Reconfigurable, non-volatile neuromorphic photovoltaics

doi:10.21203/rs.3.rs-2558516/v1

Download PDF

Article

Reconfigurable, non-volatile neuromorphic photovoltaics

https://doi.org/10.21203/rs.3.rs-2558516/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 20 Jul, 2023

Read the published version in Nature Nanotechnology →

Version 1

posted

You are reading this latest preprint version

Reconfigurable image sensors for the recognition and understanding of real-world objects are now becoming an essential part of machine vision technology. The neural network image sensor — which mimics neurobiological functions of the human retina —has recently been demonstrated to simultaneously sense and process optical images. However, highly tunable responsivity concurrently with non-volatile storage of image data in the neural network would allow a transformative leap in compactness and function of these artificial neural networks (ANNs) that truly function like a human retina. Here, we demonstrate a reconfigurable and non-volatile neuromorphic device based on two-dimensional (2D) semiconducting metal sulfides (MoS₂ and WS₂) that is concurrently a photovoltaic detector. The device is based on a metal/semiconductor/metal (M/S/M) two-terminal structure with pulse-tunable sulfur vacancies at the M/S junctions. By modulating sulfur vacancy concentrations, the polarities of short-circuit photocurrent —can be changed with multiple stable magnitudes. Device characterizations and modeling reveal that the bias-induced motion of sulfur vacancies leads to highly reconfigurable responsivities by dynamically modulating the Schottky barriers. A convolutional neuromorphic network (CNN) is finally designed for image process and object detection using the same device. The results demonstrated the two-terminal reconfigurable and non-volatile photodetectors can be used for future optoelectronics devices based on coupled Ionic-optical-electronic effects for Neuromorphic computing.

Physical sciences/Nanoscience and technology/Nanoscale devices/Nanosensors

Physical sciences/Engineering/Electrical and electronic engineering

Physical sciences/Nanoscience and technology/Nanoscale materials/Two-dimensional materials

The use of sensory nodes in neuromorphic vision networks is on the rise, but the data generated by these nodes are often unstructured and redundant for machine-learning algorithms¹. This is because the sensory units, which operate mostly in the analog domain, are physically separate from the processing units, which are typically executed digitally using von Neumann architecture. As a result, large amounts of data must be transferred to local processing units or cloud-based systems, leading to high energy consumption, latency, insufficient storage space, and communication bandwidth². In contrast, the human visual system can perform complex tasks in real-time due to its inherent parallelism, which allows for the combination of image sensing and processing functions^3–5.

To address this issue, researchers have developed retina-inspired devices based on bulk materials and two-dimensional (2D) materials, such as variable-sensitivity photodetectors (VSDP)^{6, 7}, dual-gate photodiodes (DGP)^{8, 9}, two-terminal photo-memories (TPM)^10–14, and gate-tunable vision sensors (GVS)^15–22. However, these neural network sensors often suffer from bias-dependent dark current with high power consumption (VSDP), volatile photocurrent for neural network (TPM), or complicated preparation process for integration and low photoresponsivity (DGP, GVS). In this context, 2D materials offer superior features for neuromorphic vision hardware, since they show strong light-matter interaction in a broad part of the spectrum²³, ease of fabrication and integration^{24, 25}, as well as the possibility of static and dynamic tunability of the potential profile within a device^26–28. Further, 2D materials as a technology are now advanced and mature enough to be used over wafer scales in complex integrated device and circuit level systems which can further be easily integrated with silicon readout/control electronics^29–34.

