Implementation of PPI with Nano Amorphous Oxide Semiconductor Devices for Medical Applications

Introduction

Synaptic behaviors are important to investigate and implement because neuromorphic computing has been proposed as a new computing paradigm to analyze complex information like that from the brain.^1–3 This has promulgated the development of “next-generation” nanodevices that have unique functions and characteristics.⁴ In recent years, many kinds of electronic devices have been used successfully to produce artificial synapses. For example, Kim et al reported that a three-terminal synaptic transistor based on carbon nanotubes can provide reliable synaptic functions that encodes relative timing and regulates weight change.⁵ Zou et al also reported the fabrication of tin oxide (SnO₂) nanowire synaptic transistors using polymer–electrolyte gating. The fabricated devices not only exhibited excellent performance but also can mimic important synaptic behavior.⁶ Esqueda et al also presented aligned carbon nanotube (CNT) synaptic transistors for large-scale neuromorphic computing systems. The synaptic behavior of these devices was achieved via charge-trapping effects, commonly observed in carbon-based nanoelectronics.⁷ Therefore, important artificial synapse devices are needed to realize the synaptic behaviors in neuromorphic systems. Representative synaptic behaviors include information memory and processing. Such synaptic behaviors are initiated by the dynamics of Ca²⁺ ion transport influx among synapses.^4,5 When a stimulation such as an action potential drives an influx of ions to reach a synapse, the neurotransmitters are stimulated and released. The transfer mechanism of ion dynamics is analogous to that of Ca²⁺ dynamics in a biological synapse. In this way, the use of ionic semiconductor channels in electronic devices could be a useful way to emulate the biological synapses. The devices include both resistors, capacitors, transistors and circuits. Here, nanoscale inorganic transistors are employed for neurocomputing and for improving medical research.^6–14

Amorphous oxide semiconductor (AOS) devices have attracted much attention in recent years. One of the representative materials is In-Ga-Zn-O (IGZO). IGZO has been one of the most popular AOS materials since it was published in 2004. They have some attractive properties, including transparency, flexibility, and relatively low-cost fabrication mainly attributed to the low temperature processing at no higher than 200°C.^15–19 Most of AOS are n-type materials. This is because p-type AOS materials have carrier mobilities much lower than n-type AOS materials. What is more, p-type AOS materials are different from n-type materials and are regarded to be much more difficult to obtain. AOS has been regarded as a potential candidate to use in displays and portable sensors.

Here, this paper shows some typical synaptic characteristics, based on an IGZO thin-film transistor (TFT) structure. Some representative synaptic behaviors are investigated. For instance, the artificial synapse that emulates a “plasticity” of a response is required for the development of neuromorphic computing systems. One typical transmission behavior is the excitatory postsynaptic current (EPSC).^7–12 An important plasticity behavior is paired-pulse facilitation (PPF).^20–23 In comparison, the synaptic behaviors mimicked by the three terminal transistors are easier to control than the two terminal devices.^24,25 This paper suggests the feasibility of using semiconductor devices to study memory loss dynamics, which helps in the development of the biocompatible and biodegradable artificial synapse devices for medical applications.

Materials and Methods

Fabrication of Artificial Synapse Devices. The artificial device of this study had the IGZO TFT bottom gate structure (Figure 1B). A SiO₂/Si substrate was purchased and used as the bottom gate and dielectric layer (100 nm thick). The IGZO channel was then deposited using radio-frequency (RF) magnetron sputtering. The Ar:O₂ was in the range from about 5:3 to 5:0. The deposition time was about 5 mins. The vacuum was ~6 × Torr. Afterwards, about a 25 nm-thick IGZO was deposited and the patterns with the specifically designed channel lengths and widths were defined by lithography and wet etching. The width and length were in micrometers, which could be scaled down by using more advanced lithography techniques such as e-beam lithography. Subsequently, the electrode patterns were defined by the masks made by lithography. The top electrodes were formed by thermal evaporation through a shadow mask formed by a photoresist. The 10 nm Ti adhesion layer was then deposited using an e-beam evaporator in a vacuum of ~4 ×10⁻⁶ Torr. A 10 nm Au top electrode layer was then deposited after Ti deposition in the same e-beam evaporator chamber to act as the conduction layer and protection layer for Ti. The photoresistor layer which prevents metal electrode deposition was consequently removed by ultrasonication and cleaned in acetone, ethanol, and distilled water for 20 min each. The bottom gate TFT structure was annealed at 475 K for 2 h. The length and width of the electrode lines were 30–100μm and 100μm, respectively. The length and width of the electrode in a square were 120–150μm.

