Artificial synapses and methods

Abstract

Controls the enhancement or reduction of memory using photoelectric conversion elements. [Solution] The disclosed photoresponsive artificial synapse comprises a photoelectric conversion element whose photoelectric power generation performance changes when it receives a light stimulus, the photoelectric conversion element holding the changed photoelectric power generation performance as a memory of the light stimulus, and a controller that selectively irradiates the photoelectric conversion element with one of a plurality of input lights of different wavelengths to change the photoelectric power generation performance, which is the memory of the light stimulus.

Description

[Technical Field]

[0001] This disclosure relates to artificial synapses and methods. [Background technology]

[0002] Patent Document 1 discloses an artificial neural network circuit comprising a crossbar circuit having memristors.

[0003] Patent document 2 discloses a heterojunction type photoelectric conversion element. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-57278 [Patent Document 2] Japanese Patent Publication No. 2023-154738 [Overview of the Initiative]

[0005] Artificial synapses are devices that artificially mimic biological synapses. They are attracting attention as core devices for neural-mimicking computing.

[0006] This disclosure relates to the application of a photoelectric conversion element whose photoelectric power generation performance changes when stimulated by light to a photoresponsive artificial synapse. Preferably, the artificial synapse that receives light stimulation can remember the light stimulation, and further preferably, the enhancement or reduction of that memory can be controlled.

[0007] Therefore, in one respect, the problem that this disclosure seeks to solve is to control the enhancement or reduction of memory by the photoelectric conversion element.

[0008] One aspect of this disclosure is a photoresponsive artificial synapse. The artificial synapse of the disclosure is a photoelectric conversion element whose photoelectric performance changes when it receives a light stimulus, and may include a photoelectric conversion element that holds the changed photoelectric performance as a memory of the light stimulus, and a controller that selectively irradiates the photoelectric conversion element with one of a plurality of input lights of different wavelengths to change the photoelectric performance, which is the memory of the light stimulus.

[0009] Another aspect of this disclosure is a method. The disclosed method may include a photoelectric converter element whose photoelectric performance changes when it receives a light stimulus, storing the changed photoelectric performance as a memory of the light stimulus, and selectively irradiating the photoelectric converter element with one of a plurality of input lights of different wavelengths to change the photoelectric performance, which is the memory of the light stimulus.

[0010] Further details will be described in the embodiments below. [Brief explanation of the drawing]

[0011] [Figure 1] Figure 1 is a diagram of the structure of an artificial synapse. [Figure 2] Figure 2 shows an example of artificial synapse fabrication. [Figure 3] Figure 3 is an explanatory diagram of the light soaking effect of a piezoelectric phototronic device. [Figure 4] Figure 4 is an explanatory diagram of the residual light soaking effect in a piezoelectric element. [Figure 5] Figure 5 is an explanatory diagram illustrating the difference in output signals with and without light soaking. [Figure 6] Figure 6 is an explanatory diagram of memory retention and forgetting in light soaking. [Figure 7] Figure 7 shows an example of an artificial synapse device using a piezoelectric element. [Figure 8] Figure 8 shows an example of using artificial synapses to identify numbers in an image. [Figure 9] Figure 9 shows the irradiation of light with different wavelengths. [Figure 10] FIG. 10 is a diagram showing changes in power generation performance (photocurrent) due to light of different wavelengths. [Figure 11] FIG. 11 is a diagram showing a hypothetical mechanism of selective signal enhancement / signal suppression by incident light of different wavelengths. [Figure 12] FIG. 12 is a diagram showing measurement data of STDP characteristics. [Figure 13] FIG. 13 is a diagram showing the results of an experiment demonstrating changes in output signals caused by applying stress to artificial synapses. [DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION]

[0012] <1. Overview of Artificial Synapses and Methods>

[0013] (1) The photoresponsive artificial synapse according to the embodiment is a photoelectric conversion element whose photovoltaic power generation performance changes upon receiving a light stimulus, and the photoelectric conversion element retains the changed photovoltaic power generation performance as a memory of the light stimulus, and a controller that selectively irradiates the photoelectric conversion element with any one of a plurality of input lights having different wavelengths to change the photovoltaic power generation performance that is the memory of the light stimulus. By selectively irradiating the photoelectric conversion element with any one of a plurality of input lights of different wavelengths, the photovoltaic power generation performance can be changed, and enhancement or reduction of the memory by the photoelectric conversion element can be controlled.

[0014] (2) The plurality of input lights having different wavelengths may include first wavelength light that enhances the photovoltaic power generation performance. Changing the photovoltaic power generation performance may include enhancing the photovoltaic power generation performance that is the memory of the light stimulus. In this case, the memory by the photoelectric conversion element can be enhanced.

[0015] (3) The plurality of input lights having different wavelengths may include second wavelength light that reduces the photovoltaic power generation performance. Changing the photovoltaic power generation performance may include reducing the photovoltaic power generation performance that is the memory of the light stimulus. In this case, the memory by the photoelectric conversion element can be reduced.

[0016] (4) The plurality of input light sources of different wavelengths may include a first wavelength light source that enhances the photoelectric performance and a second wavelength light source that reduces the photoelectric performance. Changing the photoelectric performance may include controlling the enhancement and reduction of the photoelectric performance, which is the memory of the photostimulus. In this case, the enhancement and reduction of the memory by the photoelectric conversion element can be controlled.

[0017] (5) The controller may be configured to adjust the characteristics of each of the plurality of input lights of different wavelengths. In this case, the memory can be further adjusted.

[0018] (6) The characteristics may include one or more characteristics selected from the group consisting of light intensity, light pulse period, light pulse width, light pulse interval, number of light pulses, and light pulse application timing.

