JP7420356B2 - Photonic neural network on silicon substrate based on tunable filter and its modulation method - Google Patents

Description

Cross-reference of related applications

This application is proposed based on a Chinese patent application with application number 202010620808.5 and filing date of July 1, 2020, and claims priority to this Chinese patent application, and all contents of this Chinese patent application are referred to. Incorporated into this application as.

The present disclosure relates to the field of optical communication technology, and in particular to a photonic neural network on a silicon substrate based on a tunable filter and a modulation method thereof.

In recent years, with the rapid development of artificial intelligence and big data technology, the requirements for the computing power of chips, especially the rate and power consumption, are increasingly increasing. Conventional electronic chips are limited by Moore's Law, which slows their growth in integration and computing power. Photonic chips are attracting increasing attention as information processing carriers because they have ultra-high transmission speeds and ultra-low power consumption. Among them, silicon photonic chips are manufactured using silicon-on-insulator (SOI) materials, are compatible with the traditional CMOS (complementary metal-oxide-semiconductor) process, and have high rates, low power consumption, and low power consumption. It has advantages such as consumption, high integration and immunity to electromagnetic interference, and is currently the main implementation form of optical chips.

The present disclosure is directed, at least to some extent, to solving one of the technical problems in the related art.

An embodiment of the present disclosure is a linear computing network including a plurality of input ends for receiving a plurality of input optical signals and a plurality of output ends for outputting a plurality of first optical signals, the network comprising: The plurality of input optical signals and the plurality of first optical signals have a one-to-one correspondence, and the optical intensity of each input optical signal and the optical intensity of the corresponding first optical signal have a linear relationship. a plurality of photoelectric detectors for receiving the plurality of first optical signals and converting the plurality of first optical signals into a plurality of electrical signals; and a plurality of second optical signals. a plurality of light sources for outputting, each connected to one photoelectric detector and one light source, and the electrical signal output from the photoelectric detector and the second optical signal output from the light source, respectively; a plurality of tunable filters for receiving the electrical signal, modulating the second optical signal based on the electrical signal, and outputting a third optical signal, the optical intensity of the third optical signal and The optical intensity of the corresponding first optical signal provides a photonic neural network on a silicon substrate based on tunable filters having a non-linear relationship.

In embodiments of the present disclosure, the linear operation network includes a Mach-Zehnder interferometer (MZI) network or a Direct Coupler (DC) network.

In an embodiment of the present disclosure, the input optical signal , the first optical signal, the second optical signal , and the third optical signal have the same wavelength.

In embodiments of the present disclosure, the tunable filter includes a Bragg reflection grating filter.

In an embodiment of the present disclosure, the Bragg reflection grating filter includes a heating layer for receiving the electrical signal, and is located below the heating layer, and a first end thereof receives the second optical signal. a Bragg grating, the second end of which is used for outputting the third optical signal .

In embodiments of the present disclosure, the tunable filter is realized by changing the effective refractive index of the Bragg grating through thermal modulation or electrical modulation.

In embodiments of the present disclosure, the tunable filter includes an FP cavity filter.

In an embodiment of the present disclosure, the FP cavity filter includes a heating layer for receiving the electrical signal, an optical waveguide located under the heating layer, and a first end of the optical waveguide. a first distributed Bragg reflector for receiving the second optical signal; and a second distributed Bragg mirror connected to the second end of the optical waveguide for outputting the third optical signal. including a reflector .

In an embodiment of the present disclosure, the optical signal is input to the heating layer to change the effective refractive index of the optical waveguide, thereby converting the second optical signal passing through the optical waveguide into the third optical signal. Change to

In an embodiment of the present disclosure, the MZI network includes a plurality of basic units, the basic units include a coupler, an external phase shifter, and an internal phase shifter, and the coupler connects the external phase shifter and the internal phase shifter. Located between the phase shifters.

In an embodiment of the present disclosure, the DC network includes a plurality of basic units, and the basic units include a modulation arm and a tunable DC coupler.

In an embodiment of the present disclosure, the center wavelength of the tunable filter is the same as the wavelength of the second optical signal in an initial or unmodulated state.

In an embodiment of the present disclosure, when the electrical signal applied to the tunable filter changes, the center wavelength of the tunable filter shifts, and the transmitted intensity of the light wave at the wavelength of the second optical signal changes, The optical intensity of the third optical signal output from the tunable filter and the optical intensity of the first optical signal output from the linear network exhibit a nonlinear relationship.

