Fast, Energy‐Efficient InGaAs Synaptic Phototransistors on Flexible Substrate

Photodetectors sensing the short‐wave infrared (SWIR) region have great potential due to their significant advantages in a variety of applications because SWIR light possesses both characteristics of visible light and infrared light. Among them, devices using photodetectors to mimic synaptic dynamics and functions have received a great deal of attention due to their capabilities to implement simplified neural systems. However, it is essential to develop synaptic devices that can operate fast with low energy consumption for more efficient implementation of neural systems. Here, a flexible InGaAs synaptic phototransistor with a fast operation speed of under 1 ms and low energy consumption in the atto joule level is developed to femto joule level which is superior to biological synapses (50 ms, 1–100 fJ). By using InGaAs which has high carrier mobility as a channel layer, weak light, and short optical pulse width, fast operation speed in the SWIR region with low energy consumption is obtained. Moreover, the devices demonstrate synaptic behaviors such as “excitatory post synaptic current”, “paired‐pulse facilitation”, “short term plasticity”, “long term plasticity”, and “learning‐experience behavior” as neuro‐synaptic applications. These results provide the possibilities for implementation of complex synaptic functions with fast speed and low power SWIR synaptic phototransistors.


Introduction
In modern society, computers are indispensable in human life and most of them adopt von-Neumann architecture. [1]However, DOI: 10.1002/aelm.202300437 as the amount of information increases explosively, this general computing system confronts remarkable limitations resulting from its inherent structure which is a separation of processing units and memory units. [2]On the other hand, the brain processes significant amounts of information very efficiently because the brain is capable of parallel computing that the information is processed and memorized at the same time. [3,4]his is possible because brain use neural network consisting of over 10 12 neurons and synapses. [5]In this neural network, external stimuli received from the "presynaptic neurons" are propagated to "postsynaptic neurons" via synapse.Thanks to this sophisticated neural network, brain could process complex actions such as "thinking", "reasoning", "memory", and "learning". [2]Therefore, researches on artificial neural networks inspired by the human brain which could process and memorize the information at the same time have been actively conducted.Especially, artificial photonic synapses have great advantages in terms of increasing computing speed and reducing energy consumption by incorporating light stimulation sensing and optical information processing. [6]9][10] However, general photodetectors that only sense information are inappropriate for effective utilization in the aforementioned applications.Therefore, the development of synaptic photodetectors capable of sensing and storing information is essential.Photodetectors are similar to the biological photoreceptors in terms of converting light stimulation into electrical signals.[13][14][15] Among various photodetectors, phototransistors have an advantage of detecting low intensity light due to internal gain compared to photodiodes.When light is illuminated on channel of phototransistor, electrons and holes pairs are generated.These photogenerated charges along with applied gate voltage enhance channel conductance, so current is increased remarkably.Furthermore, channel conductance could be modulated by these photogenerated charges despite of same applied gate voltage, therefore, it is possible to detect SWIR light in various voltage range.][18][19] In detail, organic materials or CQD materials have advantages in terms of low-cost fabrication, flexibility, and superior optical property, however, they have the disadvantage of being susceptible to performance degradation over time due to environmental factors such as temperature and humidity, therefore, a globe box may be required in order to achieve stable measurements. [18]For the 2D materials, especially MoS 2 has been mainly researched because of good carrier mobilities, reasonable optical properties, however, it is not appropriate for working in SWIR region because of bandgap energy (≈1.8 eV) and requires relatively strong incident light power due to low absorption coefficient. [19,20]For the oxide semiconductors, although indium-gallium zinc-oxide (IGZO) has gained attention for its persistent photoconductivity because of slow recombination of electrons and holes, it has slow operation speed and demands high voltage and high incident light power because of low carrier mobility and inadequate absorption coefficient. [1,16]In addition, Si nanocrystals are also researched due to wide operation range from ultraviolet to near-infrared wavelengths, which is attributed to their favorable optical properties.However, Si nanocrystals could not possess fast operation speed because of their relatively low carrier mobility compared to III-V compound semiconductors. [21,22]On the other hand, InGaAs is one of the attractive materials that could overcome the disadvantages mentioned above.First of all, InGaAs has high carrier mobility, which means that electrons and holes could move easily through the material resulting in faster operation speed suitable for highspeed applications such as optical communication systems. [21,23]econd, InGaAs has a wide wavelength range of operation due to its high absorption coefficient, covering the near-infrared to SWIR region, which is important for various applications, including spectroscopy and imaging in low light conditions. [24,25]28] Here, we report flexible InGaAs synaptic phototransistors using charge trapping and detrapping mechanism. [2,29,30][33][34] Fabricated devices are possible to operate at SWIR region, operating faster and consuming less energy than biological synapse.Furthermore, the synaptic plasticity is successfully demonstrated by transferring photogenerated charges.These results show promising possibilities for various applications of SWIR synaptic phototransistors realizing complex synaptic functions more rapidly and energyefficiently.