[0019] (7) The artificial synapse according to the embodiment may be a self-powered type in which no power is supplied to the photoelectric conversion element in order to change the photoelectric performance which is the memory of the light stimulus.

[0020] (8) The artificial synapse according to the embodiment may further include a stress application unit for applying stress to the photoelectric conversion element. The photoelectric conversion element may be a piezoelectric element. In this case, memory can also be modulated by applying stress.

[0021] (9) The method according to the embodiment may include a photoelectric conversion element whose photoelectric performance changes when it receives a light stimulus, storing the photoelectric performance that has changed in response to the light stimulus as a memory of the light stimulus, and selectively irradiating the photoelectric conversion element with one of a plurality of input lights of different wavelengths to change the photoelectric performance which is the memory of the light stimulus.

[0022] (10) The plurality of input light sources of different wavelengths may include a first wavelength light source that enhances the photoelectric performance and a second wavelength light source that reduces the photoelectric performance. Changing the photoelectric performance may include controlling the enhancement and reduction of the photoelectric performance, which is a memory of the photoelectric stimulus.

[0023] <2. Examples of artificial synapses and methods>

[0024] The embodiments will be described in more detail below with reference to the drawings.

[0025] Figure 1 shows an example of an artificial synapse 10 according to an embodiment. The artificial synapse 10 includes a photoelectric conversion element 11. Light from a light source 12 is irradiated onto the photoelectric conversion element 11. The artificial synapse 10 also includes a controller 13. The controller 13 controls the light source 12, etc.

[0026] The artificial synapse 10 is a photoresponsive type equipped with a photoelectric conversion element 11 as an artificial synapse element. The artificial synapse element is an element that artificially mimics a biological synapse. The photoelectric conversion element 11 is a photoelectric power generation device, such as a photodiode or a solar cell. The photoelectric conversion element 11 may have a pn junction formed by joining a p-type material and an n-type material. The photoelectric conversion element 11 is used as a memory to store optical input. As a memory, the photoelectric conversion element 11 outputs a signal corresponding to the stored information. An artificial synapse utilizing the photoelectric conversion element 11 can be called a "self-powered artificial synapse" if it does not substantially require external power.

[0027] The artificial synapse 10 according to this embodiment is a photo-self-powered artificial synapse capable of selectively controlling signal enhancement / inhibition using incident light of different wavelengths. Further details on this point will be described later.

[0028] The photoelectric conversion element 11 may be a piezoelectric element. If either the p-type material or the n-type material of the photoelectric conversion element 11 is piezoelectric (or ferroelectric), the piezoelectric effect occurs and it functions as a piezoelectric element. A piezoelectric element is a heterojunction type photoelectric conversion element. For example, a piezoelectric element is a photodiode (piezophototronic diode) or solar cell (piezophototronic solar cell) that produces the piezoelectric effect.

[0029] The piezoelectric effect is a phenomenon in which stress, such as strain, applied to a piezoelectric semiconductor (or ferroelectric semiconductor) causes polarization in the piezoelectric semiconductor (or strong dielectric semiconductor), generating a polarization charge at the pn junction interface. This changes the band offset between the p-type and n-type semiconductors, resulting in changes to the photovoltaic power and other parameters. In a piezoelectric element, when stress such as mechanical strain is input, the photovoltaic power and other parameters change due to the piezoelectric effect.

[0030] In one embodiment, a piezoelectric element that produces a piezoelectric effect is used as an artificial synaptic element. A piezoelectric element is suitable as an artificial synaptic element because the memory stored in the piezoelectric element can be changed (modulated or erased (reset)) by stress applied to the piezoelectric element.

[0031] Furthermore, if the artificial synaptic element is a piezoelectric element made of a piezoelectric material, the artificial synapse can be said to be not only photoresponsive but also stress-responsive, meaning it responds to stress applied under light irradiation.

[0032] The photoelectric conversion element 11 may be, for example, a Se thin-film photoelectric power generation device (Se photoelectric conversion element). The Se photoelectric conversion element comprises Se (selenium) as a p-type semiconductor. Although the material used as the p-type semiconductor is not limited to Se, Se is suitable because it can generate photoelectric power even with relatively weak light such as indoor lighting, and has a wide range of applications as an artificial synaptic element. In addition, Se has a band gap of about 1.9 eV, and its theoretical photoelectric conversion efficiency is maximized (about 55%) in the wavelength range of fluorescent lamps or LED lighting used for indoor lighting, making it suitable for photoelectric power generation for indoor power supply. Furthermore, since Se can be formed at low temperatures, it can be formed on flexible plastic films such as PET films. For this reason, it is advantageous for applications in the fields of soft electronics and soft robotics.

[0033] In the following, p-type semiconductor materials will be indicated by adding "p-" before their name, and n-type semiconductor materials will be indicated by adding "n-" before their name.

[0034] A Se photoelectric conversion element may be formed by bonding an n-type piezoelectric material such as ZnMgO, ZnO, or BaTiO3 to p-Se. That is, a Se photoelectric conversion element as a piezoelectric element may be, for example, ZnMgO / Se, ZnO / Se, BaTiO3 / TiO2 / Se, or BaTiO3 / Se.

[0035] The photoelectric conversion element 11 may be a Cu(In,Ga)Se2 thin-film solar cell (Cu(In,Ga)Se2 photoelectric conversion element). The Cu(In,Ga)Se2 photoelectric conversion element comprises p-Cu(In,Ga)Se2. The Cu(In,Ga)Se2 photoelectric conversion element may also comprise an n-type piezoelectric material similar to that of the Se photoelectric conversion element. The Cu(In,Ga)Se2 photoelectric conversion element as a piezoelectric element may be, for example, ZnMgO / Cu(In,Ga)Se2, ZnO / Cu(In,Ga)Se2, BaTiO3 / TiO2 / Cu(In,Ga)Se2, or BaTiO3 / Cu(In,Ga)Se2.

