An on-chip neuron and a design method thereof, an information processing method, and a neural network

Abstract

The application relates to an on-chip neuron and a design method, an information processing method and a neural network thereof, and solves the problem that the integrated quantity and scale of an existing on-chip neuron are still limited by the bandwidth of electrical interconnection, pulse loss and communication delay; the on-chip neuron comprises an up-down type micro-ring structure and a waveguide type directional coupler; the up-down type micro-ring structure comprises a micro-ring waveguide, an up-loading waveguide and a down-loading waveguide, the micro-ring waveguide is provided with a first phase change film made of a phase change material; the number of the waveguide type directional couplers is at least two, and the coupling end of each waveguide type directional coupler is connected with the input end of the down-loading waveguide, and the coupling area of the waveguide type directional coupler is provided with a second phase change film made of a phase change material; and the application further provides a design method, an information processing method and a neural network based on the above neuron.

Description

Technical Field

[0001] This invention relates to an on-chip integrated optical structure that can simulate the function of a neuron, specifically to an on-chip neuron and its design method, information processing method, and neural network. Background Technology

[0002] Applications of artificial intelligence technology, such as facial recognition, machine translation, and autonomous driving, are changing people's daily lives. The realization of artificial intelligence for specific tasks requires computers to train neural networks with large amounts of data. However, the throughput of current computer architectures is very limited, resulting in low training efficiency for neural networks.

[0003] Inspired by brain structure, researchers have proposed a next-generation intelligent computing system that mimics synapses and neurons. This system encodes the information to be processed as spatiotemporal pulse information from presynaptic neurons. Postsynaptic neurons accumulate these pulses, and upon reaching a stimulation threshold, fire new neuronal pulses. This process forms a nonlinear spiking neural network through the combination of numerous neurons, and information processing is achieved using these spatiotemporally encoded neuronal pulses. Intel's TrueNorth chip has already integrated millions of neurons on a single chip. When processing specific artificial intelligence tasks, its energy efficiency is two orders of magnitude higher than traditional microelectronic chips, approaching the level of the human brain. However, the number and scale of these integrated neurons are still limited by electrical interconnect bandwidth, pulse loss, and communication latency. Summary of the Invention

[0004] The purpose of this invention is to address the limitations of existing on-chip neurons in terms of the number and scale of neurons that are still restricted by electrical interconnect bandwidth, pulse loss, and communication delay, and to provide an on-chip neuron and its design method, information processing method, and neural network.

[0005] The design concept of this invention is as follows:

[0006] Optical interconnects have the characteristics of high bandwidth, low loss, and low latency, which can solve the above-mentioned problems of electrical interconnects. It has been reported that neuromorphic photonic systems can achieve a processing speed that is 6-8 orders of magnitude higher than that of neuromorphic electronic systems.

[0007] Silicon-based optoelectronics is an optoelectronic integration technology that is compatible with mature microelectronics technology. This technology combines the ultra-large-scale logic and ultra-high precision manufacturing characteristics of CMOS technology with the advantages of ultra-high speed and ultra-low power consumption of photonics technology. It is an innovative technology that can solve the contradiction between technological evolution and cost. Neuromorphic devices based on silicon-based optoelectronics have higher pulse time resolution and lower communication delay and loss, and can realize large-scale distributed architecture of neural networks.

[0008] The nonlinear input-output relationship from presynapse to postsynapse is a crucial mechanism for information processing and brain-like intelligence in neuron models. Achieving nonlinear mapping on optoelectronic integrated chips requires nonlinear optical materials, such as saturable absorbers or phase-change materials. Phase-change materials possess threshold switching and asymptotic single-crystal / amorphous transition characteristics, which can be used to simultaneously simulate synaptic plasticity and neuronal integrated firing functions. Photonic neuromorphic devices based on these materials and optoelectronic technology hold promise for solving the limitations of traditional spiking neural networks in terms of scale, response delay, and pulse transmission energy consumption, thus overcoming bottlenecks in electronic systems across computational efficiency, memory, input / output, and energy consumption in high-performance computing fields such as artificial intelligence.