Embodiments of the present disclosure provide a photonic integrated circuit including a photonic neural network on a silicon substrate based on a plurality of tunable filters as described in any one above.

An embodiment of the present disclosure includes a step of performing a linear operation on a plurality of input optical signals to obtain a plurality of first optical signals, a step of converting the plurality of first optical signals into a plurality of electrical signals, and a step of converting the plurality of first optical signals into a plurality of electrical signals. and modulating the plurality of second optical signals using the plurality of electrical signals to obtain the plurality of third optical signals, An optical signal modulation method is proposed in which the optical intensity of the signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship.

In an embodiment of the present disclosure, the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength.

Additional aspects and advantages of the disclosure will be set forth in part in the description that follows, and in part will be obvious from the description, or may be learned by practice of the disclosure.

The above and/or additional aspects and advantages of the present disclosure will become apparent and easily understood from the description of the embodiments below in conjunction with the drawings.
1 is a structural schematic diagram of a photonic neural network on a silicon substrate based on a tunable filter according to an embodiment of the present disclosure; FIG. 1 is a structural schematic diagram of an MZI network according to an embodiment of the present disclosure; FIG. 1 is a structural schematic diagram of a DC network according to an embodiment of the present disclosure; FIG. 1 is a schematic diagram of a filter according to an embodiment of the present disclosure; FIG. 1 is a schematic diagram of a filter according to an embodiment of the present disclosure; FIG. FIG. 2 is a schematic diagram of an output spectrum of a filter according to an embodiment of the present disclosure. FIG. 3 is a nonlinear relationship diagram of filters according to an embodiment of the present disclosure. 3 is a flowchart of a method for modulating an optical signal according to an embodiment of the present disclosure.

Hereinafter, embodiments of the present disclosure will be described in detail, examples of which are illustrated in the drawings, and the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. represent. The embodiments described below with reference to the drawings are illustrative and are intended to interpret the present disclosure, and should not be understood as limiting the present disclosure.

Machine learning based on large-scale data processing is an important implementation method of artificial intelligence technology, and among these, artificial neural networks (ANNs) algorithms are one of the algorithms often used for machine learning. Currently, computing is mainly implemented using electric CMOS chips. With the increase in data volume and computing complexity, traditional CMOS computing chips can no longer fully meet the computing demands of neural networks. Compared to electric CMOS chips, neural networks realized by photonic technology are significantly enhanced in terms of computing speed (the computing speed of photonic artificial intelligence chips is approximately a thousand times that of electronic chips), and at the same time Power consumption is much lower than electric chips. Artificial neural networks include an input layer, a hidden layer, and an output layer.In each layer, information is propagated by linear combinations such as matrix multiplication, and then processed by a nonlinear activation function.Specific operations are mainly matrix multiplications. Contains calculation and non-linear activation function parts. Currently, research on photonic neural network chips is focused on optically realizing the linear operation, that is, the matrix multiplication operation part, in artificial neural networks, such as silicon-based MZI network structure or micro- A micro-ring weight bank is used to perform matrix operations. However, the nonlinear activation function part is still implemented using an electrical method, and on-chip integration with the matrix calculation part cannot be realized, so there are still problems of low calculation speed and high power consumption.

In summary, in order to adapt to the rapid development of the big data era, it is necessary to realize photonic computing chips suitable for on-chip neural networks to develop optical computing instead of electrical computing. This is the main route and is the focus of current research. The realization of nonlinear activation functions by optics or on-chip integration technology is still in the exploration stage, and photonic There is an urgent need to find excitation functions that are compatible with neural networks.

FIG. 1 is a structural schematic diagram of a photonic neural network on a silicon substrate based on a tunable filter according to an embodiment of the present disclosure.

As shown in FIG. 1, the photonic neural network 12 on a silicon substrate based on the tunable filter includes a linear arithmetic network 1 and a nonlinear arithmetic on-chip integrated structure 6.

The linear operation network 1 includes a plurality of input terminals 1IN and a plurality of output terminals 1OUT. In the embodiment shown in FIG. 1, the linear operation network 1 includes four input terminals 1IN and four output terminals 1OUT to realize a 4×4 linear operation, but it is understood that this is just an example. Should. The present disclosure can also implement other linear operations such as 6×6, 8×8, etc., and is not limited to this in the embodiments of the present disclosure. Multiple input terminals 1IN are used to receive multiple input optical signals. The plurality of output ends 1OUT are used to output the plurality of first optical signals. The plurality of input optical signals and the plurality of first optical signals have a one-to-one correspondence relationship. The optical intensity of each input optical signal and the optical intensity of the corresponding first optical signal have a linear relationship.

