5 GHz at a gain of 3.5 [Fig. 12(d)]. Under these conditions, photogenerated carriers moved at saturation velocity, and the avalanche build-up time was relatively short. However, as the bias voltage increased, the avalanche build-up time lengthened, increasing the gain but decreasing the bandwidth. As the gain increased from 8 to 12, the bandwidth decreased from 26.5 GHz to 21 GHz.

Further studies revealed that the resonance effect within the multiplication layer considerably increased the GBP of the device. Zaoui et al. proposed a resonant-type SACM APD structure with a GBP as high as 845 GHz [46] [Fig. 13(a)]. Figure 13(b) depicts the frequency–response curves of the device. Figure 13(c) illustrates the relationship between bias voltage and gain, where the gain increases with bias voltage after the breakdown voltage up to −26V and then decreases with further increases in bias. The gain-bandwidth curve indicates that as the bias voltage continues to increase, the bandwidth also increases, reaching a maximum of 13 GHz at a bias of −28V, with a gain of 65 and a GBP of 845 GHz [Fig. 13(d)]. Kim et al. reported a similar structure [Fig. 13(e)], achieving a GBP of 460 GHz using an ultra-thin 65 nm Si multiplication layer, resulting in a gain of 15.3 and a bandwidth of 30 GHz [71] [Fig. 13(f)]. The enhancement behavior appears to be the result of two phenomena. Firstly, there is a decrease in the multiplication time limitation due to a drop in gain through the space-charge effect. Secondly, there is a reduction in transit time due to an increase in electric field in the Ge layer at high bias voltages. This device operates at a high speed of 50 Gbps with a substantial multiplication gain, making it suitable for high-speed optical communication.

However, absorption efficiency remains a concern. A thick absorption layer is required to ensure device responsivity; however, this also increases the carrier transit time, hindering bandwidth improvement. Waveguide-coupled photodetectors offer a solution to this problem. By coupling the signal light to the device through a waveguide, the requirements for the size of the input area can be relaxed, and the absorption layer can be made thinner.

Si is nearly transparent to light with wavelengths above 1100 nm; thus, researchers have explored using Si multiplication layers as waveguides to transmit optical signals. The incident light passes through the Si waveguide and is coupled to the Ge absorption layer via evanescent coupling. An electric field in the Si multiplication layer then drives the generated photocarriers. Wang et al. proposed an evanescent coupled waveguide APD [72] [Fig. 14(a)]. Unlike previously mentioned structures, this device utilizes a 400 nm p-type doped Ge absorption layer and a 100 nm Si multiplication layer. Figure 14(b) illustrates the electric field distribution, where the electric field in the multiplication layer is controlled at 659 kV/cm. Thus, the device achieved a maximum bandwidth of 25 GHz with a gain of 5–10, resulting in a GBP of 276 GHz. This is due to the design of the thickness of the absorption and multiplication layers and the effective control of the electric field.

SiFotonics reported a Si-Ge APD with an ultrahigh bandwidth based on the SACM structure [48] [Fig. 14(c)]. In this design, an n-type Si layer is inserted between the N-contact and multiplication layers to better control the electric field within the multiplication layer. Figure 14(d) depicts the frequency response curves of the device, where the bandwidth reaches a maximum of 56 GHz at a gain of 1.8, with a responsivity of 1.08 A/W. The dark current is 120 nA at 0.9Vbr, and the parasitic capacitance is 22 fF.

Huang et al. reported a special SACM waveguide APD to improve responsivity [73], where a groove is introduced into the intrinsic Si waveguide. A charge layer is grown within this groove to shorten the distance between the absorption layer and the optical waveguide, improving the coupling efficiency of the evanescent waveguide. The device achieved a responsivity of 5.26 A/W at 1310 nm with a bandwidth of 29.5 GHz at a gain of 3 and a GBP of 260 GHz. Subsequently, Huang et al. optimized the layer structure and doping profile to achieve a more precise control of the electric field distribution [Fig. 14(e)] [50]. The electric field in the Ge absorption layer was maintained above 7 kV/cm to ensure that the photocarriers moved at saturation velocity, whereas the electric field at the Si-Ge interface was maintained below 150 kV/cm to prevent impact ionization within the Ge layer. The dynamic response curve of the device displays a bandwidth of 52.2 GHz at a gain of 3.8 [Fig. 14(f)]. The breakdown voltage of the device was measured at −14V, with a responsivity of 3.5 A/W at 1310 nm and a parasitic capacitance of 23.1 fF.