[0009] Synapses with malleable weighting characteristics are realized by a directional coupling structure in which the coupling region is covered with a phase change film. A photoinduced phase change in the phase change film is induced by a pump pulse signal, thereby adjusting the coupling coefficient and thus changing the output optical power of the directional coupling structure. The coupling outputs of multiple directional couplers are all connected to the input of the same straight waveguide, which allows the output power of multiple directional couplers to be superimposed.

[0010] Neurons with integrated firing function are simulated by an up-and-down microring structure with phase transition units.

[0011] The top straight waveguide of the top-down microring structure operates in the probe optical signal band, while the bottom arc waveguide operates in the pump pulse signal band. A phase change film is deposited on the microring waveguide. By accumulating the input pulses through the arc waveguide, the phase change film in the microring structure can undergo a phase change when the phase change threshold is reached. This changes the output amplitude of the probe optical signal at the output end of the top straight waveguide, thereby simulating the integrated firing function of a neuron.

[0012] By connecting the outputs of multiple directional couplers to the input of the arc-shaped waveguide in the up-down micro-ring structure, the input power of the directional coupler can be selected by controlling the state of the phase change film in the directional coupler.

[0013] To achieve the above concept, the technical solution adopted by this invention is as follows.

[0014] A type of on-sheet neuron, which is special in that:

[0015] This includes uplink / downlink microring structures and waveguide-type directional couplers;

[0016] The up-and-down microring structure includes a microring waveguide and an up carrier waveguide and a down carrier waveguide respectively disposed on the upper and lower sides of the microring waveguide. A first phase change film made of phase change material is disposed on the microring waveguide.

[0017] The number of waveguide-type directional couplers is at least two, and the coupling end of each waveguide-type directional coupler is connected to the input end of the download waveguide. The coupling region of the waveguide-type directional coupler is provided with a second phase change film made of phase change material.

[0018] The input terminal of the waveguide-type directional coupler is used to receive external pump pulse signals;

[0019] The input end of the upcarrier waveguide is used to receive the probe light signal, and the output end is used to output the probe light pulse signal. The through output port and isolation port of the waveguide-type directional coupler are left unused. The output end of the download waveguide is left unused. Since the upcarrier-downcarrier micro-ring structure is used to simulate the integrated firing function of the neuron, the pump pulse signal and the probe light signal band can be separated, reducing the transmission loss of the probe light signal. Based on this low-loss on-chip neuron, the large-scale expansion and cascading of on-chip neuromorphic devices can be realized, as well as the photonic neuromorphic function can be achieved.

[0020] Furthermore, the number of directional couplers is four.

[0021] Furthermore, the carrier waveguide is a carrier straight waveguide;

[0022] The download waveguide includes an arc-shaped waveguide and an input download straight waveguide and an output download straight waveguide extending along the input and output ends of the arc-shaped waveguide, respectively; the arc-shaped waveguide is located directly below the micro-ring waveguide;

[0023] The output terminals of all four waveguide-type directional couplers are connected to the input direct waveguide; the operating band of the input-output micro-ring structure can be adjusted by adjusting the central angle of the arc waveguide and the distance between the arc waveguide and the micro-ring waveguide.

[0024] Furthermore, the four waveguide-type directional couplers are arranged sequentially along the extension direction of the input end of the input straight waveguide. Based on the structure of the present invention, the length of the input straight waveguide can be adjusted according to the number of waveguide-type directional couplers.

[0025] Furthermore, the materials of the up-and-down micro-ring structure and the waveguide-type directional coupler are silicon, silicon oxide, silicon nitride, group III-V materials, or lithium niobate.

[0026] This invention also proposes a method for designing the above-mentioned on-chip neurons, which is characterized by including the following steps:

[0027] Step 1: Set the structural parameters of the up-and-download micro-ring structure, waveguide-type directional coupler, first phase change film and second phase change film, construct on-chip neuron, and set the peak power and pulse width of the pump pulse signal. The peak power and pulse width of the pump pulse signal must be able to reach the phase change threshold of the first phase change film.

