The configuration of the all-2D APD device is schematically shown in Figure 1A, which consists of a monolayer WS₂ overlaid onto two few-layer MoTe₂ flakes. The metallic MoTe₂ flakes act as electrodes in the APD. In this way two heterojunctions were formed at the overlapped regions between the WS₂ and MoTe₂. Afterwards, titanium and gold layers were consecutively deposited on the MoTe₂ flakes as electrodes for electrical readout. Optical microscope image of the APD is displayed in Figure 1B. The channel length (i.e., separation between the two MoTe₂ flakes) of the device is measured as 2 μm. The thickness of the WS₂ flake is 1.0 nm, and those of the two MoTe₂ flakes are 4.3 and 4.5 nm, respectively (Supplementary Figure S1, Supporting Information). The monolayer nature of the WS₂ can be further confirmed by Raman spectroscopy characterizations (Figure 1C). Two strong peaks are observed at 352 cm⁻¹ and 418 cm⁻¹, corresponding to the 2LA(M) and A_1g(Γ) modes of WS₂, respectively. The peak intensity ratio of 2LA(M)/A_1g(Γ) can be determined as 6.6, which is a typical feature of monolayer WS₂ (Cong et al., 2014; Xu et al., 2015). For Raman spectrum collected from the MoTe₂ region, a small peak at 188.91 cm⁻¹ can be observed, which is the B_g mode of MoTe₂ in 1T′ phase (Kan et al., 2015; Naylor et al., 2016; Chen et al., 2017). 1T′-MoTe₂ is a semimetal and a good candidate of 2D electrode material because of its low resistance and high carrier mobility (Zhang et al., 2014; Beams et al., 2016; Chen et al., 2016). Due to its strong exciton transition at room temperature, the pristine monolayer WS₂ exhibits a strong PL peak at 614 nm (Figure 1D). In contrast, because of its semimetal nature, negligible PL signal can be observed in the 1T′-MoTe₂ region. It is noted that in the heterostructure regions both the Raman and PL signals from WS₂ are reduced (Figures 1C,D). Such a phenomenon suggests intimate contact between the monolayer WS₂ and few-layer MoTe₂. Once the intimate contact is formed, the WS₂ excitons or lattice vibrations will transfer their energies to the MoTe₂ underneath through electromagnetic coupling. Subsequently, the lattice vibration energy or exciton energy will be dissipated by the impurities, defects, and free electrons in the semimetal layer. As a result, the Raman and PL signals of the WS₂ will be quenched.

Schottky junctions can be formed at the two heterojunctions between the semiconducting WS₂ and semimetal MoTe₂ layer. Figure 2A shows the dependence of current on bias voltage in a representative device. Specifically, the current increases along with the increase of bias voltage and saturates at 10.4 V. Afterwards, the current increases notably when the bias is further increased. Such bias voltage dependence is typical of Schottky diodes. Due to the Schottky barrier, the device exhibits a dark current as low as 93 pA under a bias voltage of 59 V. Such a dark current is much lower than many photodetectors based on 2D materials and even lower than typical commercial APDs (Table 1). The Schottky barriers at the heterojunctions can trigger electron avalanche effect upon applying a large electrical field across the device. The current at 10.4 V (26 pA) is defined as the saturation current I_sat ⁵⁹. Bias voltages above 10.4 V will accelerate the charge carriers passing through the heterojunction, giving rise to ionization collisions of the lattice and generation of additional charge carriers, i.e., occurrence of electron avalanche effect. To further demonstrate the avalanche effect, a charge carrier multiplication factor, M, is defined as M = I/I _sat, with I the current above 10.4 V. When the avalanche effect occurs, parameter M will follow the behavior (Miller, 1957), M = 1 1 − ( V V b ) n (1) where n represents ionization rate, V _b is a fitting parameter. Equation (1) can be rewritten as, ln ( 1 − 1 M ) = n [ ln ( V ) − ln ( V b ) ] (2) suggesting a linear dependence of ln ( 1 − 1 M ) on ln(V). As shown in Figure 2B, for bias voltage above 10.4 V, ln ( 1 − 1 M ) varies linearly against ln(V), which is a direct indicator of avalanche carrier multiplication. By fitting the experimental data using Eq. (2), the n and V _b are determined as 7.7 and 17.6 V. It is noted that the n in our MoTe₂–WS₂–MoTe₂ is about 6 times that of a previous report (Lei et al., 2015a), where Schottky junction was formed between a layered InSe and metal electrode (n = 1.3). The larger n and smaller V _b means that the avalanche effect is easier to be initiated in the MoTe₂–WS₂–MoTe₂ heterostructure. Moreover, once the avalanche effect is triggered, more electrons will be generated due to the larger M under a certain bias voltage. We ascribe these merits to the atomic thickness of our device, where a much larger electric field can be induced across the Schottky junction under a moderate bias.