The on-chip integrated structure 6 includes a plurality of photodetectors 2 , a plurality of light sources 3 and a plurality of tunable filters 4 . The plurality of photoelectric detectors 2 are used to receive the plurality of first optical signals and convert the plurality of first optical signals into the plurality of electrical signals. That is, each photoelectric detector 2 receives one corresponding first optical signal among the plurality of first optical signals and converts the corresponding first optical signal into an electrical signal. Each light source 3 outputs one second optical signal. Each tunable filter 4 is connected to one corresponding photodetector 2 and one corresponding light source 3. Each tunable filter 4 receives the electrical signal output from the corresponding photoelectric detector 2 and the second optical signal output from the corresponding light source 3, and modulates the second optical signal based on the electrical signal. and is used to output the third optical signal 5.

In the embodiment of the present disclosure, the optical intensity of the third optical signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship, thereby realizing the function of an activation function in a neural network.

In embodiments of the present disclosure, the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength.

As shown in FIG. 1, taking a 4×4 photonic neural network as an example, the 4×4 linear operation 1 is a linear operation part, and the nonlinear operation-on-chip integrated structure 6 is a module that realizes a nonlinear activation function. , the module realizing the nonlinear activation function is composed of a photoelectric detector, a light source and a tunable filter.

In embodiments of the present disclosure, the linear computation network 1 includes a Mach-Zehnder interferometer (MZI) network or a Direct Coupler (DC) network.

For example, the plurality of input terminals of the MZI network are optical signal input terminals, and the plurality of output terminals are connected to the input terminals of a plurality of photoelectric detectors, and are used to perform linear operations on the input optical signals.

Specifically, the linear operation part of the photonic neural network is realized based on the MZI network structure, and is, for example, a linear operation of a matrix multiplication operation.

The MZI network has a matrix structure and includes multiple basic units. As shown in FIG. 2, a basic unit structure 7 of an MZI network and an expanded structure 11 of the MZI basic unit are shown. The basic unit structure 7 includes a coupler 8, an external phase shifter 9 and an internal phase shifter 10, the coupler 8 being located between the external phase shifter 9 and the internal phase shifter 10.

The DC network has a matrix structure and includes multiple basic units. As shown in FIG. 3, a basic unit structure 7 of the DC network and an expanded structure 7-3 of the DC network are shown. The basic unit structure 7 includes a modulation arm 7-1 and a tunable DC coupler 7-2.

A linear operation is performed on the input optical signal using the MZI network or the DC network, and the first optical signal obtained after the linear operation is input to the photoelectric detector. A photoelectric detector converts the first optical signal into an electrical signal. The output end of the photoelectric detector is connected to the electrode input end of the tunable filter. The photoelectric detector inputs an electrical signal to the electrode input terminal of the tunable filter, and applies it to the tunable filter as a modulation signal of the tunable filter.

Furthermore, a tunable filter is realized by changing the effective refractive index of the optical waveguide by thermal modulation or electrical modulation due to the plasma dispersion effect of the material.

Specifically, the tunable filter includes a thermal electrode or a carrier-doped layer, and the thermal electrode transfers thermal energy to the optical waveguide to change the effective refractive index of the optical waveguide, or the carrier-doped layer forms a thermal resistance. However, the thermal resistance generates heat when a current is applied, causing a change in the optical waveguide refractive index.

Furthermore, a tunable filter may be realized by changing the effective refractive index of the optical waveguide through electro-optic modulation due to the plasma dispersion effect of the material.

In embodiments of the present disclosure, the tunable filter may be a thermally modulated Bragg reflection grating filter or an FP cavity filter based on an SOI optical waveguide (thin-film silicon material (Silicon-on-insulator, SOI)).

As shown in FIG. 4, a tunable Bragg grating filter structure is shown, 13 is a silicon base, 14 is a silica buried oxide layer, 15 is a heating layer (the structure is not limited), 17 is a Bragg reflection grating (the structure is not limited).

The heating layer 15 is used to receive an electrical signal, the Bragg grating 17 is located below the heating layer 15, and the first end of the grating 17 is used to receive a second optical signal. , the second end of the grating 17 is used to output a third optical signal.

