阅读说明：本技术 一种光学神经网络全光非线性激活层及其实现方法 (Optical neural network all-optical nonlinear activation layer and implementation method thereof ) 是由 廖琨 戴天翔 胡小永 龚旗煌 于 2021-03-24 设计创作，主要内容包括：本发明公开了一种光学神经网络全光非线性激活层及其实现方法。本发明通过将MZI波导构型与石墨烯异质增强的Bi-2Te-3非线性材料相结合,利用片上波导结构设计进一步放大石墨烯异质增强的Bi-2Te-3材料的非线性响应,完成了片上集成非线性激活层的设计,实现了片上集成波导中进行光学非线性计算,解决了光学非线性材料非线性程度较弱的问题,扩展了光学神经网络的功能,为多层纯光学神经网络的使用提供了可能；本发明提出的光学非线性激活层不仅能够用于片上集成光学神经网络,还能够用于其他集成光学信号处理平台中需要使用非线性计算的场景,且响应速度极快,能够满足低能耗高速计算的需求。(The invention discloses an optical neural network all-optical nonlinear activation layer and an implementation method thereof. According to the invention, the MZI waveguide configuration and graphene heteroenhancement Bi are adopted 2 Te 3 Nonlinear material combination, and on-chip waveguide structure design is utilized to further amplify graphene heteroenhancement Bi 2 Te 3 The nonlinear response of the material completes the design of the on-chip integrated nonlinear activation layer, realizes the optical nonlinear calculation in the on-chip integrated waveguide, solves the problem of weak nonlinear degree of the optical nonlinear material, expands the functions of the optical neural network and provides possibility for the use of a multilayer pure optical neural network; the invention provides an optical nonlinear active layerThe method can be used for not only an on-chip integrated optical neural network, but also scenes needing nonlinear calculation in other integrated optical signal processing platforms, has extremely high response speed, and can meet the requirements of low energy consumption and high speed calculation.)

1. Based on Bi₂Te₃An optical neural network all-optical nonlinear activation layer of a material, which is characterized by comprising a plurality of unit structures, wherein each unit structure comprises: waveguide structure, first electric control phase shifter and graphene heterogeneous enhanced Bi₂Te₃A nonlinear material and a second electrically controlled phase shifter; the waveguide structure comprises a waveguide beam splitter, a first upper branch waveguide, a first lower branch waveguide, a directional coupler, a second upper branch waveguide and a second lower branch waveguide, the waveguide structure is divided into two branches of waveguides with equal light intensity through the output end of the waveguide beam splitter, the two branches of waveguides are respectively a first upper branch waveguide and a first lower branch waveguide, and the tail ends of the first upper branch waveguide and the first lower branch waveguide are coupled and enter the directional coupler; the output end of the directional coupler is divided into two branch waveguides which are a second upper branch waveguide and a second lower branch waveguide respectively; arranging a first electrically controlled phase shifter on the first upper branch waveguide; covering the second upper branch waveguide with single-layer graphene, and depositing a layer of bismuth telluride (Bi) on the single-layer graphene₂Te₃Realizing the heterostructure of single-layer graphene-bismuth telluride, thereby obtaining the Bi with enhanced graphene heterogeneity₂Te₃A non-linear material; graphene heteroenhancement of Bi₂Te₃The carrier relaxation in the heterostructure of the nonlinear material becomes faster than in bismuth telluride, andthe light absorption intensity can be obviously improved, the light nonlinearity is favorably enhanced, and the modulation depth is improved; arranging a second electrically controlled phase shifter on the second lower branch waveguide; the tail ends of the second upper branch waveguide and the second lower branch waveguide are combined; therefore, the whole unit structure forms a Mach-Zehnder interferometer structure; connecting the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures with the output ends of the corresponding linear computing unit waveguides respectively to form an all-optical nonlinear active layer of the optical neural network;

