Amorphous GaO_x Thin Film-Based Optoelectronic Artificial Synapses Towards Physical Reservoir Computing

Abstract

This study investigated the optoelectronic synaptic properties of amorphous gallium oxide (GaO_x) thin films for low-power physical reservoir computing (PRC) applications. The fabricated devices were irradiated with time series UV-C light to characterize the paired pulse facilitation (PPF) index, a fundamental synaptic property governed by transient photocurrent dynamics. Furthermore, the short-term memory (STM) capacity and parity check (PC) nonlinearity were quantitatively evaluated as essential PRC performance metrics, alongside a practical demonstration using a handwritten digit recognition task. The experimental results revealed a high PPF index when the width and interval of the input light pulses were comparable to or shorter than the inherent photocurrent time constants of the device. Although the evaluated nonlinearity was lower than that of conventional optoelectronic artificial synapses based on other semiconductor materials, the GaO_x device exhibited a comparable short-term memory capacity. Consequently, the reservoir layer achieved a high classification accuracy of approximately 90% in the handwritten digit recognition task. As these performance metrics were higher than those of the annealed sample, the device without annealing proved to be more suitable for PRC applications. These findings indicate that the amorphous GaO_x thin film device holds significant potential to serve as a robust, UV-C-responsive edge artificial intelligence (AI) sensor in harsh environments, such as outer space.

1. Introduction

With the recent spread of IoT and cloud computing [1], the importance of edge processing has been recognized for enhancing information communication, and, in particular, edge artificial intelligence (AI) is attracting attention as a key technology for next-generation sensor devices [2]. As an important characteristic of edge AI, devices that can simultaneously perform sensing and AI information processing without relying on large-scale data transmission and reception are required. Physical reservoir computing (PRC) is a machine learning framework that uses physical devices to store and process information [3]. PRC is expected to be applied to widely deployed edge AI sensors, as it can perform real-time processing of temporal information due to its low power consumption.

Recently, PRC that can respond to time-varying light input has been attracting attention [4]. To realize PRC, one of the essential components is an optoelectronic synaptic device. The device has been fabricated using various materials and devices, including indium gallium zinc oxide (InGaZnO) [5], polymer–zinc oxide (ZnO) nanocomposites [6], gallium oxide (GaO_x) [7], perovskite thin film [8], and dye-sensitized solar cells [9]. By adopting appropriate optoelectronic properties such as the bandgap and the highest occupied molecular orbital (HOMO)–lowest unoccupied molecular orbital (LUMO) gap, light with a specific wavelength can be processed. In space application, on space objects such as spacecraft and lunar structures, external photoelectric effects caused by UV-C and electron collisions generate strong electric fields, leading to localized discharges that may damage or destroy the electrical components [10]. To avoid failures, a sensor that can predict and classify temporal UV-C light is required, with low power consumption. Edge AI sensors based on PRC with optoelectronic artificial synapses that respond to UV-C satisfy the above requirements.

Among various candidate materials, conventional oxide semiconductors have been widely studied for synapses featuring analog plasticity [11,12,13]. However, they often lack intrinsic solar-blind characteristics. In contrast, GaO_x is a wide-bandgap semiconductor that responds to UV-C light and is a promising candidate for optoelectronic artificial synapse targeting for the applications mentioned above [14,15]. In particular, amorphous GaO_x exhibits high UV-C responsiveness and lower power consumption compared to other crystalline phases, while offering the advantage of low-cost fabrication without annealing [16,17]. While research on its application to neuromorphic computing is progressing [7,18,19], the performance metrics for its application to PRC have not been fully investigated. For instance, crucial metrics such as paired pulse facilitation (PPF), short-term memory (STM) tasks, parity check (PC) tasks, and handwriting recognition tasks have not been sufficiently explored for GaO_x.

In this study, we fabricated amorphous GaO_x thin film devices and systematically characterized their performance metrics to assess their operation as PRC reservoir layers. The findings obtained in this study are expected to significantly contribute to the development of amorphous GaO_x-based optoelectronic synaptic devices for PRC applications.

