Reservoir computation device

Abstract

To provide a reservoir computation device which offers high computation accuracy to provide many different reservoir states.SOLUTION: A reservoir computation device is provided, comprising a light source, a reservoir computation element having an active layer designed to interact with light from the light source, and light control means for controlling the light from the light source, the reservoir computation element comprising a light transmission portion to allow the light from the light source to reach at least a partial region of the active layer.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an apparatus for reservoir calculation. [Background technology]

[0002] In recent years, neural networks that mimic brain neural networks and artificial intelligence have been actively researched. For example, artificial intelligence techniques that utilize multi-layer neural networks including deep learning are known. This technology generally has high performance, but it has problems such as requiring a large amount of computational resources, high power consumption, and large-scale equipment, etc. For these reasons, there are high barriers to application to small mobile devices, especially those with limited power and volume (size).

[0003] Another approach to neural networks is called reservoir computing, which is disclosed in, for example, US Pat. Nos. 5,993,311, 5,945,211, 5,971,223, and 5,133,633. Reservoir calculations utilize the nonlinearity of physical phenomena and, in principle, can be performed using fewer resources. However, in the past, the nonlinearity (expressiveness) of the reservoir was low, and it was not possible to reduce the computational resources sufficiently to ensure the accuracy of the calculation. This meant that it was necessary to increase the size of the reservoir, making it difficult to utilize the characteristics that the reservoir was originally intended to achieve. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-204888 A [Patent Document 2] JP 2019-101635 A [Patent Document 3] Patent Publication No. 2021-60830 [Non-patent literature]

[0005] [Non-Patent Document 1] Edge-Of-Chaos Learning Achieved by Ion-Electron Coupled Dynamics in an Ion-Gating Reservoir, Daiki Nishioka, et al.,arXiv:2207.02573. [Non-Patent Document 2] Experimental Demonstration of High-Performance Physical Reservoir Computing with Nonlinear Interferred Spin Wave Multi-Detection, Wataru Namiki, et al.,arXiv:2207.03216 [Non-Patent Document 3] Reservoir Computing, Gohei Tanaka, Ryosuke Nakane, Akira Hirose (Morikita Publishing) pp91-92. Summary of the Invention [Problem to be solved by the invention]

[0006] SUMMARY OF THE DISCLOSURE The object of the present invention is to provide a reservoir calculation device which allows a large number of different reservoir states to be calculated with high accuracy. [Means for solving the problem]

[0007] The configuration of the present invention for solving the problems is shown below. (Configuration 1) A reservoir computing element having an active layer that interacts with light from the light source, and a dimming means for dimming the light from the light source, The reservoir computing element has a light transmitting portion through which light from the light source reaches at least a portion of the active layer. (Configuration 2) 2. The reservoir calculation device according to claim 1, wherein the light adjusting means is a switching means for turning the light source on and off. (Configuration 3) The reservoir calculation element includes an active layer made of a ferromagnetic material, a spin wave excitation electrode, and a detection electrode; 3. The reservoir calculation device according to claim 1 or 2, further comprising a magnetic field application means for applying a magnetic field to the reservoir calculation element. (Configuration 4) 4. The reservoir computing device of claim 3, wherein the ferromagnetic material is one or more selected from the group consisting of yttrium-iron-garnet (YIG), pure iron, nickel-iron (Permalloy), cobalt-iron-boron (CoFeB), and cobalt-manganese-iron-silicon (Heusler alloy). (Configuration 5) 5. The reservoir computing apparatus of claim 4, wherein the ferromagnetic material is yttrium iron garnet (YIG). (Configuration 6) The reservoir calculation device according to configuration 5, wherein the wavelength of the light source is 310 nm or more and 1100 nm or less. (Configuration 7) 7. The reservoir calculation device according to any one of configurations 3 to 6, wherein the distance between the spin wave excitation electrode and the detection electrode is 1 μm or more and 10 mm or less. (Configuration 8) 6. The reservoir calculation device according to configuration 5, wherein the distance between the spin wave excitation electrode and the detection electrode is 3 μm or more and 10 mm or less. (Configuration 9) The reservoir calculation device according to any one of configurations 3 to 8, wherein the spin wave excitation electrode and the detection electrode are in contact with an active layer made of the ferromagnetic material. (Configuration 10) The reservoir calculation device of any one of configurations 3 to 9, wherein the spin wave excitation electrode and the detection electrode include one or more elements selected from the group consisting of gold (Au), silver (Ag), palladium (Pd), platinum (Pt), copper (Cu), tungsten (W), titanium (Ti), tin (Sn), zinc (Zn), chromium (Cr), cobalt (Co), iron (Fe), and nickel (Ni). (Configuration 11) 11. The reservoir calculation device according to any one of configurations 3 to 10, wherein the magnetic field applying means is a permanent magnet. (Configuration 12) 11. The reservoir calculation device according to any one of configurations 3 to 10, wherein the magnetic field application means is an electromagnet. Effect of the Invention

