A photonic neuron device unit and a photonic neuron computing device

Abstract

The present invention belongs to the field of optoelectronic technology, specifically a photonic neuron device unit and a photonic neural computing device. The photonic neuron device unit of the present invention is an on-chip photonic device structure formed by a waveguide, an upper and lower combination of neuron materials; the photonic neuron material has a bidirectional transformation characteristic, which is amorphous or crystalline before operation. When the applied light intensity exceeds the threshold, part or all of the material will be transformed into a molten liquid state, resulting in a change in light absorption rate / transmittance, and realizing neuron function; the structure of the photonic neural computing device includes: a photonic neural synapse unit and a photonic neuron device unit; the photonic neural synapse unit modulates the optical signal and transmits it to the photonic neuron device unit after waveguide multiplexing; the photonic neuron device unit is directly connected to the next neural synapse unit through a waveguide; and so on, a large regular neural computing network is constructed. The present invention can meet the application requirements of high-speed and low-energy photonic neural computing.

Description

Technical Field

[0001] The present invention belongs to the field of optoelectronic technology, and in particular relates to a photonic neuron device unit and a photonic neuron computing device. Background Art

[0002] The data increment in the era of big data presents exponential characteristics, which puts forward higher requirements for future computing efficiency and computing energy efficiency. The current computing system is based on the von Neumann architecture, with computing-storage separation and frequent data transmission, which greatly hinders the improvement of computing efficiency and has defects such as energy efficiency wall. Compared with electrons, photons have advantages such as low latency and large bandwidth, which makes photon computing, especially photon neural computing, show great application potential in the development of high-speed and low-energy computing systems.

[0003] Photonic neural computing can be subdivided into free-space photonic neural computing and on-chip photonic neural computing. In on-chip photonic neural computing, light is transmitted and calculated in waveguides. It has the advantages of high integration and compatibility with traditional electronic chips, and has attracted widespread attention. There are three main components of photonic neural computing devices: waveguides for light transmission, neural synaptic units responsible for storage and calculation, and neuron units for data filtering. Neuron units are photonic devices with nonlinear responses. In order to meet the needs of high-speed photonic computing applications, they need to meet the response speed ≤10 ns, computing energy consumption ≤0.01 pJ, and the number of repeatable operations ≥10. 6 Second-rate.

[0004] Existing nonlinear optical response devices that can be used for neurons include waveguide devices based on thermo-optic effect, LiNbO3 devices based on electro-optic effect, and phase change devices based on free carrier dispersion effect. However, thermo-optic effect devices have high energy consumption and slow response speed; electro-optic effect devices are large in size and have high operating voltage; phase change devices are non-volatile and need to be refreshed frequently, which limits the development of high-speed, low-energy photonic neural computing devices. Summary of the invention

[0005] Based on the above problems, the purpose of the present invention is to provide a new type of photonic neuron device unit to meet the needs of high-speed, low-energy photonic neural computing applications, and at the same time provide a photonic neural computing device.

[0006] The photonic neuron device unit provided by the present invention is an on-chip photonic device structure formed by the combination of waveguide and neuron material on the top and bottom, which utilizes the transformation mechanism of the crystalline-liquid-crystalline phase transition of neuron material, and is different from the above-mentioned thermo-optical, electro-optical and phase change devices. Among them:

[0007] The photon neuron material is composed of one element among S, Se, Sb and Te, or a combination of at least two of these elements. Its chemical formula can be expressed as M x (S a Se b Sb c Te d ) 1-x , wherein M is a doping material, and 0≤x≤0.5, preferably 0<x≤0.5; a+b+c+d=1, 0≤a,b,c,d≤1;

[0008] The photonic neuron material has a bidirectional conversion characteristic, that is, it has no selectivity for the transmission direction of light;

[0009] The photonic neuron material can achieve transient light absorption rate / transmittance when the light intensity reaches a preset value, and spontaneously returns to the initial state after the light intensity is removed;

[0010] The photonic neuron material is in an amorphous or crystalline state before operation. When the applied light intensity exceeds a threshold, part or all of the material will be transformed into a molten liquid state, resulting in changes in light absorption rate / transmittance, thereby realizing neuron function;

[0011] The photonic neuron material, after the light intensity is removed, the liquid material will transform into the original structural state, realizing the spontaneous unit reset function, that is, the amorphous state-liquid state-amorphous state or crystalline state-liquid state-crystalline state transformation occurs;

[0012] The light response time of the photon neuron is less than 500 ns;

[0013] The photon neuron can be cycled ≥ 10 times 6 Second-rate.