[0036] The photoelectric conversion element 11 may be a perovskite solar cell (MAPbI3 or FAPbI3). Here, MA is methylammonium (CH3NH3NH3). 3+ ) and FA is formamidinium(HC(NH2) 2+ Perovskite solar cells can also be equipped with an n-type piezoelectric material similar to that of Se photoelectric elements. Examples of perovskite solar cells as piezoelectric elements are BaTiO3 / TiO2 / MAPbI3, BaTiO3 / MAPbI3, BaTiO3 / TiO2 / FAPbI3, or BaTiO3 / FAPbI3.

[0037] The photoelectric conversion element 11 may be a CdS / CdTe solar cell, which is a type of CdTe solar cell. Since Cds has piezoelectric properties, it is a piezophototronic element. The same effect can also be obtained with Cu2ZnSnS4 (CZTS), which replaces the rare metals In and Ga in the aforementioned Cu(In,Ga)Se2 photoelectric conversion element.

[0038] Each of the photoelectric conversion elements 11 described above may change its photoelectric power generation performance when it receives a light stimulus. The phenomenon in which photoelectric power generation performance changes when it receives a light stimulus is called the light soaking effect. The photoelectric conversion elements 11 of the embodiment store the received light stimulus as a change in photoelectric power generation performance. That is, the photoelectric conversion elements 11 retain the change in power generation performance due to the light soaking effect as a memory of the received light stimulus. The change in power generation performance may be, for example, an improvement or a decrease in power generation performance.

[0039] The light source 12 irradiates the photoelectric conversion element 11 with light that will be stored in the photoelectric conversion element 11 as a photostimulus. The light source 12 is, for example, an LED. The artificial synapse may be equipped with one or more light sources 12 that irradiate the photoelectric conversion element 11 with light. The light source 12 may also be an external light source for the artificial synapse 10.

[0040] The light from the light source 12 is used for writing to or reading from the photoelectric conversion element 11, which is a memory.

[0041] Furthermore, the light from the light source 12 is used to enhance or degrade the memory of the photoelectric conversion element 11. Multiple types of light with different wavelengths can be selectively irradiated onto the photoelectric conversion element 11 to enhance or degrade its memory. To irradiate with multiple types of light with different wavelengths, the artificial synapse 10 may be equipped with multiple light sources 12, each irradiating with light of a different wavelength. Alternatively, multiple types of light with different wavelengths can be irradiated by a single light source 12 whose emitted light wavelength is variable.

[0042] Furthermore, the multiple light sources 12 may each have separate light sources according to their respective uses, such as writing, reading, or enhancing or reducing memory, or a common light source may emit light for each of their uses.

[0043] Figure 2(A) shows a module of an artificial synapse 10 equipped with multiple photoelectric conversion elements 11. The artificial synapse 10 shown in Figure 2(A) is, as an example, equipped with a 4x4 two-dimensional array of photoelectric conversion elements 11.

[0044] Figure 2(B) shows the fabrication process of a ZnMgO / Se photoelectric conversion element 11, which is an example of such an element. As shown in Figure 2(B), first, an ITO electrode is formed on a glass substrate. Then, n-ZnMgO is formed on the ITO electrode by sputter deposition. Furthermore, Te / p-se is formed on the n-ZnMgO by vapor deposition. Finally, MoO3 / Au is formed on the Te / p-se by vapor deposition. Figure 2(C) shows the cross-sectional structure of the ZnMgO / Se artificial synapse photoelectric conversion element 11 fabricated in this manner. Light is irradiated onto this ZnMgO / Se artificial synapse from the upper glass substrate side.

[0045] Figure 3 shows the light-soaking effect of the photoelectric conversion element 11. Here, the light-soaking effect is explained using n-ZnMgO / p-Se as an example of the photoelectric conversion element 11. Note that ZnMgO is a piezoelectric semiconductor.

[0046] In Figure 3, (A) shows the current-voltage characteristics of n-ZnMgO / p-Se before light soaking (before light irradiation), and (B) shows the energy levels before light soaking. Also in Figure 3, (C) shows the current-voltage characteristics of n-ZnMgO / p-Se after light soaking (after light irradiation), and (D) shows the energy levels after light soaking.

[0047] Interfacial defects exist at the ZnMgO / Se junction interface. These interfacial defects are activated before light soaking. As shown in Figure 3(A), the photoelectric power generation performance is low before light soaking because the interfacial defects are activated. However, as shown in Figure 3(C), the photoelectric power generation performance improves after light soaking. This is because a specific metastable trap at the n-ZnMgO / p-Se heterointerface captures photocarriers due to light irradiation. Therefore, light soaking deactivates the interfacial defects and improves power generation performance. The metastable trap captures photocarriers for a certain period of time after light irradiation, and its carrier capture lifetime is relatively long. By utilizing this phenomenon, the degree of activation of the artificial synapse using the photoelectric conversion element 11 can be controlled depending on the conditions (illuminance, irradiation time, etc.) under which it is exposed to light irradiation. Therefore, an element capable of "memorizing" light input can be realized.

[0048] Figure 3 shows an example of a phenomenon in which the light-soaking effect occurs, and in some cases, the light-soaking effect can reduce photovoltaic performance. In other words, the light-soaking effect can improve or decrease photovoltaic performance. The light-soaking effect can also be observed as a change in the conductivity of the element.

