† Corresponding author. E-mail:
Project supported by the National Key Research and Development Program of China (Grant Nos. 2017YFA0205704 and 2018YFB2200101), the National Natural Science Foundation of China (Grant Nos. 91964107 and 61774133), the Fundamental Research Funds for the Central Universities, China (Grant No. 2018XZZX003-02), the National Natural Science Foundation of China for Innovative Research Groups (Grant No. 61721005), and the Zhejiang University Education Foundation Global Partnership Fund.
High-performance neuromorphic computing (i.e., brain-like computing) is envisioned to seriously demand optoelectronically integrated artificial neural networks (ANNs) in the future. Optoelectronic synaptic devices are critical building blocks for optoelectronically integrated ANNs. For the large-scale deployment of high-performance neuromorphic computing in the future, it would be advantageous to fabricate optoelectronic synaptic devices by using advanced silicon (Si) technologies. This calls for the development of Si-based optoelectronic synaptic devices. In this work we review the use of Si materials to make optoelectronic synaptic devices, which have either two-terminal or three-terminal structures. A series of important synaptic functionalities have been well mimicked by using these Si-based optoelectronic synaptic devices. We also present the outlook of using Si materials for optoelectronic synaptic devices.
A brain is the most powerful information processor in nature.[1] An interconnected neural network in a brain is capable of distributed parallel processing, enabling high-speed computing. Moreover, both information processing and memory can take place in a brain, avoiding the data transfer between memory and processors in the current mainstream computing based on the von Neumann structure.[2] Hence, a brain consumes rather low energy.[3–5] Now it is widely believed that neuromorphic computing (i.e., brain-like computing) is one of the most important choices for next-generation computing. Since synapses play a critical role in a neural network, artificial synapses (i.e., synaptic devices) are thought to be crucial building blocks for neuromorphic computing.
Researchers have been investigating synaptic devices for many years. Synaptic devices based on the complementary metal–oxide–semiconductor (CMOS) technology of silicon (Si) were initially developed.[6–9] These devices did not have the capability of inherent hardware-based memorizing. The imitation of synaptic functionalities was mainly achieved by designing complex analog circuits.[10–12] Memristors were then used as synaptic devices, which possessed both processing and memorizing functions.[13–16] Although they demonstrated great potential for emulating synaptic functionalities,[17–19] memristors needed to face the trade-off among bandwidth, connection, and density.[20] It has been realized that the incorporation of light into synaptic devices can offer advantages such as ultrafast computing, high bandwidth, robustness, and low power consumption.[21–26] Light may enable real-time sensing, which helps to realize the functionalities of a visual nervous system by using synaptic devices.[27–31] Hence, optoelectronic synaptic devices working with light have recently attracted great attention. Various materials such as metal oxide films,[32–38] carbon nanotubes,[39–42] perovskite,[43–48] and two-dimensional layered materials[49–54] have been employed to fabricate optoelectronic synaptic devices with two- or three-terminal structures.
Si has evolved to be the material of choice for very large-scale integration (VLSI) circuits, on which the amazing success of current mainstream computing is based. If Si is used to make optoelectronic synaptic devices, the advanced mature technology of Si together with the low cost of Si should significantly contribute to the development of optoelectronically integrated neural network, which may be seriously demanded by the large-scale deployment of neuromorphic computing based on optoelectronic integration in the future. In this work we review the progress of using Si for optoelectronic synaptic devices. Before the discussion on the two- and three-terminal Si-based optoelectronic devices, we first briefly introduce biological synapses and synaptic plasticity. The opportunities and challenges for the development of Si-based optoelectronic synaptic devices in the future are also presented.