In this study, we present a two-terminal photovoltaic detector based on metal/2D-semiconductor/metal (M/S/M) structures using layered metal sulfides (MoS2, WS2) as the 2D semiconductors. By artificially introducing sulfur vacancies (positive ions) and controlling their migration with an electric field (voltage pulses), we can change the local sulfur ion concentration and adjust the Schottky barrier height at the M/S junction regions. This scheme not only allows for the adjustment of multiple resistance states in the 2D M/S/M photodetectors, mimicking biological synapses, but also the sign of the net photocurrent. The increase/decrease of Schottky barrier height in the two-terminal photodetector can break the symmetry of electric potential in the M/S/M devices, leading to change in sign of the photocurrent and even achieving zero photocurrent at zero bias. Overall, this allows for tuning the amplitude and polarity of zero-biased photocurrent with non-volatile responsivities, which when applied to neuromorphic vision networks, greatly simplifies the architecture for object position detection.

The device fabrication process involves transferring a mechanically exfoliated 2D MoS2 flake of approximately 10 nm thickness (Supplementary Information Fig. 1) onto SiO2/p + + Si substrates. Standard microfabrication processes were then used to fabricate the source/drain (S/D) electrodes. It is important to note that the 2D MoS2 under the S/D metal electrodes is treated with argon/oxygen plasma, which leaves a high concentration of sulfur vacancies (ions) at the M/S junction regions. The MoS2 photodetector has two M/S Schottky diodes connected back-to-back, as shown in Fig. 1a. Scanning electron microscopy (SEM) images (Fig. 1b) show the 2D-MoS2 M/S/M photodetector where the MoS2 under the S/D electrodes is treated with plasma, while the MoS2 channel (~ 1 µm) is untreated. To further characterize the device, we used Raman spectrum and transmission electron microscopy (TEM) to analyze the pristine and plasma-treated MoS2. A redshift of E_2g¹ peak and a blue shift of A_1g peak, indicating a change in MoS2 thickness with plasma treatment (Supplementary Information Fig. 2) was observed. Cross-sectional TEM and corresponding energy dispersive spectroscopy (EDS) analysis confirmed the presence of rich sulfur vacancies (positive ions) introduced by plasma treatment.

Wavelength-dispersive X-ray spectroscopy (WDS) was used to analyze the distribution of sulfur atoms and molybdenum atoms in the channel and below the S/D electrode. We observe a contrast ratio of 1.95 of sulfur atoms to molybdenum atoms in the channel and a ratio of 1.42 below the S/D electrodes, indicating the presence of sulfur vacancies after plasma treatment (Supplementary Information Fig. 3). X-ray photoelectron spectroscopy also confirms that the S/Mo atomic ratio of pristine MoS2 is 1.94 while the S/Mo atomic ratio of plasma treated MoS2 is reduced to 1.50, further indicating the absence of sulfur atoms (Supplementary Information Fig. 4). Additional device fabrication and structural characterization is detailed in Methods. Figures 1e and 1f show the dynamic photocurrent characteristics of the two-terminal MoS2 photodetectors with plasma-treated M/S junctions. The MoS2 photodetector initially shows an obvious photovoltaic effect with short-circuit current (I_sc) of + 10 nA and open-circuit voltage (V_oc) of -6 mV under 532 nm light illumination (Fig. 1e). However, after 15 V pulses of 10 s duration, the MoS2 photodetector shows opposite photocurrent behaviors with I_sc of -13 nA and V_oc of -8 mV under the same biasing conditions (Fig. 1f). In Figs. 1g, 1h and Supplementary Information Fig. 5, The time-resolved photocurrent further demonstrates its tunable polarity in the visible to infrared regions (Fig. 1g, h and Supplementary Information Fig. 5). The tunable polarity in I_sc is also observed in other metal sulfides like WS2 (Supplementary Information Fig. 6). However, for the pristine MoS2 M/S/M device (untreated with plasma), the photocurrent has no obvious change upon application of same magnitude and duration of voltage pulses, (Supplementary Information Fig. 7). The tunable photocurrent by voltage pulses is mainly attributed to the migration of sulfur vacancies by the electric field and strong ion-electron coupling.