The Mechansim of the InGaAs Synaptic Phototransistors
Figure 1a represents schematic illustration of biological vision system and artificial photonic synapse, where InGaAs acts as the photo-sensing layer and channel layer.Device structure of InGaAs synaptic phototransistor with Al 2 O 3 gate oxide acting as charge trapping layer is shown in Figure 1b.A cross sectional image of this structure is obtained through transmission electron microscopy (Figure S1, Supporting Information).Figure 1c demonstrates the hole-trapping, which is main mechanism for the synaptic behaviors of the InGaAs phototransistor.When 1550 nm wavelength SWIR light which has 0.8 eV photon energy is illuminated on the In 0.53 Ga 0.47 As channel (≈0.75 eV bandgap), most SWIR light is efficiently absorbed, thereby generating electrons and holes pairs.These photogenerated electrons or holes are trapped in gate oxide and act as a built-in potential, and especially in case of n-type semiconductor, these trapped holes provoke more major carrier electrons at the channel, causing a change in threshold votage (V th ), which is called photogating effect. [35,36]Therefore, the drain current (I d )-gate voltage (V g ) curves are shifted horizontally and more details are described in Figure 2c,d.When the light is removed, the trapped photogenerated holes recombine with electrons, resulting in a gradual decrease in photocurrent, which could show synaptic behaviors.

Fabrication Process and Characteristics of the InGaAs Synaptic Phototransistors
The fabrication process of flexible InGaAs phototransistor starts with Al 2 O 3 (10 nm) deposition on InGaAs wafer using atomic layer deposition (ALD).Structure of InGaAs phototansistor wafer is shown in Figure 2a.InP is used as a substrate because lattice constant of InP matches that of In 0.53 Ga 0.47 As.In 0.53 Ga 0.47 As (200 nm) and InP (50 nm) are grown on InP continuously serving as etch stop layer.The highly n-doped In 0.53 Ga 0.47 As (50 nm) is grown to form source and drain.Finally, In 0.53 Ga 0.47 As (50 nm), which serves as a channel layer, is grown with a thickness that could efficiently absorb SWIR light.W (10 nm) and Au (60 nm) are deposited on the Al 2 O 3 /InGaAs layers and a thin polyimide film laminated with Si substrate respectively by DC sputtering.W and Au serve two functions as the gate and metal reflector.In particular, the metal reflector enhances light absorption in the InGaAs channel by recycling photons. [26]][34]37] Figure 2b shows atomic force microscope (AFM) image representing surface roughness.[40] The InGaAs phototransistor wafer is bonded successfully because of this smooth surface.After metal wafer bonding process, InP layer and InGaAs layer are eliminated using HCl based solution and H 3 PO 4 based solution respectively.To form source and drain, conventional lithography is conducted using positive photoresist (GXR 601).Wet chemical ethcing and dry gas etching follow for the isolation of devices.SiO 2 (50 nm) is deposited for passivation and metal pad using plasma enhanced chemical vapor deposition (PECVD).After via formation on source and drain where the SiO 2 is deposited Mo (30 nm) and Au (200 nm) are deposited to form source and drain using E-beam evaporator.Finally, gate is formed using wet chemical etching process.The I d -V g curves under different optical power of 1550 nm wavelength laser are shown in Figure 2c,d.As shown in Figure 2c,d, photocurrent is increased as V g , drain voltage (V d ), and optical power increases.Furthermore, I d -V g curves under different optical power of 1550 nm wavelength laser are remarkably shifted toward the left indicating a photogating effect. [35,36]Therefore, the fabricated devices are capable of operating in both depletion and accumulation mode.In this paper, synaptic behaviors are measured in accumulation mode, especially at −0.3 V gate voltage.As shown in Figure 2c,d, I d -V g curves shift toward the left also means the presence of trapped hole charges.The hole trapping is attributed to the deep state at the Al 2 O 3 /InGaAs interface and this is the main mechanism for the synaptic behavior. [29,30,41]Additionally, Figure 2e shows the characteristics of flexible InGaAs phototransistor under different bending radii, starting from 20 mm and decreasing 5 mm increments.There are no significant differences both on-current and off-current, and this is probably due to the thin InGaAs channel  thickness (50 nm).Because of this thin InGaAs channel, there is almost no change in active area where SWIR light is illuminated despite of bending state.Moreover, the mechanical stability of the flexible InGaAs phototransistor is confirmed as shown in Figure 2f.These results indicate the possibility of applying flexible synaptic devices in various fields that require flexibility, such as wearable devices and biomedical devices.