[0008] The present invention provides a reservoir calculation apparatus that provides high accuracy of calculations that can yield many different reservoir states. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is an explanatory diagram showing the configuration of a reservoir calculation device according to a first embodiment. [Diagram 2] FIG. 11 is an explanatory diagram showing the configuration of a reservoir calculation device according to a second embodiment. [Diagram 3] FIG. 13 is an explanatory diagram showing the procedure of signal processing for reservoir calculation using spin waves. [Figure 4] FIG. 1 is a diagram illustrating a configuration of an information processing device. [Diagram 5] This is an optical microscope photograph of a prototype element (element for reservoir calculation using spin waves). [Figure 6] FIG. 13 is a characteristic diagram showing the output signal of a device for calculating a reservoir utilizing spin waves. [Figure 7] (a) is a characteristic diagram showing the signal difference when light irradiation is on and off, and (b) is a characteristic diagram in which the differential signal is converted into a normalized area and mapped to visualize the condition dependency. [Figure 8] This is a characteristic diagram showing the output signal when the prototype element is turned off (light off), when it is turned on (light on), and the difference between them. [Figure 9] FIG. 13 is a characteristic diagram showing the actual measurement and prediction when a self-regressive moving average task (NARMA) is performed on a prototype element when the light is off, and the deviation therebetween. [Figure 10] FIG. 13 is a characteristic diagram showing the actual measurement and prediction when a self-regressive moving average task (NARMA) is performed on a prototype element while it is turned on, and the deviation therebetween. [Figure 11]FIG. 13 is a characteristic diagram showing the actual measurement and prediction, and the deviation, when a self-regressive moving average task (NARMA) is performed with a mixture of on and off states, for a prototype element. [Figure 12] This is a characteristic comparison diagram showing the error index NMSE when NARMA was performed on a prototype element by comparing the on state, the off state, and a mixture of the on and off states. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the description, the expression A to B means A or more and B or less.

[0011] (Embodiment 1) In reservoir calculations, signals such as time series data are input to the reservoir section, and the multiple outputs obtained from it are used as reservoir states to perform calculations such as learning and prediction. In this case, the reservoir states are obtained as multiple time series data (nodes) that are different from each other, and the more of them there are, the more advantageous it is in terms of improving the accuracy of the calculations. Therefore, attempts are made to obtain more reservoir states, but there is a trade-off between the number of reservoir states and their dissimilarity, meaning that the reservoir states tend to be similar to each other when trying to obtain many reservoir states, and there is a limit to this.

[0012] The present invention makes it possible to easily obtain many different reservoir states by switching the light irradiation state on and off, thereby improving the accuracy of calculations.

[0013] The reservoir calculation device of the present invention comprises a light source, a reservoir calculation element having an active layer that interacts with the light of the light source, and a dimming means for dimming the light of the light source, and the reservoir calculation element has a structure having a light-transmitting portion through which the light from the light source reaches at least a part of the active layer. Here, examples of the active layer include a semiconductor layer of an electric double layer transistor described later, a ferromagnetic layer of a spin-wave utilizing magnetic element described in embodiment 2, a semiconductor layer of an oxidation-reduction transistor, and a semiconductor layer of a MOS transistor. Examples of the dimming means include a switching means for turning the light source on and off, a variable light amount means that can also adjust the light intensity, a variable pulse width dimming means that dims the light to a pulsed light whose pulse width can be adjusted, and a variable wavelength means that can also control the light wavelength.

[0014] In the following, in the first embodiment, the structure and features of an electric double layer transistor using an ion conductor and a semiconductor will be described as an example. Note that the function of an electric double layer transistor as a reservoir calculation element is disclosed in Non-Patent Document 1.