The dynamic photocurrent characteristics of the device were studied using a two-terminal M/S/M configuration, with one electrode as an anode and another as a cathode. We collected multiple and non-volatile photocurrent states (I_sc) by applying a constant light intensity under short-circuit conditions. To investigate the tunable polarity and amplitude of the photocurrent state, we conducted scanning photocurrent measurements by scanning the laser spot across the anode, channel, and cathode (Fig. 2 and Supplementary Information Figure 8). As shown in Fig. 2a-i and Fig. 2a-iv, the photoresponse, without applying any voltage pulses, is locally observed at the metal/MoS2 contact (dashed square) with similar amplitude but opposite signs — where the photocurrent at the anode is positive while the photocurrent at the cathode is negative. However, after voltage pulse programming (15 V and 15 s duration), the same device shows enhanced positive photocurrent at the anode, while the negative photocurrent at the cathode almost vanishes, as shown in Fig. 2a-ii. Further, the same device after voltage pulse programming with reversed sign but same amplitude ( -15 V and 15 s duration) shows opposite photocurrent behavior i.e. the negative photocurrent at the cathode is enhanced while the positive photocurrent at the anode almost vanishes, as shown in Fig. 2a-iii. To further confirm this reconfigurability, the 15V 15s voltage pulse is again applied to the same device with a photocurrent state as shown in Fig. 2-iii, and the photocurrent is fully recovered to its initial state as shown in Fig. 2a-i (similar amplitude but opposite signs, Supplementary Information Figure 9). Fig. 2a-iv shows the corresponding photocurrent amplitude as a function of laser spot position extracted from Figs. 2a-i, ii, iii. Note that the volage pulsing can effectively modulate the amplitude and polarity of local photocurrent at the M/S junctions. Fig. 2b shows the photocurrent map of pristine MoS2 M/S/M photodetector without plasma treatment and the photocurrent amplitude at the M/S junction shows negligible changes even with multiple voltage pulse programming. Further, the amplitude of photocurrent of pristine device (~ pA) is much lower than the plasma treated device (~ mA). This experiment shows that the migration of sulfur vacancies (ions) strongly affects the photocurrent amplitude and polarity of the plasma-treated MoS2 photodetector. This is due to the ion-electron coupling effect. Fig. 2c illustrates the I-V characteristics of the plasma-treated MoS2 photodetector in the dark and under global illumination. Initially, the device shows an I_sc of -6 nA and a V_oc of 5 mV (black line, left of Fig. 2c). However, after many rounds of voltage pulse programming (-15/-20 V, 100 ms, +50/100 pulse), the device ultimately shows opposite I-V behavior with an I_sc of 50 nA and a V_oc of -25 mV. Similarly, a device with an I_sc of 50 nA and a V_oc of 10 mV (black line, right of Fig. 2c) can be ultimately modulated to an opposite I-V behavior with an I_sc of 50 nA and a V_oc of -25 mV after voltage pulse programming (15 V, 100 ms, +400/700 pulse). In contrast, for pristine MoS2 devices, no obvious photovoltaic effect was observed after many rounds of voltage pulse programming (Fig. 2d). The plasma-treated MoS2 devices further show six states of I_sc and V_oc by varying the duration of the voltage pulses (Supplementary Information Figure 10).

It is important to note that this phenomenon is not observed in a one-off device and has been highly reproducible in many devices as shown in the statistical chart (Fig. 2f and Supplementary Information Figure 11).

To understand the atomic mechanism behind the reconfigurable and non-volatile responsivity of MoS2 M/S/M devices, we conducted a series of experiments using WDS characterizations, Kelvin probe force microscope (KPFM), and Sentaurus-TCAD simulation. We observed the in-situ distribution of sulfur atoms in the MoS2 channel by applying voltage biases to the M/S contact, creating an in-plane electric field of approximately 10 MV/cm. Our results, shown in Fig. 3a, revealed that before applying voltage pulses, the sulfur atoms were uniformly distributed across the channel. The red signal in WDS map shows the distribution of sulfur atoms in the channel. From Fig. 3a-i and Fig. 3a-ii, the WDS map and corresponding elemental spectrum demonstrate uniform sulfur distribution across the MoS2 channel before electric pulse programming. However, after applying a negative voltage pulsing (amplitude, -10 V; duration, 10 s) on anode (with cathode grounded, Fig. 3b-i), the MoS2 channel shows the obvious absence of sulfur atoms near anode indicating the migration of sulfur vacancies. In Fig. 3b-ii, the intensity of sulfur atoms near the anode is lower than the intensity of sulfur atoms near the cathode.

