A photonic processor based on optical frequency combs and dual microring modulators

Abstract

The application provides a photonic processor based on an optical frequency comb and a double micro-ring modulator, and relates to the field of photonic devices, and comprises the following: an optical signal layer which generates a plurality of optical frequency comb signals containing equidistantly spaced wavelength components and uniformly distributes the generated optical signals after processing; an input signal layer which applies an electrical signal to a Mach-Zehnder modulator array so that the output optical signal of the Mach-Zehnder modulator array performs mapping of an input tensor; a weight signal layer which inputs a weight tensor as an electrical signal of a double micro-ring modulator array and interacts with the input signal layer to perform matrix operation or convolution operation; a matrix operation summation layer which is used to add optical signals of different wavelengths when performing matrix operation; and a convolution operation summation layer which is used to couple optical signals in different time periods to a combiner to perform summation operation when performing convolution operation, and convert the summed optical signals into electrical signals. The application can realize large-scale matrix and convolution operation, and the structure is scalable.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of photonic devices, and particularly relates to a photonic processor based on optical frequency comb and double micro-ring modulator. BACKGROUND

[0002] Due to the rise of big data, the performance of artificial neural networks has been greatly improved, and is widely used in natural language processing, material science and biological information and other fields. However, due to the limitation of traditional von Neumann structure, the hardware implementation of artificial neural networks based on electronic devices has certain bottlenecks. This structure is suitable for serial and digital operation, while artificial neural networks are more suitable for distributed and parallel computing because they need to perform large-scale linear operations. And in order to meet the actual needs, the processor needs to improve the processing speed and reduce the energy consumption.

[0003] Since photons are bosons and have higher frequencies, photonic device-based processors can effectively solve the above problems. Current schemes for implementing convolution operations using photonic devices include wavelength division multiplexing schemes and coherent schemes. The principle of the wavelength division multiplexing scheme is to first distribute different wavelength optical signals into different time periods for multiplication operation through dispersion effect, and then compress the optical signals distributed in different time periods into the same time period through an anomalous dispersion fiber to realize the summation operation function. The coherent scheme is based on the principle of matrix decomposition, which decomposes the matrix to be implemented into three different sub-matrices, and then uses three cascaded Mach-Zehnder interferometer networks to perform the functions of the three matrices respectively.

[0004] The above schemes are mainly aimed at matrix operations or convolution operations in neural networks. The photonic processor based on the coherent scheme is suitable for the calculation mode of matrix operation, and more photonic devices are used when performing convolution operation, making the whole system more complex. The wavelength division multiplexing scheme based on dispersion effect is used to implement the calculation mode of convolution operation, which reduces the efficiency of matrix operation and makes it difficult to implement effective dispersion effect on chip when performing matrix operation. SUMMARY

[0005] In view of the above problems, the present application provides a photonic processor based on optical frequency comb and double micro-ring modulator, which realizes large-scale matrix operation and convolution operation by combining the rich wavelength components of the optical frequency comb, and realizes convolution operation and matrix multiplication operation at different ports through the action of the double micro-ring modulator, thereby accelerating the operation of the neural network. The photonic processor comprises:

[0006] An optical signal layer for generating a plurality of optical frequency comb signals containing equally spaced wavelength components, and uniformly distributing the processed optical signals;

[0007] an input signal layer for applying an electrical signal to a Mach-Zehnder modulator array to cause the Mach-Zehnder modulator array to output an optical signal that performs a mapping of an input tensor;

[0008] a weight signal layer for inputting a weight tensor of a matrix operation or a convolution operation as an electrical signal to the dual microring modulator array to perform the matrix operation or the convolution operation with the signal of the input signal layer;

[0009] a matrix operation summation layer for adding optical signals of different wavelengths based on the first photodetector array when performing the matrix operation;

[0010] a convolution operation summation layer for coupling optical signals in different time periods to a combiner to perform a summation operation and converting the summed optical signals into electrical signals through the second photodetector array when performing the convolution operation.

