A photonic memory-computation integrated device and a photonic neural network structure

Abstract

The present invention belongs to the field of micro-nano optoelectronic technology, and specifically relates to a photonic memory-computation integrated device and a photonic neural network structure. The photonic memory-computation integrated device of the present invention includes a multi-ring resonator and a photonic storage material, and the photonic storage material is located on any one or several of the multi-rings, or on the waveguides at the through end or the drop end; electrode layers can be added above and below the photonic storage material, and high-bandwidth and high-storage-density photonic storage and computing functions are realized through optical, electrical or thermal excitation. The present invention utilizes the wide passband and high quality factor characteristics of the multi-ring resonator, and combines with the structure of the material / device on the waveguide to realize the memory-computation integrated function with high bandwidth and high storage density; the photonic neural network includes photonic neural synapses composed of low-order multi-ring resonator memories and photonic neurons composed of high-order multi-ring resonator memories. The present invention provides a basis for the development of on-chip integrated optical path applications such as high-speed, low-power consumption, high-precision, and scalable photonic neural networks.

Description

Technical Field

[0001] The present invention belongs to the field of micro-nano optoelectronic technologies, and particularly relates to a photonic memory-computation integrated device and a photonic neural network structure. Background Art

[0002] With the rapid development of the Internet of Things (IoT), big data analysis, and cloud computing, artificial neural networks (ANNs) are playing an increasingly important role in fields such as computer vision, speech recognition, and natural language processing. However, existing electronic computing systems are restricted by the von Neumann architecture and cannot fully meet the requirements of large-scale matrix-vector multiplication (MVM) in ANNs. At the same time, CMOS-based electronic computing hardware is approaching the limits of computing speed and energy efficiency. In contrast, optical technology is a better method for realizing fast data transmission and processing, with advantages such as high speed, low energy consumption, low heat dissipation, and high bandwidth. Moreover, optical interconnections for low-latency communication on multi-core chips and nanosecond-level ultrafast optical computing elements have been developed in the past decade. Optical technology provides an upgraded approach for ANNs that can significantly boost parallel data processing speed and energy efficiency, namely, optical neural networks (ONNs).

[0003] The key components of an ONN are photonic neural synapses that implement weight storage and data operation and photonic neurons that implement non-linear activation. Existing photonic neural synapse devices and photonic neuron devices mainly modulate optical signals by applying optical, electrical, or thermal excitation to the material / device structure on the waveguide or directly to the waveguide structure of a specific material (such as LiNbO 3 ). Although this structure has relatively simple manufacturing processes and regulation methods, it cannot eliminate the scattered wavelengths generated during the optical signal transmission and modulation processes, which is not conducive to the array integration of devices and the improvement of the accuracy of neural networks. Although the two output ports of a ring resonator respectively have band-pass and band-stop filtering functions and can significantly increase the storage density of the material on the resonator ring, its passband / stopband is relatively narrow and can usually only accommodate light of a single wavelength, so its application in photonic neural devices still has significant limitations. A multi-ring resonator can broaden the passband / stopband by changing the number of resonator rings, while ensuring the high coupling efficiency and high quality factor of the resonator, providing a theoretical basis and ideas for increasing the bandwidth of photonic devices and improving the storage density of photonic devices. Summary of the Invention

[0004] Based on the above problems, the purpose of the present invention is to provide a novel photonic memory-computation integrated device to significantly increase the storage density and working bandwidth of the device; at the same time, to provide a photonic neural network to reduce crosstalk between different working wavelengths, improve the training efficiency, test accuracy, and scalability of the neural network.

[0005] The novel photon storage and computing integrated device provided by the present invention has the following structure from bottom to top: a multi-ring resonator and a photon storage material. By utilizing the wide passband and high quality factor characteristics of the multi-ring resonator and combining with the structure of materials / devices on the waveguide, the integrated function of high bandwidth and high storage density is realized. This device can also be used alone as an optical computing device or an optical storage device.

