Integrated optical 90-degree optical mixer based on optical neural network

Abstract

The present application provides an integrated optical 90-degree optical mixer based on an optical neural network, including an optical neural network chip configured with an optimal transmission matrix, which is used to process received coherent light, generate simulated signal light and reference light, and mix the simulated signal light and reference light to form four paths of light with different phases; wherein, a gradient descent algorithm is used to train the initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix. The present application uses a gradient descent algorithm to fit the transmission matrix of the optical neural network chip to the transmission matrix of the 90-degree optical mixer to achieve the 90-degree optical mixer task. Due to its trainability, the optical 90-degree optical mixer of the present application no longer relies on complex structural design and low-error processing methods, has better robustness, and can compensate for the imbalance of output coupling of different optical paths.

Description

Technical Field

[0001] The present application relates to the field of photonic integration technology, and in particular to an integrated optical 90-degree optical mixer based on an optical neural network. Background Art

[0002] 90° optical mixers (90° hybrids) have important applications in a variety of fields, including coherent communications and coherent detection based on I / Q demodulation. During the demodulation process, the 90° optical mixer at the receiving end mixes the received signal light with the reference light, generating two sets of orthogonal signal components. The mixed signal is then demodulated to obtain the amplitude and phase information carried by the signal light. Current implementations of 90° optical mixers are all passive devices, such as 2×4 multimode interferometers (MMIs) and hybrid multimode interferometers.

[0003] However, due to the inevitable errors caused by simulation during processing and design, the degradation of performance parameters such as phase error and imbalance after the actual 90° optical mixer mixes the signal light and the reference light will lead to distortion of the demodulated signal, thereby significantly increasing the complexity of subsequent digital signal processing. Summary of the Invention

[0004] In view of the above-mentioned shortcomings of the prior art, the purpose of this application is to provide an integrated optical 90-degree optical mixer based on an optical neural network, which is used to solve the problem of demodulated signal distortion caused by the degradation of performance parameters of the existing 90-degree optical mixer.

[0005] To achieve the above-mentioned purpose and other related purposes, the first aspect of the present application provides an integrated optical 90-degree optical mixer based on an optical neural network, comprising: an optical neural network chip configured with an optimal transmission matrix, used to process the received coherent light, generate simulated signal light and reference light, and mix the simulated signal light and reference light to form four paths of light with different phases; wherein, a gradient descent algorithm is used to train the initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix.

[0006] In some embodiments of the first aspect of the present application, the optical neural network chip includes a beam splitting structure and a multimode interference cascade network; wherein the beam splitting structure is used to generate simulated signal light and reference light; and the multimode interference cascade network is used to mix the simulated signal light and reference light.

[0007] In some embodiments of the first aspect of the present application, the beam splitting structure includes a first multimode interferometer; the first multimode interferometer is connected to two first output waveguides, and one of the first output waveguides is equipped with a first modulator; wherein, one path of light output by the first multimode interferometer is controlled by the first modulator to form simulated signal light, and the other path of light is used as reference light.

[0008] In some embodiments of the first aspect of the present application, the width of the first multimode interferometer is 8 microns and the length is 58.248 microns.

[0009] In some embodiments of the first aspect of the present application, the multimode interference cascade network is provided with a plurality of forward propagation training layers in sequence along the propagation direction of light.

[0010] In some embodiments of the first aspect of the present application, each forward propagation training layer is composed of a second multimode interferometer and four second modulators; wherein the second multimode interferometer is connected to four second output waveguides, and each second output waveguide is equipped with a second modulator.

[0011] In some embodiments of the first aspect of the present application, the gradient descent algorithm is used to train the initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix. The specific process includes: obtaining the phase loaded by each second modulator in the initial multi-mode interference cascade network to obtain an initial transmission matrix; based on the initial transmission matrix, using the gradient descent algorithm to obtain the optimal transmission matrix; and performing phase control on each second modulator in the initial multi-mode interference cascade network according to the optimal transmission matrix to obtain an optical neural network chip configured with an optimal transmission matrix.

[0012] In some embodiments of the first aspect of the present application, each second multi-mode interferometer has a width of 16 microns and a length of 466.52 microns.

[0013] In some embodiments of the first aspect of the present application, the tapered length of each port of each second multi-mode interferometer is 30 microns and the width is 3 microns.