[0011] optionally, a semiconductor laser for providing a continuous wave optical signal;

[0012] a microring resonator for generating a plurality of optical frequency comb signals containing equally-spaced wavelength components based on a nonlinear Kerr effect to accelerate the matrix operation and the convolution operation;

[0013] a waveform shaper for flattening the uneven optical frequency comb signals into optical signals with different wavelength components but with the same power;

[0014] a beam splitter for uniformly distributing the generated optical signals to output ports;

[0015] the beam splitter is implemented by a cascaded 1x2 multimode interferometer, a Y-branch, or a directional coupler.

[0016] optionally, the input signal layer comprises:

[0017] the Mach-Zehnder modulator array composed of a plurality of Mach-Zehnder modulators for inputting an input tensor in the matrix operation or the convolution operation as an electrical signal to the Mach-Zehnder modulators to implement the mapping of the input tensor.

[0018] optionally, the weight signal layer comprises the dual microring modulator array composed of a plurality of dual microring modulators, each dual microring modulator comprising a cross waveguide, a microring modulator MRM_n, and a microring modulator MRM_s, for:

[0019] In performing the convolution operation, the resonance wavelength of MRM_n is moved to the wavelength of the input optical signal by thermo-optic effect, then the resonance wavelength of MRM_n is deviated from the wavelength of the input optical signal by electro-optic effect, the transmittance of the input optical signal through MRM_n is changed by adjusting the deviation value, to correspond to the weight value in the weight tensor, while the resonance wavelength of MRM_s is far away from the wavelength of the input optical signal, in the dormant state;

[0020] In performing the matrix operation, the resonance wavelength of MRM_s is moved to the wavelength of the input optical signal by thermo-optic effect, then the resonance wavelength of MRM_s is deviated from the wavelength of the input optical signal by electro-optic effect, the transmittance of the input optical signal through MRM_s is changed by adjusting the deviation value, to correspond to the weight value in the weight tensor, while the resonance wavelength of MRM_n is far away from the wavelength of the input optical signal, in the dormant state.

[0021] Optionally, the matrix operation summation layer comprises:

[0022] The first photodetector array composed of a plurality of photodetectors is used to perform summation operation in the matrix operation process, which is realized by adding optical signals of different wavelengths. Wherein, the optical power received by each photodetector in the first photodetector array satisfies:

[0023] P1=W11*X1+W21*X2+W31*X3;

[0024] P2=W12*X1+W22*X2+W32*X3;

[0025] P3=W13*X1+W23*X2+W33*X3;

[0026] Wherein, P1, P2, P3 are respectively the optical power received by each photodetector in the first photodetector array, X1, X2, X3 are respectively the optical signals output by each Mach-Zehnder modulator in the Mach-Zehnder modulator array, and W11, W21, W31, W12, W22, W32, W12, W23 and W33 are respectively the transmittance of MRM_s in each dual-micro-ring modulator in the dual-micro-ring modulator array for the optical signal near the corresponding resonance wavelength.

[0027] Optionally, the convolution operation summation layer comprises:

[0028] The optical delay line array is used to separate optical signals at different time instants;

[0029] The directional coupler array is used to couple optical signals at different time instants into the combiner array;

[0030] The wave combiner array is used for combining the optical signals coupled out by the directional coupler array into the same waveguide to realize summation.

[0031] The second photodetector array is used for converting the optical signals into electrical signals when performing convolution operation.

[0032] The technical scheme provided by the embodiment of the present application at least brings the following beneficial effects:

[0033] Through the combination of the rich wavelength components of the optical frequency comb, large-scale matrix operation and convolution operation are realized, and through the action of the double micro-ring modulator, the convolution operation and the matrix multiplication operation are realized at the north end and the south end of the double micro-ring modulator respectively, so as to accelerate the operation of the neural network, and the structure proposed in the present application has scalability and can be monolithically integrated, and can adapt to the demand of large-scale linear operation.