2. Materials and Methods

Figure 1a shows a schematic illustration of the GaO_x-based synaptic device. First, a squared α-Al₂O₃ (0001) wafer (Furuuchi Chemical, Tokyo, Japan) with a lateral dimension of 10 mm was organically cleaned with toluene, acetone, and isopropyl alcohol, followed by a mixture of sulfuric acid and phosphoric acid. Subsequently, the substrate was introduced into a fine-channel mist chemical vapor deposition (CVD) system for GaO_x thin film deposition. The source precursor solution was prepared by dissolving gallium (III) acetylacetonate (14405-43-7, Sigma-Aldrich Co. LLC, St. Louis, MO, USA) in ultrapure water with a precursor concentration of 0.05 M. The precursor mist, which was generated by an ultrasonic transducer from the solution, was transferred into the fine-channel furnace with flowing air as the carrier gas and the dilution gas at rates of 1 LPM and 1 LPM, respectively. The furnace temperature was maintained at 260 °C, and the deposition time was 60 min. The resultant thin film had a thickness of approximately 170 nm. After deposition, post-annealing at 400 °C was performed using a tubular electric furnace (ARF-50KC, Asahi-rika Co., Ltd., Chiba, Japan). The annealing time and the air flow rate were 60 min and 1 LPM, respectively. The detailed fabrication process has been already described elsewhere [20].

The crystallinity and optical transmittance of the obtained GaO_x thin film were characterized using X-ray diffraction (XRD: SmartLab, Rigaku Corp., Tokyo, Japan) and a UV-vis-IR spectrometer (UV-3600Plus, SHIMADZU CORPORATION, Kyoto, Japan), respectively. Then, to observe the temporal variation in the photocurrent, an interdigitated Ni/Au multilayer electrode, with thicknesses of 50 and 100 nm, respectively, was formed on the surface of the GaO_x thin film. The channel width and the channel length were approximately 35.6 mm and 0.2 mm, respectively, and the number of combs was 9. The time constant of the photocurrent was measured. A power supply (GPD-2303S, GWINSTEK, Kanagawa, Japan) was used to apply 10 V between the electrodes while irradiating the sample with a square wave of deep-UV light (UXM-500SX, Ushio Inc., Tokyo, Japan). Then, we measured a synaptic property such as PPF. Pulsed UV light, which was formed using the deep-UV light with a mechanical shutter (LS6, Vincent Associates, Rochester, NY, USA), was irradiated to the device. The total illuminated area was approximately 7.1 mm². The photocurrent was measured using a DAQ device (USB-6366, National Instruments, Austin, TX, USA) via a current–voltage amplifier (T-IVA001MAC, Turtle Industries, Ibaraki, Japan). Furthermore, to characterize the performance as a reservoir layer of PRC, the STM task and PC task were adopted [21]. The STM task measures the ability to reproduce past inputs and is used as a benchmark task to characterize short-term memory. On the other hand, the PC task characterizes nonlinearity by assessing tasks such as exclusive OR (XOR) and the ability to reproduce past inputs, serving as a benchmark task for nonlinearity assessment. The training phase utilized transient responses to a sequence of 1000 random binary UV pulses to determine the readout weights. In this sequence, the binary states of ‘1’ and ‘0’ correspond to the presence and absence of UV irradiation, respectively. Each pulse had a width of 0.5 s, and the number of virtual nodes was set to 5. To evaluate the correlation, these trained weights were multiplied by the transient responses from a separate sequence of 1000 pulses, and the resulting output was compared with a time-delayed version of the input sequence. The details are described in a previous study [6]. In addition, we demonstrated a hand-written digit recognition task using MNIST. The detailed experimental conditions related to the above tasks are described elsewhere [21,22,23,24,25].

3. Results and Discussion

3.1. Structural and Optical Properties of GaO_x Thin Film

Figure 1b shows the XRD patterns of the films without and with post-annealing. The peak observed at 41.68° is attributed to α-Al₂O₃ (0006) (PDF 04-007-5143). Regardless of annealing, no peaks corresponding to Ga₂O₃ were observed. Figure 1c shows optical transmittance as a function of the wavelength. Both samples exhibited transparency in the visible region, with reduced transmittance in the UV region below 300 nm. Figure 1d shows the Tauc plots of the films. The optical bandgap was determined from the intersection of the extrapolated linear region of (αhν)² and the horizontal axis. The optical bandgaps of the samples without and with post-annealing were 4.86 and 4.75 eV, respectively. These values were smaller than that of GaO_x thin films, which was reported to be 5.17 eV [26]. The narrowing bandgap might be due to the formation of intra-gap states at shallow positions in the gap, which reduced the apparent optical bandgap [27].