Characterization. The measurement of the electrical properties and curves were all performed at room temperature and atmospheric pressure. The current−voltage (I−V) curve including transfer curves (I_DS vs V_GS) and the time dependence of the discharge current and diffusion potential were measured by using a semiconductor parameter analyzer (4200-SCS, Keithley). The current responses under the pulses were obtained by using the pulse mode of a Keithley 4200. During the electrical measurements, the Ti/Au source electrode was electrically grounded, and an external bias was applied to the Si bottom gate electrode. A top view of IGZO TFT was captured using high-resolution field-emission scanning electron microscopy (FE-SEM, JSM-7401F,JEOL) at a 5 kV acceleration voltage (Figure 1C).

Results and Discussion

Figure 1 shows the IGZO sample structure. Figure 1A is a representative synapse. Figure 1B shows a typical TFT bottom gate structure. Here, the bottom gate TFT was used and fabricated which is more convenient in the lab. However, the top gate TFT could also be used to realize similar synaptic behaviors.^8,9,16-18 The IGZO transistor samples were fabricated using the following steps which are very common for lab study. The fabrication was based on a wafer with n-type Si as the bottom gate layer and a 100-nm thick SiO₂ as an insulating layer on top. This could be purchased from the fabrication company. The IGZO channel layer and electrodes were deposited in the lab by us. The IGZO semiconductor layer was then deposited by radio-frequency (RF) magnetron sputtering. The thickness and quality of the IGZO film could be controlled and adjusted by deposition time, power and gas flow ratio. Here, Ar: O₂ was around 3:0. The 25–50 nm-thick IGZO semiconductor channel was then patterned by lithography, with the width/length ranging from 3 to 400 μm. The electrodes for the source and drain were then deposited and patterned. These electrodes were deposited by electron beam evaporation. The thickness and quality of the films could be controlled also by the deposition time and speed. Here, the electrodes were Ti/Au with 80 nm/14 nm thickness, respectively.

In this artificial synapse, the Ti/Au electrodes are used as the drain and source which are connected with the IGZO channel, as shown in Figure 1B. The gate voltage is defined as V_GS and the drain voltage is V_DS. The source voltage is always grounded to be 0 V. The channel current is the drain current measured from the drain and defined as I_DS. The IGZO channel has metal ions which could provide mobile carriers, including electrons and holes, to transport from the source to the drain (or vice versa), and to inject into the insolating layer SiO₂, when given proper V_GS and V_DS. The transport of the mobile ions is similar to the way ions move in the synapses. Therefore, we could use the IGZO TFT to mimic the representative artificial synaptic transmission behaviors, including EPSC and PPF.^13,14,21,22 The procedure of mimicking is as follows: A pulse with an amplitude V_O is applied as V_GS on the bottom gate terminal. As Figure 1B shows, when the pulse is applied on the gate, I_DS could be stimulated, measured and demonstrated as a representative transmission behavior, EPSC. Figure 1C shows the SEM image of the sample TFT top view with an IGZO channel. Figure 1D shows the cross-section image of the TFT structure, with different layers from the bottom gate, dielectric layer, and channel layer to the metal layers.

Figure 1 Schematic of the IGZO-based artificial bio synapses, i.e., IGZO TFT. (A) The typical structure of synapses, including the synaptic cleft: pre- and postsynaptic neuron parts. (B) A IGZO TFT device bottom gate structure and the proposed measurement of representative synaptic transmission behavior in IGZO TFT. (Voltage is applied to the bottom gate, and the drain current is the bionic neural signal). (C) SEM top image of IGZO TFTs with an IGZO channel between the source and the drain top electrodes (scale bar=300μm). (D) SEM image of IGZO TFTs layers cross section (scale bar = 100 nm).