[0006] Optionally, in the photon storage and computing integrated device, the photon storage material can be a non-volatile reversible material (such as phase change material, two-dimensional material, ferroelectric material, magnetic material, etc.) for realizing matrix-vector multiplication and the storage of (weight) matrix; it can also be a reversible material with nonlinear optical response or optoelectronic response (such as phase change material, two-dimensional material, ferroelectric material, magnetic material, etc.) for realizing the nonlinear activation function.

[0007] Optionally, in the photon storage and computing integrated device, the radius and number of the resonant rings of the multi-ring resonator can be adjusted according to the requirements of the working wavelength band and bandwidth, and usually 2 to 5 rings or more are used.

[0008] Optionally, for the multi-ring resonator, a drop waveguide can be configured. Under the action of the resonant ring, the throughport and the drop port respectively exhibit the functions of a band-stop filter and a band-pass filter; or a drop waveguide can be not configured. Under the action of the resonant ring, light in a specific wavelength band can be coupled into it, but the throughport does not exhibit the filter function.

[0009] Optionally, the photon storage material can be located on any one or several of the multi-ring resonators, or on the waveguides of the throughport or the drop port.

[0010] Optionally, the upper part of the photon storage and computing integrated device is covered with a protective layer material, such as dielectric materials like ITO, SiO 2 、Al 2 O 3 、HfO 2 etc. The protective layer material can be prepared by methods such as PVD, PECVD, ALD, etc.

[0011] Optionally, the waveguide material of the photon storage and computing integrated device must have a transparent window in the used optical wavelength band (such as C band and L band), and the window material can be one of Si 3 N 4 、Si、LiNbO 3 etc.

[0012] Optionally, in the structure of the photonic memory - in - computing device, electrode materials can also be added to form a multi - ring resonator / lower electrode / photonic storage material / upper electrode structure, and the material state is changed by electrical or thermal excitation; alternatively, electrode materials can be not added, and the material state is directly changed by the photothermal effect generated by the input light, thereby completing the modulation of light and realizing the photonic memory - in - computing function.

[0013] Optionally, the lower electrode material and the upper electrode material are electrode materials that are transparent in the used optical wavelength band, such as ITO, graphene, carbon nanotubes, etc., or electrode materials with weak absorption and less affected by light intensity and temperature.

[0014] Optionally, the photonic storage material can achieve reversible state transition of the material through at least one of the following ways: heat generated by light absorption, Joule heat generated by electricity, direct heating by a thermal field, etc.

[0015] The present invention also provides a photonic neural network implemented by using the photonic memory - in - computing device, and its structure includes: a pre - placed photonic neural synapse device array and a post - placed photonic neuron device; where:

[0016] In the pre - placed photonic neural synapse device array, the photonic neural synapse device is composed of a low - order multi - ring resonator and a photonic storage material / device on the resonator ring, and the photonic storage material / device must have non - volatile storage characteristics. Here, " / " means "or", the same hereinafter;

[0017] The post - placed photonic neuron device is composed of a high - order multi - ring resonator and a photonic neuron material / device on the resonator ring, and the photonic neuron material / device must have a non - linear optical response or a photoelectric response. Here, the "order" refers to the number of resonator rings in the multi - ring resonator, and "low - order" and "high - order" are relative. Specifically, the number of resonator rings of the multi - ring resonator in the post - placed photonic neuron device is greater than the number of resonator rings of the multi - ring resonator in the pre - placed photonic neural synapse device.

[0018] Here, the "photonic neuron material / device" is specifically the "photonic storage material / device".

[0019] The pre - placed photonic neural synapse device array and the post - placed photonic neuron device are both connected by a bus waveguide, and the working wavelength bands of the devices in the synapse device array are adjacent to each other and do not overlap, and the working wavelength band of the neuron device includes the working wavelength bands of each synapse device.

[0020] The pre - placed photonic neural synapse device array obtains the optical pulse signals of the corresponding working wavelength band from the bus waveguide for modulation, and transmits them to the post - placed photonic neuron device through the bus waveguide; the post - placed photonic neuron device obtains the optical pulse signals of all working wavelength bands from the bus waveguide for modulation, and transmits them to the output waveguide.