[0028] The parameters include the waveguide widths of the upper carrier waveguide, lower carrier waveguide, and microring waveguide, the radius of the microring waveguide, and the waveguide spacing of the coupling region in the upper and lower carrier waveguide microring structure; the waveguide width, coupling length, and waveguide spacing of the coupling region in the waveguide-type directional coupler; and the width and length of the first phase change film and the second phase change film.

[0029] Step 2: Input the probe light signal to the input end of the upper carrier guide; input the pump pulse signal to the input end of each waveguide-type directional coupler, and output the probe light pulse signal to the output end of the upper carrier guide. If the probe light pulse signal meets the requirements, the design of the on-chip neuron is completed; otherwise, proceed to Step 3.

[0030] Step 3: Pause the input of the probe light signal and the pump pulse signal, and adjust at least one of the parameters mentioned in Step 1, and then proceed to Step 4;

[0031] Step 4: Input a probe light signal to the input end of the upper carrier guide; input a pump pulse signal to the input end of each waveguide-type directional coupler, and output a probe light pulse signal at the output end of the upper carrier guide; if the probe light pulse signal meets the requirements, the fabrication of the on-chip neuron is completed; otherwise, return to step 3 until the probe light pulse signal meets the requirements.

[0032] Further, in step 1, the download waveguide includes an arc-shaped waveguide and an input download straight waveguide and an output download straight waveguide extending along the input and output ends of the arc-shaped waveguide, respectively; the arc-shaped waveguide is located directly below the micro-ring waveguide;

[0033] Among the parameters, the waveguide width of the download waveguide includes the waveguide width of the arc waveguide, the input download straight waveguide, and the output download straight waveguide. The parameters also include the central angle and radius of curvature of the arc waveguide. By adjusting the central angle and radius of curvature of the arc waveguide, the working band can be adjusted.

[0034] This invention also proposes an information processing method for the above-mentioned on-chip neurons, which is characterized by including the following steps:

[0035] Step 1: Input the probe light signal to the input terminal of the upper carrier guide; input the pump pulse signal to the input terminal of each waveguide-type directional coupler;

[0036] If the second phase change film undergoes a phase change, the pump pulse signal is not output; if the second phase change film does not undergo a phase change, the pump pulse signal is output from the coupling end of the waveguide-type directional coupler and coupled to the input end of the download waveguide.

[0037] Step 2: The pump pulse signal coupled to the download waveguide is coupled to the micro-loop waveguide;

[0038] If the first phase change film undergoes a phase change, the output terminal of the upper carrier guide will output a probe light pulse signal; if the first phase change film does not undergo a phase change, the output terminal of the upper carrier guide will not output a probe light signal.

[0039] This invention also proposes a neural network based on optoelectronic integration, which is unique in that:

[0040] It includes multiple interconnected on-chip neurons;

[0041] The on-chip neurons are those proposed above.

[0042] The input of the waveguide-type directional coupler of the first on-chip neuron is used to receive external pump pulse signals;

[0043] The input terminal of the upcarrier conduction of each on-chip neuron is used to receive the probe light signal, and the output terminal is used to output the probe light pulse signal;

[0044] The upcarrier output of each on-chip neuron is connected to the input of a waveguide-type directional coupler in another neuron.

[0045] The beneficial effects of this invention are:

[0046] 1. In this invention, by using an on-chip micro-ring structure to simulate the integrated firing function of a neuron, the pump pulse signal and the probe light signal band can be separated, reducing the transmission loss of the probe light signal. Based on this low-loss on-chip neuron, large-scale expansion and cascading of on-chip neuromorphic devices can be realized, as well as photonic neuromorphic function can be achieved.

[0047] On-chip neurons based on phase change materials and optoelectronic technology solve the problems of scale limitation, response delay and pulse transmission energy consumption of existing spiking neural networks.

[0048] 2. In this invention, the device manufacturing process based on this structure is compatible with the traditional complementary metal-oxide-semiconductor process for very large-scale integrated circuits, requiring no special processes, which is beneficial for large-scale device manufacturing and cost reduction.