[0049] The light-soaking effect persists even after light irradiation. In other words, the altered photoelectric power generation performance remains even after light irradiation. By utilizing the residual light-soaking effect, the photoelectric conversion element 11 can be used as a memory for storing light input. That is, the photoelectric conversion element 11 operates as a memory that stores the residual light-soaking effect (photoelectric power generation performance) as memory information during the period in which the light-soaking effect caused by light stimulation persists. In other words, the photoelectric conversion element 11 operates as a non-volatile memory that indicates the stored value (memory value) based on the power generation performance altered by light-soaking. Since the changing power generation performance takes a continuous value, the photoelectric conversion element 11, as a memory, operates as an analog memory.

[0050] The photoelectric conversion element 11, acting as memory, holds a value corresponding to the input light (history of light) as a memory value. During the period in which the light soaking effect persists, the photoelectric conversion element 11 can output (read out) an electrical signal (output signal) corresponding to the light soaking effect. In other words, the photoelectric conversion element 11 can output a memory value.

[0051] In the photoelectric conversion element 11, which functions as an artificial synaptic element, the photovoltaic power or photocurrent generated when light is irradiated is used as the weight (signal value) transmitted by the synapse.

[0052] Figure 4 shows an example of the light-soaking effect. Light is shone onto the element exhibiting the light-soaking effect (light-soaking) as shown in Figure 4(A), and then the element is left undisturbed in a dark environment. In this case, as shown in Figure 4(B), the open-circuit voltage V of the element... oc (V) is relatively low in the initial stages of light irradiation because the interface defects are in an active state. However, as light irradiation (light soaking) continues, the charge generated by the light irradiation is trapped in the interface defects, and the interface defects become inactive. As a result, the open-circuit voltage V increases with continued light irradiation. oc This improves (change in power generation performance). Then, when light irradiation is removed and the element is left in a dark environment, the open-circuit voltage V oc It does not decrease sharply and immediately, but rather decreases gradually, reaching a certain open-circuit voltage V oc This state persists for a period of time. In other words, the light soaking effect remains even in a dark environment, and the light input is stored in the element's memory.

[0053] As shown in Figures 4(C) to (E), the photocurrent (short-circuit current J per unit area) sc ), as well as the curve factor FF and photovoltaic efficiency Eff, the open-circuit voltage V oc It changes in a similar manner, indicating that changes in power generation performance and memory of the optical input are occurring.

[0054] Figure 5 shows the reading of the light irradiation history (memory value) stored in the element (photovoltaic device) by light soaking. In Figure 5, Figure 5(A) shows the case with light soaking (with light irradiation history), and Figure 5(B) shows the case without light soaking (without light irradiation history).

[0055] As shown in Figure 5(A), with light soaking, the absolute value of the photovoltaic device current gradually increases due to light soaking, improving power generation performance. When light soaking ends and the environment becomes dark, the current of the photovoltaic device becomes almost zero because there is no light irradiation. However, even without light irradiation, the improvement in power generation performance (light soaking effect) is maintained for a while. Therefore, when light irradiation for memory reading is performed while the improvement in power generation performance is maintained, the photovoltaic device, which is the memory, immediately outputs a large output signal (output current).

[0056] On the other hand, as shown in Figure 5(B), without light soaking, even if light is irradiated for readout, the photovoltaic device, having been in a dark environment prior to the irradiation, has (almost) no memory of the light irradiation history, and therefore outputs a small output signal or no output at all.

[0057] As is clear from the comparison between Figure 5(A) and Figure 5(B), even with the same intensity of light irradiation for readout, the magnitude of the output signal differs depending on whether or not there is light soaking (light irradiation history). Therefore, it can be seen that the photovoltaic device remembers the light input prior to the readout.

[0058] Figure 6 illustrates the memory and forgetting of light irradiation history. The memory of light irradiation history decreases over time. As shown in Figure 6(A), when the dark time after light soaking is short, the photovoltaic device, which acts as memory, stores the light irradiation (light soaking), and therefore outputs a relatively large output signal (output current) when light irradiation is performed for readout.

[0059] On the other hand, as shown in Figure 6(B), if the dark period after light soaking is long, the photovoltaic device, which acts as memory, forgets the light irradiation (light soaking). Forgetting means that the change in power generation performance returns to its original state. Because the photovoltaic device has forgotten its memory, when light irradiation is performed for readout, it outputs a small output signal, or no output at all.

[0060] The fact that artificial synapses temporarily remember and then forget light stimuli is similar to how light stimuli remain for a while in the retina of living organisms after being exposed to strong light.

[0061] By arranging elements that produce the light-soaking effect (artificial synapses) in a two-dimensional array, a photoelectric conversion memristor (artificial visual synapse device) for image recognition can be formed. Figure 7 shows the writing and reading to each element (each memory) when a shadow image of the number "2" is projected onto a photoelectric conversion element 11 (photoelectric conversion memristor) which is made up of 20 × 20 elements arranged in a two-dimensional array.

[0062] Here, as shown in Figure 7(A), a shadow of the number "2" is projected during writing. During writing, the part of the number "2" is in shadow and not illuminated by light, while the rest of the device is illuminated. Therefore, as shown in Figure 7(B), even though light is shone for writing, the power generation performance of the element (cell) corresponding to the part of the number "2" does not increase. However, as shown in Figure 7(B), the power generation performance of the elements (cells) other than the part of the number "2" increases (memory of the write-soaking history).