By inputting an optical signal into the heating layer 15 and changing the effective refractive index of the grating 17, the second optical signal passing through the grating 17 is changed into a third optical signal.

As shown in FIG. 5, a tunable FP cavity filter structure is shown, 13 is a silicon base, 14 is a silica buried oxide layer, 15 is a heating layer (the structure is not limited), 16 is an optical waveguide, and 17 is a Bragg reflection grating (the structure is not limited).

Specifically, the heating layer 15 is used to receive electrical signals, the optical waveguide 16 is located below the heating layer 15, and the first distributed Bragg reflector 17 is connected to the first distributed Bragg reflector 17 of the optical waveguide 16. The second distributed Bragg reflector 17 is connected to the second end of the optical waveguide 16 and is used to receive a second optical signal, and the second distributed Bragg reflector 17 is connected to the second end of the optical waveguide 16 and used to output a third optical signal. used for.

In the embodiment of the present disclosure, a first distributed Bragg reflector 17 and a second distributed Bragg reflector 17 are added on both sides of the FP cavity to constitute a DBR (Distributed Bragg Reflector) FP cavity filter. The second optical signal enters the DBR FP cavity filter via the optical waveguide, first enters the first distributed Bragg reflector 17, and the light wave with the center wavelength is transmitted between the first distributed Bragg reflector 17 and the second distributed Bragg reflector 17. The light waves interfere in the FP cavity between the two distributed Bragg reflectors 17 to form a standing wave, and the transmitted light wave is output as a third light signal via the waveguide. The reflectance is set using gratings on both sides, and the length of the FP cavity is rationally set so that there is a certain variation between the transmitted center wavelength and the wavelength of the second optical signal. If the FP cavity is thermally or electrically modulated, the center wavelength will shift to the wavelength of the second optical signal, and the transmission at the wavelength of the second optical signal will be higher, resulting in a suitable transmission curve. It will be done. Thereby, it is realized that the third optical signal light intensity and the first optical signal light intensity exhibit a nonlinear relationship.

As shown in FIG. 1, the light source is provided at the optical signal input end of the tunable filter and solely provides the optical signal to the tunable filter. The tunable filter modulates the electrical signal and the optical signal provided by the light source, and outputs a third optical signal, a light wave 5, at its output end.

The light source is an input optical signal of the tunable filter, and the input optical signal and the optical signal after linear calculation have the same wavelength. In the initial non-modulated state, the center wavelength of the tunable filter and the wavelength of the input optical signal are equal, and as shown in Figure 6, at this time, the reflectance at the center wavelength of the tunable filter is the largest and the transmittance is the smallest. , the output light intensity is close to 0. As the optical intensity of the output optical signal of the linear operation changes, the modulated electrical signal applied to the tunable filter changes, the center wavelength of the tunable filter shifts to the right, and the transmitted intensity of the optical wave at the wavelength of the input optical signal becomes Accordingly, when the center wavelength of the filter shifts, the transmittance of light waves at the wavelength of the input optical signal changes. As a result, the output light intensity of the filter and the output light intensity of the MZI matrix exhibit a nonlinear relationship.

As shown in Figure 7, if the bandwidth and free spectral range (FSR) of the tunable filter are designed within a certain range, and the output light intensity of the MZI matrix is small, the output light intensity of the filter will also be small. , when the output light intensity of the MZI matrix is large, the output light intensity of the filter is also large. Therefore, for the output optical signal of the MZI network, the transmission of light waves in the tunable filter naturally realizes a function approximating tan h , thereby realizing a nonlinear function and increasing the speed and energy efficiency ratio of calculation. be able to.

Furthermore, in one embodiment of the present disclosure, a photonic neural network composed of an MZI network, a photodetector, a light source, and a tunable filter is cascaded to obtain a multilayer photonic neural network.

Specifically, the output optical signal of the tunable filter can be directly input into the next layer MZI network structure to realize a cascade of multilayer networks. The input optical signal for the nonlinear activation function part of each layer is input from a single light source, and there is a positive correlation between the output optical intensity and the input optical intensity. The attenuation problem is also solved.