The measured I-T is shown in Figure 2. When a DC voltage sweeps from 0V to 30V and decreases to 5V after holding 5s. (0 V → 30 V → 5 V), the status of IGZO TFT is from turn-off to turn-on, i.e., with I_DS reaching a maximum current I_o in a transient way and decreasing slowly back to the off state. In general, under an electric field, the active metal ion functional groups and insulating layer could have carrier charging interactions. The channel could then move from the semiconductor to conductive layer. When the pulse is applied, the channel current increases abruptly to the compliance current (CC) level at the V_o level, which means the channel changes from insulating to conducting. When the pulse is released, the channel current decreases back, which corresponds to the channel changing from conducting back to insulating. The measured curve in Figure 2 fits a simple exponential function which is similar to synaptic transmission behavior, such as EPSC.^13,14,26 This simple exponential function is described as:

(1)

where Tau and A are the relaxation time and the prefactor, I₀ is the current level when the IGZO TFT voltage is at 5V steady state, and I(t) is the current level at the time t.

Figure 2 Typical and representative electrical transmission behavior of the IGZO-based artificial bio synapses. (A) Relaxation characteristic of an artificial synapse. (B) Current level measurement and fitting of the relaxation of an artificial synapse based on the IGZO TFTs.

This transmission behavior is due to the charging and discharging of the mobile carriers stimulated by the gate pulse. The process is illustrated in Figure 3. The pulse consists of an amplitude of V_o and a base voltage of 0 V. Before the V_o pulse is applied, at stage ①, the carriers are not attracted to the interface between the IGZO channel and the insulating layer SiO₂. Afterwards, a pulse with an amplitude of V_o is applied, it could bring a change in the IGZO TFT which includes functional groups such as Zn-O, In-O, and Ga-O. These functional groups could interact with metal ions and persist in the channel and in metal ions and oxide carriers. At stage ② when V_GS increases from 0 to V_o, the mobile carriers in the channel are activated and attracted to the interface between the IGZO and the SiO₂ layer. The mobile carriers could continue to stay within the interface of IGZO-SiO₂, even further injecting into the defects and get trapped, during stage ③ when V_GS is kept at V_o. At stage ④ when the pulse with an amplitude V_o is released and V_GS decreases from V_o to 0 V, the mobile carriers which are attracted to the interface channel are driven back into the source and the drain. At stage ⑤ when V_GS is returned to 0 V, the trapped charges are injected in the defects which respond more slowly than the mobile carriers in the channel and are driven back to the source and the drain.

Figure 3 Overview of a voltage pulse profile and the corresponding discharge current in the TFTs with the IGZO channel in red (trapped electrons, mobile carriers trapped by interface between the IGZO and the SiO2 layer or defects). The number for the sub-images on the right correspond to each respective numbered region in the schematic on the left.

A previous study mainly focussed on the materials, device structure, and pulse number impact on artificial synaptic behaviors. Here, we focus on the impact on pulse amplitude on the paired-pulse regulations. In this way, oxide devices could be used to implement the functionalities of PPF and PPI.^21–23 Therefore, the oxide electronic devices have the potential to mimic biological synapse characteristics, not only in information processing (such as writing and erasing data), but also in both representative synaptic behaviors including memorizing and processing information and memory loss ways.^17,18

As shown in Figure 4, the PPF could be obtained by applying a V_GS pulse with V_o = 30 V which lasts for 5ms.^23,24 The CC corresponding to the former pulse is smaller than that for the latter pulse. This shows the CC have an accumulation effect after each pulse. This is reasonable, because the trapped charges are moving slowly so that some might remain in the defects during stage ⑤, and increase the CC in the next pulse. This suggests that when V_GS is no higher than 30 V, the PPF could be obtained and the related mechanism dominant. The mechanism could be mainly attributed to the reason that the mobile charges trapped in the defects are not released completely during stage ⑤. The CC for the second pulse could include the unreleased charges for the first pulse and increase. This means the time space between the two pulses, i e, Δt is not large enough.

Figure 4 Experimental demonstration of paired-pulse facilitation (PPF) through the discharge current stimulated by voltage pulse on the IGZO.