[0049] 3. In this invention, based on the non-volatile phase transition characteristics of phase change materials, the plastic weight of synapses and the integrated firing function of neurons can be simulated, and a simple recognition function can be realized.

[0050] 4. The on-chip neuron based on optoelectronic integration proposed in this invention can operate in different wavebands by changing the shape, width, and distance of the download waveguide located directly below the micro-ring waveguide.

[0051] 5. In this invention, the on-chip neuron with the above-described structure solves the problem that the number and scale of neurons integrated on existing chips are still limited by electrical interconnect bandwidth, pulse loss and communication delay.

[0052] 6. In this invention, initial parameters are first set to construct on-chip neurons, and then the parameters of the on-chip neurons are adjusted according to the output probe light pulse signal until the required probe light pulse signal is obtained. The probe light pulse signal obtained by this method is more accurate. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall structure of an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram illustrating the principle structure of the download-type micro-ring structure in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram illustrating the principle structure of the waveguide-type directional coupler coupled to the download waveguide in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the embodiment of the present invention with the weights of the four waveguide directional coupler branches set to "1010", where "1" indicates that W1 and W3 are in a high transmission state, and "0" indicates that W2 and W4 are in a low transmission state.

[0057] Figure 5 This is a histogram showing the strength of the output probe optical pulse signal under different weight states of the four waveguide directional coupler branches in an embodiment of the present invention. The horizontal axis represents the transmission state of each branch, and the vertical axis represents the strength of the probe optical pulse signal. "1" indicates a high transmission state, and "0" indicates a low transmission state.

[0058] In the figure, 1. Up-carrier waveguide; 11. Micro-ring waveguide; 12. Up-carrier waveguide; 13. Down-carrier waveguide; 131. Arc waveguide; 132. Input down-carrier straight waveguide; 133. Output down-carrier straight waveguide; 2. Waveguide-type directional coupler; 3. First phase change film; 4. Second phase change film;

[0059] A1, Input terminal of the arc waveguide; A2, Output terminal of the arc waveguide; A3, Input terminal of the upcarrier waveguide; A4, Output terminal of the upcarrier waveguide; A5, Input terminal of the waveguide-type directional coupler; A6, Coupler terminal of the waveguide-type directional coupler. Detailed Implementation

[0060] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0061] The "up" and "down" mentioned in the text refer to... Figure 1 Described from the perspective of [the author / organization].

[0062] This invention proposes an on-chip neuron, such as Figure 1 As shown, it includes an up-and-down microring structure 1 and four waveguide-type directional couplers 2;

[0063] The uploading / downloading microring structure 1 includes a microring waveguide 11 and an uploading waveguide 12 and a downloading waveguide 13 respectively disposed on both sides of the microring waveguide 11; the uploading waveguide 12 is an uploading straight waveguide; the downloading waveguide 13 includes an arc-shaped waveguide 131 disposed below the microring waveguide 11 and an input downloading straight waveguide 132 and an output downloading straight waveguide 133 extending along the input and output ends of the arc-shaped waveguide 131 respectively; the arc-shaped waveguide 131 is located directly below the microring waveguide 11, and a first phase change film 3 made of phase change material is disposed on the microring waveguide 11; the advantage of using the arc-shaped waveguide 131 is that the operating band of the uploading / downloading microring structure 1 can be adjusted by adjusting the central angle of the arc-shaped waveguide 131 and the distance between the arc-shaped waveguide 131 and the microring waveguide 11. The larger the central angle, the shorter the operating wavelength, because the shorter the wavelength, the more the coupling length needs to be increased.

[0064] Four waveguide-type directional couplers 2 are arranged sequentially along the input straight waveguide 132, and the coupling ends of all four waveguide-type directional couplers 2 are connected to the input straight waveguide 132, thereby enabling the coupling ends A6 of the four waveguide-type directional couplers to be connected to the input end A1 of the arc-shaped waveguide. Each waveguide-type directional coupler 2 has a second phase change film 4 made of phase change material in its coupling region; for example... Figure 2 As shown, the waveguide-type directional coupler 2 is used to simulate a variable weighted synaptic structure, and the input pulse optical power at the input end A1 of the arc waveguide is controlled by the state of the second phase change film 4 set on it.

