A flexible InGaAs nanomembrane PhotoFET with tunable responsivities in near- and short-wave IR region for lightweight imaging applications

Photodetectors are a key enabler for next generation imaging and communication technology. In particular, hyperspectral sensing and imaging through haze, fog, rain, and other atmospheric conditions are of special interests for a widening range of surveillance applications.¹ Low-power light-weight Near- and Short-Wave Infra-Red (IR) imagers with high sensitivities are desired for future high-mobility robotic applications such as autonomous drones and smart vehicles. For instance, thin membrane-like devices, which enable large-area-array implementation of surface imagers and sensors, may enable ultra-light machines with wide-field imaging capability. In order to achieve a strong signal response, a high photon to electron conversion rate and an inherent gain mechanism that generates multiple carriers per absorbed photon is desired.^2–5 Photodetectors can exist in two terminals form such as photodiodes and photoconductors, or three terminals in the form of a photo-BJT/FET. Photodiodes that have extremely fast response do not possess a photoconductive gain. On the other hand, photoconductive gain can be realized in a photoFET by lowering the transit time through increasing E-field (V_ds biasing) and provides photo amplification through channel carrier modulation (V_g biasing).^2,5

In the active research field, photodetectors ranging from ultraviolet to infrared wavelengths are being investigated with various novel materials and structures.^2–25 For example, conventional semiconductors such as Si, Ge, III-V,^3,4,10–13 to organic molecules, metal oxides, and emerging two-dimensional (2D) materials such as graphene, transition metal dichalcogenides (TMD), and monochalcogenides.^{5–8,14–26} Among the novel 2D materials, TMD and monochalcogenides materials boast advantages in the wide range of bandgaps that are available. Nonetheless, conventional InGaAs can be a more attractive proposition with it having a tunable direct bandgap depending on the stoichiometry and that the thickness can be accurately tailored epitaxially to increase the absorption cross sectional area, while not affecting its bandgap.^13,27 However, conventional transistors made with top metallization structure block out most incident light and are not ideal as a photodetector.

In this work, we show a novel flexible photoFET architecture based on an inverted thin-body (15 nm) InGaAs (In = 52%) nanomembrane with a theoretical direct bandgap of ∼0.7 eV.²⁸ The photoFET responds to short-wave IR radiation 1.8 µm and below wavelengths, tunable with external biasing. The transistor gain allows us to boost the response for weaker radiation when the intrinsic response wanes, while the InGaAs channel can be epitaxially engineered to maximize absorption cross section. The InGaAs photoFET architecture has a fully exposed channel through epitaxial lifting-off (ELO) of the InGaAs MOSHEMT onto a flexible substrate. In this process, the top gate-source-drain metal contacts were also transferred intact that provides the bottom electrical access for the photoFET. Hence, blockage of incident light by the contact metal is completely avoided as compared to the same device with top metallization. As a result, a large photon absorption area to device size is realized, minimizing the footprint of the photoFET. Figure 1 shows the schematic of the MOSHEMT before and transfer, illustrating the device inverted and the rigid InP substrate removed. This allows the device’s channel to be completely exposed, allowing all incident light to be absorbed without any obstruction.

Through biasing of the three terminals, the photoFET photocurrent can be tuned by ∼5 orders over incident light wavelengths of 0.6 µm–1.8 µm. With an estimated incident power of ∼65 nW on the photo-FET, photo-responsivities of 350 A/W at 0.6 µm–15 A/W at 1.8 µm in the on-state is calculated. In addition, we show that flexing the sample (down to 10 cm radius) does not produce detrimental effect on its electrical performance, paving way for such device to be used for applications with non-conformal surfaces.