[0014] In the photon neuron material, the doping material M can be non-metallic and metallic elements such as C, H, N, Al, or oxides, nitrides, etc.

[0015] Optionally, a protective layer may be added above the photonic neuron material, and the material of the protective layer is selected from dielectric materials such as ITO, SiO2, Al2O3, HfO2, etc.

[0016] Optionally, the waveguide material is one of SiN, Si, LiNbO3, etc., and has a transparent window in the used optical band.

[0017] Optionally, electrodes are added to the structure to form a waveguide / lower electrode / neuron material / upper electrode structure, and the material state is changed by electrical or thermal excitation to regulate light and realize the photonic neuron function.

[0018] Optionally, the lower electrode / upper electrode material is an electrode material that is transparent in the used light band, or an electrode material that has weak absorption and is not affected by light intensity / temperature, and can be specifically selected from ITO, graphene, carbon nanotubes, and the like.

[0019] Optionally, the photonic neuron material can selectively realize amorphous-liquid-amorphous or crystalline-liquid-crystalline volatile transition through electrical and thermal stimulation.

[0020] Optionally, the photonic neuron material can achieve the transformation of the material from an amorphous / crystalline state to a liquid state through at least one of heat generated by light absorption, Joule heat generated by electricity, and direct heating by a thermal field.

[0021] The present invention also provides a photonic neural computing device implemented by the above-mentioned photonic neural device unit, which is composed of photonic neural computing units connected in sequence; the photonic neural computing unit, its structure includes: a front photonic neural synapse unit and a photonic neural device unit; wherein:

[0022] The front-end photonic neural synapse unit is a phase change photonic device unit, a thermo-optic modulator or an electro-optic modulator, etc., and may also be composed of a combination of several of the above-mentioned photonic devices.

[0023] The front-end photonic neural synapse unit is used to modulate the optical signal and transmit it to the photonic neuron device unit after waveguide multiplexing.

[0024] The number of the front photonic neural synapse units is at least one, and may be two, three, or more.

[0025] The photonic neuron device unit is directly connected to the pre-synaptic unit in the next photonic neural computing unit through a waveguide; and so on, a large regular neural computing network is constructed.

[0026] In the photonic neural computing device, a photon signal is transmitted through an optical waveguide, modulated when passing through a preceding photonic neural synapse unit, and then calculated in the photonic neuron device unit to obtain a photon signal, which is then transmitted through the optical waveguide to the preceding photonic neural synapse unit of the next photonic neural computing unit, thereby realizing large-scale information computing.

[0027] The material in the photonic neuron device unit is crystalline before and after operation and has high temperature tolerance; the photonic neuron device unit realizes its function through a crystalline-liquid-crystalline phase change process, and its operation speed is only limited by the device heating rate, and can achieve ultra-fast operation of nanoseconds or even picoseconds; the photonic neuron device unit can use single-element materials with uniform elements and simple preparation, and can completely avoid performance differences between devices caused by component inhomogeneity and device failure caused by element segregation.

[0028] The photonic neural computing device can realize direct transmission of optical signals between the front-end photonic neural synaptic unit and the photonic neuron device unit without the assistance of devices such as photoelectric converters, which greatly reduces the complexity of a single photonic neural computing unit and reduces the speed limit and energy consumption limit caused by conversion. Similarly, optical signals can also be directly transmitted between different photonic neural computing units. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the cross-sectional structure of the photonic neuron device unit of the present invention.

[0030] Figure 2 The graph is a curve of incident light energy and normalized absorption change rate of the photonic neuron device based on single-element Te of the present invention.

[0031] Figure 3 This is a curve of the number of repetitive operations of the photonic neuron device based on single-element Te according to the present invention.

[0032] Figure 4 It is a schematic diagram of the interface structure of the photonic neuron device unit based on electrical / thermal stimulation of the present invention.

[0033] Figure 5 This is a schematic diagram of the structure of the photonic neural computing unit of the present invention.