[0065] Definition: Phase change materials are materials that have multiple stable states and can non-volatilely switch between these stable states.

[0066] like Figure 2 As shown, under the action of the upper and lower micro-ring structure 1, the waveguide-type directional coupler 2, and the first phase change film 3 and the second phase change film 4 made of phase change material, the pump pulse signal is input at the input end A1 of the arc waveguide. The pump pulse signal is coupled to the micro-ring waveguide 11. Based on the evanescent wave coupled to the first phase change film 3 on the micro-ring waveguide 11, when the power absorbed by the first phase change film 3 exceeds the phase change threshold of the phase change material, the first phase change film 3 will undergo a phase change.

[0067] The input terminal A3 of the upper carrier waveguide receives the probe light signal, and the wavelength of the probe light signal is the resonant wavelength of the micro-ring. The probe light signal is coupled into the micro-ring waveguide 11. When the first phase change film 3 on the micro-ring waveguide 11 does not undergo a phase change, the probe light signal is at the resonant wavelength of the micro-ring, and the output probe light signal power is low. When the first phase change film 3 on the micro-ring waveguide 11 undergoes a phase change due to the pump pulse signal input from the input terminal A1 of the arc waveguide, the resonant state of the micro-ring changes. That is, when the wavelength of the input probe light signal no longer resonates, the power of the probe light signal output from the output terminal A4 of the upper carrier waveguide suddenly increases, thus generating an output pulse.

[0068] The materials of the uplink / downlink microring structure 1 and the waveguide-type directional coupler 2 are silicon, silicon oxide, silicon nitride, group III-V materials, or lithium niobate.

[0069] like Figure 3 The diagram shows the working principle of the waveguide-type directional coupler 2 (presynaptic neuron). Specifically, the pump pulse signal is input through the input terminal A5 of the waveguide-type directional coupler. When the power of the input pump pulse signal does not reach the phase transition threshold of the second phase transition film 4 on the waveguide-type directional coupler 2, the pump pulse signal input through the input terminal A5 can be coupled to the output terminal A6 of the waveguide-type directional coupler, i.e., input to the input terminal A1 of the arc-shaped waveguide. When the power of the input pump pulse signal reaches the phase transition threshold of the second phase transition film 4 on the waveguide-type directional coupler 2, it can cause a phase transition in the second phase transition film 4 covering the coupling region. After the second phase transition film 4 undergoes a phase transition, the pump pulse signal input through the input terminal A5 of the waveguide-type directional coupler 2 will not be able to be output from the output terminal A6 of the waveguide-type directional coupler, and therefore cannot be input to the input terminal A1 of the arc-shaped waveguide.

[0070] The method for preparing on-chip neurons in this invention is as follows: Figure 2 As shown, both the top-down microring structure 1 and the waveguide-type directional coupler 2 can be fabricated using semiconductor processes. The device structure is defined on the substrate using photolithography or electron beam lithography, and the substrate is etched using inductively coupled plasma etching to finally realize the corresponding structure on the substrate. Then, a first phase change film 3 and a second phase change film 4 made of phase change material of a specified shape can be deposited on the coupling region of the microring waveguide 11 and the waveguide-type directional coupler 2 using physical vapor deposition or chemical vapor deposition and photoresist stripping, respectively. This yields the top-down microring structure 1 with the first phase change film 3 and the waveguide-type directional coupler 2 with the second phase change film 4.