When pulse intervals are short, because of plasticity behavior, synaptic will produce a PPI effect. In order to observe a significant PPF effect, we increased the pulse intervals from 0.75s to 3s, so that the PPI effect could be weakened. As shown in Figure 5, PPI could be implemented by applying a V_GS pulse with V_o = 40 V for 5ms. The first stimulated CC is a maximum. This means that the mobile charges stimulated by the former pulse do not remain to increase the CC in the next pulse.^20–23 This has two possible reasons. 1. The mobile charges have released completely during stage ⑤. 2. The mobile charges trapped in the defects would lead to a decrease of CC. As for reason 1, because the mobile charges during the 30 V pulse are not released between the pulses, it is unlikely the charges are released during a pulse of 40 V. The 40 V pulse is higher than 30 V and should attract more charges and would release completely during the same Δt in Figures 4 and 5.

Figure 5 Experimental demonstration of paired-pulse inhibition (PPI) through the discharge current stimulated by voltage pulse on the IGZO TFT.

To investigate the effect of voltage stress on the device insulators, different voltage conditions were applied in order to obtain a transfer curve and the loops are compared.^14–16 As can be seen in Figure 6, a large hysteresis window of transfer curves should be attributed to trapped carriers. When V_GS increases up to 30 V, as shown in Figure 6A, the hysteresis window is small. This means there were no defects during the 30 V V_GS stress. When V_GS increased up to 40 V, as shown in Figure 6B, the hysteresis window was significant. This suggests that 40V V_GS, unlike 30 V V_GS, had an impact on device film quality and performance. The hysteresis direction is clock-likewise and it suggests electron injection into the insulator. This reason is feasible because when electrons are injected into the TFT insulator by 40 V V_GS, the actual V_GS effect applied on the channel could be reduced so that I_DS and thus CC is reduced at the same V_GS. Because, when enough V_GS is applied on the gate and a proper V_DS applied on drain, a strong electric field would be generated near the drain. Electrons get enough energy in such a field on the drain edge could be injected into the gate oxide and cause damages at the interface and even within the oxide, which is called a hot carrier injection (HCI). Some TFT parameters would decrease caused by HCI, such as the threshold voltage, transconductance, channel mobility, and the drain current I_DS.^27–29 Therefore, the transfer curve of the sample under 40 V showed a decreased I_DS when V_GS is returned from 40 V to 0 V, which is consistent with the decrease of the current amplitude as shown in Figure 5. This clock-likewise hysteresis and thus the electron injection mechanism are consistent with the observations that CC decreases at pulses of 40 V V_GS.

Figure 6 Hysteresis of transfer curve for the IGZO TFT. (A) Transfer curve for the IGZO TFT, during the 30 V VGS stress. (B) Transfer curve for the IGZO TFT, during the 40V VGS stress.

When a series of voltage pulses at a frequency of 667 Hz and an amplitude of 30 V are applied to the IGZO TFT, a short-term plasticity (STP) can be observed as shown in Figure 7. In the first six pulses, the CC continues to increase. The reason explained here is similar to the explanation of PPF, since the trapped charges are moving slowly, there may be some residual in the defect during stage ⑤ and CC, therefore, rises in the next pulse. When a pair of pulses continues to be applied, CC tends to stabilize. This may be due to the trapped charges injected into the defects reaching saturation. Trapped charges spontaneously returned to their original position and led to STP after the weak stimuli (10 pulses) ceased.

Figure 7 Verification of short-term plasticity (STP) of the IGZO-based artificial bio synapses under a 10 pulses mode.

Conclusions

In summary, PPI was observed for the first time on nano oxide electronic transistors. Representative artificial synaptic behaviors, such as synaptic transmission, PPF and PPI were processed and mimicked here. These functions were realized by the metal dynamics in the nano oxide artificial synapse. When an electrical stimulus was applied, the artificial bio synapse had a temporary increase and then spontaneous decay of conductance along time. The increase in conductance could be amplified or reduced by applying successive stimuli at short intervals. Repeated stimuli made the transition have a residual current for the next round of CC, or could have electrons injected to decrease the next round of CC due to the large amplitude and V_o impact on the device insulator. This research could, thus, help improve neuromorphic electronics for neurocomputing based on biocompatible artificial synapses as well as the development of medical research.