The avalanche effect and small dark currents of the MoTe₂–WS₂–MoTe₂ heterostructure can greatly benefit photodetection. Figure 2C illustrates the operation principle of the heterostructure APD. Electron−hole pairs will be generated in both of the 2D layers and Schottky junctions upon light illumination. When a bias voltage is applied to initiate the avalanche effect, carrier multiplications will occur, giving rise to a rapid increase of the photocurrent with the applied bias. Additionally, a larger S/N ratio will be obtained as well. Figure 2D shows the current responses of the device measured in the dark and under 532-nm laser excitations of different intensities (green: 0.13 W/cm²; cyan: 0.38 W/cm²; blue: 0.64 W/cm²). The dark current remains below 100 pA even when the bias voltage is above 60 V. In contrast, the photocurrent increases slowly with applied bias smaller than 10.4 V, then dramatically grows when the bias voltage become larger due to the avalanche effect. Moreover, the photocurrent increases against illumination intensity. By plotting the photocurrent as a function of the laser intensity, linear dependences can be observed at two typical bias voltages (30 and 40 V, Figure 2E) for laser intensity upto 0.64/cm².

R, EQE, and AG are three important parameters evaluating the performance of an APD. Specifically, R, EQE, and AG can be calculated according to the following formulae (Yu et al., 2013; Long et al., 2019), { R = I p h P i n E Q E = I p h h c P i n e λ A G = I p h − I d a r k I p h 0 − I d a r k 0 (3) where I _ph, I _dark, and P _in represent photocurrent, dark current, and incidence light intensity, respectively. Parameter h is the Planck constant (h = 6.62607015 × 10⁻³⁴ J s), c is the speed of light (c = 2.99792458 × 108 m/s), e is quantity of a unit electric charge (e = 1.602176634 × 10⁻¹⁹ C), λ is wavelength of incidence light. Parameters I _ph0 and I _dark0 are the photocurrent and dark current before the occurrence of avalanche effect, respectively. In our analyses, I _ph0 and I _dark0 were taken at the bias of 10.4 V.

We evaluate the R, EQE, and AG of the heterojunction device with an illumination intensity of 0.64 W/cm². As shown in Figure 2D, the AG increases distinctly as a function of the bias voltage. At a bias of 59 V, the R and EQE are calculated to be 6.02 A/W and 1,406%, respectively, corresponding to an AG of 587. Table 1 summarizes the performances of various 2D photodetectors, including APD and non-APD types, as well as typical APDs that are commercially available. It is seen that although the R of the MoTe₂–WS₂–MoTe₂ heterojunction device is moderate, its EQE and AG are among the best ones. In particular, the AG of the heterojunction device is much better than the listed commercial APDs can exhibit. Moreover, the dark current of our device is merely 93 pA at a bias of 59 V, which is at the lowest level among both of the 2D and commercial photodetectors. The normalized photocurrent-to-dark current ratio (NPDR) can be further calculated as NPDR = R/I _dark = 6.47 × 10¹⁰ W⁻¹, which is better than most of the 2D photodetectors can provide (Table 1). These results suggest that our 2D APD can provide an excellent S/N ratio and favor the detection of low-level signals.