According to embodiments of the present disclosure, a linear operation is performed on the input light wave by the MZI network or the DC network, and the photoelectric detector is used to convert the optical signal outputted after the MZI network or the DC linear operation into an electrical signal. , the light source provides an optical signal to the tunable filter, and the tunable filter modulates the electrical signal and the optical signal provided by the light source and outputs a light wave. As a result, a photonic neural network chip with low power consumption, low delay, high energy efficiency ratio, and high precision is realized, and the input optical signal of the nonlinear activation function part of each layer is input from a single light source, and the output light There is a positive correlation between the intensity and the input light intensity, and at the same time, the problem that the optical signal intensity is attenuated due to insertion loss and transmission loss when cascading multiple layers is also solved.

Hereinafter, a modulation method using a photonic neural network on a silicon substrate based on a tunable filter proposed according to an embodiment of the present disclosure will be described with reference to the drawings.

FIG. 8 is a flowchart of a modulation method using a photonic neural network on a silicon substrate based on a tunable filter according to one embodiment of the present disclosure.

As shown in FIG. 8, the modulation method using a photonic neural network on a silicon substrate based on the tunable filter includes the following steps.

S1: Perform a linear operation on a plurality of input optical signals to obtain a plurality of first optical signals.

S2, converting the plurality of first optical signals into a plurality of electrical signals.

S3, obtain a plurality of second optical signals.

S4, modulating the plurality of second optical signals using the plurality of electrical signals to obtain a plurality of third optical signals.

Specifically, a linear operation is performed on a plurality of input optical signals using an MZI network, a plurality of first optical signals obtained after the linear operation are inputted to a plurality of photoelectric detectors, and the optical signal is converted into an electric signal by the photoelectric detector. A tunable filter is realized by converting it into a signal, inputting the electrical signal to a tunable filter, and changing the effective refractive index of the optical waveguide by thermal modulation or electrical modulation. A second optical signal having the same wavelength as that of the linear operation is applied to the second optical signal, which is modulated based on the electrical signal and the second optical signal provided by the light source, and a third optical signal is output at the output end.

In the embodiment of the present disclosure, the optical intensity of the third optical signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship, thereby realizing the function of an activation function in a neural network.

In embodiments of the present disclosure, the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength.

More specifically, after the linear calculation part is finished, the light wave is output from the MZI matrix, the optical signal is converted into an electrical signal by the on-chip integrated photoelectric detector, and the optical signal is converted into an electrical signal as a modulation signal of the tunable filter. In addition, the effective refractive index of the optical waveguide is changed by thermal or electrical modulation. An optical signal having the same wavelength as that of the linear calculation is input to the tunable filter, and is used as the input optical signal of the tunable filter. The tunable filter is designed so that the center wavelength of the tunable filter and the wavelength of the input optical signal are equal in the initial non-modulated state.

It should be noted that the explanation of the above embodiment of the photonic neural network also applies to the method of the embodiment, and will not be repeated here.

In the modulation method using a photonic neural network on a silicon substrate based on a tunable filter proposed according to an embodiment of the present disclosure, a linear operation is performed on an input light wave by an MZI network, and an optical signal obtained after the linear operation is used as a photonic neural network. By inputting the optical signal into a detector, converting the optical signal into an electrical signal by a photoelectric detector, and inputting the electrical signal into a tunable filter, and changing the effective refractive index of the optical waveguide by thermal modulation or electrical modulation, A tunable filter is realized, and a light source applies an optical signal with the same wavelength as the linear calculation to the tunable filter, modulates it based on the electrical signal and the optical signal provided by the light source, and outputs a light wave at the output end. do. As a result, a photonic neural network chip with low power consumption, low delay, high energy efficiency ratio, and high precision is realized, and the input optical signal of the nonlinear activation function part of each layer is input from a single light source, and the output light There is a positive correlation between the intensity and the input light intensity, and at the same time, the problem of attenuation of the optical signal intensity due to insertion loss and transmission loss when cascading multiple layers is also solved.

Note that the terms "first" and "second" are for descriptive purposes only, and imply the meaning of indicating or suggesting relative importance or indicating the number of technical features indicated. Don't understand that. Thereby, the features defined as "first" and "second" may explicitly or implicitly include at least one such feature. In the description of this disclosure, unless clearly and specifically limited otherwise, "plurality" means at least two, such as two, three, etc.

In the description herein, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" refer to the description of the embodiment. or means that a specific feature, structure, material, or characteristic described in conjunction with an example is included in at least one embodiment or example of the present disclosure. As used herein, the general descriptions of the above terms are not necessarily referring to the same embodiment or example. Additionally, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, those skilled in the art can make combinations and combinations of the different embodiments or examples and features of the different embodiments or examples described herein, unless mutually contradictory.