[0015] <Device Structure> As shown in Fig. 1, which shows the main configuration, the reservoir calculation device 1 of the present invention is an electric double-layer transistor (EDLT) that has a gate electrode 14, a source electrode 12, and a drain electrode 13, with an ion conductor 16 and a semiconductor layer 11 having a channel layer 15 as a core, and uses the drain current of the EDLT as a reservoir. By performing the reservoir calculation on hardware that allows for high integration, it becomes possible to reduce the calculation resources. Here, the source electrode 12 and the drain electrode 13 are in electrical contact with the semiconductor layer 11. In particular, it is preferable for them to be in ohmic contact.

[0016] The ion conductor 16 is made of a material that conducts ions, and may be in any of the following forms: solid, gel, or liquid. When the ion conductor 16 is solid, it has the advantages of being easy to integrate by using a semiconductor process, being easy to handle, and having no risk of liquid leakage. When it is gel, it has the advantages of being able to suppress the risk of liquid leakage and being relatively easy to handle. When it is liquid, it has the advantages of being easy to bring the ion conductor 16 and the semiconductor layer 11 into close contact with each other. The liquid may be organic or inorganic. The thickness of the ion conductor 16 is preferably 1 nm or more and 1 cm or less in order to obtain reservoir characteristics. A specific example of the ion conductor 16 is Li-SiO 2 -ZrO 2 (LZSO), PEO (polyethylene oxide) / LiClO 4 and 1-ethyl-3-methylimidazolium-bis(trifluoromethanesulfonyl)imide (EMI-TFSI). Among these, LZSO is preferred because it is relatively stable when in contact with a reducing electrode such as metallic Li.

[0017] The conductor ions of the ion conductor 16 are not particularly limited, but Li + Ion, H + Ion, Na + Ion, Ag + Ion, Cu + Ion, Cl - Ion, Br - Ion, I - Ion, SO 4 - Ion, FeCl 4 - ion, OH - Ion, BF 4 - Ion, PF 6 - Ion, CF 3 SO - ion, (FSO 2 ) 2 N - ion, (CF 3 SO 2 ) 2 N - Ion, C6 H 11 N 2 + Ion, C 10 H 20 F 6 N 2 O 5 S 2 + Among these, Li ions are widely used, have a large amount of accumulated know-how, and have the advantage of showing relatively high ionic conductivity among solid materials. + Ions (Li ions) can be preferably used.

[0018] The semiconductor layer 11 may be a hydrogen-terminated diamond semiconductor, silicon, GaAs, Ga 2 O 5 , SiC, GaP, InP, ZnSe, CdS, GaN, SiGe, CuInSe 2 , graphene, single-layer MoS 2 , or ZnO. Among these, hydrogen-terminated diamond semiconductors are preferred because they are resistant to degradation caused by ion implantation (chemical stability). Here, single crystal diamond is preferred. (111) diamond and (100) diamond are preferred. Furthermore, silicon can also be preferably used as the semiconductor layer 11 because it is widely used, has high quality, is low cost, and is suitable for integration.

[0019] The channel layer 15 is formed in a surface portion of the semiconductor layer 11. Specifically, examples of the channel layer 15 include a surface portion (hydrogen-terminated portion) of a hydrogen-terminated diamond semiconductor and a layer in which an impurity is doped into the surface portion of a semiconductor such as silicon. Here, the channel layer 15 may have a channel with one channel width or may have multiple channels with different channel widths. In the former case where one channel is arranged, it is necessary to convert the time series data into multiple different input pulse voltage signals and repeatedly input them to obtain multiple reservoir states, whereas in the case where channels with multiple channel widths are arranged, it is possible to simultaneously obtain multiple reservoir states from one input pulse voltage signal. A typical channel width is 100 μm or more and 800 μm or less. Although the channel layer 15 is in contact with the ion conductor 16 in FIG. 1, it is not necessarily in direct contact with the ion conductor 16. 2 A thin insulating film such as a thin insulating film may be formed on the channel layer 15. When a thin insulating film is formed, a chemical reaction between the channel layer 15 and the ion conductor 16 can be suppressed, and stability over time can be improved.