[0021] Optionally, in the array of pre - photon synaptic devices, there are at least 2 synaptic devices.

[0022] Optionally, the neural network structure unit can be directly connected to the bus waveguide of the next neural network structure unit through the output waveguide of this unit to achieve the scale expansion of the neural network.

[0023] Optionally, the photon storage material / device of the pre - photon synaptic device can be one or several of non - volatile phase change materials / devices, two - dimensional materials / devices, ferroelectric materials / devices, magnetic materials / devices, thermo - optic effect materials / devices, electro - optic effect materials / devices, etc.

[0024] Optionally, the photon neuron material / device of the post - photon neuron device can be one or several of phase change materials / devices, two - dimensional materials / devices, ferroelectric materials / devices, magnetic materials / devices, thermo - optic effect materials / devices, electro - optic effect materials / devices, etc. with non - linear response.

[0025] The photon memory - computing integrated device of the present invention has the following characteristics and advantages:

[0026] Compared with traditional electrical devices, the photon memory - computing integrated device has the characteristics of high speed, low power consumption, and high bandwidth. At the same time, it breaks through the limitations of the traditional von Neumann architecture and further improves the upper limit of memory - computing speed and energy efficiency. Compared with other photon memory - computing integrated devices, the photon memory - computing integrated device of the present invention significantly improves the bandwidth and storage density, and can realize the memory - computing functions of more channels and more storage states.

[0027] The photon neural network of the present invention has the following characteristics and advantages:

[0028] Compared with the traditional electrical neural network architecture, the photon neural network of the present invention has the performance advantages of high speed, low power consumption, and high bandwidth. Compared with other photon neural network architectures, the photon neural network of the present invention significantly improves the bandwidth and storage density of the memory - computing unit, thus realizing the improvement of training efficiency and test accuracy; at the same time, the photon neural network can expand the scale of the neural network through the interconnection of bus waveguides and has strong scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the waveguide structure of the photon memory - computing integrated device of the present invention.

[0030] Figure 2 It is a relationship curve graph of the insertion loss of the device (three - ring resonator) in Embodiment 1 and the input optical wavelength.

[0031] Figure 3Schematic diagram of the waveguide structure of the photonic neural network of the present invention.

[0032] Reference numerals in the figure: 10 is a resonator ring (three rings), 11 is a through waveguide, 12 is a drop waveguide, 13, 14, 15 are photon storage materials (optional) corresponding to the three resonators, 16 is a photon storage material (optional) on the through waveguide 11, 17 is a photon storage material (optional) on the drop waveguide 12, 18 is an input optical pulse, 19 is a through output optical pulse, 20 is a drop output optical pulse, 21, 22, 23 are photonic neural synapse devices, 24, 25, 26 are non-volatile materials / devices (optional) corresponding to the photonic neural synapse devices 21, 22, 23, 27 is a photonic neuron device, 28 is a non-linear response material / device (optional) corresponding to the photonic neuron device 27, 29 is a bus waveguide, 30 is an output waveguide, 31 is an input optical pulse, 32 is an output optical pulse. Detailed implementation mode

[0033] Example 1

[0034] As an example, a photonic computing-in-memory device, the waveguide structure of which is composed of a three-ring resonator and a photon storage material layer (13 - 17), as Figure 1 shown. Among them, the three-ring resonator is composed of a resonator ring (three rings) 10, a through waveguide 11, and a drop waveguide 12.

[0035] The three-ring resonator of the photonic computing-in-memory device can be configured with a drop waveguide 12. Under the action of the resonator ring 10, the through and drop ends respectively present the functions of a band-stop filter and a band-pass filter; it can also not be configured with a drop waveguide 12. Under the action of the resonator ring, light in a specific wavelength band can be coupled into it, but the through end does not present the filter function. Preferably, in this embodiment, the three-ring resonator is configured with a drop waveguide 12.