an optical signal input from the input end of the waveguide beam splitter is divided into two paths with equal light intensity through the output end of the waveguide beam splitter, one path of the optical signal passes through a first upper branch waveguide provided with a first electric control phase shifter, the other path of the optical signal passes through a first lower branch waveguide, a phase difference is introduced into the first upper branch waveguide by adjusting the voltage of the first electric control phase shifter, and the two paths are coupled through a directional coupler, so that the splitting ratio of any light intensity in the two coupled paths is realized; bi with two paths respectively enhanced by being provided with graphene heterogeneity₂Te₃The nonlinear material second upper branch waveguide and the second lower branch waveguide are provided with a second electric control phase shifter, the voltage of the second electric control phase shifter is adjusted, so that the optical signal of the second upper branch waveguide and the optical signal of the second lower branch waveguide have a phase difference pi, the linear absorption part of bismuth telluride on the optical signal is offset, the nonlinear absorption part of bismuth telluride on the optical signal is amplified, and finally the optical signal is combined into one path, so that the nonlinear transmission part of the optical signal is obtained; the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures are connected with the output ends of the waveguides of the linear calculation unit to be used as nonlinear active layers, and therefore optical nonlinear calculation is introduced into the on-chip optical neural network.

2. The all-optical nonlinear active layer of the optical neural network as claimed in claim 1, wherein the thickness of the waveguide structure is 70 to 220nm, and the width is 400 to 500 nm.

3. The all-optical nonlinear active layer of the optical neural network of claim 1, wherein the portions of the output ends of the waveguide splitters connected to the first upper branch waveguide and the first lower branch waveguide are respectively formed by two arcs so as to reduce bending loss introduced during turning.

4. The all-optical nonlinear activation layer of the optical neural network as claimed in claim 1, wherein the bismuth telluride has a thickness of 10-20 nm and an area of not less than 2 μm x 2 μm.

5. The all-optical nonlinear activation layer of the optical neural network as claimed in claim 1, wherein the waveguide structure is a ridge waveguide or a slit waveguide formed by depositing a waveguide material on an insulating substrate and etching the waveguide material; etching off other parts of the waveguide material on the insulating substrate, and forming a ridge waveguide by the remaining parts; alternatively, the waveguide material on the insulating substrate is etched, and the etched portion forms a slot waveguide.

6. The all-optical nonlinear active layer of the optical neural network as claimed in claim 5, wherein the waveguide material is a communication band low-loss semiconductor material compatible with Complementary Metal Oxide Semiconductor (CMOS) process.

7. The Bi-based material of claim 1₂Te₃An implementation method of an all-optical nonlinear activation layer of an optical neural network of materials is characterized by comprising the following steps:

1) structure preparation:

forming a waveguide structure on an insulating substrate, wherein the waveguide structure comprises a waveguide beam splitter, a first upper branch waveguide, a first lower branch waveguide, a directional coupler, a second upper branch waveguide and a second lower branch waveguide, the waveguide structure is divided into two branches of waveguides with equal light intensity through the output end of the waveguide beam splitter, the two branches of waveguides are respectively the first upper branch waveguide and the first lower branch waveguide, and the tail ends of the first upper branch waveguide and the first lower branch waveguide are coupled and enter the directional coupler; the output end of the directional coupler is divided into two branch waveguides, namely a second upper branch waveguide and a second lower branch waveguideA waveguide; arranging a first electrically controlled phase shifter on the first upper branch waveguide; covering the second upper branch waveguide with single-layer graphene, and depositing a layer of bismuth telluride (Bi) on the single-layer graphene₂Te₃Realizing the heterostructure of single-layer graphene-bismuth telluride, thereby obtaining the Bi with enhanced graphene heterogeneity₂Te₃A non-linear material; graphene heteroenhancement of Bi₂Te₃The carrier relaxation in the heterostructure of the nonlinear material becomes faster than that in bismuth telluride, and the optical absorption intensity can be obviously improved, thereby being beneficial to enhancing the nonlinearity of light and improving the modulation depth; arranging a second electrically controlled phase shifter on the second lower branch waveguide; the tail ends of the second upper branch waveguide and the second lower branch waveguide are combined; therefore, the whole unit structure forms a Mach-Zehnder interferometer structure; connecting the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures with the output ends of the corresponding linear computing unit waveguides respectively to form an all-optical nonlinear active layer of the optical neural network;