[0034] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter in the description. BRIEF DESCRIPTION OF DRAWINGS

[0035] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0036] Figure 1 is a structural diagram of a photonic processor based on an optical frequency comb and a double micro-ring modulator according to an embodiment of the present application;

[0037] Figure 2 is a structural diagram of a double micro-ring modulator according to an embodiment of the present application;

[0038] Figure 3 is a structural diagram of a photonic processor performing convolution operation according to an embodiment of the present application

[0039] Figure 4 is a structural diagram of a photonic processor performing matrix operation according to an embodiment of the present application;

[0040] Figure 5 is a schematic diagram of a convolution operation input tensor according to an embodiment of the present application;

[0041] Figure 6 is a schematic diagram of a convolution operation weight tensor according to an embodiment of the present application;

[0042] Figure 7 is a schematic diagram of a corresponding optical signal in a convolution operation according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] Embodiments of the present application are described below in detail, examples of which are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0044] The data processed in the convolution operation is mainly high-dimensional tensors, wherein the input tensor is a three-dimensional tensor, and the height, width and input channel number are H1, W1 and C1 respectively, and the weight tensor is a four-dimensional tensor, and the height, width, input channel number and output channel number are H2, W2, C1 and C2 respectively, and the matrix multiplication is mainly the multiplication operation of a one-dimensional vector and a two-dimensional matrix.

[0045] Figure 1 is a structure diagram of a photonic processor based on optical frequency comb and double micro-ring modulator according to an embodiment of the present application, as shown in Figure 1 The photonic processor is divided into:

[0046] an optical signal layer for generating a plurality of optical frequency comb signals containing equally spaced wavelength components, and uniformly distributing the processed optical signals;

[0047] an input signal layer for applying an electrical signal to a Mach-Zehnder modulator array to make the output optical signal perform input tensor mapping;

[0048] a weight signal layer for inputting the weight tensor of matrix operation or convolution operation as an electrical signal of a double micro-ring modulator array, and acting with the signal of the input signal layer to perform matrix operation or convolution operation;

[0049] a matrix operation summation layer for adding optical signals of different wavelengths based on a first photodetector array when performing matrix operation;

[0050] a convolution operation summation layer for coupling optical signals in different time periods to a combiner to perform summation operation when performing convolution operation, and converting the summed optical signals into electrical signals through a second photodetector array.

[0051] As shown in Figure 1 The optical signal layer includes a semiconductor laser LD, a micro-ring resonator MRR, a waveform shaper WS and a beam splitter Splitter.

[0052] Specifically, the semiconductor laser LD is used to provide a continuous wave optical signal, the micro ring resonator MRR is used to generate a plurality of optical frequency comb signals containing equidistantly separated wavelength components according to the nonlinear Kerr effect, to accelerate matrix operation and convolution operation, the waveform shaper WS is used to flatten the uneven optical frequency comb signal into an optical signal with different wavelength components but the same power, and the beam splitter Splitter is used to uniformly distribute the generated optical signal to the output ports.

[0053] In the embodiment of the application, the model of the semiconductor laser LD includes a distributed feedback semiconductor laser, a distributed Bragg reflection semiconductor laser and a vertical cavity surface emitting laser, and the beam splitter Splitter is completed by cascading a 1x2 multimode interferometer or a Y-branch binary tree structure or a directional coupler.

[0054] In a possible embodiment, the model of the semiconductor laser LD is a feedback semiconductor laser, and the beam splitter Splitter is realized by cascading a 1x2 multimode interferometer.

[0055] As shown in Figure 1 , the input signal layer includes a Mach-Zehnder modulator array MZM1-MZM3, which is used to perform mapping of the input tensor by applying an electrical signal to the Mach-Zehnder modulator array to make the output optical signal of the Mach-Zehnder modulator array perform the mapping of the input tensor.