Synaptic Plasticity of the InGaAs Synaptic Phototransistors
Artificial photonic synapse is cable of transforming the optical signal including frequency, intensitiy, and quantity of the light into electrical signal.[44][45][46][47][48][49][50] In detail, the EPSC is the post synaptic current which is attributed to stimulation by pre synpatic neuron, the PPF is the ratio of consecutive EPSCs produced by continuous stimulation and the plasticity is the property of memorizing stimulation.SWIR light (1550 nm) is used to show that fabricated devices could mimic synaptic behavior (Figure 3a-c).These synaptic operations were achieved under V g = −0.3V and V d = 0.05 V which can achieve large V th shift using photogating effect.I d increases sharply when optical pulse with 250 ms light width stimulates devices.The I d decreases gradually when optical pulse is removed and it shows typical EPSC characteristics.As shown in Figure 3b,c, the ratio between first photocurrent and second photocurrent decreases as the time inverval between the two optical pulses increases.This can be defined by the following equation: where A 1 is the peak value of first photocurrent and A 2 is the peak value of second photocurrent.

SWIR Image Learning and Memorizing
A 4 × 4 array is used to show the recognition ability under STP and LTP, and each phototransistor is stimulated in the shape of the letter "A" for image memory training as shown in Figure 4a-c.As shown in Figure 4a, when weak SWIR light pulses of 0.2 μW cm −2 (1 ms peak width, 500 Hz) stimulate the array, the intensity of the color decreases after only 1 s training representing short-term memory.On the other hand, as optical power increases, the intensity of the color remains longer representing long term memory (Figure 4b,c).These results show that these InGaAs phototransistors can be used for SWIR based active night vision system as they are capable of image learning under SWIR light and maintaining memory.Furthermore, the energy consumed per synaptic event of a single synaptic phototransistor could be defined by the following equation: where S is the area of the device and P is the power density of the input light at a spike duration of t (time). [51,52]Judging from Figure 4, the energy consumption of one device could be calculated as 0.675, 6.75, and 67.5 fJ respectively at the different optical power density of the input light.The energy consumption is decreased by using light with low optical power density and short optical pulse width of 1 ms, moreover, fast operation speed corresponding 1 ms is achieved by using InGaAs as a channel layer due to its high electron mobility. [21]We expect that actual operation speed is about to a few microseconds, achieving zepto level energy consumption. [28]In addition, only 1 ms optical pulse width enables to stimulate the phototransistors which is much faster than speed of information trasfer in biological synapse. [53]dditionally, comparison of SWIR photodetectors is shown as Table S1 (Supporting Information).The fabricated device has a strong potential for applications in fast, energy-efficient neuromorphic computing system.

Imitating Human Behaviors
Figure 5a-d shows different EPSC value under various SWIR light pulses (22.5 μW cm −2 , 2.5 ms peak width, 200 Hz).As shown in Figure 5e, the photocurrent increases when the number of photons increases, which translates that enhancement of synaptic behavior results from increasing number of photons, indicating that repeated stimulation can successfully modulate synapse behavior.Especially, as shown in Figure 5e, when the more SWIR light pulses are irradiated, the more photocurrent is generated and the more slowly the photocurrent is decreased suggesting STP and LTP.][56] The brain requires less energy and time to perceive information during the "relearning" process compared to the initial "learning" process. [57]Here, this learningexperience behavior is mimicked by illuminating InGaAs synaptic phototransistor with SWIR light of 1550 nm wavelength.Initially, the InGaAs phototransistor is illuminated with SWIR light (22.5 μW cm −2 , 2.5 ms peak width, 200 Hz) for 1s.The value of EPSC increases very rapidly during the illumination and decreases slowly during the darkness, and this corresponds to process of learning and forgetting respectively.After that only 0.5 s of light is required for the phototransistor to have its first EPSC value, and this corresponds to process of relearning.Therefore, this result shows that fabricated devices successfully mimic the biological brain.

Conclusion
In summary, we have fabricated flexible InGaAs synaptic phototransistor operating in SWIR region, which is faster and more energy efficient than biological synapse.The InGaAs synaptic phototransistors are fabricated on polyimide film as a flexible substrate using a metal wafer bonding technique.We confirm photogating effect attributed to trapped holes generated by SWIR light, as a result, InGaAs channel conductance is modulated.Moreover, the fabricated devices successfully mimic synaptic behaviors and have mechanical flexibility.Overall, these results demonstrate strong potentials for SWIR synaptic phototransistors to be used in a variety of applications, utilizing superior optoelectronic properties of III-V semiconducting materials.