[0071] Based on the fabrication of the aforementioned devices, a supervised approach can be used to simulate the basic functions of neurons. This involves changing the weights of each waveguide-type directional coupler branch through external training. Taking a four-branch structure as an example, W1, W2, W3, and W4 represent the weights of the transmission states of the four waveguide-type directional couplers. Figure 4 and Figure 5 As shown, the weights of the four branches are set as "1010", "0010", "0100", "0110", "1000", and "1100", where "0" represents a low transmission state and "1" represents a high transmission state. "1010" indicates that W1 and W3 are in a high transmission state, meaning that the second phase change film 4 of these two branches has not undergone a phase change. The pump pulse signal input from the input terminal A5 of the waveguide directional coupler can be efficiently output from the output terminal A6 of the waveguide directional coupler. W2 and W4 are both... In the low transmission state, that is, the second phase change films 4 of the two paths undergo a phase transition, and the pump pulse signal input from the input terminal A5 of the waveguide directional coupler 2 cannot be output from the output terminal A6 of the waveguide directional coupler 2. This is because the second phase change films 4 of paths W2 and W4 undergo a phase transition under the action of the pump pulse signal, causing the pump pulse signals of these two paths to be unable to couple to the input direct waveguide 132 of the upper and lower micro-ring structure 1 through the waveguide directional coupler 2. When there is a pump pulse signal input to paths W1 and W3, such as Figure 4 As shown, through the waveguide-type directional coupler 2, the pump pulse signal can be coupled to the input download straight waveguide 132 of the up-down micro-ring structure 1, and then coupled into the micro-ring waveguide 11, causing the first phase change film 3 on the micro-ring waveguide 11 to undergo a phase change, thereby realizing the drift of the micro-ring resonance curve, causing a sudden change in the output power of the probe light signal input from the input terminal A3 of the up-carrier waveguide, and generating an output pulse at the output terminal A4 of the up-carrier waveguide; when the pump pulse signals W2 and W4 are input, under the premise of low output state, the pump pulse signal cannot be coupled to the input download straight waveguide 132 of the up-down micro-ring structure 1, and thus cannot cause the output terminal A4 of the up-carrier waveguide to output a pulse. Therefore, the probe light pulse signal can only be generated when the input pulse is "1010". The above process is the process of the on-chip neuron recognizing "1010" set in this invention.

[0072] The process of recognizing “0110” and “1100” is the same as the process of recognizing “1010”.

[0073] This invention also proposes a design method for the above-mentioned on-chip neurons, comprising the following steps:

[0074] Step 1: Set the structural parameters of the up-and-download micro-ring structure 1, waveguide-type directional coupler 2, first phase change film 3 and second phase change film 4, construct on-chip neurons, and set the peak power and pulse width of the pump pulse signal. The peak power and pulse width of the pump pulse signal must be able to reach the phase change threshold of the first phase change film 3.

[0075] The parameters include the waveguide widths of the upper carrier waveguide 12, the lower carrier waveguide 13, and the micro-ring waveguide 11, the radius of the micro-ring waveguide 11, and the waveguide spacing of the coupling region in the upper and lower carrier waveguide 1; the waveguide width, coupling length, and waveguide spacing of the coupling region in the waveguide-type directional coupler 2; and the widths and lengths of the first phase change film 3 and the second phase change film 4.

[0076] The download waveguide 13 includes an arc-shaped waveguide 131 and an input download straight waveguide 132 and an output download straight waveguide 133 extending along the input and output ends of the arc-shaped waveguide 131, respectively; the arc-shaped waveguide 131 is located directly below the micro-ring waveguide 11;

[0077] Therefore, in the parameters, the waveguide width of the download waveguide 13 includes the waveguide width of the arc waveguide 131, the input download straight waveguide 132, and the output download straight waveguide 133. The parameters also include the central angle and radius of curvature of the arc waveguide 131. By adjusting the central angle and radius of curvature of the arc waveguide 131, the working band can be adjusted.

[0078] Step 2: Input a probe light signal to the input terminal of the upper carrier guide 12; input a pump pulse signal to the input terminal of each waveguide-type directional coupler 2, and output a probe light pulse signal to the output terminal of the upper carrier guide 12. If the probe light pulse signal meets the requirements, the design of the on-chip neuron is completed; otherwise, proceed to step 3.