[0056] As shown in Figure 1 , the weight signal layer includes a double micro ring modulator array R1-R9, and as shown in Figure 2 , each double micro ring modulator includes a cross waveguide, a micro ring modulator MRM_n and a micro ring modulator MRM_s, and the weight signal layer performs different processing when performing convolution operation and matrix operation.

[0057] Specifically, taking the input tensor (H1=3, W1=3, C1=3) of the convolution operation as shown in Figure 5 and the weight tensor (H2=2, W2=2, C1=3, C2=3) of the convolution operation as shown in Figure 6 , when performing convolution operation, as shown in Figure 3 , the resonant wavelength of MRM_n in R1-R9 is moved to the wavelength of the input optical signal by the thermo-optic effect, and then the resonant wavelength of MRM_n is deviated from the wavelength of the input optical signal by the electro-optic effect, the transmittance of the input optical signal through MRM_n is changed by adjusting the deviation value, to correspond to the weight value in the weight tensor, and the optical signal is transmitted to the north end, which is shown in black in Figure 3 , and the resonant wavelength of MRM_s in R1-R9 is far away from the wavelength of the input optical signal, which does not affect the input optical signal, and is in a dormant state, which is shown in gray in Figure 3 .

[0058] wherein, Figure 3 The middle dotted arrow is the direction of light transmission.

[0059] It is worth mentioning that, Figure 3 The input signal waveform of MZM1-MZM3 corresponds to the value of the input channel of the input tensor, and the number of rows of the dual-microring modulator array corresponds to the number of input channels in the weight tensor, and the number of columns corresponds to the number of output channels of the weight tensor. Taking the dual-microring modulator R1 in the first row and the first column of the dual-microring modulator array as an example, the input electrical signal waveform of MRM_n in the modulator corresponds to the value of the first input channel of the first output tensor in the weight tensor, and the specific electrical signal waveform is as shown in Figure 7 The middle figure shows.

[0060] The formula of convolution calculation is as follows:

[0061]

[0062] wherein, p, m, n, k, i, j are positive integers.

[0063] In the scheme proposed in the present application, the first convolution calculation result of the first input channel is taken as an example, and the calculation results of other input channels can be obtained in the same way. The output of the first convolution calculation is W 1,1, 1 , 1*X 1, 1 , 1+W 1,1, 1 , 2*X 1, 1 , 2+W 1,1, 2 , 1*X 1, 2 , 1+W 1,1, 2 , 2*X 1, 2 , 2, the calculation result is realized by R1 and MZM1. The input electrical signal of MZM1 is the waveform after flattening the elements of the first input channel of the input tensor. The input electrical signal of R1 is the waveform after flattening the elements of the first input channel of the first output channel of the weight tensor, as shown in Figure 7 Since the convolution operation needs to add the signals in the four different time periods in Figure 7 , a delay line array is used to unify the optical signals at different times to the same time, and then it is taken as the input port of the combiner for calculation. Among them, Figure 3 The input optical signals of Com1 in Figure 7As shown, s is the fourth input port of Com1. The sum of the signals during the time interval 3t0-4t0 in the figure is the output value of the first convolution calculation.

[0064] Specifically, with Figure 4 Taking the input vector and weight matrix of matrix operations represented by elements on a medium Mach-Zehnder modulator array and a dual micro-ring modulator array as an example, when performing matrix operations, such as... Figure 4 As shown, the resonant wavelength of MRM_s in R1-R9 is shifted to the wavelength of the input optical signal through the thermo-optic effect. Then, the electro-optic effect causes a certain deviation between the resonant wavelength of MRM_s and the wavelength of the input optical signal. By adjusting this deviation value, the transmittance of the input optical signal through MRM_s is changed, corresponding to the weight values ​​in the weight tensor, and the optical signal is transmitted to the south end. Figure 4 The image is displayed in black, and the resonant wavelength of MRM_n in R1-R9 is far from the wavelength of the input optical signal, so it will not affect the input optical signal and is in a dormant state. Figure 4 It is displayed in gray.