[0014] In some embodiments of the first aspect of the present application, each second modulator has a length of 250 microns and a width of 3 microns.

[0015] As described above, the integrated optical 90-degree optical mixer based on optical neural network of the present application has the following beneficial effects:

[0016] This application uses a gradient descent algorithm to fit the transmission matrix of an optical neural network chip to the transmission matrix of a 90-degree optical mixer to achieve the 90-degree optical mixer task. Due to its trainability, the optical 90-degree optical mixer in this application no longer relies on complex structural design and low-error processing methods, has better robustness, and can compensate for imbalances in the output coupling of different optical paths. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Shown is a structural schematic diagram of an integrated optical 90-degree optical mixer based on an optical neural network in one embodiment of the present application.

[0018] Figure 2 Shown is a structural schematic diagram of a first multi-mode interferometer in a specific embodiment of the present application.

[0019] Figure 3 Shown is a schematic diagram of a forward propagation training layer in one embodiment of the present application.

[0020] Figure 4 Shown is a schematic diagram of a forward propagation training layer in another embodiment of the present application.

[0021] Figure 5 Shown is a schematic diagram comparing the expected results of the output intensity change with the input phase of the optical neural network chip after training at a wavelength of 1550nm in a specific embodiment of the present application.

[0022] Figure 6 Shown is a schematic diagram of the phase error output within the wavelength range of 1530nm to 1570nm after the optical neural network chip is trained at a wavelength of 1550nm in a specific embodiment of the present application.

[0023] Figure 7 Shown is a schematic diagram of the common mode rejection ratio of an optical neural network chip in a specific embodiment of the present application after training at a wavelength of 1550nm and outputting within the wavelength range of 1530nm to 1570nm. DETAILED DESCRIPTION

[0024] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0025] In the embodiments of this application, terms such as "first" and "second" are used to distinguish identical or similar items with substantially the same functions and effects. For example, the terms "first" and "second" are used solely to distinguish different multimode interferometers and do not define their order. Those skilled in the art will understand that terms such as "first" and "second" do not define the number or order of execution, and do not necessarily imply that they are different.

[0026] It should be noted that in the embodiments of this application, words such as "exemplary" or "for example" represent examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0027] In the embodiments of the present application, "at least one" refers to one or more, and "more" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: the existence of A alone, the existence of A and B at the same time, and the existence of B alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can represent: a, b, c, ab, ac, bc or abc, where a, b, c can be single or multiple.

[0028] To facilitate understanding of the embodiments of this application, first Figure 1 Detailed description. Figure 1 A schematic block diagram of an integrated optical 90-degree optical mixer based on an optical neural network in an embodiment of the present invention is shown. The integrated optical 90-degree optical mixer based on an optical neural network in this embodiment includes:

[0029] An optical neural network chip configured with an optimal transmission matrix is ​​used to process received coherent light to generate simulated signal light and reference light, and to mix the simulated signal light and reference light to form four paths of light with different phases;

[0030] Among them, the gradient descent algorithm is used to train the initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix.

[0031] It should be noted that on-chip optical neural networks (ONNs) are an application of integrated optics. They are a new computing architecture that uses a variety of optical devices (such as optical waveguides, beam splitters, phase modulators, and optical resonant rings) to implement neural network calculations. By encoding information into physical properties such as light intensity, phase, or polarization, this technology can complete the core matrix operations in neural networks in a high-speed, parallel, and low-power manner.

[0032] Integrated photonics technology can integrate high-density, compact optical links onto a single chip. Its high compatibility with complementary metal oxide semiconductor (CMOS) processes significantly reduces manufacturing costs and improves production efficiency and reliability. Integrated photonics has broad potential for applications in future data centers, high-speed interconnect systems, optical communications, and quantum computing.

[0033] In one embodiment, the four paths of light with different phases have relative phases of 0 degrees, 90 degrees, 180 degrees, and 270 degrees, respectively.

[0034] In one embodiment, the optical neural network chip includes a silicon substrate, a silicon dioxide lower cladding layer, a silicon device layer, a silicon dioxide upper cladding layer, a modulator layer and an aluminum wire layer arranged in sequence from bottom to top.

[0035] In one embodiment, an optical neural network chip includes an input grating, a beam splitting structure, a multi-mode interference cascade network, and an output grating.