2) an optical signal input from the input end of the waveguide beam splitter is divided into two paths with equal light intensity through the output end of the waveguide beam splitter, one path of the optical signal passes through a first upper branch waveguide provided with a first electric control phase shifter, the other path of the optical signal passes through a first lower branch waveguide, a phase difference is introduced into the first upper branch waveguide by adjusting the voltage of the first electric control phase shifter, and the two paths are coupled through a directional coupler, so that the splitting ratio of any light intensity in the two coupled paths is realized;

3) bi with two paths respectively enhanced by being provided with graphene heterogeneity₂Te₃The nonlinear material second upper branch waveguide and the second lower branch waveguide are provided with a second electric control phase shifter, the voltage of the second electric control phase shifter is adjusted, so that the optical signal of the second upper branch waveguide and the optical signal of the second lower branch waveguide have a phase difference pi, the linear absorption part of bismuth telluride on the optical signal is offset, the nonlinear absorption part of bismuth telluride on the optical signal is amplified, and finally the optical signal is combined into one path, so that the nonlinear transmission part of the optical signal is obtained;

4) the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures are connected with the output ends of the waveguides of the linear calculation unit to be used as nonlinear active layers, and therefore optical nonlinear calculation is introduced into the on-chip optical neural network.

8. The method of claim 7, wherein in step 1), the waveguide structure is formed by depositing waveguide material on an insulating substrate, and etching the waveguide material to form a ridge waveguide or a slot waveguide; etching off other parts of the waveguide material on the insulating substrate, and forming a ridge waveguide by the remained parts; alternatively, the waveguide material on the insulating substrate is etched, and the etched portion forms a slot waveguide.

9. The method of claim 7, wherein in step 1), the waveguide material is a semiconductor material with low transmission loss in communication band compatible with Complementary Metal Oxide Semiconductor (CMOS) process.

Technical Field

The invention relates to an optical signal processing technology, in particular to a design and implementation method of an all-optical nonlinear activation layer in an on-chip optical neural network.

Background

The main structures in most existing designs of optical neural networks with waveguides integrated on chip and other on-chip optical computing platforms are a beam splitting waveguide unit and cascaded Mach-Zehnder interferometers (MZIs), which are only suitable for linear computing. Because the superposition of linear computation is still linear computation, the computation result of the design can be equal to one-time matrix multiplication operation no matter how the total layer number is, the parameter range is limited, and the requirement of a neural network for fitting data cannot be met. Therefore, the optical neural network relies on further nonlinear calculation in an electronic circuit in the subsequent information processing process, and the function of the complete neural network cannot be integrated on an optical platform. The nonlinear response of the existing optical nonlinear material is weaker, and the modulation depth of signals is not enough; in addition, an inherent tradeoff between ultra-fast response time and large non-linearity is often present, such that large non-linearity coefficients can generally only be traded for slower response times. Therefore, efficient and feasible nonlinear calculation is difficult to realize in the transplanting process of a large-scale optical hardware platform, so that the optical nonlinear activation layer only stays in a theoretical concept and cannot realize practical application.