To demonstrate the reversible migration of sulfur vacancies, positive (amplitude, + 10 V; duration, 10 s) electric pulses are applied to anode in sequence; the MoS2 channel restoring to its initial uniform sulfur distribution indicating the reversible migration of sulfur ions (Fig. 3c-i). And the uniform distribution of sulfur atoms is observed in elemental spectrum as shown in Fig. 3c-ii. The device schematics illustrate the sulfur atom migration near anode and cathode (Fig. 3a-iii, Fig. 3b-iii and Fig. 3c-iii). After the positive/negative electric pulses, the distribution of sulfur atom changes the ions density at anode and cathode. Since the layered MoS2 offers ionic pathways with low activation energy, the sulfur ions have a high in-plane diffusivity^{35, 36}. Therefore, the sulfur ion migration in the MoS2 channel, driven by an electric field, modifies the local sulfur concentration. The increase/decrease in the local S ion concentration leads to the dynamic decrease/increase of Schottky barrier height^{37, 38}. The asymmetric Schottky barrier of M/S contact can facilitate the separation of photo-generated electron-hole pairs due to the potential gradient leading to a large I_sc. Moreover, like memristors, the multiple photocurrent states are non-volatile because the barrier is modulated by ion concentration which remains fixed in the absence of a large electric field. The asymmetric distribution of sulfur ions in the MoS2 channel results in a strong photovoltaic effect with a maximum I_sc of -795 nA and V_oc of 60 mV (Supplementary Information Fig. 12). Also, the maximum optical responsivity of 369.22 mA/W was achieved with ionic migration.

We further conducted KPFM measurements (to map the magnitude and potential of MSM device, Fig. 3d) verifying the dynamic modulation of the Schottky barrier at the metal/MoS2/metal junction with various voltage programming conditions. The results, shown in Fig. 3e, indicated that the Schottky barrier height at the anode region (high sulfur vacancy concentration) was lower than that at the cathode region (low sulfur vacancy concentration), with a difference of approximately 50 meV. In contrast, the pristine MoS2 device displayed a symmetric Schottky barrier (Supplementary Information Fig. 13). These results were consistent with the WDS maps, which showed that a negative voltage applied to the anode attracts sulfur vacancies towards the anode and reduces the number of vacancies near the cathode, leading to the formation of Schottky or Ohmic contact at the metal/MoS2 interface and affecting the observed I_sc and V_oc. To further understand the impact of the asymmetric Schottky barrier on photocurrent, we used Sentaurus-TCAD to simulate the band structure, electric field distribution, and photocurrent response of the metal/MoS2/metal device. The simulated band structure, shown in Fig. 3f, displayed the potential gradient induced by the asymmetric Schottky barrier. Figure 3g also showed an obvious I_sc due to the asymmetric Schottky barrier. The simulated electron density distribution with different barrier at anode and cathode under illumination are shown in Supplementary Information Fig. 14. The lower anode barrier corresponds to the larger I_sc, confirming the experiment results.

Overall, our theoretical and experimental results discussed above demonstrate that the reconfigurable and non-volatile responsivity of M/S/M photodetectors is due to the electric field-driven redistribution of sulfur vacancies, which dynamically changes the Schottky barrier and doping profile.