A brain consists of approximately 1011 neurons. Each neuron is connected to other neurons via approximately 103–104 synapses.[3,4] Figure
The EPSC/IPSC is of significance for information processing, learning, and memory. The magnitude of EPSC/IPSC is determined by the connection strength (i.e., synaptic weight) of a synapse. The synaptic weight is closely related to the activity history of either or both sides of the synapse. The change (potentiation or depression) of the synaptic weight is called synaptic plasticity.[24] There are short-term plasticity (STP) and long-term plasticity (LTP). The STP usually occurs in the time frame of milliseconds to minutes, while the LTP lasts several hours or longer. Paired-pulse facilitation (PPF)/depression (PPD) is a typical STP, which is signified by the potentiation/depression of the postsynaptic current for a second stimulus. Figure
Spike-rate-dependent synaptic plasticity (SRDP) is a typical type of LTP.[64,65] The firing frequency of the presynaptic spikes readily affects the LTP. For example, low frequency (1–5 Hz) spikes may lead to long-term depression, while high frequency (20–100 Hz) spikes may render long-term potentiation.[66] Figure
In optoelectronic synaptic devices, photoelectric conversion is efficiently utilized. Both optical and electrical spikes can be used for optoelectronic synaptic devices. Optical spikes are generally regarded as the stimulations from presynaptic neurons or external environment. Recently, optical signals have also been demonstrated to act as postsynaptic outputs.[70,71] When optoelectronic synaptic devices are fabricated, they have either two-terminal[72,73] or three-terminal structures.[74–76] Various operations may be realized by using these devices, which should help the construction of complex artificial neural networks (ANNs). Here we review Si-based optoelectronic synaptic devices by categorizing them into two-terminal and three-terminal ones.
Two-terminal synaptic devices are usually fabricated to build a crossbar array. The simple structure of two-terminal synaptic devices is quite suitable for the high-density integration of ANNs. This has encouraged the recent hardware implementation of ANNs with outstanding on-chip training performance.[77] Two-terminal synaptic devices are often purely electronic.[78] However, light has already been incorporated into two-terminal synaptic devices, enabling the integration of real-time sensing, processing, and memory functions. The resulting two-terminal optoelectronic devices now hold great promise for the development of neuromorphic computing based on optoelectronically integrated artificial neural networks.[73]
In late 1980s, AT&T Bell Laboratories demonstrated optically programmable neural networks based on photoconductive hydrogenated amorphous Si (a-Si:H) arrays.[79–82] The first generation of photoconductive a-Si:H arrays had the vertical (or sandwich-type) structure.[79] As shown in Fig.
Due to the fact that the vertical arrays could seriously suffer from the point and line defects of a-Si:H, the AT&T Bell Laboratories subsequently developed the second generation of optically programmable synaptic arrays based on a planar (or gap-cell type) structure.[80–82] As schematically shown in Fig.
Chen et al.[83] have recently demonstrated optoelectronic synaptic devices based on the heterostructure of hafnium metal oxide (HfO2) and bulk Si. The two-terminal device structure is schematically shown in Fig.
While the Pt/HfO2/p-Si device showed stateful photoresponses and nonvolatile memory behavior via transient optoelectronic stimulus, the emulation of STP or LTP behaviors was not realized. However, the STDP of the device made it applicable to the existing ANNs based on memristors.[84] The incorporation of optical signals may generalize a broader artificial synaptic system combining sensing and neural processing, which is expected to bear advanced functionalities in light-controlled cognitive and optical neuromorphic hardware.
Ultra-thin memristors with low reverse current are normally required to build high-density 3D stacking neuromorphic chips. He et al.[85] chose the 2D layered material of MoS2 to from a heterostructure with bulk Si. Optoelectronic synaptic devices based on this MoS2/Si structure were fabricated. The schematic of such a typical device is shown in Fig.
For the W/n-MoS2/p-Si device, the existence of a native SiO2 layer with trap sites played a key role in the photoelectric memory. This should enlighten the development of hybrid synaptic devices based on Si. To take advantage of the interfacial traps, ultrathin materials, e.g., monolayer 2D materials, are good candidates to build heterostructures. Surface traps of Si also promise to play a significant role in synaptic devices when the size of Si is reduced into the nanoscale regime.
Given their remarkable optical properties,[86–92] Si nanocrystals (NCs) have recently been used for optically stimulated synaptic devices. Tan et al.[93] first proposed two-terminal optoelectronic synaptic devices with the structure of ITO/Si NCs/Al, as illustrated in Fig.