[0020] Materials for the gate electrode 14, source electrode 12, and drain electrode 13 can be any material that has low electrical resistance and does not react with the ion conductor 16, such as corrosion, and can include metals selected from the group consisting of aluminum (Al), tungsten (W), gold (Au), titanium (Ti), platinum (Pt), palladium (Pd), copper (Cu), silver (Ag), magnesium (Mg), zinc (Zn), iron (Fe), cobalt (Co), nickel (Ni), and molybdenum (Mo), alloys containing one or more metals selected from the group, doped polysilicon, carbon, graphene, and graphite. Also, a single layer film made of the above materials or a laminated film made of a combination of the above materials can be used. Although the gate electrode 14, the source electrode 12 and the drain electrode 13 are in contact with the ion conductor 16 in FIG. 1, they do not necessarily have to be in direct contact with each other. 2A thin insulating film such as a thin insulating film may be formed on the gate electrode 14, the source electrode 12, the drain electrode 13, and the ion conductor 16. When a thin insulating film is formed, chemical reactions between the gate electrode 14, the source electrode 12, and the drain electrode 13 and the ion conductor 16 can be suppressed, and stability over time can be improved.

[0021] A light irradiation device (light source) 18 is disposed on the upper surface on which the channel layer 15 is formed, and light 21 of a wavelength that interacts with carriers in the channel layer 15, which is the active layer, can be turned on (irradiated) and off in a controllable manner via a dimmer 19. The dimmer 19 turns on and off the light irradiation device 18 and adjusts the amount of light using power supplied from a power source 20. In some cases, the light emitted by the light irradiation device 18 is made into pulsed light, and the pulse width and frequency are adjusted as necessary.

[0022] The upper surface of the channel layer 15 of the reservoir computing device 1 is provided with a light-transmitting portion 17 so that the light 21 is irradiated onto the channel layer 15, which is a light-interactive active layer. Generally, since a metal with high light-shielding properties is often used for the gate electrode 14, it is preferable to provide an opening in the gate electrode 14. However, when the gate electrode material is a light-transmitting material such as ITO or FTO, it is not necessarily required to provide the opening 17 of the gate electrode 14 on the upper surface of the channel layer 15 of the gate electrode 14. Also, when polysilicon is used as the gate electrode material and infrared light is used as the light 21, it is not necessarily required to provide the opening 17 of the gate electrode 14 on the upper surface of the channel layer 15 of the gate electrode 14.

[0023] In the reservoir calculation device 1, highly nonlinear characteristics are obtained from the complex response of the drain current flowing through the semiconductor channel, which is caused by both the change in the electronic carrier density of the semiconductor induced by the electric double layer at the interface between the ion conductor and the semiconductor in response to the application of a pulse voltage, and the charge / discharge behavior of the electric double layer. In addition, photons generated by turning on / off or adjusting the brightness of the light irradiation device 18 interact with the carriers present in the channel layer 15, making it possible to easily obtain many different reservoir states, improving the calculation accuracy. The reservoir computing device 1 is suitable for high integration because it utilizes a phenomenon that occurs in a very thin region called the electric double layer.

[0024] <Manufacturing method> The reservoir calculation device 1 can be manufactured by the following steps. First, a semiconductor layer 11 is prepared, and a source electrode 12 and a drain electrode 13 are formed thereon. Methods for forming the source electrode 12 and the drain electrode 13 include a lift-off method, and a method in which a conductive material constituting the source electrode 12 and the drain electrode 13 is deposited by a sputtering method or a vapor deposition method, and then processed into electrodes by lithography and etching. Thereafter, the ion conductor 16 is formed by a CVD (Chemical Vapour Deposition) method, a pulsed laser deposition method, a sputtering method, a coating method, or the like. Thereafter, the gate electrode 14 is formed. Examples of methods for forming the gate electrode 14 include a lift-off method, and a method in which a conductive material constituting the gate electrode 14 is deposited by a sputtering method, a vapor deposition method, or the like, and then processed into an electrode by lithography and etching. Thereafter, a light irradiation device 18 using an LED or the like and a dimmer 19 are arranged, and power can be supplied from a power source 20, so that the reservoir calculation device 1 can be manufactured. The manufactured reservoir calculation device 1 has the functions essential for reservoir calculation, namely (a) nonlinear characteristics, (b) short-term memory, and (c) high dimensionality, and has high calculation accuracy.

[0025] (Embodiment 2) In the second embodiment, a nonlinear interference type spin wave reservoir calculation device (SWRC, reservoir calculation device) 2 having a function of adjusting light including on / off and irradiating the ferromagnetic surface will be described.