Figure 2F shows a typical 2D photocurrent map of the MoTe₂–WS₂–MoTe₂ heterojunction APD. To avoid electrical breakdown of the device, a relatively small voltage of 25 V is applied to reversely bias the left Schottky junction. According to Figure 2A, such a bias can already trigger the avalanche effect. In addition, to induce strong enough photocurrent for the 2D mapping, a relatively large incidence intensity of 305.73 W/cm² was used. As shown in Figure 2F, photocurrent was visible near the left Schottky junction (reversely bias) and the channel region. However, negligible photocurrent can be found at the right Schottky junction which is forward biased. These observations can be understood by considering that in the avalanche regime, the photocurrent is proportional the magnitude of collision ionization. The photo-generated electrons in the reversely-biased Schottky junction and regions nearby will experience a longer acceleration path, which will therefore undergo more collision events. These additional collisions will generate more electrons, giving rise to stronger photocurrents.

Figure 3A illustrates the switching characteristic of the broadband photoresponse of MoTe₂–WS₂–MoTe₂ heterojunction APD under a bias voltage of 50 V at 532-nm excitation. The three ON/OFF cycles are similar with each other, suggesting that photodetection performance of our APD is repeatable. Response time (RT), which is another important parameter characterizing a photodetector, can be deduced from one typical cycle. As shown in Figure 3B, the RTs are revealed as 1,260 and 475 ms for the laser-on and–off processes, respectively. As an APD, the RTs of our heterojunction photodetector is ordinary among the 2D photodetectors (Table 1). The origin of such long RTs can be possibly due to introduction of carrier trapping centers during the device manufacturing processes. It is known that impurities and defects will be generated at the interface of the heterojunction by stacking different 2D materials via wet-transfer method (Rooney et al., 2017). These impurities and defects will act as trapping states for electrons and holes, which will increase the photocurrent gain but largely compromise the RT of the photodetectors (Hu et al., 2012; Lopez-Sanchez et al., 2013; Lei et al., 2015a; Lei et al., 2015b). To fasten the device responses, improvement of the device manufacturing processes is required.

We further study the broadband photocurrent performance of the MoTe₂–WS₂–MoTe₂ APD by illuminating the device at different wavelengths. To that end, the bias voltage was fixed at 30 V, the laser spot was focused onto the center of the left Schottky junction shown in Figure 2F. The incidence wavelength was selected by placing a specific narrowband optical filter in front of the exit of the supercontinuum laser. Representative switching behaviors of the APD illuminated at 480, 516, 532, 580, 610, and 633 nm are shown in Figure 3C. The device exhibits similar ON/OFF behaviors and same dark currents at different excitation wavelengths. Moreover, two clear photocurrent maxima were observed at 516 and 610 nm, respectively. This can be seen more clearly by plotting the R against illumination wavelength, i.e., the photocurrent spectrum (Figure 3D). The peak at 610 nm should correspond to the exciton-A transition in the monolayer WS₂, as corroborated with the PL spectrum shown in Figure 1D. The peak at 516 nm with a larger R can be ascribed to exciton-B in monolayer WS₂. The separation between the two photocurrent maxima is 0.37 eV, which is consistent with the splitting energy of the valence band minimum in WS₂ arising from the spin-orbit coupling at K (K′) valley (Zeng et al., 2013; Zhao et al., 2013; Zhu et al., 2015). Previous studies have demonstrated that the recombination rate of exciton-B is much smaller than that of exciton-A, giving rise to a much lower PL quantum yield of exciton-B (Zeng et al., 2013; Zhao et al., 2013; Zhu et al., 2015). Accordingly, more photo-generated carriers will undergo avalanche multiplications under optical excitation associated with exciton-B, leading to a stronger photocurrent response. Another important observation is that the photocurrent response can extend to wavelengths larger than 614 nm, i.e., the exciton transition energy of WS₂. Such an effect can be ascribed to photo-generated electron transmitting over the Schottky barrier from the Fermi energy of the metallic MoTe₂ ⁵⁹. On the basis of the photocurrent spectrum shown in Figure 3D, the Schottky barrier height can be extracted as 1.93 eV (Supplementary Figure S2, Supporting Information), corresponding to a wavelength of 642 nm. This low Schottky barrier height guarantees photocurrent response of the APD under optical excitation with energy smaller than the exciton transition energy of the monolayer WS₂.