[0034] The numbers in the figure are: 10 is the photonic neuron material layer, 11 is the optical waveguide, 12 is the protective layer, 13 is the lower electrode, 14 is the upper electrode, 15 is the photonic neural synapse unit, 16 is the photonic neuron device unit, and 17 is the pulsed laser. DETAILED DESCRIPTION

[0035] Example 1

[0036] This embodiment provides a photonic neuron device unit, whose structure is as follows: Figure 1 As shown, it is an on-chip photonic device composed of an optical waveguide 11, a neuron material 10 and a protective layer 12 from bottom to top. Among them:

[0037] The photon neuron material layer 10 is composed of at least one element of S, Se, Sb and Te, or a compound of at least any two of these elements, and its general chemical formula can be expressed as M x (S a Se b Sb c Te d ) 1-x , wherein M is a doping material, and 0≤x≤0.5, a+b+c+d=1. Preferably, in this embodiment, the photon neuron material is a single substance Te, that is, Te1.

[0038] The photonic neuron material layer 10 can be prepared by sputtering, evaporation, chemical vapor deposition (CVD), plasma enhanced chemical vapor deposition (PECVD), low pressure chemical vapor deposition (LPCVD), metal compound vapor deposition (MOCVD), molecular beam epitaxy (MBE), atomic vapor deposition (AVD) or atomic layer deposition (ALD).

[0039] The thickness, width and height of the photonic neuron material layer 10 can be set according to actual application requirements. Preferably, the thickness of the photonic neuron material layer 10 can be 0.4-100 nm, the width can be 0.05-50 μm, and the length can be 0.05-50 μm. More preferably, in this embodiment, the thickness of the photonic neuron material layer 10 is 5-40 nm, the width is 0.5-10 μm, and the length is 1.3 μm.

[0040] The optical waveguide layer material 11 may be one of Si3N4, Si, LiNbO3 and the like.

[0041] The protective layer material 12 may be a transparent material such as ITO, Al2O3, HfO2, graphene, etc. Preferably, in this embodiment, the protective layer material 12 is ITO.

[0042] The thickness of the protective layer material 12 may be 0.4-1000 nm, and the width and length are not less than the photonic neuron material layer 10. Preferably, in this embodiment, the thickness of the protective layer material is 5-40 nm.

[0043] The protective layer material 12 can be prepared by sputtering, evaporation, chemical vapor deposition (CVD), plasma enhanced chemical vapor deposition (PECVD), low pressure chemical vapor deposition (LPCVD), metal compound vapor deposition (MOCVD), molecular beam epitaxy (MBE), atomic vapor deposition (AVD) or atomic layer deposition (ALD).

[0044] Specifically, Figure 2 The graph is a graph of the input light energy and normalized absorption change rate of the photon neuron unit (single-element Te neuron unit). When the input light energy of the photon neuron unit is ~1 pJ, the absorption change rate starts to increase continuously from ~0; when the input light energy continues to increase, the absorption change rate continues to increase; until the input light energy increases to 10 pJ, the absorption change rate reaches ~1, and there is no obvious change when the light energy continues to increase, indicating that the device has reached the absorption saturation state.

[0045] Specifically, Figure 3 is the relationship between the number of repeated operations and the normalized absorption change rate of the single-element Te neuron unit under different input light energy conditions. The absorption change rate of the single-element Te neuron unit can be increased by 10 under both low input light energy and high input light energy. 8 After that, it remains basically unchanged, that is, the neuron unit can be repeated 10 times. 8 More than times.

[0046] Example 2

[0047] In the photon neuron unit described in Example 1, electrode layers are added above and below the photon neuron material layer 10, respectively, and the photon neuron function is realized by electrical / thermal excitation. The structure is as follows: Figure 4 As shown, it includes an optical waveguide 11, a lower electrode layer 13, a photonic neuron material layer 10 and an upper electrode layer 14.

[0048] The materials of the lower electrode 13 and the upper electrode layer 14 can be one or more of ITO, graphene, carbon nanotubes, etc., which are transparent electrode materials in the used light band, or electrode materials with weak absorption and not affected by light intensity / temperature.

[0049] The photonic neuron material can selectively realize a volatile transition from amorphous state to liquid state to amorphous state or from crystalline state to liquid state to crystalline state through at least one of light, electricity and heat stimulation.

[0050] The photonic neuron material can achieve the transformation of the material from an amorphous / crystalline state to a liquid state through at least one of the following methods: heat generated by light absorption, Joule heat generated by electricity, and direct heating by a thermal field.

[0051] Example 3

[0052] This embodiment provides a solution for implementing a photonic neural computing device, the unit structure of which is as follows: Figure 5 As shown, it includes an optical waveguide 11, a photonic neural synapse unit 15 and a photonic neuron unit 16.

[0053] The number of the photonic neural synapse unit 15 is not less than one.

[0054] The photon neuron unit 16 can be connected to the preceding photon neural synapse unit of the next photon neuron computing unit to construct a large regular neural network.