The work of Tan et al. expanded the application of Si NCs into optoelectronic synaptic devices for the first time. This was encouraging for the realization of large-scale neuromorphic computing. However, their work failed to emulate the synaptic depression, which could be realized by modifying the device structure or introducing electrical stimuli.[98]
Although optoelectronic synaptic devices often have optical inputs and electrical outputs, Zhao et al. have managed to fabricate two-terminal optoelectronic synaptic devices with electrical inputs and optical outputs.[70,71] These devices were fabricated by using light-emitting Si NCs, which could be called Si quantum dots (Si QDs) as well. Figure
The Si-NC-based electroluminescent synaptic devices were capable of emulating synaptic plasticity by taking the optical power of the EL as the synaptic weight. Typical PPF was emulated by applying paired electrical spikes (Fig.
Zhao et al.[71] subsequently reported the second device structure of glass/ITO/PEDOT:PSS/P3HT/PFN/Si-QDs/ZnO/Ag for the Si-NC-based electroluminescent synaptic devices. Figure
The Si-NC-based electroluminescent synaptic devices successfully output optical signals with the wavelength of 850 nm. Since the NIR light is vital in nowadays optical communication, the output of NIR light is intriguing for synaptic devices. The realization of electroluminescent synaptic devices may facilitate the bidirectional conversion between optical and electrical signals, which is a critical step in optoelectronic integration of ANNs.
Yin et al.[99] developed synaptic phototransistors by using boron-doped Si NCs as the channel. The schematic structure of the synaptic Si-NC phototransistors is shown in Fig.
With the stimulation of combined optical spikes, negative electrical spikes, and positive electrical spikes, symmetric/asymmetric Hebbian/anti-Hebbian STDP learning rules were realized by the synaptic Si-NC phototransistors. Figure
Feasible energy consumption is still a major obstacle for nowadays hardware-based ANNs to compete with the biological neural systems.[28] Although augmentative functionalities can be realized by the synaptic Si-NC phototransistors, the energy consumption (∼ 10 nJ) of the device is much higher than that of a human brain (tens of fJ).[101] Ni et al.[102] demonstrated synaptic phototransistors based on the hybrid structure of 2D WSe2 and Si NCs (Fig.
Synaptic phototransistors based on Si NCs facilitated the development of energy-efficient neuromorphic computing. Studies on Si-based three-terminal synaptic devices are ongoing these days. Inspired by phototransistors based on Si NCs, efforts can be also devoted to synaptic transistors based on other Si nanostructures. Given the broad applications of Si and the existing CMOS technology, the fabrication process of ANNs may be simplified in the future.
Neuromorphic computing, which is an emerging computing paradigm, has gained numerous attention ever since the 1980s for its potential in artificial intelligence.[103] Inspired by biological neural networks, there have been a variety of ANNs including convolutional neural networks, recurrent neural networks, and SNNs. Hardware implementation of these ANNs is mainly based on CMOS integrated circuits, as evidenced by the IBM’s TrueNorth[104] and the Intel’s Loihi.[105] In recent years, memristors have been extensively studied for the development of electronic ANNs.[106,107] Research on optoelectronically integrated artificial networks is now in the early stage. Farhat et al.[108] proposed an optoelectronic neural network in 1985. The neural network was based on the Hopfield model. In the network, a light-emitting diode (LED) array was used to supply the input vector of light signals. An array of photodiodes was used to detect the output vector and convert the optical signals to electrical signals. The synaptic weight was controlled by the optical power which could be programmed through a shadow mask of metal, imaged mask, light valve, or raster-scanned laser beam.[109] The principle of a-Si:H based neural networks was similar. The a-Si:H synaptic devices were only used as the photoconductors to realize a matrix multiplier.
The optoelectronic ANNs proposed by Farhat et al. are far from current demands in terms of complexity, scalability, and energy-efficiency. Intriguing efforts have been recently focused on using synaptic devices to emulate the retina of the biological visual system. The retina is capable of optical sensing, data storage, and real-time processing. Novel optoelectronic synaptic devices with proper sensing-memory-processing functions are booming with diverse materials and device structures. In this context, Si is playing a critical role. For example, in the Pt/HfO2/Si device Si was not only the photoreceptor but also the intermediary agent modulating the device resistance. In the Si NCs-based devices, the dangling bonds or absorbed oxygen molecules at the Si-NC surface were utilized to realize the synaptic plasticity. In the W/n-MoS2/p-Si device, the native oxide on Si substrate supplied trap sites for both the photogenerated and electrically injected carriers.