With the above described reconfigurable and non-volatile photovoltaic response in the 2D M/S/M device, a convolutional neuromorphic network (CNN) is designed and demonstrated for image processing and classification^{15, 20}. The responsivity and conductance of our device are programmed to perform the multiply-accumulation (MAC) operation for the CNN. Here, a more complex CNN task – object detection – has been realized. In order to achieve high accuracy, we utilized eleven non-volatile responsivity states with varying I_sc polarities as depicted in Fig. 4a. And the linear interval of short-circuit current reaches 4000 mW/cm² .The temporal photocurrent response and calculation of the responsivity states are illustrated in Supplementary Information Figs. 15 and 16. Additionally, the I_sc was found to change linearly with the optical power density, which is crucial for updating the weight in the neuromorphic network, as shown in Fig. 4b. The specific pulse numbers and duration for the eleven responsivity states can be found in Fig. 4c. These responsivity states were found to be stable and maintain their integrity for longer than 1000 seconds, as demonstrated in Supplementary Information Fig. 17. Using these multiple responsivity states, we were able to successfully demonstrate image processing for "Cameraman," "Lena," and "Peppers" and image classification for MNIST "0–9" as seen in Supplementary Information Figs. 18 and 19. Furthermore, we examined the tunable conductance in 2D M/S/M devices by applying multiple electric pulses, which resulted in reconfigurable conductivity for up to 168 conductivity states, as shown in Supplementary Information Fig. 20.

Object detection is a more challenging task for convolutional neural networks (CNNs) than image classification, as it requires precise computation of localization parameters³⁹. Conventional object detection models typically require direct regression of the coordinates of the top-left corner (r,c) or the object center (x,y) and its height and width (h,w)^{40, 41}. However, these models can struggle when applied to systems with limited programmed states (both optically and electrical states), which can lead to inaccurate regression results or failures. To address this problem, we proposed a unified discrete weight object detection framework for object center detection that is able to accommodate the limited discrete levels of the programmable states in our devices, as shown in Fig. 4d. Our networks only contain convolution layers and activation operations, without BatchNorm or InstanceNorm operations. Discretization operations were applied to the weights in the offline-trained network to adapt to the limited discrete levels of our programmable states. The training strategies and the OSU Thermal Pedestrian Database used for object detection are presented in Supplementary Information Figs. 21 and 22. Instead of directly regressing the coordinates of the object center, our unified discrete weight object detection network generates a heatmap for object detection, with peaks in the heatmap corresponding to object centers. The position of the maximum value in the heatmap will not be affected after weight discretization. After training, we obtained a heatmap with 24x24 pixels for the calculation of the target position. We then employed a 24x24 memristor array to display the 24x24-pixel image. Through Ohm's law and multiply-accumulate processing⁴², the maximum row and column voltage values can be obtained (Fig. 4e). The detected (x,y) coordinate location of the object image can be displayed by the maximum voltage in the memristor array (Supplementary Information Fig. 23).

Finally, we compared the performance of object detection network with three types of weights, namely float-16 datatype weights, weight discretization only in the optical network, weight discretization in both the optical and electronic networks. The accuracy of object detection, for our 11 responsivity states in optical front-end with float-16 weights in electronic processing is ~ 97% and for our 11 responsivity states in optical front-end with our conductivity states in electronic processing is ~ 96% as shown in Fig. 4f. And the loss of three types of weights decreases in ~ 10 epochs and converges to ~ 50 epochs in Fig. 4g. These experimental results showed that the performance of our discrete weight framework is comparable with the float-16 datatype network.

In summary, we have presented reconfigurable and non-volatile photovoltaic-detector devices for intelligent image processing and object detection. The devices are based on a simple two-terminal M/S/M architecture with 2D semiconducting metal sulfides as the channel. The migration of sulfur vacancies driven by electric pulses lead to the modulation of M/S Schottky barrier due to the electron-ion coupling effect. The magnitude and polarity of device responsivity are highly tunable and non-volatile which can be used to build neuromorphic hardware for intelligent perception as demonstrated. Furthermore, scaling the neuromorphic hardware to larger dimensions is conceptually feasible due to the simple device structure and provides various training possibilities for neuromorphic vision applications.