[0065] It should be noted that, during matrix operations, the input electrical signals of the Mach-Zehnder modulator arrays MZM1-MZM3 correspond to the elements of the input vector in the matrix operation, such as... Figure 4 The elements X1, X2, and X3 marked above MZM1-MZM3 represent the electrical signals of MRM_s in the dual micro-ring modulator array, corresponding to each element in the weight matrix. Figure 4 The elements W11, W21, W31, W12, W22, W32, W12, W23, and W33 marked below MRM_s in R1-R9.

[0066] in, Figure 4 The dashed arrow in the middle indicates the transmission direction of the optical signal that satisfies the resonance condition.

[0067] like Figure 1 As shown, the matrix operation summation layer includes a first photodetector array PD1-PD3, used to perform the summation operation in the matrix operation process. This calculation is achieved by adding optical signals of different wavelengths. The optical power received by each photodetector in the first photodetector array satisfies:

[0068] P1 = W11*X1 + W21*X2 + W31*X3;

[0069] P2 = W12*X1 + W22*X2 + W32*X3;

[0070] P3 = W13*X1 + W23*X2 + W33*X3;

[0071] Wherein, P1, P2, and P3 are the optical powers received by PD1, PD2, and PD3, respectively; X1, X2, and X3 are the optical signals output by MZM1, MZM2, and MZM3, respectively; and W11, W21, W31, W12, W22, W32, W12, W23, and W33 are the transmittances of MRM_s in R1-R9 for optical signals near the corresponding resonant wavelengths.

[0072] It should be added that the above formula can be transformed into matrix multiplication form to obtain:

[0073]

[0074] Therefore, it can be seen that the value of the optical power received by the first photodetector satisfies the result of matrix multiplication of the input vectors X1, X2, X3 and the weights W11, W21, W31, W12, W22, W32, W12, W23 and W33.

[0075] like Figure 1 As shown, the convolution operation summation layer includes optical delay line arrays DL1-DL9, directional coupler arrays DC1-DC6, multiplexer arrays Com1-Com3, and a second photodetector array PD4-PD6.

[0076] Specifically, the optical delay line arrays DL1-DL9 are used to separate optical signals located at different times, the directional coupler arrays DC1-DC6 are used to couple optical signals located at different times to the multiplexer array, the multiplexer arrays Com1-Com3 are used to combine the optical signals coupled from the directional coupler array into the same waveguide to achieve summation, and the second photodetector arrays PD4-PD6 are used to convert optical signals into electrical signals when performing convolution operations.

[0077] In addition, such as Figure 1 As shown, s1-s6 are the optical signals coupled out by the directional coupler.

[0078] This application's embodiments achieve large-scale matrix operations and convolution operations by combining the rich wavelength components of an optical frequency comb. Through the action of a dual micro-ring modulator, convolution operations and matrix multiplication operations are implemented at the north and south ends of the dual micro-ring modulator, respectively, thereby accelerating the operation of the neural network. Furthermore, the structure proposed in this application is scalable and can be monolithically integrated, adapting to the needs of large-scale linear operations.

[0079] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0080] The foregoing detailed description has set forth various embodiments of the devices and / or processes via the use of specific terminology. However, embodiments thereof can be practiced with the exact description not being set forth but with the same essence; the principles set forth herein can be practiced with plasticity in a manner leading to structurally equivalent devices and / or processes. Therefore, this description is not to be construed as limiting; the scope of the disclosure is to be measured by the breadth of the following claims.

Claims

1. A photonic processor based on optical frequency combs and dual- microring modulators, characterized in that, The application relates to a photonic tensor processing unit, comprising: a light signal layer for generating a plurality of optical frequency comb signals containing equidistantly spaced wavelength components and uniformly distributing the generated light signals; an input signal layer for applying an electrical signal to a Mach-Zehnder modulator array to make the output light signal of the Mach-Zehnder modulator array perform mapping of an input tensor; a weight signal layer for inputting a weight tensor of a matrix operation or a convolution operation as an electrical signal of a dual-microring modulator array to act on the signal of the input signal layer to perform the matrix operation or the convolution operation; a matrix operation summation layer for adding light signals of different wavelengths based on a first photodetector array when performing the matrix operation; a convolution operation summation layer for coupling light signals in different time periods to a combiner to perform summation operation and converting the summed light signals into electrical signals through a second photodetector array when performing the convolution operation.