InGaAs MOSHEMT is first fabricated on a rigid InP substrate, with mesa isolation. The stack of the InGaAs MOSHEMT and the detailed process flow were reported.²⁸ Subsequently, the entire transfer and bonding processes of the devices fabricated on rigid onto the flexible substrate are illustrated as in Fig. 2. Prior to performing the epitaxial lift off process, the Source/Drain/Gate (S/D/G) electrodes contacts are re-routed to the field area via an additional patterning and Ti/Au (10/90 nm) lift off process. This is done in order to allow electrical access to the buried contacts after the devices are transferred and inverted [Fig. 1(b)]. Thereafter, a layer of 10 nm thick Al₂O₃ is deposited on the entire surface of the die to protect the devices’ crevices during InP substrate removal, and to provide a uniform adhesion layer with the flexible layer. The flexible layer used is SU-8 (SU-8 3050, Microchem), which is spin-coated on the die surface (5000 rpm, 30 s, 50 µm), followed by flood exposure of light with wavelength 405 nm to fully cross-link the SU-8. The fully cross-linked SU-8 serves as the flexible backing of the lifted-off devices as well as protecting the devices from the wet etchant together with the Al₂O₃ passivation layer. In the lift off process, the InP substrate is completely dissolved using a wet chemistry of HCl:H₂O (3:1) in ∼2 h. The lift-off process does not damage the InGaAs active layer owing to the excellent selectivity over InP.²⁹ The improved adhesion between the SU-8 and the Al₂O₃ layers is necessary to yield functional lifted-off devices. The addition of the Al₂O₃ layers enhanced the yield of the transferred devices to ∼30% from visual inspection, for which without the layer of Al₂O₃, the yield of the devices is trivial where the strip of gate metal tends to break off due to poor adhesion (Fig. S1). As a result of the weaker adhesion between the transferred layer and the flexible substrate, extra care has to be taken not to delaminate the transferred contact pads during electrical probing. Further optimization of the ELO process to improve the current yield will be further investigated.

The transferred sample is then mounted onto a larger piece of polydimethylsiloxane (PDMS, Sylgard 184) ∼500 µm thick to facilitate subsequent handling and characterizations. The SU-8 backing wets extremely well with the PDMS surface and thus requires no further adhesive. The transfer process lifts off entirely the devices on a single die and allows large area devices and circuitry transfer. Figures 3(a) and 3(b) show the photo images of the flexible sample, and microscope image of the transferred and inverted device, respectively. The completed sample is tested to be functional in both the unflex mode and the flex mode down to a bending radii of 10 cm. The characterized device as described in the following sections has a channel length of 18 µm and a channel width of 100 µm. The transferred device is stored in a humidity controlled dry cabinet, and we did not observe any obvious air-stability issue of the transferred device over the duration of our measurements. However, for practical use of the device in a non-controlled environment, passivation of the exposed active area will be recommended.

The transferred device is first characterized using Raman and Photoluminescence (PL) on the exposed channel area to understand the effect of the lift-off process, and to ensure that the material quality is maintained. The Raman spectra in Fig. 4(a) clearly shows that the InGaAs nanomembrane is preserved after the wet etching process, displaying InAs-like and GaAs-like peaks at 234 cm⁻¹ and 267 cm⁻¹, respectively. The left-shifting of the peaks by ∼3 cm⁻¹ suggests a re-distribution of stress of the materials stack after transfer, translating to a net tensile stress of ∼1% induced on the InGaAs nanomembrane. The PL spectra in Fig. 4(b) show a broad peak after transfer over a range of 0.3 eV. A test on the surrounding flexible substrate yields no peak, confirming the peak due to the InGaAs material. In comparison, typical PL spectra of InGaAs on rigid substrate show much narrower and sharper peaks.³⁰ The peak broadening observed here is likely attributed to the poor thermal conductivity of the SU-8 layer (0.2 W/mK) that causes the heating up of the nanomembrane during measurement.³¹ This is further corroborated by the need to use a much lower power and larger integration time to avoid the risk of burning the nanomembrane while obtaining the PL spectra [Fig. S1(c)].

The I_d-V_g curves of the device with varying V_ds after transfer are shown in Fig. 5(a). Table I shows the extracted parameters [from Fig. 5(b)] of the device before and after transfer, where the details of the contact resistance (R_con) extraction are provided in the supplementary material.³² 

It is seen that the reduction in both on and off current, and consequently the field effect mobility (μ_FE), is the result of an increase in R_con of the transferred device (Table I). It is understood that the increase in R_con is due to residual Al₂O₃ remaining on the contact pads. Further optimization of the process for future iterations will be needed to reduce the R_con. Apart from that, the subthreshold slope (SS) and on/off ratio remain unchanged.