Phase change materials are used in another part of on-chip optical neural network designs to introduce non-linear computations, such as germanium antimony tellurium alloy (GST). The specific implementation is that the GST material is covered on the waveguide, because the amorphous light transmittance of the GST material is higher, the crystalline light transmittance is lower, and the GST material can be recovered from the amorphous state to the crystalline state by using a weak pulse heating material under the influence of the thermal effect of light, or can be rapidly heated by using a strong pulse and then cooled to be converted from the crystalline state to the amorphous state, a critical light intensity exists between the two states, the conversion process from the crystalline state to the amorphous state of the GST material can be carried out only above the critical light intensity, and the conversion rate is positively correlated with the light intensity. The critical intensity of this phase change of the material can thus be used to set a threshold for the intensity of light propagating in the waveguide, which passes through the material if and only if the total intensity of the pulse exceeds this threshold, thereby performing a non-linear calculation. However, the phase-change material is nonvolatile, and if the last pulse exceeds a threshold value to enable the phase-change material to enter an amorphous state, extra input energy is needed to reset to the crystalline state, so that continuous operation cannot be performed, and the response time and energy consumption cannot meet the requirement of high-speed calculation.

The reason that nonlinear calculation is difficult to perform on an optical platform is that the all-optical nonlinear effect of the material is weak, the material lacks of nonlinear materials with enough strength, and the strong nonlinear effect is difficult to realize in an on-chip integrated device; GST materials that can be used for non-linear calculations have non-volatile characteristics and are also not suitable for efficient calculations with fast response. Therefore, at present, for an optical neural network with an ultra-fast time response and ultra-low energy consumption, which is intended to be realized by a hardware platform, the real introduction of the all-optical nonlinear activation layer still remains a problem to be solved urgently.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a Bi-based material₂Te₃The all-optical nonlinear activation layer of the optical neural network solves the problem of weak nonlinear response of the optical nonlinear material, expands the functions of the optical neural network and provides possibility for realizing a multilayer pure optical neural network.

One object of the present invention is to provide a Bi-based material₂Te₃An optical neural network all-optical nonlinear activation layer of the material.

The invention is based on Bi₂Te₃The optical neural network all-optical nonlinear activation layer of the material comprises a plurality of unit structures, and each unit structure comprises: waveguide structure, first electric control phase shifter and graphene heterogeneous enhanced Bi₂Te₃A nonlinear material and a second electrically controlled phase shifter; the waveguide structure comprises a waveguide beam splitter, a first upper branch waveguide, a first lower branch waveguide, a directional coupler, a second upper branch waveguide and a second lower branch waveguide, the waveguide structure is divided into two branches of waveguides with equal light intensity through the output end of the waveguide beam splitter, the two branches of waveguides are respectively a first upper branch waveguide and a first lower branch waveguide, and the tail ends of the first upper branch waveguide and the first lower branch waveguide are coupledEntering a directional coupler; the output end of the directional coupler is divided into two branch waveguides which are a second upper branch waveguide and a second lower branch waveguide respectively; arranging a first electrically controlled phase shifter on the first upper branch waveguide; covering the second upper branch waveguide with single-layer graphene, and depositing a layer of bismuth telluride (Bi) on the single-layer graphene₂Te₃Realizing the heterostructure of single-layer graphene-bismuth telluride, thereby obtaining the Bi with enhanced graphene heterogeneity₂Te₃A non-linear material; graphene heteroenhancement of Bi₂Te₃The carrier relaxation in the heterostructure of the nonlinear material becomes faster than that in bismuth telluride, and the optical absorption intensity can be obviously improved, thereby being beneficial to enhancing the nonlinearity of light and improving the modulation depth; arranging a second electrically controlled phase shifter on the second lower branch waveguide; the tail ends of the second upper branch waveguide and the second lower branch waveguide are combined; therefore, the whole unit structure forms a Mach-Zehnder interferometer structure; connecting the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures with the output ends of the corresponding linear computing unit waveguides respectively to form an all-optical nonlinear active layer of the optical neural network;