2. The optical frequency comb and dual microring modulator based photonic processor of claim 1, wherein, The light signal layer comprises: a semiconductor laser for providing continuous wave light signals; a microring resonator for generating a plurality of optical frequency comb signals containing equidistantly spaced wavelength components according to the nonlinear Kerr effect to accelerate the matrix operation and the convolution operation; a waveform shaper for flattening the uneven optical frequency comb signals into light signals with different wavelength components but the same power; a beam splitter for uniformly distributing the generated light signals to output ports; The beam splitter is realized by cascaded 1x2 multimode interferometers, Y branches or directional couplers.

3. The optical frequency comb and dual-microring modulator based photonic processor of claim 1, wherein, The input signal layer comprises: The Mach-Zehnder modulator array composed of a plurality of Mach-Zehnder modulators is used for inputting an input tensor in the matrix operation or the convolution operation as an electrical signal of the Mach-Zehnder modulator to realize mapping of the input tensor.

4. The optical frequency comb and dual-microring modulator based photonic processor of claim 1, wherein, The weight signal layer comprises the dual-microring modulator array composed of a plurality of dual-microring modulators, each of which comprises a cross waveguide, a microring modulator MRM_n and a microring modulator MRM_s, and is used for: When performing the convolution operation, the resonant wavelength of the MRM_n is moved to the wavelength of the input light signal through the thermo-optic effect, then the resonant wavelength of the MRM_n is deviated from the wavelength of the input light signal through the electro-optic effect, the transmittance of the input light signal through the MRM_n is changed by adjusting the deviation value to correspond to the weight value in the weight tensor, and the resonant wavelength of the MRM_s is far away from the wavelength of the input light signal and is in a dormant state; When performing the matrix operation, the resonant wavelength of the MRM_s is moved to the wavelength of the input light signal through the thermo-optic effect, then the resonant wavelength of the MRM_s is deviated from the wavelength of the input light signal through the electro-optic effect, the transmittance of the input light signal through the MRM_s is changed by adjusting the deviation value to correspond to the weight value in the weight tensor, and the resonant wavelength of the MRM_n is far away from the wavelength of the input light signal and is in a dormant state.

5. The optical frequency comb and dual-microring modulator based photonic processor of claim 1, wherein, The matrix operation summation layer comprises: The first photodetector array composed of a plurality of photodetectors is used for performing summation operation in the matrix operation process, and the operation is realized by adding light signals of different wavelengths; wherein the light power received by each photodetector in the first photodetector array satisfies: P1 = W11 * 1 + 21 * 2 + 31 * 3; P2 = W12 * 1 + 22 * 2 + 32 * 3; P3 = W13 * 1 + 23 * 2 + 33 * 3; Wherein, P1, P2, P3 are the light power received by each photodetector in the first photodetector array respectively, X1, X2, X3 are the light signals output by each Mach-Zehnder modulator in the Mach-Zehnder modulator array respectively, W11, W21, W31, W12, W22, W32, W12, W23 and W33 are the transmittances of MRM_s in each dual-micro-ring modulator in the dual-micro-ring modulator array for the light signals near the corresponding resonant wavelength.

6. The optical frequency comb and dual-microring modulator based photonic processor of claim 1, wherein, The convolution operation sum layer comprises: An optical delay line array for separating the light signals at different time instants; A directional coupler array for coupling the light signals at different time instants into a combiner array; The combiner array for combining the light signals coupled out by the directional coupler array into the same waveguide to realize summation; A second photodetector array for converting the light signals into electrical signals when performing convolution operation.

Images