Yuan et al. proposed a Si-Ge waveguide avalanche photodiode enhanced by a loop reflector [Fig. 14(g)], achieving high responsivity and demonstrating the potential of hybrid Si-Ge structures for high-performance photodetection [74].

Multiple epitaxies in the vertical APDs complicate this process. Lateral structures can simplify this. Srinivasan et al. reported a lateral waveguide APD [49,75] [Fig. 15(a)]. The device achieved a bandwidth of 27 GHz at a gain of 11, resulting in a GBP of 300 GHz. The responsivity at 1310 nm was 0.64 A/W. Similarly, Wang et al. introduced a lateral PIN-structured Si-Ge APD [51] [Fig. 15(b)]. Intrinsic Ge layers with widths of 280 nm and 400 nm were employed as the absorption and multiplication layers, respectively. When a bias of −4V was applied, the electric field in the Ge layer reached 92.4 kV/cm, sufficient for impact ionization. At low gain, a maximum bandwidth of 67 GHz was achieved. The bandwidth was 15 GHz at a gain of 9.2, with a GBP of 135 GHz.

Liu et al. utilized a focusing grating coupler to couple light from the optical fiber into the waveguide with a lateral SAM APD, improving the coupling efficiency of waveguide devices [Fig. 15(c)] [76]. The device exhibited a low breakdown voltage of 12 V and a high responsivity of 15.1 A/W at an input optical power of −22.49dBm. Unlike traditional vertical structures, which use a charge layer to control the electric field of the absorption and multiplication layers within reasonable levels [Fig. 15(d)], this structure leverages the dielectric constant difference between Si/Ge and the background doping within the Ge layer, achieving a bandwidth of 20.7 GHz at a bias of −10.6V, with a GBP of 217 GHz.

Xiang et al. introduced shallow trenches on both sides of the Ge absorption layer to form a ridge waveguide [77], effectively confining light propagation within the active area [Fig. 16(a)]. Figure 16(b) depicts the simulation results, demonstrating that the shallow trench design can increase the Ge absorption efficiency from 81.5% to 97.5%. This design improves the absorption efficiency and helps confine the electric field, reducing the dark current of the device. The device is fully depleted at a reverse bias of −13V (0.9Vbr), achieving a 3 dB bandwidth of over 27 GHz with a gain of 14.4. Figure 16(c) illustrates that the eye diagram remains clear at 50 Gbps when the device is operated at a reverse bias of −13.5V with an input optical power of −15dBm.

Xiang et al. reported an improved structure [Fig. 16(d)] [52]. With a gain of 12.8 at a bias voltage of −14V, a bandwidth of 48 GHz can be achieved. The electric field in the device was tuned by adjusting the Si layer structure. The electric field must be set between 100 kV/cm and 200 kV/cm to ensure that the carriers inside the Ge move at a saturation velocity and to avoid impact ionization. Figure 16(e) depicts that as the bias voltage increases from −5 to −14V, the bandwidth of the device increases from 1.7 to 48 GHz. When the bias voltage continues to increase, the bandwidth tends to stabilize, indicating that the device is not completely depleted before −14V. With further increases in bias voltage, the gain decreases, whereas the bandwidth increases. Figure 16(f) illustrates the gain-bandwidth curve. This behavior is attributed to space-charge effects caused by high current density, which weakens the built-in electric field, resulting in decreased gain and increased bandwidth up to 48 GHz. Similarly, Cao et al. reported a waveguide Ge/Si avalanche photodetector with an ultra-high gain-bandwidth product of 1440 GHz, showcasing the effectiveness of Ge/Si heterostructures in achieving both high gain and bandwidth [78].

Shi et al. reported a similar structure [Fig. 17(a)], in which the device adopts a lateral p+-i-p-i-n+ configuration [79]. The intrinsic Ge absorption layer was designed with a length of 14 μm. In contrast, the intrinsic Si multiplication layer was designed with a width of 100 nm to further improve device bandwidth, ensuring adequate quantum efficiency. The electric field in the Ge absorption layer was controlled below 200 kV/cm to allow the photogenerated carriers to drift at saturation velocity and prevent impact ionization. The device exhibited a dark current of 20 μA at −8.4V (0.9Vbr) and a responsivity of 0.77 A/W at −4.8V. Regarding the frequency response curve, the bandwidth and gain increased simultaneously as the reverse bias was raised until the gain began to decrease at −9.5V while the bandwidth continued to increase. The maximum bandwidth of 67 GHz was achieved at −10.6V, with a gain of 6.6. The GBP reached its peak at 442 GHz with a gain of 8.6 [Fig. 17(b)]. This phenomenon of decreasing gain with increasing bias is similar to the above results. The built-in electric field is weakened by an excessive current density, leading to a reduced gain.