Please note that working mechanism varied among these Si-based devices. The Pt/HfO2/Si device realized nonvolatile memory behavior by using the HfO2, which was one of the mature metal oxides used in oxygen-filament-based memristors. With the change of the voltage drop on the HfO2, the resistance of the device could be regulated. The nonvolatile memory with multiple analog states was desired for current ANN frameworks, e.g., hardware-based convolutional neural networks[110,111] and recurrent neural networks.[112] In such networks, the synaptic weights are directly programmed to the conductance of the synaptic devices. The conductance of the synaptic devices should be kept with a long retention time so that the synaptic weights could be maintained before the next update. With optoelectronic spikes the Pt/HfO2/Si device could also exhibit STDP-like characteristics, which were required for the implementation of the next-generation SNNs. Strictly speaking, the Pt/HfO2/Si device was only electrically stimulated with the assistance of light. Further work is need to enable the optical stimulation for the Pt/HfO2/Si device.
In the W/n-MoS2/p-Si and ITO/Si NCs/Al devices, volatile memory with the retention time of dozens of seconds was induced by the charge trapping/detrapping. The Si-NC-based electroluminescent devices also exhibited volatile memory with the retention time of milliseconds due to the charge trapping/detrapping. Although the retention time in these devices could be tuned by the parameters of the optical stimulation, the retention time significantly differed among the devices. We would like to mention that an optimal retention time is hard to define at this moment. Biologically, STP usually lasts from several milliseconds to minutes, while LTP can last for several hours or days. In an ANN, the required spatiotemporal dynamics may not be exactly the same as that of a biological neural network. The retention time for synaptic devices is intimately related to neuromorphic algorithms and architectures for a specific application scenario.[113] The working mechanism of charge trapping/detrapping has been widely used in optoelectronic synaptic devices although the role of traps has not been completely clarified nowadays. Traps are usually avoided in photodetectors and solar cells in order to maximize the photoelectric conversion efficiency. However, for synaptic devices, traps can help emulate the synaptic plasticity. We should note that traps may lead to device variation. However, it seems that an ANN may well tolerate the device variation.[114]
Different from the Si-NC-based two-terminal devices, the synaptic Si-NC phototransistors could emulate IPSC by taking advantage of photogating. It is usually hard to emulate the synaptic depression for optoelectronic synaptic devices. Electrical spikes are normally used for the emulation of synaptic depression.[32,115] Photogating induced negative photoconductance has been observed in several nanostructures such as carbon nanotubes,[116] ZnSe nanowires,[117] and InAs nanowires.[118] Inspired by the synaptic Si-NC phototransistors, employing these devices with negative photoconductance to realize the synaptic depression may be a promising approach. Researches on Si-NC-based synaptic devices have demonstrated the superiority of nanomaterials for the fabrication of synaptic devices. In the past decades, methods of tuning the surface, size, and doping have been developed for Si NCs, which rendered Si NCs diverse optical and electrical properties.[119] This facilitated the exploration of Si NCs for optoelectronic synapses. Please note that the optical and electrical properties of other Si nanostructures such as Si nanowires,[120] Si nanosheets,[121] and Si nanomembranes[122] also strongly depend on the surface/interfacial traps. Investigation on the use of these Si nanostructures for the fabrication of optoelectronic synaptic devices should be interesting as well.
Optoelectronic synaptic devices usually have two-terminal or three-terminal structures. When integrated into a network, the two-terminal synaptic devices form a crossbar array. The simplicity of such an array can ensure high-density integration. However, it is still challenging for the use of two-terminal synaptic devices in artificial vision systems. On one hand, it is difficult for the two-terminal devices to realize a nondestructive weight update because the electrical reading and writing operations are done by one shared terminal.[77,123] On the other hand, the coexistence of electro- and photo-active functionalities in a single material that is usually used for a two-terminal device rarely occurs.[38,72] On the contrary, the three-terminal synaptic devices (e.g., synaptic transistors) can incorporate augmentative functionalities by accommodating versatile control parameters.[124,125] Gate dielectric or photogating effect can be utilized to modulate the electronic properties, facilitating the combination of the physical properties of different assembled elements in an individual structure.[29,126] Moreover, multiple gates can be used for a synaptic transistor, which can emulate the spatial summation of postsynaptic current in a biological neural network.[3,127,128] Therefore, the three-terminal synaptic devices are more appealing for the hardware-based neuromorphic computing despite the relatively high circuit complexity.