[0036] In one embodiment, if Figure 1 As shown, the beam splitting structure includes a first multimode interferometer 1; the first multimode interferometer 1 is connected to two first output waveguides 2, and one of the first output waveguides 2 is equipped with a first modulator 3; wherein, one path of light output by the first multimode interferometer 1 is controlled by the first modulator 3 to form a simulated signal light, and the other path of light is used as a reference light.

[0037] In a specific embodiment, the first multimode interferometer 1 is a 1×2 multimode interferometer (MMI), which has one input port and two output ports. Preferably, in order to reduce insertion loss, Figure 2 As shown, the width W_splitter of the first multimode interferometer 1 is 8 um, the length L_splitter of the first multimode interferometer 1 is 58.248 um, and the insertion loss is 0.046 dB after simulation.

[0038] In one embodiment, if Figure 1As shown, the two output ports of the first multimode interferometer 1 are respectively connected to a first output waveguide 2 .

[0039] In one embodiment, the first multimode interferometer 1 and the two first output waveguides 2 are disposed in a silicon device layer, and the first modulator 3 is disposed in a modulator layer.

[0040] In one embodiment, the first modulator may be a thermo-optic modulator or an electro-optic modulator, which is not limited in the present invention.

[0041] In one embodiment, if Figure 3 as well as Figure 4 As shown, the multimode interference cascade network is provided with multiple forward propagation training layers in sequence along the propagation direction of light. Specifically, the number of forward propagation training layers is generally 4, namely forward propagation training layer A, forward propagation training layer B, forward propagation training layer C, and forward propagation training layer D.

[0042] In one embodiment, if Figure 1 、 Figure 3 as well as Figure 4 As shown, each forward propagation training layer is composed of a second multimode interferometer 4 and four second modulators 5; wherein, the second multimode interferometer 4 is connected to four second output waveguides 7, and each second output waveguide 7 is equipped with a second modulator 5.

[0043] In one embodiment, Figure 4 As shown, the second multimode interferometer 4 utilizes a 4×4 multimode interferometer (MMI); wherein the 4×4 multimode interferometer has four input ports and four output ports. Preferably, to reduce insertion loss, the width W_MMI of the second multimode interferometer 4 is 16 μm, and the length L_MMI of the second multimode interferometer 4 is 466.52 μm. The taper length L_taper of each port of the second multimode interferometer 4 is 30 μm, and the taper width W_taper of each port of the second multimode interferometer 4 is 3 μm. Simulations show an insertion loss of 0.11 dB.

[0044] In one embodiment, for the convenience of description, the second multimode interferometer in the forward propagation training layer A is referred to as the second multimode interferometer A, the second multimode interferometer in the forward propagation training layer B is referred to as the second multimode interferometer B, the second multimode interferometer in the forward propagation training layer C is referred to as the second multimode interferometer C, and the second multimode interferometer in the forward propagation training layer D is referred to as the second multimode interferometer D.

[0045] like Figure 1 、 Figure 3 as well as Figure 4As shown, the two first output waveguides 2 are connected in one-to-one correspondence to any two input ports of the second multimode interferometer A, and the four output ports of the second multimode interferometer A are respectively connected to a second output waveguide 7. The four second output waveguides 7 connected to the second multimode interferometer A are respectively connected in one-to-one correspondence to the four input ports of the second multimode interferometer B. The four output ports of the second multimode interferometer B are respectively connected to a second output waveguide 7. The four second output waveguides 7 connected to the second multimode interferometer B are respectively connected to the four input ports of the second multimode interferometer C. The four output ports of the second multimode interferometer C are respectively connected to a second output waveguide 7. The four second output waveguides 7 connected to the second multimode interferometer C are respectively connected to the four input ports of the second multimode interferometer D.

[0046] In a specific implementation, the length L_heater of the second phase modulator is 250 μm, and the width W_heater of the second phase modulator is 3 μm. π About 5V.

[0047] In one embodiment, the four second multimode interferometers in the multimode interference cascade network and the second output waveguides connected to the four second multimode interferometers are all disposed in the silicon device layer, and each second modulator in the multimode interference cascade network is disposed in the modulator layer.

[0048] In one embodiment, the second modulator may be a thermo-optic modulator or an electro-optic modulator, which is not limited in the present invention.