[0079] Step 3: Pause the input of the probe light signal and the pump pulse signal, adjust at least one of the parameters in Step 1, and proceed to Step 4;

[0080] Step 4: Input a probe light signal to the input end of the upper carrier guide 12; input a pump pulse signal to the input end of each waveguide-type directional coupler 2, and output a probe light pulse signal at the output end of the upper carrier guide 12; if the probe light pulse signal meets the requirements, the fabrication of the on-chip neuron is completed; otherwise, return to step 3 until the probe light pulse signal meets the requirements.

[0081] The present invention also proposes the above-mentioned information processing method for on-slice neurons, comprising the following steps:

[0082] Step 1: Input the probe light signal to the input terminal of the upper carrier guide 12; input the pump pulse signal to the input terminal of each waveguide-type directional coupler 2;

[0083] If the second phase change film 4 undergoes a phase change, the pump pulse signal will not be output. If the second phase change film 4 does not undergo a phase change, the pump pulse signal will be output from the coupling end of the waveguide-type directional coupler 2 and coupled to the input end of the download waveguide 13.

[0084] Step 2: The pump pulse signal coupled to the download waveguide 13 is coupled to the micro-ring waveguide 11;

[0085] If the first phase change film 3 undergoes a phase change, the output terminal of the upper carrier waveguide 12 will output a probe light pulse signal; if the first phase change film 3 does not undergo a phase change, the output terminal of the upper carrier waveguide 12 will not output a probe light signal.

[0086] The present invention also proposes a neural network based on the above-mentioned on-chip neurons, comprising multiple interconnected on-chip neurons;

[0087] The input of the waveguide-type directional coupler 2 of the first on-chip neuron is used to receive external pump pulse signals;

[0088] The input of the upcarrier conductor 12 of each on-chip neuron is used to receive the probe light signal, and the output is used to detect the light pulse signal;

[0089] The output of the upcarrier 12 of each on-chip neuron is connected to the input of a waveguide-type directional coupler 2 in another neuron.

[0090] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An on-chip neuron, characterized in that: It includes a top-down microring structure (1) and a waveguide-type directional coupler (2); The up-and-down micro-ring structure (1) includes a micro-ring waveguide (11) and an up-carrying waveguide (12) and a down-carrying waveguide (13) respectively disposed on the upper and lower sides of the micro-ring waveguide (11). A first phase change film (3) made of phase change material is deposited on the micro-ring waveguide (11). The upper carrier waveguide (12) is an upper carrier straight waveguide, which operates in the probe optical signal band; the input end of the upper carrier waveguide (12) is used to receive the probe optical signal, and the output end is used to output the probe optical pulse signal; The download waveguide (13) includes an arc-shaped waveguide (131) and an input download straight waveguide (132) and an output download straight waveguide (133) extending along the input and output ends of the arc-shaped waveguide (131), respectively; the arc-shaped waveguide (131) is located directly below the micro-ring waveguide (11); the arc-shaped waveguide (131) operates in the pump pulse signal band; the output end of the download waveguide (13) is unused; The number of waveguide-type directional couplers (2) is at least two, and the coupling end of each waveguide-type directional coupler (2) is connected to the input straight waveguide (132). The waveguide-type directional couplers (2) are arranged sequentially along the extension direction of the input end of the input straight waveguide (132) to realize the superposition of the output power of multiple waveguide-type directional couplers (2). The coupling region of the waveguide-type directional coupler (2) is provided with a second phase change film (4) made of phase change material. The input terminal of the waveguide-type directional coupler (2) is used to receive external pump pulse signals; the through output port and isolation port of the waveguide-type directional coupler (2) are left unused. When the power of the input pump pulse signal does not reach the phase transition threshold of the second phase transition film (4), the pump pulse signal can be output from the coupling end of the waveguide-type directional coupler (2) to the download waveguide (13); when the power of the input pump pulse signal reaches the phase transition threshold of the second phase transition film (4), the pump pulse signal cannot be output from the coupling end of the waveguide-type directional coupler (2).