Experimental Section
Device Fabrication: Polyimide film (Kapton, DuPont, ≈13 μm) was used as a flexible substrate and it was laminated onto Si spin-coated with polydimethylsiloxane (PDMS, 10 μm) to enable handling.InGaAs phototransistor wafer was cleaned with acetone, methanol, DI water and native oxide was removed using NH 4 OH:DI water (1:6) solution.The Al 2 O 3 (10 nm) serving as a gate oxide was deposited on the In-GaAs phototransistor wafer using ALD.W (10 nm) and Au (60 nm) serving as both photon reflector and gate metal were deposited on the Al 2 O 3 /InGaAs layers and a polyimide film respectively by DC sputtering.A metal wafer bonding technique which has strong bonding strength was used to bond the wafer and the polyimide film (Figure S2a, Supporting Information). [31,32,37]The surface treatment was performed on the metal deposited wafer and the polyimide film using Ar plasma for a metal wafer bonding technique.The metal deposited wafer and the polyimide film were bonded using a wafer bonder at 200 °C and a pressure of 40 kgf cm −2 after the surface treatment.Then, wet chemical etching was used to remove InP substrate, InGaAs layer, and InP layer using HCl:DI water (2:1), H 3 PO 4 :H 2 O 2 :DI water (1:1:10), and H 3 PO 4 :HCl (3:1) solution respectively.N + -InGaAs layer was patterned using photolithography with a positive photoresist (AZ GXR 601, MicroChemicals) and etched to form source and drain (Figure S2b, Supporting Information).Then, InP layer and In-GaAs channel layer were etched to isolate channel using H 3 PO 4 :HCl (3:1) solution and H 3 PO 4 :H 2 O 2 :DI water (1:1:10) respectively (Figure S2c, Supporting Information).Al 2 O 3 , W, Au, and W were etched to isolate gate using BOE solution, CF 4 gas, HCl:H 3 NO 4 (3:1) solution and CF 4 gas respectively (Figure S2d, Supporting Information).SiO 2 (50 nm) was deposited on the device for passivation and metal pad using PECVD (Plasma-Therm 790 Series, Plasma-Therm, U.S.) (Figure S2e, Supporting Information).The SiO 2 was patterned using photolithography with a negative photoresist (AZ 5214, MicroChemicals) and then it was etched using BOE solution for formation of source and drain contact via.After patterning the source and drain contact pad using photolithography with a negative photoresist (AZ 5214, MicroChemicals), Mo (30 nm) and Au (200 nm) were deposited for source and drain contact pad using E-beam evaporator (Figure S2f, Supporting Information).Al 2 O 3 layer was patterned using photolithography with a negative photoresist (AZ 5214, MicroChemicals) and then etched using BOE solution to form gate (Figure S2g, Supporting Information).Finally, fabricated device was peeled away from the handling substrate (Figure S2h, Supporting Information).
Device Measurement: The optoelectronic performance was measured by semiconductor parameter analyzer (4156A, Keithley) at a temperature of 23 °C, and a relative humidity of 55%.SWIR laser (81632B, Agilent) was used to control various optical pulse width and intensity.Oscilloscope (DSOX 1102A, KEYSIGHT) was used to measure synaptic behavior of the phototransistor.Function generator (AFG 31 000, Tektronix) was used to apply voltage to drain and gate.

Figure 1 .
Figure 1.a) Schematic illustration of biological vision system and designed artificial photonic synapse.b) Device structure of InGaAs phototransistor with Al 2 O 3 as a charge trapping layer.c) Schematics of hole-trapping behavior attributed to photogating effect.

Figure 2 .
Figure 2. a) Schematic illustration of the InGaAs phototransistor wafer grown on InP substrate.b) AFM image showing surface roughness of InGaAs phototransistor wafer.Drain current (I d )-gate voltage (V g ) curves under different optical power of 1550 nm wavelength laser, c) V d = 0.05 V, d) V d = 0.5 V e) I d on-current and I d off-current at different bending radius where V d = 0.05 V and V g = −0.3V. f) Mechanical stability of flexible InGaAs phototransistor bended up to 1000 cycles.

Figure 3 .
Figure 3. a) The EPSC of the flexible InGaAs synaptic phototransistor induced by an optical pulse of 250 ms.b) The EPSC of the flexible InGaAs synaptic phototansistor induced by two continuous optical pulses.c) PPF index at various time intervals.

Figure 5 .
Figure 5. a) The EPSC of the flexible InGaAs synaptic phototransistor induced by 200 SWIR light pulses (22.5 μW cm −2 , 2.5 ms peak width, 200 Hz).b) The EPSC of the fabricated devices induced by 400 SWIR light pulses.c) The EPSC of the fabricated devices induced by 1000 SWIR light pulses.d) The EPSC of the fabricated devices induced by 2000 SWIR light pulses.e) Comparison EPSCs of the fabricated devices under various SWIR light pulses.f) Learning-experience behavior of the fabricated devices under 1550 nm wavelength laser (22.5 μW cm −2 , 2.5 ms peak width, 200 Hz).