3.2. Optoelectronic Synaptic Properties of GaO_x Thin Film

Figure 2a,b show transient responses of the samples without and with post-annealing during UV pulse irradiation. The photocurrent gradually increased during UV-C irradiation. The current values just before stopping UV light irradiation, after continuously irradiating devices without and with annealing for a certain period, were 34.48 µA and 0.91 µA, respectively. When UV irradiation was stopped, the photocurrent gradually decreased for both samples. This result is likely due to persistent photoconductivity (PPC) [17,28,29].

For further analysis, the transient response curves of the photocurrent were fitted by the following equations:

$$I_{\mathrm{r}}\left(t\right)=I_0+{\displaystyle \sum _{n=1}^N}{A}_n\left\{1-\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \frac{t-t_0}{{\tau }_{\mathrm{r}n}}}\right)\right\}$$

(1)

$$I_{\mathrm{d}}\left(t\right)=I_0+{\displaystyle \sum _{n=1}^N}{B}_n\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \frac{t-t_0}{{\tau }_{\mathrm{d}n}}}\right)$$

(2)

where I_r(t), I_d(t), and I₀ are the transient photocurrent with light irradiation, the transient photocurrent without light irradiation, and the steady-state photocurrent, respectively. A_n and B_n are the constants, t is the time, t₀ is the time when the change in photocurrent begins, and τ_rn and τ_dn are the time constants. Fitting was performed for N = 3, which adequately described each transient response. τ_r1, τ_r2, and τ_r3 for the as-grown sample were 0.192, 1.55, and 16.69 s, respectively, and τ_d1, τ_d2, and τ_d3 were 0.160, 1.62, and 14.39 s, respectively. τ_r1, τ_r2, and τ_r3 for the 400 °C annealed sample were 0.07, 1.91, and 43.61 s, respectively, and τ_d1, τ_d2, and τ_d3 were 0.24, 8.78, and 54.15 s, respectively. Some previous reports defined the time required for a 10% to 90% increase and a 90% to 10% decrease in photocurrent as τ_r and τ_d, respectively [16,30]. Other reports used the equations as mentioned above, with N of 1 [31,32,33] and N of 2 [34,35,36]. While the shorter time constants in our experiment were similar to those reported previously, the longer time constants observed were notably higher.

It is known that the photocurrent is composed of two components [37]. The transport of photoexcited carriers can mainly be classified into two types: drift current due to the strong electric field between the electrodes, and slow transport mediated by trap–detrapping at gap states such as V_o defects. The former is considered the origin of the short time constant in the transient response of the photocurrent, while the latter accounts for the long time constant. The reason for fitting with three time constants is likely due to the presence of two slow transport paths. Since the precise origins of these time constants remain speculative, further investigation into the carrier transport phenomena will be conducted in the future.

Based on the slow transient photocurrent dynamics originating from the gap states, we further investigated the potential of the GaO_x films for artificial optoelectronic synapse applications. The PPF index is defined as the ratio of the intensities of two peaks when two pulses are applied to the device [6], as shown in Figure 3a. Figure 3b,c show the PPF indices of samples without and with annealing, respectively, for different pulse durations T_p and ΔT. At T_p = 0.01 s and ΔT = 0.01 s, the PPF values for the samples without and with annealing were 164.5% and 119.9%, respectively. As shown in Figure 3b,c, the without-annealing sample had a PPF index of over 130% when T_p was less than 0.1 s and ΔT was less than 0.5 s. However, the PPF index did not exceed 130% for the sample with annealing. When T_p < τ_r1 and ΔT< τ_d1, the PPF tends to be large because the photocurrent increases or decreases significantly. Previous reports suggested that the transient properties of artificial synapses can be reproduced using a series circuit of a resistor (R) and a capacitor (C) [38]. However, the present results cannot be reproduced by a simple RC circuit. The reason why might be that the time constants of photocurrents in the first and second pulses were different. It is considered that the time constant of the photocurrent during the second light irradiation differed from that of the first light irradiation because the gap states were partially occupied by photoexcited carriers from the first irradiation. In the future, precise defect control will be necessary to match the operating time range of the device to the time scale of the application because the gap states are crucial to control the time constants.