[0036] The optical waveguide layer material of the three-ring resonator can be Si 3 N 4 , Si, LiNbO 3 and so on. Preferably, in this embodiment, the optical waveguide layer material of the three-ring resonator is Si 3 N 4 .

[0037] The waveguide structure dimensions of the three-ring resonator can be set according to actual application requirements; preferably, the waveguide structure can be one or several of a strip waveguide, a ridge waveguide, a slot waveguide, etc. The total height of the waveguide structure can be 0.05 - 3 μm, the width of the waveguide structure can be 0.05 - 10 μm, the radius of the resonant ring can be 0.5 - 100 μm, and the width of the structural gap can be 0.01 - 0.5 μm; more preferably, in this embodiment, the total height of the waveguide structure is 0.34 μm, the height of the ridge waveguide structure is 0.23 μm, the width of the through / descending waveguide is 0.7 μm, the width of the resonant ring waveguide is 0.9 μm, the radius of the resonant ring is 15 μm, the ring-ring spacing is 0.37 μm, and the ring-waveguide spacing is 0.12 μm.

[0038] The three-ring resonator can use a Si 3 N 4 wafer and is prepared by means such as ultraviolet lithography, electron beam lithography (EBL), reactive ion etching (RIE), inductively coupled plasma etching (ICP), etc.

[0039] The photon storage material layer (13 - 17) can be located on any one or several of the three rings (13 - 15), or can be located on the waveguide of the through end or the descending end (16 - 17). Preferably, in this embodiment, the photon storage material layer is located on the second-order resonant ring (14).

[0040] The photon storage material layer (13 - 17) can be one or several of a phase change material, a two-dimensional material, a ferroelectric material, a magnetic material, etc. Preferably, in this embodiment, the photon storage material layer is crystalline GST, that is, crystalline Ge 2 Sb 2 Te 5 。

[0041] The thickness, width, and height of the photon storage material layer (13 - 17) can all be set according to actual application requirements; preferably, the thickness of the photon storage material layer (13 - 17) 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 photon storage material layer (13 - 17) is 10 nm, the width is 0.5 μm, and the length is 0.7 - 0.9 μm.

[0042] The photon storage material layer (13-17) can be formed by sputtering, evaporation, chemical vapor deposition (CVD), plasma-enhanced chemical vapor deposition (PECVD), low-pressure chemical vapor deposition (LPCVD), metalorganic chemical vapor deposition (MOCVD), molecular beam epitaxy (MBE), atomic vapor deposition (AVD), or atomic layer deposition (ALD), etc.

[0043] A protective material layer can be covered above the photon storage material layer (13-17), which can be ITO, Al 2 O 3 , HfO 2 , graphene and other transparent materials. Preferably, in this embodiment, the protective layer material is ITO.

[0044] The thickness of the protective layer material can be 0.4 to 1000 nm, and the width and length are both not less than those of the photon storage material layer (13-17); preferably, in this embodiment, the thickness of the protective layer material is 10 nm.

[0045] The protective layer material can be prepared by sputtering, evaporation, chemical vapor deposition (CVD), plasma-enhanced chemical vapor deposition (PECVD), low-pressure chemical vapor deposition (LPCVD), metalorganic chemical vapor deposition (MOCVD), molecular beam epitaxy (MBE), atomic vapor deposition (AVD), or atomic layer deposition (ALD), etc.

[0046] Specifically, Figure 2 is a relationship curve graph of the device insertion loss of the triple-ring resonator (10-12) and the input optical wavelength. The descending-end output optical pulse 20 of the triple-ring resonator (10-12) can maintain an insertion loss of more than -10 dB in the wavelength band of 1620.6 to 1622.4 nm, showing a good band-pass filter function; while the through-end output optical pulse 19 can maintain an insertion loss of more than -3 dB outside the wavelength band of 1620.6 to 1622.4 nm, showing a good band-stop filter function. Therefore, the input optical pulse 18 within the working wavelength band of 1620.6 to 1622.4 nm can be effectively coupled into the resonator ring 10 and interact with the photon storage material layer 14 and then output from the descending end (20) with low loss; the input optical pulse 18 outside the working wavelength band of 1620.6 to 1622.4 nm can be directly output from the through-end (19) with low loss.