an optical signal input from the input end of the waveguide beam splitter is divided into two paths with equal light intensity through the output end of the waveguide beam splitter, one path of the optical signal passes through a first upper branch waveguide provided with a first electric control phase shifter, the other path of the optical signal passes through a first lower branch waveguide, a phase difference is introduced into the first upper branch waveguide by adjusting the voltage of the first electric control phase shifter, and the two paths are coupled through a directional coupler, so that the splitting ratio of any light intensity in the two coupled paths is realized; bi with two paths respectively enhanced by being provided with graphene heterogeneity₂Te₃The nonlinear material second upper branch waveguide and the second lower branch waveguide are provided with a second electric control phase shifter, the voltage of the second electric control phase shifter is adjusted, so that the optical signal of the second upper branch waveguide and the optical signal of the second lower branch waveguide have a phase difference pi, the linear absorption part of bismuth telluride on the optical signal is offset, the nonlinear absorption part of bismuth telluride on the optical signal is amplified, and finally the optical signal is combined into one path, so that the nonlinear transmission part of the optical signal is obtained; constructed by a plurality of Mach-Zehnder interferometersThe input end of the waveguide beam splitter is connected with the output end of the linear computing unit waveguide and used as a nonlinear active layer, so that optical nonlinear computation is introduced into the on-chip optical neural network.

The waveguide structure is a ridge waveguide or a slit waveguide formed by depositing a waveguide material on an insulating substrate and etching the waveguide material; etching off other parts of the waveguide material on the insulating substrate, and forming a ridge waveguide by the remaining parts; alternatively, the waveguide material on the insulating substrate is etched, and the etched portion forms a slot waveguide.

The waveguide material is a semiconductor material with low transmission loss of communication wave band compatible with Complementary Metal Oxide Semiconductor (CMOS) process, the transmission loss is less than 3dB/cm, such as silicon or silicon nitride.

The thickness of the waveguide structure is 70-220 nm, and the width is 400-500 nm. The parts of the output end of the waveguide beam splitter, which are connected to the first upper branch waveguide and the first lower branch waveguide, are respectively formed by splicing two sections of circular arcs so as to reduce bending loss introduced during turning.

The thickness of the bismuth telluride is 10-20 nm, and the area is not less than 2 mu m multiplied by 2 mu m.

As a novel saturable absorption material, bismuth telluride has very high modulation depth, and the modulation depth reaches 70% at the wavelength of 1570 nm. However, the transient carrier dynamics of the material are researched, and the relaxation time of electrons between bands is very long and exceeds 500fs and 5ps respectively. As a typical saturable absorber, graphene can generate saturable absorption in a wide working wavelength range and has ultra-fast recovery time, in-band relaxation time is less than 150fs, and band-to-band electron relaxation time is 1.5 ps. However, due to the relatively low absorption intensity of one atomic layer, the optical modulation depth of single-layer graphene is very low, typically around 1%. Although increasing the number of graphene layers may increase the modulation depth, unnecessary unsaturation loss may also increase, which is less desirable for practical applications.

Bismuth telluride is deposited on single-layer graphene in a similar fashion to metal-semiconductor contacts to form heterostructures, with electrons transferred from the graphene to the bismuth telluride at their interface. Based on this particular energy structure, some optical properties are predicted, i.e. due to the zero band gap of graphene and the small band gap of bismuth telluride, the heterostructure should have broadband absorption characteristics from visible to infrared; and the carrier dynamics in this heterostructure are altered. The appearance of graphene provides a fast channel for relaxation of photoexcited carriers, that is, photoexcited carriers in bismuth telluride are injected into graphene as "dirac fermions" due to energy band bending caused by interlayer interaction. The direct result is that carrier relaxation in the heterostructure will be faster compared to pure bismuth telluride, and the light absorption intensity can be significantly improved, which is beneficial to enhancing light nonlinearity and increasing modulation depth.

The combination of the band gap of bismuth telluride and the optical characteristics of graphene shows that the graphene heterogeneously enhanced Bi₂Te₃The light transmittance α (I) of a nonlinear material is expressed by the following equation:wherein alpha is_SAnd alpha_NSRespectively saturable and non-saturable absorption rates, I and I_SRespectively output light intensity and saturated light intensity; the material has saturation characteristic for light absorption, the absorption rate of the material is reduced along with the increase of light intensity and finally tends to a constant, the relaxation time of the material is in picosecond magnitude, the material can be approximately considered to have no memory characteristic, the material can be automatically and quickly reset after being used every time, the energy does not need to be continuously input, and therefore the material is suitable for being used in high-speed calculation. However, the degree of nonlinearity of the material is stronger than that of the common material, but the deviation degree from the linearity is small, so that the time for directly applying the material to on-chip calculation is not obvious, and the calculation process is not facilitated.