[0026] The main structure of the SWRC(2) is shown in Figure 2. The device core part (reservoir calculation element) of the SWRC (2) is composed of a ferromagnetic layer 31 which serves as an active layer, a detection electrode 32, and a spin wave excitation electrode 33. In addition to the device core portion, the SWRC (2) is equipped with a light irradiation device (light source) 18 and a magnetic field application means, a magnetic field generating device 34, and has a light transmitting portion 17 through which light 21 from the light irradiation device 18 reaches at least a portion of the ferromagnetic layer 31.

[0027] The ferromagnetic layer 31 may be a thin film or may be a substrate having sufficient rigidity to stand on its own. The ferromagnetic layer 31 can preferably be made of one or more selected from the group consisting of yttrium-iron-garnet (YIG), pure iron, nickel-iron (Permalloy), cobalt-iron-boron (CoFeB), and cobalt-manganese-iron-silicon (Heusler alloy). In particular, YIG is preferably used because it has a long attenuation length due to the band gap and is easy to handle. In order to stabilize the spin wave interference, it is preferable that the surface of the ferromagnetic layer 31 is flat and smooth.

[0028] The detection electrode 32 and the spin wave excitation electrode 33 may be arranged on and away from the ferromagnetic layer 31, but in order to improve sensitivity and stability, it is preferable that they are formed in contact with the ferromagnetic layer 31.

[0029] The spin wave excitation electrodes 33 are composed of two or more electrodes so as to cause spin wave interference. The distance between the electrodes of the spin wave excitation electrodes 33 is set to be equal to or greater than the coherence length, in other words, equal to or less than the attenuation length. Specifically, when a YIG (yttrium iron garnet) single crystal is used for the ferromagnetic layer 31, the distance between the electrodes of the spin wave excitation electrodes 33 is preferably 1 μm to 300 μm. A high-frequency signal generator is electrically connected to the spin wave excitation electrode 33 so as to send a predetermined pulse signal. The width, height and length of the spin wave excitation electrode 33 are not particularly limited, but may be 3 μm to 70 μm, 50 nm to 10 μm, and 3 μm to 1 mm, respectively.

[0030] The detection electrode 32 is made up of one or more electrodes and obtains an output signal. There are no particular limitations on the width, height, and length of the detection electrodes 32, but examples of such width, height, and length are 3 μm to 70 μm, 50 nm to 100 μm, and 3 μm to 1 mm, respectively. The positional relationship between the detection electrode 32 and the spin wave excitation electrode 33 is not particularly limited as long as the distance d between them is equal to or less than the attenuation distance, but it is preferable that they are arranged on the same plane in order to increase detection sensitivity and stability. The distance d between the detection electrode 32 and the spin wave excitation electrode 33 is preferably equal to or less than the attenuation distance, and is preferably equal to or less than 10 mm, more preferably equal to or less than 300 μm, and even more preferably equal to or less than 100 μm. From the viewpoints of detection sensitivity and stability, the lower limit of the distance between the detection electrode 32 and the spin wave excitation electrode 33 is preferably equal to or more than 1 μm, more preferably equal to or more than 2 μm, and even more preferably equal to or more than 3 μm. In order to stabilize the spin wave interference, it is preferable that the detection electrodes 32 are sandwiched between the spin wave excitation electrodes 33 .

[0031] The spin wave excitation electrode 33 and the detection electrode 32 may be made of a material containing one or more elements selected from the group consisting of gold (Au), silver (Ag), palladium (Pd), platinum (Pt), copper (Cu), tungsten (W), titanium (Ti), tin (Sn), zinc (Zn), chromium (Cr), cobalt (Co), iron (Fe), and nickel (Ni). Here, the spin wave excitation electrode 33 and the detection electrode 32 may be made of the same material or different materials. In order to stabilize the spin wave excitation and detection, it is preferable that the surfaces of the spin wave excitation electrode 33 and the detection electrode 32 are flat and smooth, and further, that the contact resistance with the detector terminal and the excitation source terminal is selected as small as possible. The film formation method of these electrodes can be sputtering, MOCVD, vapor deposition, plating, or the like.