[0055] Specifically, in the photonic neural computing unit, Figure 5 As shown in the figure, the pulse laser 17 is input from the left side, transmitted in the waveguide 11 along the direction of the arrow, and first modulated by the photon neural synapse unit 15; then at the intersection of the waveguides, the modulated pulse laser 17 is multiplexed into a laser beam; it continues to be transmitted in the waveguide 11 along the direction of the arrow, and is filtered by the photon neuron unit 16; finally, the pulse laser 17 that reaches the threshold will continue to be transmitted along the waveguide 11 to the pre-synapse unit of the next photon neural computing unit. In the above manner, the photon neural computing unit realizes the information processing and transmission process.

Claims

1. A photonic neuron device unit, characterized in that: It is an on-chip photonic device structure formed by the combination of waveguide, neuron material, and upper and lower parts, where: The photon neuron material is selected from S, Se, Sb, Te, and its chemical formula is M x (S a Se b Sb c Te d ) 1-x , where M is the doping material, and 0≤x≤0.5; a+b+c+d=1; The photonic neuron material has a bidirectional conversion characteristic, that is, it has no selectivity for the transmission direction of light; The photonic neuron material can achieve transient light absorption rate / transmittance when the light intensity reaches a preset value, and spontaneously returns to the initial state after the light intensity is removed; The photonic neuron material is in an amorphous or crystalline state before operation. When the applied light intensity exceeds a threshold, part or all of the material will be transformed into a molten liquid state, resulting in changes in light absorption rate / transmittance, thereby realizing neuron function; The photonic neuron material, after the light intensity is removed, the liquid material is transformed into the initial structural state, realizing a spontaneous unit reset function, that is, an amorphous-liquid-amorphous or crystalline-liquid-crystalline transformation occurs; The light response time of the photon neuron is less than 500 ns; The photon neuron can be cycled ≥ 10 times 6 Second-rate.

2. The photonic neuron device unit according to claim 1, characterized in that: In the photonic neuron material, the doping material M is selected from C, H, N, Al, or selected from oxides and nitrides.

3. The photonic neuron device unit according to claim 1 or 2, characterized in that: A protective layer is added above the photon neuron material, and the material of the protective layer is selected from ITO, SiO2, Al2O3, and HfO2 dielectric materials.

4. The photonic neuron device unit according to claim 1 or 2, characterized in that: The waveguide material is one of SiN, Si and LiNbO3 and has a transparent window in the used optical band.

5. The photonic neuron device unit according to claim 1 or 2, characterized in that: Electrodes are also added to form a structure in which a waveguide, a lower electrode, a neuron material, and an upper electrode are arranged in sequence. The material state is changed through electrical or thermal excitation to regulate light and realize the function of a photonic neuron.

6. The photonic neuron device unit according to claim 5, characterized in that: The lower electrode and upper electrode materials are transparent electrode materials in the used light band, or electrode materials with weak absorption and not affected by light intensity / temperature; specifically, they are selected from ITO, graphene, and carbon nanotubes.

7. The photonic neuron device unit according to claim 1, 2 or 6, characterized in that: Through electrical or thermal stimulation, the volatile transition from amorphous state to liquid state to amorphous state or from crystalline state to liquid state to crystalline state is achieved.

8. The photonic neuron device unit according to claim 1, 2 or 6, characterized in that: The material is transformed from an amorphous state / crystalline state to a liquid state by at least one of heat generated by light absorption, Joule heat generated by electricity, and direct heating by a thermal field.

9. A photonic neural computing device implemented by the photonic neural device unit according to any one of claims 1 to 8, characterized in that: It is composed of sequentially connected photon neural computing units; the photon neural computing unit has a structure including: a front photon neural synapse unit and a photon neuron device unit; wherein: The front-end photonic neural synapse unit is a phase-change photonic device unit, a thermo-optic modulator or an electro-optic modulator, or is composed of a combination of the above photonic devices; The front-end photonic neural synapse unit is used to modulate the optical signal and transmit it to the photonic neuron device unit after waveguide multiplexing; The number of the front photonic neural synaptic unit is at least one; The photonic neuron device unit is directly connected to the pre-synaptic unit in the next photonic neural computing unit through a waveguide; and so on, a large regular neural computing network is constructed.

10. The photonic neural computing device according to claim 9, characterized in that: The photonic neural synapse is one or more of a non-volatile phase-change photonic synapse, a thermo-optical effect synapse and an electro-optical effect synapse.

Images