[0063] As shown in Figure 7(C), after writing, light irradiation for reading is applied to all cells. If the light irradiation for reading occurs immediately after or relatively shortly after write soaking, elements (cells) with a write soaking history will generate a larger photocurrent or photovoltaic power than elements (cells) without a write soaking history. As shown in Figure 7(C), if reading can be performed by light irradiation, no power supply for reading is required, which is preferable. Note that after a long period of time has elapsed since write soaking, the difference in output between elements with and without a write soaking history becomes small even if light irradiation for reading is performed.

[0064] The readout of the stored light soaking history (changes in power generation performance; memory information) is not limited to light irradiation; it may also be done by simultaneously applying a bias voltage to all elements (all cells). In this case as well, if readout is performed immediately after or relatively shortly after light soaking, a difference in output will occur depending on whether or not there is a light soaking history. When readout is performed by applying a bias voltage, light irradiation for readout is not required, making it suitable when it is desirable to avoid light irradiation for readout.

[0065] The applied bias voltage may be supplied by another photovoltaic device. In this case, the light irradiation for reading is applied to the other photovoltaic device, not the photovoltaic device (photoelectric memristor) on which the data was written. The other photovoltaic device generates the bias voltage for reading, and this generated bias voltage is applied to the photovoltaic device (photoelectric memristor) on which the data was written for reading. In this case, since the power for reading is obtained by light irradiation, an external power supply for reading is unnecessary. Furthermore, since the photovoltaic device on which the data was written is not irradiated with light for reading, this is suitable when it is desirable to avoid light irradiation for reading.

[0066] The stored write-soaking history may be read out by methods other than those described above. For example, the readout may be performed by reading the open-circuit voltage of the photovoltaic device.

[0067] As described above, the photoelectric conversion element 11 serving as an artificial synapse stores input light, and outputs an output signal corresponding to the stored content as a synapse transmission weight (signal value) within the period during which the stored content is maintained.

[0068] A two-dimensional array element (photoelectric conversion memristor) as shown in Figure 7 can be used as an artificial visual synapse element as an input layer of a neural network. Figure 8 shows an example of classifying numbers in an image by a neural network. The multiple (400) nodes X1 to X that constitute the input layer in Figure 8 400 are each a single photoresponsive artificial synapse element.

[0069] The neural network is, for example, machine-trained to output the probability (Y0 to Y9 [%]) that the number in the image is 0 to 9 from the image applied to the input layer. The machine-trained neural network can classify numbers from images. For example, when an image of the number "2" is input to the neural network, the neural network outputs a high value of Y2 [%]. Note that the input of the image of the number "2" here can be performed by reading information in Figure 7(C). The reading of information in Figure 7(C) is the reading of the information written in Figure 7(B). Further, in machine learning, the weights and the like of the neural network are adjusted so as to reduce the error of the output data obtained from the input data applied to the input layer, and the input data applied to the input layer is used during the adjustment. For this reason, the artificial synapse of the input layer is required to have a memory function. That is, it is important that the artificial synapse used in the neural network has a function of storing (memorizing) the light irradiation history.

[0070] Note that the memory and output of the artificial synapse may vary depending on the intensity of the light stimulus applied to the artificial synapse. For example, when the intensity of the light applied to the artificial synapse is low, the memory value becomes small. When the light intensity is medium, the memory value also becomes medium. When the light intensity is high, the memory value also becomes large.

[0071] Furthermore, the intensity of the memory (the degree of change in photovoltaic performance) can also change depending on the properties of the light. These properties of light include the intensity of the light (light pulse), as well as the frequency of the light pulse, the width of the light pulse (pulse input time), the interval between light pulses, the number of light pulses, the timing of light pulse application (the length of time from the writing timing), and the wavelength of the light (energy). The light irradiation history is stored in the artificial synapse according to these properties of light.

[0072] Artificial synapses can retain memory of light input for several seconds or more after receiving a strong light stimulus (long-term plasticity: LTP). The unit of retention can be minutes, hours, days, or years.

[0073] Artificial synapses can forget memories relatively quickly when exposed to weak light stimulation (short-term plasticity: STP). In the case of short-term plasticity, once the light pulse stops, the artificial synapses can return to their original conductivity in just a few seconds or less.

[0074] Whether an artificial synapse exhibits short-term or long-term plasticity can be determined by the light stimulation. For example, with light pulses that are very high in frequency but have a very small number of pulses, the artificial synapse may only exhibit short-term plasticity.

[0075] Furthermore, light irradiation may be performed multiple times. Repeated light irradiation can gradually increase conductivity. Multiple irradiations of weak light produce a weak signal (memory of small values), while multiple irradiations of strong light produce a strong signal (memory of large values).

[0076] As mentioned above, in the photoelectric conversion element 11, a specific metastable trap at the junction interface traps photocarriers for a certain period of time after light irradiation, and its carrier trapping lifetime is long. According to the inventors' experiments, it was found that the photoelectric power generation performance can be nonvolatilically enhanced with specific wavelengths of light, while the photoelectric power generation performance can be nonvolatilically reduced with other specific wavelengths.

[0077] Figures 9 and 10 show the changes in photovoltaic power generation performance due to input light of different wavelengths. Figure 9(A) shows the photoelectric conversion element 11 being irradiated with first wavelength light (first pulse light) with a wavelength of 365 nm (=3.4 eV), and Figure 9(B) shows the photoelectric conversion element 11 being irradiated with second wavelength light (second pulse light) with a wavelength of 455 nm (=2.7 eV).

[0078] As shown in Figure 9(C), the first wavelength light and the second wavelength light are optical pulses. In Figure 9(C), the period of the optical pulse is 50 ms. Four to five photocurrent measurements were taken during the input of one pulse, and four to five photocurrent data points were obtained for one pulse input, and their average was calculated.