[0049] The following will describe the light propagation path with reference to the accompanying drawings: Figure 1 as well as Figure 2 As shown, coherent light is input from the input grating and then split by the first multimode interferometer 1. The first modulator 3 phase modulates one of the two light paths output by the first multimode interferometer 1 to form simulated signal light, and the other light path is used as reference light.

[0050] For the convenience of description, the second multimode interferometer in the forward propagation training layer A is referred to as the second multimode interferometer A, the second multimode interferometer in the forward propagation training layer B is referred to as the second multimode interferometer B, the second multimode interferometer in the forward propagation training layer C is referred to as the second multimode interferometer C, and the second multimode interferometer in the forward propagation training layer D is referred to as the second multimode interferometer D.

[0051] Furthermore, the simulated signal light can be input from the first port of the second multimode interferometer A, and the reference light can be input from the third port of the second multimode interferometer A. The second multimode interferometer A interferes with the simulated signal light and the reference light. The four-path light output from the second multimode interferometer A is phase-modulated by the four second modulators of the forward propagation training layer A and then transmitted to the second multimode interferometer B. The second multimode interferometer B interferes with the four-path light entering the second multimode interferometer B. The four-path light output from the second multimode interferometer B is phase-modulated by the four second modulators of the forward propagation training layer B and then transmitted to the second multimode interferometer C. The second multimode interferometer C interferes with the four-path light entering the second multimode interferometer C. The four-path light output from the second multimode interferometer C is phase-modulated by the four second modulators of the forward propagation training layer C and then transmitted to the second multimode interferometer D. The second multimode interferometer D interferes with the four lights entering the second multimode interferometer D. The four lights output by the second multimode interferometer D are phase-modulated by the four second modulators of the forward propagation training layer D, and then coupled out through the output grating 6 to form four lights with different phases.

[0052] In one embodiment, the specific process of using a gradient descent algorithm to train an initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix includes: obtaining the phase loaded by each second modulator in the initial multimode interference cascade network to obtain an initial transmission matrix; using the gradient descent algorithm based on the initial transmission matrix to obtain an optimal transmission matrix; and controlling the phase of each second modulator in the initial multimode interference cascade network according to the optimal transmission matrix to obtain an optical neural network chip configured with the optimal transmission matrix. It should be noted that an untrained optical neural network chip is defined as an initial optical neural network chip.

[0053] Specifically, before training the optical neural network chip, the wavelength of the coherent light input to the chip is determined, and the phase loaded by each second modulator in the initial multimode interference cascade network is obtained. The phases loaded by each second modulator constitute the initial transmission matrix. Based on this initial transmission matrix, the optimal transmission matrix is ​​obtained using the ADAM gradient descent algorithm.

[0054] The ADAM gradient descent algorithm uses the loss function shown in the following formula 1:

[0055]

[0056] Among them, a represents the custom parameter for controlling the equivalent common mode rejection ratio, b represents the custom parameter for controlling the phase error, and I imax I represents the normalized maximum value of the i-th output of the optical neural network chip after input phase sweep; iminRepresents the normalized minimum value of the i-th output of the optical neural network chip after input phase sweep. i It represents the phase of the i-th output of the optical neural network chip after the input phase is scanned and the fast Fourier transform (FFT) is performed.

[0057] Furthermore, the phase of each second modulator in the initial multimode interference cascade network is controlled based on the optimal transmission matrix to obtain an optical neural network chip configured with the optimal transmission matrix. For example, the second modulator is a thermo-optical modulator. By changing the temperature of the thermo-optical modulator, the phase of light in the corresponding second transmission waveguide is changed.

[0058] Furthermore, after all the second modulators are regulated, no further regulation is performed on the second modulators during use. Those skilled in the art can obtain different outputs of the optical neural network chip by regulating the first modulators to meet usage requirements.

[0059] In one embodiment, Figure 5 The figure shows the results after the optical neural network chip is trained at a wavelength of 1550nm. After changing the phase of the input signal, the intensity of the four output lights of the optical neural network chip changes with the phase of the input signal and is compared with the theoretical target of an ideal 90° optical mixer.