Device fabrication. Multi-layer MoS₂ flakes were mechanically exfoliated from the high- quality bulk crystal (HQ Graphene, Netherlands) onto the polydimethylsiloxane (PDMS) and transferred onto a silicon substrate with 300nm SiO₂ (Universitywafer, America). Then, the electrode patterns were defined by the electron-beam lithography (EBL). After that, the positions of electrode patterns were exposed to O₂ and Ar plasma in the Precision Soft Etching System (nanoETCH), while the MoS₂ channel was protected by a photolithography-patterned photoresist. For O₂ or Ar plasma treatment process, the etch power was set to 10 W, the chamber pressure is 1×10^− 7 Torr, the gas flow rate was 30 sccm. O₂ and Ar processing times were 10 s and 30 s, respectively. Finally, the 1 nm Cr and 50 nm Au layers were deposited on MoS2 by thermal evaporation with a base pressure of 3 × 10^− 6 Torr and a deposition rate of 0.3 Å s^− 1, followed by the standard lift-off process.

WDS measurements. WDS measurements of the M/S/M devices were carried out by a unique wavelength-dispersive soft X-ray emission spectrometer in an electron probe micro analyzer (EPMA) equipped with a Schottky field emission (FE) electron gun, which is designed to enhance the analytical area of nanometers in size.

TEM imaging and EDS mapping. Samples were treated with plasma and evaporated 10nm Au for protection immediately. For TEM, samples were prepared using a Thermo Scientific Helios G4 HX dual-beam system. The samples were covered by a carbon layer deposited using electron-beam evaporation and a thicker platinum protective layer was deposited using sputtering. TEM imaging was conducted in a Thermo Scientific Tecnai F20 (200kV) transmission electron microscope. We also got the distribution of elements from the EDS mapping in our materials.

Raman and XPS measurements. Raman spectra were obtained by using a Lan Ram HR800 with a 532 nm laser as an excitation source. XPS characterizations were using a PHI5000 VersaprobeIII system that base pressure was 4 × 10^− 8 Torr. The source was operated at 15 kV with an emission current of 4.5 mA.

AFM and KPFM measurements. The thickness and surface potential of our device were characterized by a scanning probe microscope (Cypher S).The KPFM measurements were using a tip coated with Ti/Ir thin film (ASYELEC-01-R2) in the SKPM mode at 0 bias voltage. And the AFM imaging was acquired from the height retrace mode in the KPFM testing process.

Device characterization. The electrical characterizations of the devices were conducted using a commercial Keysight B1500A in a Lake Shore probe station. The devices were wire-bonded onto a 28-pin printed circuit board (PCB) and 520 nm, 830 nm laser was used for photoresponse measurements. The devices was positioned at the central of the light spot to make sure a uniform illumination. The data of response time was acquired using a Textronix MDO3014 oscilloscope. All the measurements were carried out at room temperature in an ambient environment.

TCAD simulations. Simulations of metal/MoS₂/metal device under illumination conditions were performed using Sentaurus TCAD from Synopsys Inc. The model of the device and additional details were discussion in the Supplementary Information.

Data availability

The data that support the conclusions of this study are available from the corresponding authors upon reasonable request. Source data are provided with this paper.

Code availability

The codes used for simulation and data plotting are available from the corresponding authors upon reasonable request.

Acknowledgements

This work was supported by the National Key Research and Development Program of China (2021YFA0715602), the National Natural Science Foundation of China (62261136552, 62005303, and 62134001), the International Partnership Program of Chinese Academy of Sciences (Grant No. 181331KYSB20200012), the Science and Technology Commission of Shanghai Municipality (21JC1406100), the Alfred P. Sloan Foundation (Sloan Fellowship in Chemistry) (D.J.), the Open Research Projects of Zhejiang Lab (2022NK0AB01).