[0079] Figure 10(A) shows the changes in photocurrent as responses to first-wavelength and second-wavelength light, respectively. In Figure 10(A), the vertical axis represents the photocurrent, and the horizontal axis represents the number of pulses and time input to the photoelectric conversion element 11. Figure 10(B) shows the ratio of the first photocurrent as a response to first-wavelength light to the second photocurrent as a response to second-wavelength light.

[0080] As shown in Figure 10(A), in the case of first-wavelength light, it can be seen that the photocurrent (first photocurrent) increases as the number of input pulses increases. On the other hand, in the case of second-wavelength light, it can be seen that the photocurrent (second photocurrent) decreases as the number of input pulses increases. Also, as shown in Figure 10(B), as the number of input pulses increases, the ratio of the first photocurrent to the second photocurrent increases.

[0081] Thus, selective signal enhancement / suppression is possible using incident light of different wavelengths. Therefore, by using light of two or more wavelengths, it is possible to realize a device that can "memorize" and "erase" optical input through self-generation (self-power supply). In other words, since the photoelectric power generation performance, which is the memory of the optical stimulus, can be changed by irradiation with light, there is no need to supply external power to the photoelectric conversion element, such as by applying voltage, in order to change the memory. Therefore, changes in photoelectric power generation performance (changes in memory) can be performed using a self-power supply system.

[0082] By distinguishing between different wavelengths of light, the photoelectric conversion element 11 becomes an optical memristor capable of both non-volatile enhancement and non-volatile suppression of signals. Specifically, for non-volatile enhancement, after stimulation with light (short-wavelength light, e.g., wavelength 365 nm), the photoelectric conversion element 11 generates a higher short-circuit current for subsequent light pulses compared to the initial light stimulation. This represents memory of having received a previous light pulse. Specifically, for non-volatile suppression, after stimulation with light (longer-wavelength light than the former, e.g., wavelength 455 nm), the photoelectric conversion element 11 generates a lower short-circuit current for subsequent light pulses compared to the initial light stimulation. This represents forgetting of having received a previous light pulse. Furthermore, the magnitude and temporal characteristics of enhancement (memory) and suppression (forgetting) can be adjusted depending on the characteristics of the incident light, such as frequency, intensity, pulse width, and energy, which modulate the response of the photoelectric conversion element 11. This photoelectric conversion element 11 can also function as a spiking synapse.

[0083] It is not necessary to perform both signal enhancement and signal suppression. For example, the degree of signal enhancement may be selectively controlled using light of two or more wavelengths with different degrees of signal enhancement, or the degree of signal reduction may be selectively controlled using light of two or more wavelengths with different degrees of signal suppression. Furthermore, signal enhancement / suppression may be controlled more precisely by using light of two or more wavelengths with different degrees of signal enhancement and light of two or more wavelengths with different degrees of signal suppression.

[0084] Figure 11 illustrates the hypothetical mechanism of selective signal enhancement / suppression by incident light of different wavelengths. Figure 11(A) shows the energy-band diagram in the dark, Figure 11(B) shows the energy-band diagram during synaptic enhancement (incident light with a wavelength of 365 nm), and Figure 11(C) shows the energy-band diagram during synaptic suppression (incident light with a wavelength of 455 nm).

[0085] As shown in the diagram, the band gap Eg of ZnMgO is 3.4 eV. Therefore, when light with energy greater than this (in this case, light with a wavelength of 365 nm) is incident, electrons in the valence band can be excited to the conduction band. Consequently, the photocurrent increases, and the signal is amplified. On the other hand, when light with energy less than the band gap Eg (in this case, light with a wavelength of 455 nm) is incident, electrons in the valence band cannot be excited, so relaxation becomes dominant over electron excitation. Consequently, the photocurrent decreases, and the signal is suppressed.

[0086] As shown in Figure 1, when signal enhancement is desired, the controller 13 selectively irradiates the photoelectric conversion element 11 with, for example, a first wavelength light of 365 nm from the light source 12. This results in non-volatile enhancement. Conversely, when signal suppression is desired, the controller 13 selectively irradiates the photoelectric conversion element 11 with, for example, a second wavelength light of 455 nm from the light source 12. This results in non-volatile suppression.

[0087] The controller 13 may adjust the characteristics of each of the multiple input light sources of different wavelengths. The characteristics may include, for example, one or more characteristics selected from the group consisting of light intensity, light pulse period, light pulse width, interval between light pulses, number of light pulses, and light pulse application timing. By adjusting these characteristics, it is preferable to adjust the degree to which the signal (photovoltaic performance) from light of each wavelength is enhanced or reduced.

[0088] Now, in order to more accurately mimic biological nerve cells than conventional formal neuron models that recognize information using matrices of analog numerical weights, spiking neural networks (SNNs), which process digital signals of pulses called spike trains, become important.

[0089] A spiking neural network (SNN) is a type of artificial neural network designed to mimic the function of biological neurons in the brain. Unlike conventional artificial neural networks that process information using continuous signals and static time steps, SNNs rely on discrete events called "spikes" or "pulses." A spike is a short, sharp burst of electrical activity emitted by a neuron when the input exceeds a certain threshold. This event-driven mechanism allows SNNs to process information dynamically and asynchronously, making them efficient and biologically realistic.

[0090] The defining characteristic of SNNs lies in their ability to replicate the energy-efficient and highly adaptive learning processes of the brain. In real biological nervous systems, neurons do not fire continuously, but only in response to significant stimuli. This reduction in resource usage allows the brain to perform complex calculations with minimal energy. SNNs are neural networks that aim to replicate this highly efficient system. This is particularly important for applications requiring low-power operation, such as neuromorphic computing.