3.3. Reservoir Computing Tasks and Handwriting Recognition Demonstration

The results of the STM and PC tasks are shown in Figure 4. Figure 4a shows that the determination coefficients with a delay of 1 for the without- and with-annealing samples were estimated to be 0.97 and 0.82, respectively. The correlation coefficients at delay 3 for the samples without and with annealing were 0.12 and 0.03, respectively. The memory capacity (C_STM) is the sum of the determination coefficients for all delays in the STM task [21]. The C_STM of the without- and with-annealing samples was 1.62 and 1.05, respectively. The C_STM of GaO_x has not been investigated yet. Compared with the C_STM of the ZnO-CNF-based optoelectronic artificial synapse, which was 1.8 [6], the C_STM of GaO_x was slightly smaller than this value. Figure 4b shows that the determination coefficients for samples without and with annealing in the PC task were 0.70 and 0.063, respectively. When delay was 2, the determination coefficients of the former and the latter samples were 0.02 and 0.001, respectively. The parity check capacity (C_PC) is the sum of the determination coefficients for all delays in the PC task. The C_PC for the samples without and with annealing were 0.731 and 0.066, respectively. Since the PC task requires recall of past inputs as well as the STM task, it is likely that the PC tends to be small when the C_STM is small. C_STM and C_PC were found to decrease after annealing. In other words, annealing the sample reduces the ability to store information.

We hypothesize that the annealing process improves the crystallinity of the films and reduces the number of trap sites, leading to a decrease in the recombination rate. This leads to shorter time constants and quick saturation or relaxation, resulting in lower C_STM and C_PC. Future structural analyses will be necessary to experimentally verify this hypothesis. Furthermore, from the perspective of computing performance, it should be noted that the required C_PC value depends on the complexity of the target task. A low C_PC indicates limited nonlinearity, which restricts the computational capacity of the reservoir and makes it difficult to process complex calculations. Therefore, when applying the reservoir to tasks that inherently demand high nonlinearity, it is necessary to utilize devices with a higher C_PC. To improve the nonlinearity, possible methods include optimizing the deposition conditions, controlling the defect states, and inducing ionic conductivity [39,40].

For the handwriting recognition task, samples without and with annealing were irradiated with 16 patterns of 4-bit pulses. Figure 5a shows the current values at 2 ms after the end of each irradiation. Different current values were obtained for each pulse for the without-annealing sample, while no significant change was observed for each pulse for the with-annealing sample. Figure 5b shows the accuracy of the number recognition for each epoch for the samples without and with annealing. For the without-annealing sample, the accuracy was 80% and 90% for epochs of 1 and 10, respectively. For the with-annealing sample, the accuracy was 58% and 87% at epochs of 1 and 10, respectively. The annealing process decreased the accuracy at any number of epochs. The accuracies of the β-Ga₂O₃-based optical synaptic device and the ZnO-CNF-based optoelectronic artificial synapse were 78% and 88%, respectively [6,36]. The accuracy of the sample without annealing, at an epoch of 10, was higher than these values. Figure 5c,d show the confusion matrix for the samples without and with annealing, respectively, at epochs of 10. Both samples were shown to exhibit a high probability of recognizing and classifying numbers.

Although this initial proof-of-concept evaluation is currently limited to a single representative device, the findings obtained in this study demonstrate the performance of the amorphous GaO_x reservoir layer. These insights are expected to significantly contribute to the further development of amorphous GaO_x-based optoelectronic synaptic devices for PRC applications.

4. Conclusions

In conclusion, we investigated the optoelectronic synaptic properties of amorphous GaO_x thin films. It was revealed that the transient response curve of the photocurrent during UV light irradiation consists of a linear combination of three exponential functions with different time constants: sub-seconds, seconds, and tens of seconds. By annealing the samples, most of the time constants were extended. The characterization of the PPF index for various T_p and ΔT for two consecutive light pulses revealed that, when T_p was equal to or shorter than the fast time constant, it exhibited higher values compared to other pulse conditions. To characterize the performance as a PRC, STM and PC tasks were conducted, and our device showed a C_STM of 1.62 and a C_PC of 0.731. Although these values are lower compared to ZnO-based synaptic devices [6], there is room for improvement through the control of the shape, crystallinity, and oxygen vacancies in GaO_x thin films. Furthermore, using our device as the reservoir layer in PRC, we conducted a handwritten digit recognition task and found that it achieved a high accuracy of approximately 90%. Therefore, the performance comparison revealed that the device without annealing is more suitable for PRC applications than the device with annealing, demonstrating that the amorphous GaO_x thin films have potential for utilization as PRC devices.

In the future, we aim to develop synaptic devices that operate within the desired time range and explore their application in PRC. For example, to realize an AI optoelectronic sensor that uses time series light related to biological signals as input data, it is necessary to develop a device with the highest C_STM in the sub-second to second range. This will require controlling the material properties such as the density of gap states and interface states. By envisioning various applications, we believe that it will be possible to achieve more versatile or specialized devices for PRC.