[0032] A light irradiation device (light source) 18 is disposed on the upper surface of the ferromagnetic layer 31, and light 21 having a wavelength that interacts with carriers in the ferromagnetic layer 31, which is the active layer, can be turned on (irradiated) and off controllably via a dimmer 19 to the ferromagnetic layer 31. When YIG is used as the ferromagnetic layer 31, the wavelength of the light 21 is preferably 310 nm or more and 1100 nm or less. The wavelength band of the light 21 varies depending on the band gap of the ferromagnetic layer 31. The dimmer 19 turns on and off the light irradiation device 18 and adjusts the light intensity using power supplied from the power source 20. In some cases, the light emitted by the light irradiation device 18 is pulsed light, and the pulse width and frequency are adjusted as necessary.

[0033] The SWRC (2) has a magnetic field line generator 34 that generates magnetic field lines 35 that penetrate the ferromagnetic layer 31 in the thickness direction. The magnet may be an electromagnet or a permanent magnet. When an electromagnet is used, it has the advantage that the specified magnetic field conditions can be changed responsively without changing the position of the magnet, and when a permanent magnet is used, it has the advantage that the energy consumption for generating the magnetic field can be reduced, resulting in energy saving.

[0034] In SWRC(2), when a pulse wave is applied to the spin wave excitation electrode 33, spin wave interference is induced, and since this has nonlinear characteristics, reservoir calculations can be performed. Figure 3 shows an example of the procedure for performing reservoir calculations, and the reservoir state can be obtained from the spin wave waveform measured by the detection electrode 32. In this case, n reservoir states are obtained from each point (1 to n) of the spin wave waveform for a single input pulse, and a virtual node X 1 ~X n These can be used to perform reservoir calculations. Furthermore, by combining turning the light irradiation device 18 on, off, and dimming it, it contributes to diversifying the spin wave signal that detects the interaction between the irradiated light and the carriers, making it possible to easily obtain many different reservoir states and improving the calculation accuracy.

[0035] <Manufacturing method> The SWRC (2) can be manufactured by the following process. First, a ferromagnetic layer 31 is prepared, and then the spin wave excitation electrode 33 and the detection electrode 32 are formed thereon. Methods for forming the spin wave excitation electrode 33 and the detection electrode 32 include a lift-off method, and a method in which a conductive material constituting the spin wave excitation electrode 33 and the detection electrode 32 is deposited by a sputtering method, a vapor deposition method, or the like, and then processed into electrodes by lithography and etching. Thereafter, a light irradiation device 18 using an LED or a lamp, a dimmer 19, and power supply from a power source 20 are arranged, and a magnetic field generator 34 is arranged, and the SWRC (2) is manufactured. The manufactured reservoir calculation device 1 has the functions essential for reservoir calculation, namely (a) nonlinear characteristics, (b) short-term memory, and (c) high dimensionality, and has high calculation accuracy.

[0036] 4, an input signal from an input device 301 for processing input information may be processed by a reservoir calculation device 302 of the present invention (the reservoir calculation device according to any one of the above configurations 1 to 12), and the output signal may be transmitted to an output device 303, which processes and outputs the signal. The information processing device 3 requires few calculation resources, consumes little power, and is a small-sized information processing device. EXAMPLES

[0037] Example 1 In Example 1, a nonlinear interference type spin-wave reservoir calculation device (SWRC) with an on-off light irradiation function was fabricated and its characteristics were evaluated.

[0038] <Apparatus and its structure> The main structure of the SWRC(2) is shown in Figure 2, and an optical microscope photograph focusing on the electrode part of the prototype chip is shown in Figure 5. As the ferromagnetic layer 31, a (111) oriented single crystal YIG substrate (manufactured by MTI) with one side polished, formed by the floating zone method, was used. On the ferromagnetic layer 31, a detection electrode 32 and a spin wave excitation electrode 33 are formed in such a manner that the detection electrode 32 is sandwiched between the spin wave excitation electrodes 33. Here, the detection electrode 32 has a line width of 10 μm (signal line) to 20 μm (ground line), the spin wave excitation electrode 33 has a line width of 10 μm (signal line) to 20 μm (ground line), the distance d between the detection electrode 32 and the spin wave excitation electrode 33 is 30 μm, and the length is 800 μm. On the surfaces of the detection electrode 32 and the spin wave excitation electrode 33 that contact the ferromagnetic layer 31, a 10 nm thick titanium (Ti) layer is formed as an adhesion reinforcing layer, and on the Ti layer, a 90 nm thick gold (Au) layer is formed as an electrode material. The method for forming the Ti / Au electrodes is an electron beam deposition method.