[0047] Specifically, the photon storage material layer 14 can strongly absorb the high-power input optical pulse 18 within the working wavelength range of 1620.6 - 1622.4 nm, and undergo crystallization or amorphous phase change under the action of the photothermal effect, thereby changing the absorption coefficient of the photon storage material layer 14 and realizing the photon storage function.

[0048] Specifically, the photon storage material layer 14 can weakly absorb the low-power input optical pulse 18 within the working wavelength range of 1620.6 - 1622.4 nm, change the power of the optical pulse and output it, and is not sufficient to trigger crystallization or amorphous phase change, thereby realizing the optical computing function. In the above manner, the photon memory and computing integrated device has the memory and computing integrated function with high bandwidth and high storage density.

[0049] Embodiment 2

[0050] Provide a solution for implementing a photon neural network using the photon memory and computing integrated device, the structure of which is as Figure 3 shown, including 3 photon neural synapse devices (21 / 24, 22 / 25, 23 / 26), 1 photon neuron device (27 / 28), a bus waveguide 29, and an output waveguide 30.

[0051] The photon neural synapse devices (21 / 24, 22 / 25, 23 / 26) are composed of second-order multi-ring resonators (21 - 23) and photon storage materials / devices (24 - 26) on the resonant rings, and the photon storage materials / devices (24 - 26) must have non-volatile storage characteristics.

[0052] The photon storage materials / devices (24 - 26) can be one or several of non-volatile phase change materials / devices, two-dimensional materials / devices, ferroelectric materials / devices, magnetic materials / devices, thermo-optical effect materials / devices, electro-optical effect materials / devices, etc. Preferably, the photon storage materials / devices (24 - 26) are Ge 2 Sb 2 Te 5 (GST) thin films.

[0053] The photon neuron device (27 / 28) is composed of a fourth-order multi-ring resonator 27 and a photon neuron material / device 28 on the resonant ring, and the photon neuron material / device 28 must have a non-linear optical response or a photoelectric response.

[0054] The photon neuron device (27 / 28) can be one or several of phase change materials / devices, two-dimensional materials / devices, ferroelectric materials / devices, magnetic materials / devices, thermo-optic effect materials / devices, electro-optic effect materials / devices, etc. with non-linear responses. Preferably, the photon neuron material / device 28 is a graphene device or Te 1 (elemental Te) thin film.

[0055] The output waveguide 30 of the neural network structure unit can be directly connected to the bus waveguide 29 of the next neural network structure unit.

[0056] Specifically, the training / test set data can be encoded into input optical pulses 31 of different wavelengths simultaneously and transmitted to each photon synapse device (21 / 24, 22 / 25, 23 / 26) through the bus waveguide 29; each photon synapse device (21 / 24, 22 / 25, 23 / 26) obtains the optical pulse signals in the corresponding working band from the bus waveguide 29 for matrix-vector operations or weight matrix storage, and transmits them to the post photon neuron device (27 / 28) through the bus waveguide 29; the post photon neuron device (27 / 28) obtains all the optical pulse signals in the working band from the bus waveguide 29 for non-linear modulation and transmits them to the output waveguide 30 to obtain the output optical pulse 32; the output optical pulse 32 is used as the input optical pulse 31 of the next neural network structure unit, or the training / test result is obtained after decoding. In the above way, the neural network structure unit can effectively reduce the crosstalk between different working wavelengths and significantly improve the accuracy and scalability of the neural network.