2. The on-chip neuron according to claim 1, characterized in that: The number of waveguide-type directional couplers (2) is four.

3. An on-chip neuron according to claim 1 or 2, characterized in that: The materials of the uplink / downlink microring structure (1) and the waveguide-type directional coupler (2) are silicon, silicon oxide, silicon nitride, group III-V materials, or lithium niobate.

4. A method for designing on-chip neurons according to any one of claims 1-3, characterized in that, Includes the following steps: Step 1: Set the structural parameters of the on-chip neuron, waveguide directional coupler (2), first phase change film (3) and second phase change film (4), and set the peak power and pulse width of the pump pulse signal. The peak power and pulse width of the pump pulse signal must be able to reach the phase change threshold of the first phase change film (3). The parameters include the waveguide widths, radius, and waveguide spacing of the upper carrier waveguide (11), lower carrier waveguide (13), and micro-ring waveguide (11) in the upper and lower carrier micro-ring structure (1); the waveguide width, coupling length, and waveguide spacing of the coupling region of the waveguide-type directional coupler (2); and the widths and lengths of the first phase change film (3) and the second phase change film (4). Step 2: Input a probe light signal to the input end of the upper carrier guide (12); input a pump pulse signal to the input end of each waveguide-type directional coupler (2), and output a probe light pulse signal to the output end of the upper carrier guide (12). If the probe light pulse signal meets the requirements, the design of the on-chip neuron is completed; otherwise, proceed to step 3. Step 3: Pause the input of the probe light signal and the pump pulse signal, and adjust at least one of the parameters mentioned in Step 1, and then proceed to Step 4; Step 4: Input a probe light signal to the input end of the upper carrier guide (12); input a pump pulse signal to the input end of each waveguide-type directional coupler (2), and output a probe light pulse signal at the output end of the upper carrier guide (12); if the probe light pulse signal meets the requirements, the fabrication of the on-chip neuron is completed; otherwise, return to step 3 until the probe light pulse signal meets the requirements.

5. The design method according to claim 4, characterized in that: In step 1, the download waveguide (13) includes an arc waveguide (131) and an input download straight waveguide (132) and an output download straight waveguide (133) extending along the input and output ends of the arc waveguide (131), respectively; the arc waveguide (131) is located directly below the micro-ring waveguide (11); Among the parameters, the waveguide width of the download waveguide (13) includes the waveguide width of the arc waveguide (131), the input download straight waveguide (132), and the output download straight waveguide (133). The parameters also include the central angle and radius of curvature of the arc waveguide (131). By adjusting the central angle and radius of curvature of the arc waveguide (131), the working band can be adjusted.

6. A method for information processing of on-chip neurons according to any one of claims 1-3, characterized in that, Includes the following steps: Step 1: Input the probe light signal to the input end of the upper carrier guide (12); input the pump pulse signal to the input end of each waveguide-type directional coupler (2); If the second phase change film (4) undergoes a phase change, the pump pulse signal will not be output. If the second phase change film (4) does not undergo a phase change, the pump pulse signal will be output from the coupling end of the waveguide-type directional coupler (2) and coupled to the input end of the download waveguide (13). Step 2: The pump pulse signal coupled to the download waveguide (13) is coupled to the micro-ring waveguide (11) again; If the first phase change film (3) undergoes a phase change, the output terminal of the upper carrier waveguide (12) outputs a probe light pulse signal; if the first phase change film (3) does not undergo a phase change, the output terminal of the upper carrier waveguide (12) does not output a probe light signal.

7. A neural network, characterized in that: It includes multiple interconnected on-chip neurons; The on-chip neuron is one of the on-chip neurons described in any one of claims 1-3; The input of the waveguide-type directional coupler (2) of the first on-chip neuron is used to receive external pump pulse signals; The input end of the upcarrier conduction (12) of each on-chip neuron is used to receive the probe light signal, and the output end is used to output the probe light pulse signal; The output of the upcarrier (12) of each on-chip neuron is connected to the input of a waveguide-type directional coupler (2) in another on-chip neuron.

Images