[0039] A light irradiation device 18 is disposed on the upper surface on which the detection electrode 32 and the spin wave excitation electrode 33 are formed, and light 21 with wavelengths of 310 nm to 1100 nm, which interacts with the carriers of YIG, can be controllably turned on (irradiated) and off onto the YIG substrate 31 via a dimmer 19 that controls on / off. The light irradiation device 18 is an AL-320 (manufactured by Asahi Spectroscopy Co., Ltd.) with an output of 300 W. The magnetic field generating device 34, which consists of an electromagnet that generates magnetic field lines 35 that penetrate the ferromagnetic layer 31 along its thickness direction, i.e., along the (111) direction of the YIG single crystal, is a leakage magnetic field type uniaxial special shape electromagnet (manufactured by Toei Scientific Industry Co., Ltd.) and can generate a magnetic field of up to 500 mT.

[0040] <Characteristics evaluation> A high-frequency signal generator consisting of an RF probe manufactured by Toei Scientific Industries Co., Ltd. and an arbitrary waveform generator (AWG5202, manufactured by Tektronix) was used as the equipment for sending a pulse signal for evaluating characteristics to the spin wave excitation electrode 33. A mixed signal digital oscilloscope (MS068B, manufactured by Tektronix) was used as the device for monitoring the output from the detection electrode 32. Here, the input signal and the output signal were amplified to 30 dB and 38 dB, respectively, and 500 signals were acquired and integrated and averaged to improve the S / N ratio. In addition, an interval of 4 μs was provided before the next measurement to prevent the influence of the previous measurement from remaining. The measurements were carried out at room temperature (25°C).

[0041] First, we confirmed that SWRC(2) causes spin wave interference under the following conditions. With the light 21 from the light irradiation device 18 turned off, five magnetic fields ranging from 172 mT to 250 mT were applied. Here, the input signal for exciting the spin waves was a 2.5 V pulse signal. The results are shown in Figure 6. It can be seen that the signal waveform is caused by interference of spin waves, and that the interference changes depending on the strength of the magnetic field, causing the signal waveform to change. It should be noted that Non-Patent Document 2 discloses that the SWRC (2) functions as a reservoir computing element in the absence of light irradiation, i.e., in the off-light condition.

[0042] Next, measurements were taken with the light 21 from the light irradiation device 18 turned on and off, and the difference in signal waveform between when it was on and when it was off was plotted. An example is shown in FIG. 7(a). This figure shows the result of applying a magnetic field of 172 mT and integrating 500 measurements with an input pulse of 300 mV. This waveform was indexed as a normalized area (NA) according to the following formula (1), and measured under five levels of magnetic field (172, 180, 186, 200, and 250 mT) and seven levels of spin wave excitation input signal intensity (300 to 450 mV in 25 mV increments). The results are shown in the form of a map in FIG. 7(b). Here, NA is the area of ​​the graph of (waveform when on) - (waveform when off) 1 , the absolute value of the off waveform is squared and the value is A 2 When NA=A 1 / A 2 ...(Formula 1) is given by: The change in the light on / off output signal difference relative to the change in the spin wave excitation input signal intensity was largest when a magnetic field of 172 mT was applied, and was smallest when a magnetic field of 250 mT was applied. For reference, Figure 8 shows the waveforms of the output signal and its difference when the light is off and on, i.e., when the light is off and on, under two levels of magnetic field application conditions: 172 mT, which produces a large output change, and 250 mT, which produces a small output change.

[0043] Next, under conditions where a magnetic field of 172 mT, which has a large amount of output change, was applied, the NARMA TASK (nonlinear autoregressive moving average task), a benchmark test for reservoir evaluation, was solved using three methods: light off, light on, and light off & on mixed, and the three methods were compared. Here, for light off, 50 nodes were extracted from the output signal waveform with light off, for light on, 50 nodes were extracted from the output signal waveform with light on, and for light on & off mixed, 25 nodes were extracted from the waveform with light off and 25 nodes were extracted from the waveform with light on, for a total of 50 nodes. In addition, to give an overview, the NARMA TASK predicts near-future values ​​that have a non-linear relationship with past values, and is disclosed in, for example, Non-Patent Document 3. Figures 9, 10, and 11 show (a) actual measurements, (b) predictions, (c) the deviation between actual measurements and predictions in training, and (d) actual measurements, (e) predictions, and (f) the deviation between actual measurements and predictions in testing, measured or obtained with light off, light on, and light on & off mixed, respectively. As shown in FIGS. 11(c) and (f), it is clear that the light on & off mixed has smaller deviations and higher accuracy than the light off and light on mixed.