Claims

1. A photonic memory and computing integrated device, characterized in that, from bottom to top: a multi-ring resonator and a photonic storage material. By utilizing the wide passband and high quality factor characteristics of the multi-ring resonator and combining with the structure of the material / device on the waveguide, the integrated function of high bandwidth and high storage density is realized; this device can be used alone as an optical computing device or an optical storage device; the photonic storage material is a non-volatile reversible material for realizing matrix-vector multiplication and matrix storage; or a reversible material with nonlinear optical response or optoelectronic response for realizing nonlinear activation function; the radius and number of the resonant rings of the multi-ring resonator are adjusted according to the requirements of the working band and bandwidth; the multi-ring resonator is configured with a descending waveguide. Under the action of the resonant ring, the through-end and the descending-end respectively present the functions of a band-stop filter and a band-pass filter; or without a descending waveguide, under the action of the resonant ring, light in a specific band can be coupled into it, but the through-end does not present the filter function.

2. The photonic memory and computing integrated device according to claim 1, characterized in that, the photonic storage material is located on any one or several of the multi-ring resonators, or on the waveguides at the through-end or the descending-end.

3. The photonic memory and computing integrated device according to claim 2, characterized in that, A protective layer is covered above the photon storage and computing integrated device, and the protective layer material is selected from ITO, SiO 2 , Al 2 O 3 , HfO 2 dielectric materials.

4. The photonic memory and computing integrated device according to claim 3, characterized in that, Its waveguide material has a transparent window in the optical wavelength band used, and the waveguide material uses Si 3 N 4 , Si, LiNbO 3 One of them.

5. The photonic memory and computing integrated device according to any one of claims 1-4, characterized in that, an electrode material is further added to form a structure including a multi-ring resonator, a lower electrode, a photonic storage material and an upper electrode, and the material state is changed by electrical or thermal excitation; the modulation of light is completed to realize the integrated function of photonic memory and computing.

6. The photonic memory and computing integrated device according to claim 5, characterized in that, the lower electrode material and the upper electrode material are electrode materials transparent in the used optical band, or electrode materials with weak absorption and less affected by light intensity and temperature, selected from ITO, graphene, and carbon nanotubes.

7. A photonic neural network implemented by using the photonic memory and computing integrated device according to any one of claims 1-6, characterized in that, comprising: a preposed photonic neural synapse device array and a postposed photonic neuron device; wherein: in the preposed photonic neural synapse device array, the photonic neural synapse device is composed of a low-order multi-ring resonator and a photonic storage material / device on the resonant ring, and the photonic storage material / device must have non-volatile storage characteristics; here, " / " means "or", the same below; the postposed photonic neuron device is composed of a high-order multi-ring resonator and a photonic neuron material / device on the resonant ring, and the photonic neuron material / device has a nonlinear optical response or optoelectronic response; here, the "order" refers to the number of resonant rings in the multi-ring resonator, and "low-order" and "high-order" are relative; the preposed photonic neural synapse device array and the postposed photonic neuron device are both connected by a bus waveguide, and the working bands of the devices in the synapse device array are adjacent to each other and do not overlap, and the working band of the neuron device includes the working bands of each synapse device; The pre - front photon neural synapse device array obtains the optical pulse signals in the corresponding working band from the bus waveguide for modulation, and transmits them to the post - front photon neuron device through the bus waveguide; the post - front photon neuron device obtains all the optical pulse signals in the working band from the bus waveguide for modulation and transmits them to the output waveguide.

8. The photon neural network according to claim 7, wherein, the neural network structure unit is directly connected to the bus waveguide of the next neural network structure unit through the output waveguide of this unit to realize the scale expansion of the neural network.

9. The photon neural network according to claim 8, wherein, the photon storage material / device of the pre - front photon neural synapse device is one or several of non - volatile phase - change materials / devices, two - dimensional materials / devices, ferroelectric materials / devices, magnetic materials / devices, thermo - optical effect materials / devices, and electro - optical effect materials / devices; the photon neuron material / device of the post - front photon neuron device is one or several of phase - change materials / devices, two - dimensional materials / devices, ferroelectric materials / devices, magnetic materials / devices, thermo - optical effect materials / devices, and electro - optical effect materials / devices with non - linear response.

Images