The optical properties of the device are measured on a customized probe-station where the light of wavelengths 0.6/1.5/1.8 µm is guided through an optical fiber mounted onto an external manipulator arm. The output of the optical fibre is then fixed to ∼5 mm above the device prior to measurements. The power of the light incident on the device’s active area is described as a spread of the light angle from where it exits the optical fibre and reaches the device. In our measurements, an incident power of ∼65 nW to the exposed InGaAs is applied. The details of the incident power calculation are shown in Equation S1 of the supplementary material. In our device, the entire exposed InGaAs above the Source/Drain/Gate region plays the active role in the photocurrent generation.

The illumination pulse train response of the device over a time duration of 120 s shown in Fig. 6 is measured by fixed biasing both the V_g and the V_ds to 1 V. Figures 6(a)–6(c) show the device’s response comparison of the three different light wavelengths. Longer time response curve up to 300 s can be found in Fig. S3. The spike and drop in the photocurrent (I_{d-illuminated} − I_{d-non-illuminated}) corresponds to turning on and off the light. From the response graph with increased duration, we observe a decrease in the photocurrent with time that is the largest for the 0.6 µm illumination, followed by 1.5 µm and 1.8 µm. There is minimal change in photocurrent for 1.5 µm illumination and no noticeable change in photocurrent for 1.8 µm illumination. The transferred device sits on a flexible substrate that has a >2 orders lower thermal conductivity (0.2 W/mK) as compared to its original rigid substrate—InP (68 W/mK).³¹ The reduction of heat spreading substrate can result in localized heating as the device is powered on, resulting in a drift in the photocurrent generated. The effect appears to be most prominent for the 0.6 µm light wavelength tested. This could be related to the larger change in light absorption depth when localized heating is occurring.²⁶ The effect of substrate thermal conductivity on the device performance will be further investigated. The strongest response comes from the 0.6 µm light illumination, and the weakest from the 1.8 µm, as it comes close to the theoretical bandgap of the InGaAs material at ∼0.7 eV. The photocurrent generated by the device at 1.8 µm is also comparably noisier as compared to the two smaller wavelengths, but is still readily detected. Furthermore, from the time response plot, we extracted the response time of the device with the three different wavelengths shown in Fig. 7. We extracted the 80% rise and fall time (t_rise, t_fall) and tabulated in Fig. 7.¹⁹ The response time is the shortest for 0.6 µm, followed by 1.5 µm and 1.8 µm. It is noted that all are of the order of 1s, comparable to the existing hybrid 2D materials based photodetectors with enhanced responsivities.^19,23 However, as compared to a photodiode, the response time is considerably slower with a longer rise time, as a trade-off to achieving a larger gain/responsivity.³³ 

The I_d-V_g response of the device as a function of different incident power at the same wavelength of 0.6 µm is shown in Fig. 8(a), with a fixed V_ds of 1 V. A clear rise in the I_d of the illuminated device can be observed as the V_g bias is swept from −3 V to 3 V, corresponding to the off region to the on region of the device, respectively. The same testing methodology is used for all three different wavelengths (0.6/1.5/1.8 µm) at an incident power of 65 nW, with a fixed V_ds of 1 V, and across a V_g bias of −3 V–3 V. Figure 8(b) shows the plot of the resultant photocurrent as a function of V_g to the three different light wavelengths.