Although the embodiments of the present disclosure have been shown and described in the preamble, the above embodiments are illustrative and should not be understood as limiting the present disclosure, and those skilled in the art can Changes, modifications, substitutions and variations may be made to the embodiments described above.
The invention described in the original claims of the present invention will be added below.
[C1]
a plurality of input ends for receiving a plurality of input optical signals;
a plurality of output ends for outputting a plurality of first optical signals, the plurality of input optical signals and the plurality of first optical signals having a one-to-one correspondence relationship; a linear calculation network having a linear relationship between the optical intensity of each input optical signal and the optical intensity of the corresponding first optical signal;
a plurality of photoelectric detectors for receiving the plurality of first optical signals and converting the plurality of first optical signals into a plurality of electrical signals;
a plurality of light sources for outputting a plurality of second optical signals;
each connected to one photoelectric detector and one light source, and receiving the electrical signal outputted from the photoelectric detector and the second optical signal outputted from the light source, respectively, and based on the electrical signal; a plurality of tunable filters for modulating the second optical signal and outputting a third optical signal,
A photonic neural network on a silicon substrate based on a tunable filter, wherein the optical intensity of the third optical signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship.
[C2]
The photonic neural network on a silicon substrate based on a tunable filter according to C1 , wherein the linear operation network includes a Mach-Zehnder interferometer (MZI) network or a direct coupler (DC) network.
[C3]
On a silicon substrate based on a tunable filter according to C1 or 2, wherein the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength. photonic neural network.
[C4]
A photonic neural network on a silicon substrate based on a tunable filter according to any one of C1 to C3, wherein the tunable filter includes a Bragg reflection grating filter.
[C5]
The Bragg reflection grating filter is
a heating layer for receiving the electrical signal;
a Bragg grating located below the heating layer, a first end used to receive the second optical signal and a second end used to output the third optical signal; A photonic neural network on a silicon substrate based on the tunable filter described in C4, comprising:
[C6]
A photonic neural network on a silicon substrate based on the tunable filter according to C5, characterized in that the tunable filter is realized by changing the effective refractive index of the Bragg grating by thermal modulation or electrical modulation. .
[C7]
A photonic neural network on a silicon substrate based on a tunable filter according to any one of C1 to C3, wherein the tunable filter includes an FP cavity filter.
[C8]
The FP cavity filter is
a heating layer for receiving the electrical signal;
an optical waveguide located under the heating layer;
a first distributed Bragg reflector connected to a first end of the optical waveguide for receiving the second optical signal;
a second distributed Bragg reflector connected to the second end of the optical waveguide and for outputting the third optical signal; a silicon substrate based on the tunable filter according to C7; Photonic neural network above.
[C9]
The second optical signal passing through the optical waveguide is changed into the third optical signal by inputting the optical signal into the heating layer and changing the effective refractive index of the optical waveguide. A photonic neural network on a silicon substrate based on the tunable filter described in C8.
[C10]
The MZI network includes a plurality of basic units, the basic units including a coupler, an external phase shifter and an internal phase shifter, the coupler being located between the external phase shifter and the internal phase shifter. A photonic neural network on a silicon substrate based on the tunable filter according to any one of C2 to C9.
[C11]
On a silicon substrate based on a tunable filter according to any one of C2 to C9, the DC network includes a plurality of basic units, and the basic units include a modulation arm and a tunable DC coupler. photonic neural network.
[C12]
The silicon based tunable filter according to any one of C1 to C11, wherein the center wavelength of the tunable filter is the same as the wavelength of the second optical signal in an initial or non-modulated state. Photonic neural network on board.
[C13]
When the electric signal applied to the tunable filter changes, the center wavelength of the tunable filter shifts, and the transmission intensity of the light wave at the wavelength of the second optical signal changes, so that the electric signal from the tunable filter shifts. Any one of C1 to C12, characterized in that the optical intensity of the output third optical signal and the optical intensity of the first optical signal output from the linear network exhibit a nonlinear relationship. A photonic neural network on a silicon substrate based on the tunable filter described in .
[C14]
A photonic integrated circuit comprising a photonic neural network on a silicon substrate based on the tunable filter according to any one of a plurality of C1-13.
[C15]
performing a linear operation on the plurality of input optical signals to obtain a plurality of first optical signals;
converting the plurality of first optical signals into a plurality of electrical signals;
obtaining a plurality of second optical signals;
modulating the plurality of second optical signals using the plurality of electrical signals to obtain a plurality of third optical signals,
A method for modulating an optical signal, wherein the optical intensity of the third optical signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship.
[C16]
The method according to C15, wherein the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength.