[0044] Next, the measurement accuracy (error) of the on & off mixed light was quantified using NMSE (normalized mean square error). The results are shown in Figure 12. By using mixed on and off light, the error was significantly reduced, for example by more than 78% compared to when the light was off. This demonstrates that the device and method of the present invention, which involve turning light irradiation on and off, can significantly improve measurement accuracy.

[0045] The NMSE is calculated by the following formula (2): where T, d(k), and yp(k) are the training phase (T=3500) or the test phase (T=500), the target signal at discrete time k, and the predicted signal at discrete time k.

[0046]

number

[0047] The present invention provides a reservoir calculation device capable of obtaining many different reservoir states with high calculation accuracy, and is therefore believed to be a great contribution to the development of the industry. [Explanation of symbols]

[0048] 1:Reservoir calculation device 2: Reservoir calculation device, SWRC 3: Reservoir calculation system 11: Semiconductor layer (hydrogen-terminated diamond) 12: Source electrode (source) 13: Drain electrode (Drain) 14: Gate electrode (gate) 15: Channel layer 16: Ionic conductor (LSZO) 17: Light transmission part (opening) 18: Light source (light irradiation device) 19: Dimmer (switch) 20: Power supply 21: Light 31: Ferromagnetic layer (YIG) 32: Detection electrode 33: Electrode for spin wave excitation 34: Magnetic field generator 35: Magnetic field lines 36: Electrical resistance measurement terminal (spare) 41: Detector terminal A 42: Detector terminal B 43: Excitation source terminal A 44: Excitation source terminal B 45: Detection electrode (spare) 301: Input device 302:Reservoir calculation device 303: Output device

Claims

1. a reservoir computing element having an active layer that interacts with light from the light source; and a dimming means for dimming the light from the light source; The reservoir computing element has a light-transmitting portion that allows light from the light source to reach at least a portion of the active layer.

2. 2. The reservoir calculation device of claim 1, wherein said dimming means is a switching means for turning said light source on and off.

3. The reservoir computing element includes an active layer made of a ferromagnetic material, a spin wave excitation electrode, and a detection electrode; 3. The reservoir calculation device according to claim 1, further comprising a magnetic field applying means for applying a magnetic field to said reservoir calculation element.

4. 4. The reservoir computing device of claim 3, wherein the ferromagnetic material is one or more selected from the group consisting of yttrium iron garnet (YIG), pure iron, nickel iron (Permalloy), cobalt iron boron (CoFeB), and cobalt manganese iron silicon (Heusler alloy).

5. 5. The reservoir computing device of claim 4, wherein the ferromagnetic material is yttrium iron garnet (YIG).

6. 6. The reservoir calculation device of claim 5, wherein the wavelength of the light source is between 310 nm and 1100 nm.

7. 4. The reservoir calculation device according to claim 3, wherein the distance between the spin wave excitation electrode and the detection electrode is 1 μm or more and 10 mm or less.

8. 4. The reservoir calculation device according to claim 3, wherein the distance between the spin wave excitation electrode and the detection electrode is 3 μm or more and 10 mm or less.

9. The reservoir computing device according to claim 3 , wherein the spin wave excitation electrode and the detection electrode are in contact with the active layer made of the ferromagnetic material.

10. 4. The reservoir computing device of claim 3, wherein the spin wave excitation electrode and the detection electrode contain one or more elements selected from the group consisting of gold (Au), silver (Ag), palladium (Pd), platinum (Pt), copper (Cu), tungsten (W), titanium (Ti), tin (Sn), zinc (Zn), chromium (Cr), cobalt (Co), iron (Fe), and nickel (Ni).

11. 4. The reservoir calculation device according to claim 3, wherein said magnetic field applying means is a permanent magnet.

12. 4. The reservoir calculation device according to claim 3, wherein said magnetic field applying means is an electromagnet.

Images