[0091] One of the key functions of SNNs is their ability to learn and adapt based on spike timing. This is a mechanism called spike timing-dependent plasticity (STPD), a learning rule inspired by how biological synapses strengthen or weaken over time. If a spike from one neuron consistently occurs immediately before another neuron's spike, their connection is strengthened, reinforcing the idea that the two neurons are correlated. Conversely, if the first neuron's spike follows the second neuron's, their connection weakens, indicating a decrease in the importance of their relationship.

[0092] This timing-dependent adjustment of synaptic weights allows SNNs to capture temporal patterns within data, making them ideally suited for tasks involving time-sensitive information, such as speech recognition, motion detection, and sequential decision-making.

[0093] As in the embodiment, STDP can be realized by using input light of different wavelengths. In other words, the controller 13 can adjust the weight (signal value; photovoltaic performance) of the artificial synapse 10 according to the timing of the spike (pulse).

[0094] Figure 12 shows the experimental results of STDP using optical input. For STDP measurement, input pulse light for writing was shone onto the photoelectric conversion element 11 for a period of 0 ms to 500 ms. Subsequently, input pulse light for reading was applied with a pulse interval of 100 ms and an optical irradiation time (pulse width) of 20 ms. From 100 ms after writing was complete to 900 ms, the short-circuit current generated by photoelectric generation from each input pulse light was measured as the output signal of the artificial synapse. An LED was used as the light source.

[0095] The above procedure was performed at both 365 nm and 455 nm wavelengths.

[0096] In Figure 12, the horizontal axis represents time, and the vertical axis represents the difference in current (ΔCurrent). The difference in current is the difference between the current generated by a given input pulse and the current generated by the input pulse immediately preceding it. As shown in Figure 12, information about synaptic timing-dependent plasticity can be obtained from the change in timing between the write pulse and the read pulse.

[0097] Therefore, the controller 13 can implement STDP by adjusting the pulse timing.

[0098] Furthermore, similar to other memristors, the memory can be reset to its state before the optical pulse was input by applying a reverse bias voltage (for example, -1V) to the photoelectric conversion element 11, which acts as the memory. Therefore, the controller 13 can control the photoelectric conversion element 11 to apply a reverse bias and reset the memory.

[0099] Furthermore, the controller 13 can dynamically modulate or completely erase memories by applying stress, such as mechanical bending, by utilizing the piezoelectric effect of the photoelectric conversion element 11, which is a piezoelectric element. As shown in Figure 1, the artificial synapse 10 can be equipped with a stress application unit 14 that applies stress, such as mechanical bending, to the photoelectric conversion element 11. The controller 13 can control the stress application unit 14 to control the stress applied to the photoelectric conversion element 11.

[0100] The controller 13 can modulate (including erase) the memory by one of the following: light irradiation, reverse bias application, and stress application, or a combination of two or more of these. The magnitude of the effect on modulation or erasure can also be controlled by the magnitude of the applied bending or stress.

[0101] Figure 13 shows experimental results of changes in the memory of artificial synapses due to stress. In the experiment, artificial synapses using n-ZnMgO / p-Se elements as piezoelectric elements were used. In the experiment, the artificial synapses were first placed in a dark environment ("Initial (Dark)" in Figure 13). In the initial (dark) state, a bias voltage of 0.8V was applied to the piezoelectric element and the dark current (vertical axis in Figure 13) was read.

[0102] Then, multiple light pulses (HALS width 455 nm, pulse frequency 100 Hz) were irradiated onto the artificial synapse for approximately 5 seconds to perform light soaking (Figure 13(A), "Light Irradiation"). The light soaking effect alters the photovoltaic performance or conductivity. In the experiment, the increase in dark current due to the light soaking effect was measured. The dark current during light irradiation was measured by setting the bias voltage to 0.8 V at the timing when the pulsed light was turned OFF, as shown in Figure 13(B), and reading the diode current in the dark. Note that the application of the bias voltage itself does not cause a change in photovoltaic performance or conductivity. A lower bias voltage is preferable for reading, as a lower bias voltage can reduce power consumption.

[0103] As shown in Figure 13(A), the dark current gradually increases from 2.3 μA to 4 μA upon light irradiation. This means that the conductivity state of the artificial synapse changes upon light irradiation. In other words, the artificial synapse remembers the light irradiation.

[0104] As shown in Figure 13(A), even after light irradiation is stopped ("Memory (Dark)" in Figure 13(A)), a state of high dark current (high conductivity) is maintained for several seconds. In other words, the memory of light irradiation is maintained even after light irradiation. In memory (dark), as in the initial (dark) state, a bias voltage of 0.8V was applied to the piezoelectric element and the dark current was read. The period for which the memory is maintained varies depending on the light irradiation conditions and the material of the device, and can be several seconds, several minutes, several hours, or several days.

[0105] In the experiment, while memory was being maintained, stress was applied to an artificial synapse using a piezoelectric element by mechanically bending it significantly, as shown in Figure 13(C) ("bending deformation" in Figure 13(A)). As shown in Figure 13(A), the dark current decreases after the initial bending deformation. In other words, the conductivity change is reset by the application of stress (memory reset).

[0106] The application of stress is not limited to mechanical bending, but may also be through vibration or other external forces. For the application of stress, the artificial synapse according to this embodiment may be equipped with a stress application unit 14. The stress application unit may be, for example, a mechanism for mechanically bending the artificial synapse (bending drive device), a mechanism for pressing the artificial synapse (pressing device), or a mechanism for applying vibration to the artificial synapse (vibration device). However, the artificial synapse does not need to be equipped with a stress application unit, and may utilize external forces applied to the artificial synapse.