Author contributions

J. M., D. J., and W. H. conceived the idea/concept and directed the collaboration and execution. T.L. and X.F. fabricated the devices and performed the measurements. T.L., J. M., P.Z., X. W., D.J., and W. H. analyzed the experimental data. T. L. X. Z., and J. M. did the TCAD device simulation. B.S. and B.C. performed the AI object detection. T. L., J.M., W. H., and D.J. co-wrote the manuscript with contributions from all the authors. All authors discussed the results and implications and commented on the manuscript at all stages.

Competing interests

The authors declare no competing interests.

Sze, V., Yu-Hsin, C., Emer, J., Suleiman, A.&Zhengdong, Z. Hardware for machine learning: challenges and opportunities. In 2017 IEEE Custom Integrated Circuits Conference (CICC) 179–186 (IEEE, 2017)
Zhou, F.&Chai, Y. Near-sensor and in-sensor computing Nat. Electron. 3, 664–671 (2020).
Gollisch, T.&Meister, M. Eve Smarter than Scientists Believed: Neural Computations in Circuits of the Retina Neuron 65, 150–164 (2010).
Kyuma, K., et al. ARTIFICIAL RETINAS - FAST, VERSATILE IMAGE-PROCESSORS Nature 372, 197–198 (1994).
Kolb, H. How the retina works - Much of the construction of an image takes place in the retina itself through the use of specialized neural circuits Am. Scientist 91, 28–35 (2003).
Funatsu, E., et al. An artificial retina chip with current-mode focal plane image processing functions IEEE Trans. Electron Devices 44, 1777–1782 (1997).
Nitta, Y., Ohta, J., Tai, S.&Kyuma, K. Variable-sensitivity photodetector that uses a metal-semiconductor-metal structure for optical neural networks Opt. Lett. 16, 611–613 (1991).
Jang, H., et al. In-sensor optoelectronic computing using electrostatically doped silicon Nat. Electron. 5, 519–525 (2022).
Mennel, L., et al. Ultrafast machine vision with 2D material neural network image sensors Nature 579, 62–66 (2020).
Chen, S., Lou, Z., Chen, D.&Shen, G. An Artificial Flexible Visual Memory System Based on an UV-Motivated Memristor Adv. Mater. 30, (2018).
Cui, B., et al. Ferroelectric photosensor network: an advanced hardware solution to real-time machine vision Nat. Commun. 13, (2022).
Sun, L., et al. In-sensor reservoir computing for language learning via two-dimensional memristors Sci. Adv. 7, eabg1455 (2021).
Zhou, F., et al. Optoelectronic resistive random access memory for neuromorphic vision sensors Nat. Nanotechnol. 14, 776–782 (2019).
Ahmed, T., et al. Optically Stimulated Artificial Synapse Based on Layered Black Phosphorus Small 15, (2019).
Lee, S., Peng, R., Wu, C.&Li, M. Programmable black phosphorus image sensor for broadband optoelectronic edge computing Nat. Commun. 13, 1485 (2022).
Liao, F., et al. Bioinspired in-sensor visual adaptation for accurate perception Nat. Electron. 5, 84–91 (2022).
Liu, K., et al. An optoelectronic synapse based on α-In2Se3 with controllable temporal dynamics for multimode and multiscale reservoir computing Nat. Electron. 5, 761–773 (2022).
Pi, L., et al. Broadband convolutional processing using band-alignment-tunable heterostructures Nat. Electron. 5, 248–254 (2022).
Seo, S., et al. Artificial optic-neural synapse for colored and color-mixed pattern recognition Nat. Commun. 9, 5106 (2018).
Wang, C.-Y., et al. Gate-tunable van der Waals heterostructure for reconfigurable neural network vision sensor Sci. Adv. 6, eaba6173 (2020).
Yu, J., et al. Bioinspired mechano-photonic artificial synapse based on graphene/MoS2 heterostructure Sci. Adv. 7, eabd9117 (2021).