With the knowledge of dark current, AG, RT, and M, the S/N ratio can be readily calculated according to (Lei et al., 2015a), { S N = I p h 2 σ 2 σ 2 = 2 e × I d a r k × B W × M 2 (4) where BW is the bandwidth. In a specific calculation, BW is set as inverse of the RT (Lei et al., 2015a). Under an illumination intensity of 0.64 W/cm² and a bias voltage of 59 V, the S/N ratio is calculated as 71 dB, which is 10 times larger than that of a 2D APD (60 dB) consisted of layered InSe and bulk metal electrode (Lei et al., 2015a). The enhanced S/N ratio in our device is attributed to its ultralow dark current even in the avalanche regime.

Finally, we characterized the current noise density of the MoTe₂–WS₂–MoTe₂ APD, whereby the noise equivalent power (NEP) and normalized detectivity (D*) can be calculated. Figure 4 shows the low-frequency noise power spectra of the device under different applied bias voltages. The noise increases against the bias voltage. In addition, all of the four noise power spectra can be well fitted using the equation S n ( f ) = K ( I d a r k β f α ) (Chang et al., 2011), where S _n(f) is the spectral density of the noise power, K is a constant, and α and β are two fitting parameters. The fitting results suggest that the 1/f noise prevails at low frequencies (1–100 Hz) for our APD. The 1/f noise usually exists in 2D photodetectors (Balandin, 2013; Na et al., 2014), which is induced by the disorder or defects (Clément et al., 2010). By optimizing the fabrication processes, especially the stacking of the 2D crystals, it is expected that the 1/f noise can be further reduced.

By integrating S _n(f) within a given bandwidth B (usually B is set as 1 Hz), the total noise current power can be expressed as (Chang et al., 2011), 〈 i n 2 〉 = ∫ 0 B S n ( f ) d f (5)

The NEP and D* can thereafter be obtained as (Chang et al., 2011), { N E P = 〈 i n 2 〉 R D ∗ = A B N E P (6) where A is the effective APD area used to normalized the noise, which is 294 μm² according to the triangular WS₂ region shown in Figure 1B. Parameter B is the test bandwidth which is set as 100 kHz. Therefore, for a bias voltage of 40 V, the corresponding NEP and D* are calculated as 7.49 × 10^–11 W/Hz^0.5 and 7.24 × 10⁹ Jones, respectively.

In summary, we successfully demonstrate an all-2D APD structure with ultralow dark current. By stacking a monolayer semiconducting WS₂ onto two few-layer semimetal MoTe₂ flakes, two back-to-back Schottky barriers were formed at the two heterojunctions between the WS₂ and MoTe₂. Due to the double Schottky barriers and good crystallinity of the 2D crystals, the fabricated device structure can exhibit excellent electrical avalanche effect. When operating in the avalanche regime, the heterojunction structure can act as an APD with improved photodetection performances and a remarkably low dark current. The EQE of our APD is 1,406%, with an AG of 587 and dark current as low as 93 pA. Moreover, due to the small Schottky barrier height, the 2D APD can operate in a broadband spectrum range from 400 to 700 nm. Further optimization of the APD performances is possible. For example, by selecting 2D crystals with favorable energy band structures and alignments, it is possible to establish proper Schottky barriers to further minimize the dark currents and expand the operation wavelength ranges. Additionally, by improving the processing techniques of the 2D stacked heterostructures, one can reduce the RT and current noise. We therefore believe that the results obtain in the current study can pave the way for design and fabrication of miniaturized all-2D optoelectronic devices with supreme performances.