Wang et al. proposed an APD-compensated spiral inductor to maximize the device bandwidth [54]. Figure 18(a) illustrates the schematic of the device structure with the spiral inductor. This approach achieved an ultrahigh bandwidth of 78 GHz and a GBP of 609 GHz at a gain of 11.7. When devices were matched with large inductors (622 pH), small inductors (375 pH), and no inductor, their DC characteristics exhibited no substantial differences. Figure 18(b) illustrates small-signal measurements of the devices, with an ultrahigh bandwidth of 78 GHz achieved by the APD with a small inductor.

Shi et al. reported an APD with a similar structure. This design achieved a bandwidth of up to 53 GHz with a gain ranging from 9 to 19.5 and an ultrahigh GBP of 1033 GHz, with a responsivity of 0.87 A/W [53]. The main structure of the device comprises a lateral SACM structure and spiral inductors on the electrodes [Fig. 18(c)]. The absorption efficiency of the 20 μm Ge absorption layer exceeds 90% owing to the incident light coupling through a Si-tapered waveguide. By adjusting the horizontal distance between i-Ge and i-Si, the electric field at the Ge interface was maintained below 100 kV/cm to prevent impact ionization, reducing the effective k value and improving the GBP.

The inductance-bandwidth simulation indicates that when the inductance is less than 540 pH, the bandwidth increases slowly, as Lp is much smaller than the dominant inductance (Lm≈3.5nH) [Fig. 18(d)]. Without external inductors, the bandwidth is 25 GHz. When Lp approaches 540 pH, the maximum bandwidth exceeds 62 GHz, and the GBP reaches 1200 GHz. At this point, the external and internal inductors complement each other, achieving optimal resonance with the APD capacitor, resulting in the maximum bandwidth and GBP. Further increases in Lp led to reduced bandwidth, likely owing to the rapid attenuation of the intermediate- and high-frequency components with relatively large inductors.

Figure 18(e) depicts the bandwidths at different biases. As the reverse bias increases from −3 to −7V, the bandwidth increases from 1.7 to 46 GHz. When bias increases to −8.6V, the bandwidth slowly rises to 53 GHz. Further increases in the reverse bias do not improve bandwidth and gain decreases. The reduction in gain is attributed to the combined effects of phonon scattering and space-charge effects. Figure 18(f) illustrates the gain-bandwidth curve. Ge/Si APDs have also been extensively studied, offering low breakdown voltages and suitability for low-power applications. For instance, Assefa et al. demonstrated a Ge/Si APD with a breakdown voltage below 10 V, making it suitable for nanophotonic on-chip interconnects [80–83].

In recent years, the GBP of Si-based APDs has continuously increased owing to mature processes, and the development of InP-based APDs has slowed. Table 1 summarizes the performance of recently published APDs based on III-V and IV group materials. Among III-V-based APDs, InGaAs/InAlAs devices have traditionally been favored for their high bandwidth and gain-bandwidth products, making them suitable for high-speed optical communication. However, recent advancements in Si-Ge APDs have demonstrated comparable or even superior bandwidth and gain-bandwidth products, challenging the dominance of III-V materials in high-speed applications. Additionally, Si-Ge APDs benefit from the compatibility of silicon with CMOS technology, offering lower dark currents and cost-effective fabrication.Table 1.

Performance Comparison of Recently Published APDs Based on III-V and IV Group Materials

Recent advancements in pure-Si APDs have demonstrated their potential for high-speed optical communication. For example, Peng et al. reported an all-silicon microring resonator (MRR)-based APD with a responsivity exceeding 65 A/W, showcasing the potential of Si-based devices for high-sensitivity detection at the O–C band [89]. Furthermore, an 8×160Gbps all-silicon APD chip was demonstrated, highlighting the advantages of monolithic integration and CMOS compatibility [90].