Zhang, Z., et al. All-in-one two-dimensional retinomorphic hardware device for motion detection and recognition Nat. Nanotechnol. 17, 27–32 (2022).
Lien, D.-H., et al. Engineering Light Outcoupling in 2D Materials Nano Lett. 15, 1356–1361 (2015).
Wang, Q. H., Kalantar-Zadeh, K., Kis, A., Coleman, J. N.&Strano, M. S. Electronics and optoelectronics of two-dimensional transition metal dichalcogenides Nat. Nanotechnol. 7, 699–712 (2012).
Shim, J., et al. Controlled crack propagation for atomic precision handling of wafer-scale two-dimensional materials Science 362, 665–670 (2018).
Khan, M. A., et al. The non-volatile electrostatic doping effect in MoTe2 field-effect transistors controlled by hexagonal boron nitride and a metal gate Sci Rep 12, 12085 (2022).
Wang, M., et al. Robust memristors based on layered two-dimensional materials Nat. Electron. 1, 130–136 (2018).
Zhu, X., Li, D., Liang, X.&Lu, W. D. Ionic modulation and ionic coupling effects in MoS2 devices for neuromorphic computing Nat. Mater. 18, 141–148 (2019).
Chen, H., et al. Logic gates based on neuristors made from two-dimensional materials Nat. Electron. 4, 399–404 (2021).
Liu, W., et al. Graphene charge-injection photodetectors Nat. Electron. 5, 281–288 (2022).
Tong, L., et al. 2D materials-based homogeneous transistor-memory architecture for neuromorphic hardware Science 373, 1353–1358 (2021).
Wang, Y., et al. An in-memory computing architecture based on two-dimensional semiconductors for multiply-accumulate operations Nat. Commun. 12, (2021).
Miao, J., et al. Heterojunction tunnel triodes based on two-dimensional metal selenide and three-dimensional silicon Nat. Electron. 5, 744–751 (2022).
Choi, C., et al. Human eye-inspired soft optoelectronic device using high-density MoS2-graphene curved image sensor array Nat. Commun. 8, (2017).
Duy, L., Rawal, T. B.&Rahman, T. S. Single-Layer MoS2 with Sulfur Vacancies: Structure and Catalytic Application J. Phys. Chem. C 118, 5346–5351 (2014).
Komsa, H.-P., Kurasch, S., Lehtinen, O., Kaiser, U.&Krasheninnikov, A. V. From point to extended defects in two-dimensional MoS2: Evolution of atomic structure under electron irradiation Phys. Rev. B 88, (2013).
Li, D., et al. MoS2 Memristors Exhibiting Variable Switching Characteristics toward Biorealistic Synaptic Emulation Acs Nano 12, 9240–9252 (2018).
Sangwan, V. K., et al. Multi-terminal memtransistors from polycrystalline monolayer molybdenum disulfide Nature 554, 500–504 (2018).
Xingyi, Z., Koltun, V.&Krahenbuhl, P. Tracking Objects as Points. Computer Vision - ECCV 2020. 16th European Conference 474–490 (2020)
Redmon, J., Divvala, S., Girshick, R., Farhadi, A.&Ieee You Only Look Once: Unified, Real-Time Object Detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 779–788 (IEEE, 2016)
Ren, S., He, K., Girshick, R.&Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In 29th Annual Conference on Neural Information Processing Systems (NIPS) (2015)
Ielmini, D.&Wong, H. S. P. In-memory computing with resistive switching devices Nat. Electron. 1, 333–343 (2018).

There is NO Competing Interest.

SupplementaryInformationfinal.docx

Download PDF

Journal Publication

published 20 Jul, 2023

Read the published version in Nature Nanotechnology →

Version 1

posted

You are reading this latest preprint version

Reconfigurable, non-volatile neuromorphic photovoltaics

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Device Fabrication And Architecture

Dynamic Photocurrent Characteristics

Ion Modulation And Ionic Coupling Effect

Reconfigurable Neuromorphic Sensing

Conclusion

Methods

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1