[0060] In one embodiment, Figure 6 As shown, the relative phase error performance of the optical neural network chip trained at 1550nm wavelength in the 1530nm-1570nm band is demonstrated. i The difference between the phase output at port i and the expected phase at port i is less than ±5° in the range of 1530nm-1560nm, and the phase error is less than ±2° at the target wavelength of 1550nm.

[0061] In one embodiment, Figure 7 The figure shows the equivalent common mode rejection ratio performance of the optical neural network chip trained at 1550nm in the 1530nm-1570nm band. At 1550nm, it is less than -40dB, and less than -20dB in the 1530nm-1570nm range.

[0062] It should be noted that this application is compatible with a variety of semiconductor process platforms, does not rely on modulation methods, is conducive to cost reduction, and has strong scalability. Compared with the existing 90° optical mixer, this application has better processing robustness. By training the optical neural network chip, parameters such as phase error and common mode rejection ratio can be optimized, and it does not rely on complex structural design and low-error processing methods. This application can compensate for the imbalance caused by the coupling of the four output ports by regulating the on-chip light, without relying on subsequent signal processing.

[0063] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0064] In summary, the present application provides an integrated optical 90-degree optical mixer based on an optical neural network, including an optical neural network chip configured with an optimal transmission matrix, which is used to process the received coherent light, generate simulated signal light and reference light, and mix the simulated signal light and reference light to form four paths of light with different phases; wherein, a gradient descent algorithm is used to train the initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix. The present application uses a gradient descent algorithm to fit the transmission matrix of the optical neural network chip to the transmission matrix of the 90-degree optical mixer to achieve the 90-degree optical mixer task. Due to its trainability, the optical 90-degree optical mixer of the present application no longer relies on complex structural design and low-error processing methods, has better robustness, and can compensate for the imbalance of output coupling of different optical paths. Therefore, the present application effectively overcomes the various shortcomings of the prior art and has high industrial utilization value.

[0065] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical concepts disclosed in this application shall be covered by the claims of this application.

Claims

1. An integrated optical 90-degree optical mixer based on an optical neural network, characterized in that: include: An optical neural network chip configured with an optimal transmission matrix is ​​used to process received coherent light to generate simulated signal light and reference light, and to mix the simulated signal light and reference light to form four paths of light with different phases; The initial optical neural network chip is trained using a gradient descent algorithm to obtain an optical neural network chip configured with an optimal transmission matrix; The optical neural network chip includes a beam splitting structure and a multimode interference cascade network; wherein the beam splitting structure is used to generate simulated signal light and reference light; the multimode interference cascade network is used to mix the simulated signal light and reference light; the beam splitting structure includes a first multimode interferometer; the first multimode interferometer is connected to two first output waveguides, and one of the first output waveguides is equipped with a first modulator; wherein one path of light output by the first multimode interferometer is controlled by the first modulator to form simulated signal light, and the other path of light serves as reference light; The multimode interference cascade network is provided with a plurality of forward propagation training layers in sequence along the propagation direction of light; each forward propagation training layer is composed of a second multimode interferometer and four second modulators; wherein the second multimode interferometer is connected to four second output waveguides, and each second output waveguide is equipped with a second modulator; The specific process of using the gradient descent algorithm to train the initial optical neural network chip to obtain an optical neural network chip configured with an optimal transmission matrix includes: Obtaining the phase loaded by each second modulator in the initial multi-mode interference cascade network to obtain an initial transmission matrix; Based on the initial transmission matrix, an optimal transmission matrix is ​​obtained using an ADAM gradient descent algorithm; The phase of each second modulator in the initial multi-mode interference cascade network is controlled according to the optimal transmission matrix to obtain an optical neural network chip configured with the optimal transmission matrix.

2. The integrated optical 90-degree optical mixer based on optical neural network according to claim 1, characterized in that: The first multi-mode interferometer has a width of 8 microns and a length of 58.248 microns.

3. The integrated optical 90-degree optical mixer based on optical neural network according to claim 1, characterized in that: Each second multimode interferometer has a width of 16 micrometers and a length of 466.52 micrometers.

4. The integrated optical 90-degree optical mixer based on optical neural network according to claim 1, characterized in that: The tapered length of each port of each second multi-mode interferometer is 30 microns and the width is 3 microns.

5. The integrated optical 90-degree optical mixer based on optical neural network according to claim 1, characterized in that: Each second modulator has a length of 250 microns and a width of 3 microns.

Images