The effective k value of a device is related to the electric field and material properties. Zheng et al. calculated the band structure of a six-layer InAs/AlAs digital alloy using a tight-binding method [91]. Compared with traditional random alloys, the energy distribution of holes in digital alloys is compressed, whereas the energy distribution of electrons remains relatively unchanged. Suppressing hole impact ionization results in a lower k value for the InAlAs digital alloy, which can reduce noise. Figure 19(a) illustrates the I–V curves of devices fabricated using random and digital alloys. The excess noise factors are 2.7 for the random alloy and 1.6 for the digital alloy. Figure 19(b) depicts that the effective k of the digital alloy was close to 0.01.

Wang et al. proposed a special multiplication layer material based on the digital alloy InAlAs for an SACM APD structure [92] [Fig. 19(c)]. Measurements reveal that the device has a dark current density of 4.1mA/cm2 and a responsivity of 0.48 A/W at 0.9Vbr. From the noise measurements presented in Fig. 19(d), the k value was approximately 0.15, lower than 0.2 in the traditional random alloy InAlAs. Ultimately, at a gain of 6.5, the device achieved a responsivity of 3.1 A/W and a bandwidth of 8.7 GHz. Yi et al. used AlAs0.56Sb0.44 as the absorption and multiplication layers for a PIN structure APD [93] [Fig. 20(a)]. For the 1550 nm AlAsSb layers, the effective k was 0.005 [Fig. 20(b)]. This material demonstrated extremely low noise at larger thicknesses, resulting in an extremely low effective k value. Chen et al. reported an InAs quantum dot waveguide APD [94] [Fig. 20(c)]. At 1310 nm and with a bias of −5V, the responsivity was 0.234 A/W. The device had a very low dark current of 1.3 nA at 0.99Vbr and a GBP of 240 GHz. At −6V, the device achieved a bandwidth of 2.26 GHz. Blain et al. fabricated a planar InAs APD structure [95]. Good linearity was observed at 1550 nm. At a gain of 54, the device achieved a noise-equivalent power of 45fW/Hz, with an excess noise factor of only 1.6. Lim et al. reported an InAs APD [96]. The dark current of the device was reduced by passivating SU-8 (a negative-tone epoxy-based photoresist) on the surface of a planar InAs APD. Figure 20(d) illustrates the device structure. At a bias of −0.2V, the dark current density was measured to be 0.52A/cm2. At −0.3V and a wavelength of 1520 nm, the device exhibited a quantum efficiency of 51%. The emergence of these low-k-value materials offers various possibilities for the future development of low-dark-current detectors.

In addition to traditional bulk materials, low-dimensional materials have recently emerged as promising candidates for constructing high-performance APDs. Two-dimensional (2D) materials, such as graphene and transition metal dichalcogenides (TMDs), offer unique advantages, including high carrier mobility, broadband absorption, and mechanical flexibility. For instance, 2D-material-based APDs have demonstrated ultrahigh responsivity and fast response time, making them suitable for applications in flexible optoelectronics and integrated photonics [97,98].

Similarly, zero-dimensional (0D) materials, such as colloidal quantum dots (CQDs), have shown great potential in infrared detection and high-gain applications. CQDs exhibit tunable bandgaps and high quantum efficiency, enabling the development of APDs with ultrahigh gain and broadband spectral coverage. Recent studies have demonstrated CQD-based APDs with a gain-bandwidth product exceeding that of traditional III-V and IV group materials, highlighting their potential for next-generation photodetectors [99].

This study reviewed different materials and structures of APDs for high-speed optical communication systems. The main issue is the trade-off between bandwidth and responsivity. Solutions include increasing responsivity with additional mirrors, increasing bandwidth through waveguide design, decreasing carrier transit time by a hybrid absorption structure, and reducing adverse effects at the heterojunction by using simpler lateral structures. Some methods to reduce the effective k include thinning the multiplication layer to reduce k through the dead-space effect. Some materials with lower k values are increasingly used for APD. Achieving a perfect APD is impossible. However, certain characteristics can be improved to suit different applications through rational design.

Optical communication is evolving towards higher bandwidths, greater integration, lower costs, and reduced power consumption. Therefore, the demand for high-bandwidth, low-dark-current APDs is increasing. The key to improving bandwidth and reducing dark current lies in using materials with sufficiently low k values for the multiplication layers. Considering the current design and research status of communication system APDs, the full potential of many materials has yet to be realized. Further research is required to understand the material properties and rationally design device structures, overcoming technological bottlenecks and improving future device performance.

Acknowledgment. The authors are grateful to Lequan Zhang, Peng Liu, and Pinyao Wang for the helpful and enlightening discussions.