Biological synapses operate at extremely low energy levels with the energy consumption of ∼ 20 fJ per event.[129] However, energy consumption of synaptic devices remains rather high up to now.[28] For an optoelectronic synaptic device, the energy consumption consists of the electrical and optical components. The electrical energy consumption can be calculated by using the integration of UI over time, where U and I are the voltage and current across the device, respectively. The optical energy is determined by the light source, which can be natural or artificial depending on the application scenarios. In fact, the optical energy may be also used to evaluate the optical sensitivity of the device. To minimize the energy consumption, the choice of materials and device structures for synaptic devices is critical.
Although the potential of Si-based optoelectronic synaptic devices for neuromorphic computing has been well witnessed, more efforts are definitely needed to devote to the development of highly intelligent and energy-efficient Si-based optoelectronic synaptic devices. In the meantime, the achievements of Si photonics[130] should be exploited to incorporate Si-based optoelectronic synaptic devices into ANNs.
We have presented advances in Si-based optoelectronic synaptic devices. Amorphous Si, bulk Si, and Si NCs have been used in optoelectronic synaptic devices. This should encourage the exploration of other Si materials for the fabrication of optoelectronic synaptic devices. Although the two-terminal and three-terminal device structures are popular for Si-based optoelectronic synaptic devices, novel device structures deserve investigation in the context of optoelectronic integration for ANNs. One of the direct applications of Si-based optoelectronic synaptic devices is for the simulation of a visual neural system. Further improvement of the synaptic functionalities and further reduction of the energy consumption for Si-based optoelectronic synaptic devices will greatly contribute to the realization of the visual neural system based on these devices.
[1] | |
[2] | |
[3] | |
[4] | |
[5] | |
[6] | |
[7] | |
[8] | |
[9] | |
[10] | |
[11] | |
[12] | |
[13] | |
[14] | |
[15] | |
[16] | |
[17] | |
[18] | |
[19] | |
[20] | |
[21] | |
[22] | |
[23] | |
[24] | |
[25] | |
[26] | |
[27] | |
[28] | |
[29] | |
[30] | |
[31] | |
[32] | |
[33] | |
[34] | |
[35] | |
[36] | |
[37] | |
[38] | |
[39] | |
[40] | |
[41] | |
[42] | |
[43] | |
[44] | |
[45] | |
[46] | |
[47] | |
[48] | |
[49] | |
[50] | |
[51] | |
[52] | |
[53] | |
[54] | |
[55] | |
[56] | |
[57] | |
[58] | |
[59] | |
[60] | |
[61] | |
[62] | |
[63] | |
[64] | |
[65] | |
[66] | |
[67] | |
[68] | |
[69] | |
[70] | |
[71] | |
[72] | |
[73] | |
[74] | |
[75] | |
[76] | |
[77] | |
[78] | |
[79] | |
[80] | |
[81] | |
[82] | |
[83] | |
[84] | |
[85] | |
[86] | |
[87] | |
[88] | |
[89] | |
[90] | |
[91] | |
[92] | |
[93] | |
[94] | |
[95] | |
[96] | |
[97] | |
[98] | |
[99] | |
[100] | |
[101] | |
[102] | |
[103] | |
[104] | |
[105] | |
[106] | |
[107] | |
[108] | |
[109] | |
[110] | |
[111] | |
[112] | |
[113] | |
[114] | |
[115] | |
[116] | |
[117] | |
[118] | |
[119] | |
[120] | |
[121] | |
[122] | |
[123] | |
[124] | |
[125] | |
[126] | |
[127] | |
[128] | |
[129] | |
[130] |