Claims (10)

a plurality of input ends for receiving a plurality of input optical signals;
a plurality of output ends for outputting a plurality of first optical signals, the plurality of input optical signals and the plurality of first optical signals having a one-to-one correspondence relationship; a linear calculation network having a linear relationship between the optical intensity of each input optical signal and the optical intensity of the corresponding first optical signal;
a plurality of photoelectric detectors for receiving the plurality of first optical signals and converting the plurality of first optical signals into a plurality of electrical signals;
a plurality of light sources for outputting a plurality of second optical signals;
each connected to one photoelectric detector and one light source, and receiving the electrical signal outputted from the photoelectric detector and the second optical signal outputted from the light source, respectively, and based on the electrical signal; a plurality of tunable filters for modulating the second optical signal and outputting a third optical signal,
The optical intensity of the third optical signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship,
The tunable filter includes an F -P cavity filter,
By changing the effective refractive index of the optical waveguide of the FP cavity filter by thermal modulation or electrical modulation, the second optical signal passing through the optical waveguide is converted into the third optical signal. A photonic neural network chip on a silicon substrate based on a tunable filter , characterized in that :
The FP cavity filter is
a heating layer for receiving the electrical signal;
an optical waveguide located under the heating layer;
a first distributed Bragg reflector connected to a first end of the optical waveguide for receiving the second optical signal;
a second distributed Bragg reflector connected to the second end of the optical waveguide for outputting the third optical signal; a photonic on a silicon substrate based on a tunable filter; neural network chip . The photonic neural network chip on a silicon substrate based on a tunable filter as claimed in claim 1, wherein the linear operation network includes a Mach-Zehnder interferometer (MZI) network or a direct coupler (DC) network. The silicon based tunable filter according to claim 1 or 2, wherein the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength. Photonic neural network chip on board. The MZI network includes a plurality of basic units, the basic units including a coupler, an external phase shifter and an internal phase shifter, the coupler being located between the external phase shifter and the internal phase shifter. A photonic neural network chip on a silicon substrate based on a tunable filter according to any one of claims 2 to 3 . Tunable filter based silicon according to any one of claims 2 to 3 , characterized in that the DC network comprises a plurality of basic units, the basic units comprising a modulation arm and a tunable DC coupler. Photonic neural network chip on board. The tunable filter according to any one of claims 1 to 5 , wherein the center wavelength of the tunable filter is the same as the wavelength of the second optical signal at an initial stage or in a non-modulated state. Based on a photonic neural network chip on a silicon substrate. When the electrical signal applied to the tunable filter changes, the center wavelength of the tunable filter shifts, the transmission intensity of the light wave at the wavelength of the second optical signal changes, and the output from the tunable filter changes. The optical intensity of the third optical signal outputted from the linear calculation network and the optical intensity of the first optical signal outputted from the linear calculation network exhibit a nonlinear relationship. A photonic neural network chip on a silicon substrate based on the tunable filter described in Section. A photonic integrated circuit comprising a photonic neural network chip on a silicon substrate based on a tunable filter according to any of claims 1 to 7 . performing a linear operation on the plurality of input optical signals to obtain a plurality of first optical signals;
converting the plurality of first optical signals into a plurality of electrical signals;
obtaining a plurality of second optical signals;
modulating a plurality of second optical signals using the plurality of electrical signals to obtain a plurality of third optical signals,
The optical intensity of the third optical signal and the corresponding optical intensity of the first optical signal have a nonlinear relationship,
The second optical signal passing through the optical waveguide is changed into the third optical signal by changing the effective refractive index of the optical waveguide of the F -P cavity filter by thermal modulation or electrical modulation. A method for modulating an optical signal, the method comprising:
The FP cavity filter is
a heating layer for receiving the electrical signal;
an optical waveguide located under the heating layer;
a first distributed Bragg reflector connected to a first end of the optical waveguide for receiving the second optical signal;
a second distributed Bragg reflector connected to a second end of the optical waveguide for outputting the third optical signal . 10. The method of claim 9 , wherein the input optical signal, the first optical signal, the second optical signal, and the third optical signal have the same wavelength.

Images