In agreement with the illumination pulse train response (Fig. 6), the largest photocurrent is generated by light wavelengths of 0.6 µm followed by 1.5 µm and 1.8 µm. In all cases, the photocurrent generated by the device is observed to be tunable across ∼5 orders from V_g biasing, boosting response when signal wanes. This observation is analogous to that observed by Huang et al.,⁵ where the photocurrent generated increases when the gate bias and drain bias are increased. The mechanism of the photoFET is a result of two effects. First, this is due to the photoconductive effect due to the increasing relationship between the output current and drain voltage. Second, in the on state, the channel conductance is enhanced, thus allowing photo generated carriers to traverse the channel multiple times before recombination, thereby enhancing the gain of the device. The device response to varying V_ds from 0.25 V to 1.5 V over a V_g range of 6 V is measured, and the results are plotted in the form of a photocurrent contour map in Figs. 9(a)–9(c) for the three different light wavelengths. The photocurrent contour map reveals the dynamic range of the device response, and its tunable responsivity. It provides an indication of the specific response of the device with the different biases applied to V_g and V_ds, The amplification of weak signals can be useful for detection where signal intensity is low or when the wavelength of the light wavelengths desired to be detected moves closer to the bandgap of the material. The photo-responsivity at appropriate biases is estimated by the following equations,

where,

The maximum incident power of 65 nW is taken, and maximum photo-responsivity is calculated to be 15 A/W, 250 A/W, and 350 A/W for light wavelengths of 1.8 µm, 1.5 µm, and 0.6 µm, respectively.

The device can be flexed down to bending radii of 10 cm before the SU-8 layer starts to delaminate from the PDMS. The device under a flex state [10 cm bending radii—inset of Fig. 10(a)] is supported by a slight left shift in the Raman spectra [Fig. 10(a)] of the InGaAs channel. The left shift of the Raman peaks is attributed to a further tensile stress imparted to the InGaAs nanomembrane. However, the device shows no sign of degradation electrically from the I_d-V_g plot [Fig. 9(b)], where the overlay of the two curves before and after flexing agrees with each other. Hence, this opens up opportunities for this kind of devices to be implemented in applications where non-conformal surface attachment is required.

The dynamic detection wavelengths and photo-responsivity of the inverted InGaAs photoFET make it an attractive candidate for imaging applications. The benchmark plot in Fig. 11 shows that the channel-exposed InGaAs photoFET can be a competitive alternatives to 2D TMDs and organics, while being superior to conventional Si/Ge/III-V based photodetectors in the region of <1.8 µm, for both research and commercial devices.^2–26,34 Although 2D TMDs are well studied to have a wide range of bandgaps for wide wavelengths range of photo detection, most 2D TMDs are also known to be sensitive to the environment and thus requires some form of passivation. Furthermore, there is a trade-off in increasing the absorption cross sectional area to maintaining a constant bandgap. On the other hand, the InGaAs nanomembrane photoFET exhibits several advantages. First, mature technology exists to epitaxially tune the thickness of the InGaAs layer to achieve maximum photon absorption while maintaining the bandgap. Second, by exploiting the key advantage of thin-film devices by lifting off and inverting it, a novel architecture with a fully exposed channel for enhancing photosensitive area while allowing it to be flexible is achieved. Last, through biasing of the three terminals of the active exposed photoFET, the photo-responsivity of the device can be dynamically tuned to amplify and to detect weak signals.

In summary, we have demonstrated an ELO and inverted InGaAs photoFET device with buried metallization on a flexible substrate. The ELO process allows us to transfer complete devices with metallization, leading to increased density and throughput for high volume manufacturing. The novel inverted InGaAs photoFET structure with exposed channel presented in this work allows for large photon absorption area. Coupled with the three terminals of the photoFET, the device shows large dynamic response range and tunability of ∼5 orders for near and short-wave IR (0.6 µm–1.8 µm). With appropriate V_g and V_ds biasing, a high photo-responsivity of up to 350 A/W is measured for light with wavelength down to 0.6 µm. Such a device can be a competitive alternative to 2D TMD/organics photoFET, while being superior to conventional Si/Ge/III-V based photodetectors in the region of <1.8 µm. In addition, the flexible device is shown to be functional down to 10 cm bending radii without electrical performance degradation, thus making it suitable for lightweight surface mounted sensors applications.

See supplementary material for additional information on the material and electrical characterization of the transferred device.

The work is supported in part by Singapore National Research Foundation’s Returning Singapore Scientist Scheme (Grant No. NRF-RSS2015-003), Hybrid Integrated Flexible Electronic Systems (HiFES) Program (hifes.nus.edu.sg), and E6Nanofab at the National University of Singapore (NUS).