[0107] When stress is applied, the piezoelectric effect of a piezoelectric piezophototronic element induces a potential gradient across the pn junction. This potential gradient exerts a force on charge carriers confined at the interface, causing them to move away from defect levels. As electrons and holes move away from defect levels, recombination of these charge carriers or release of charge from trapping levels occurs, effectively reducing the power generation performance (conductivity) of the device.

[0108] If the stress applied to a piezoelectric element is sufficiently large, its power generation performance (conductivity) can be returned to its initial state (the state before light irradiation) (reset). Conversely, if the applied stress is relatively weak, it is possible to change the magnitude of the power generation performance (conductivity) rather than resetting it. Therefore, by applying stress, it becomes possible to modulate or erase (reset) the memory in a piezoelectric element as an artificial synaptic element. This process in an artificial synaptic element is similar to synaptic inhibition in biological systems, where the strength of the synaptic connection decreases and the synapse forgets or resets the stored information.

[0109] Finally, let me add a comment on the outlook from synapses (basic elements) to networks (finished products). The development of spiking neural networks begins with their fundamental building blocks, synapses. Se / ZnMgO heterojunctions function as artificial synapses, the basic elements in the network, and can mimic key features of biological synapses. These devices exhibit memory and learning behaviors, such as temporarily or permanently storing information and adapting responses based on external stimuli, such as light pulses. When exposed to a series of light pulses, the artificial synapses can modulate their electrical responses in the form of changes in photocurrent.

[0110] To transition from a single synapse to a network, multiple artificial synapses can be arranged in an array (matrix). This artificial synapse array forms the first layer of the network, where each synapse interacts with its neighbors, collectively processing input signals. In a visual system, this means organizing photodiodes into a grid, with each device capturing a small portion of the field of view. By combining responses from all devices, the array can process spatial and temporal information, such as recognizing shapes, edges, and motion patterns in a scene. This network is the input layer. Further additions (hidden layers, like those in a deep neural network) allow the system to evolve into a fully functional spiking neural network. The initial layer (ZnMgO / Se synapses) detects simple patterns like edges, while deeper layers identify more complex structures such as objects and motion trajectories. The network is trained using STDP or similar timing-based rules to adjust the strength of connections between synapses. This training allows the network to learn specific tasks, such as object recognition, detecting environmental changes, and tracking time series.

[0111] The finished product could take the form of an artificial vision system in which the Se / ZnMgO-based device functions as both an input layer sensor and a processing unit. Such a system would mimic the biological eye-brain collaboration mechanism, capturing visual stimuli, processing information in real time, and generating appropriate responses. It would be something like an artificial visual cortex. The main advantage of this approach is energy efficiency. Because the Se / ZnMgO heterojunction responds only to large amounts of light input and processes information asynchronously, the system's power consumption is expected to be minimal.

[0112] The present invention is not limited to the above embodiments, and various modifications are possible. [Explanation of symbols]

[0113] 10: Artificial synapses 11: Photoelectric conversion element 12:Light source 13: Controller 14: Stress application section

Claims

1. It is a light-responsive artificial synapse, A photoelectric conversion element whose photoelectric power generation performance changes when it receives a light stimulus, and which stores the changed photoelectric power generation performance as a memory of the light stimulus, A controller that selectively irradiates the photoelectric conversion element with one of several input lights of different wavelengths to change the photoelectric power generation performance, which is a memory of the light stimulation, An artificial synapse equipped with this feature.

2. The plurality of input light sources of different wavelengths include a first wavelength light source that enhances the photovoltaic performance. Changing the aforementioned photovoltaic performance includes enhancing the aforementioned photovoltaic performance, which is the memory of the aforementioned light stimulus. The artificial synapse according to claim 1.

3. The plurality of input light sources of different wavelengths include a second wavelength light that reduces the photovoltaic performance. Changing the aforementioned photovoltaic performance includes reducing the aforementioned photovoltaic performance, which is the memory of the aforementioned light stimulus. The artificial synapse according to claim 1.

4. The plurality of input light sources of different wavelengths include a first wavelength light source that enhances the photoelectric power generation performance and a second wavelength light source that reduces the photoelectric power generation performance. Changing the aforementioned photoelectric performance includes controlling the enhancement and reduction of the photoelectric performance, which is the memory of the photostimulation. The artificial synapse according to claim 1.

5. The controller is configured to adjust the characteristics of each of the multiple input lights having different wavelengths. The artificial synapse according to claim 1.

6. The aforementioned characteristics include one or more characteristics selected from the group consisting of light intensity, light pulse period, light pulse width, light pulse interval, number of light pulses, and light pulse application timing. The artificial synapse according to claim 5.

7. This is a self-powered type in which no power is supplied to the photoelectric conversion element in order to change the photoelectric power generation performance, which is the memory of the light stimulus. The artificial synapse according to claim 1.

8. The system further comprises a stress application unit for applying stress to the aforementioned photoelectric conversion element, The aforementioned photoelectric conversion element is a piezoelectric element. The artificial synapse according to claim 1.

9. A photoelectric conversion element whose photoelectric power generation performance changes when it receives a light stimulus is used to store the changed photoelectric power generation performance as a memory of the light stimulus. By selectively irradiating the photoelectric conversion element with one of several input lights of different wavelengths, the photoelectric power generation performance, which is the memory of the light stimulation, is changed. A method for preparing for something.

10. The plurality of input light sources of different wavelengths include a first wavelength light source that enhances the photoelectric power generation performance and a second wavelength light source that reduces the photoelectric power generation performance. Changing the aforementioned photoelectric performance includes controlling the enhancement and reduction of the photoelectric performance, which is the memory of the photostimulation. The method according to claim 9.

Images