Another object of the present invention is to provide a Bi-based material₂Te₃An optical neural network all-optical nonlinear activation layer of the material.

The invention is based on Bi₂Te₃The method for realizing the all-optical nonlinear activation layer of the optical neural network comprises the following steps:

1) structure preparation:

forming waves on an insulating substrateThe waveguide structure comprises a waveguide beam splitter, a first upper branch waveguide, a first lower branch waveguide, a directional coupler, a second upper branch waveguide and a second lower branch waveguide, the waveguide structure is divided into two branches of waveguides with equal light intensity through the output end of the waveguide beam splitter, the two branches of waveguides are respectively a first upper branch waveguide and a first lower branch waveguide, and the tail ends of the first upper branch waveguide and the first lower branch waveguide are coupled and enter the directional coupler; the output end of the directional coupler is divided into two branch waveguides which are a second upper branch waveguide and a second lower branch waveguide respectively; arranging a first electrically controlled phase shifter on the first upper branch waveguide; covering the second upper branch waveguide with single-layer graphene, and depositing a layer of bismuth telluride (Bi) on the single-layer graphene₂Te₃Realizing the heterostructure of single-layer graphene-bismuth telluride, thereby obtaining the Bi with enhanced graphene heterogeneity₂Te₃A non-linear material; graphene heteroenhancement of Bi₂Te₃The carrier relaxation in the heterostructure of the nonlinear material becomes faster than that in bismuth telluride, and the optical absorption intensity can be obviously improved, thereby being beneficial to enhancing the nonlinearity of light and improving the modulation depth; arranging a second electrically controlled phase shifter on the second lower branch waveguide; the tail ends of the second upper branch waveguide and the second lower branch waveguide are combined; therefore, the whole unit structure forms a Mach-Zehnder interferometer structure; connecting the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures with the output ends of the corresponding linear computing unit waveguides respectively to form an all-optical nonlinear active layer of the optical neural network;

2) an optical signal input from the input end of the waveguide beam splitter is divided into two paths with equal light intensity through the output end of the waveguide beam splitter, one path of the optical signal passes through a first upper branch waveguide provided with a first electric control phase shifter, the other path of the optical signal passes through a first lower branch waveguide, a phase difference is introduced into the first upper branch waveguide by adjusting the voltage of the first electric control phase shifter, and the two paths are coupled through a directional coupler, so that the splitting ratio of any light intensity in the two coupled paths is realized;

3) bi with two paths respectively enhanced by being provided with graphene heterogeneity₂Te₃A second upper branch waveguide of non-linear material and a second electrically controlled phase shifterThe voltage of the second electric control phase shifter is adjusted, so that the optical signal of the second upper branch waveguide and the optical signal of the second lower branch waveguide have a phase difference pi, the linear absorption part of the bismuth telluride on the optical signal is offset, the nonlinear absorption part of the bismuth telluride on the optical signal is amplified, and finally the optical signal is combined into one path, so that the nonlinear transmission part of the optical signal is obtained;

4) the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures are connected with the output ends of the waveguides of the linear calculation unit to be used as nonlinear active layers, and therefore optical nonlinear calculation is introduced into the on-chip optical neural network.

In the step 1), the waveguide structure is formed by depositing a waveguide material on an insulating substrate and then etching the waveguide material to form a ridge waveguide or a slit waveguide; etching off other parts of the waveguide material on the insulating substrate, and forming a ridge waveguide by the remained parts; alternatively, the waveguide material on the insulating substrate is etched, and the etched portion forms a slot waveguide.

The waveguide material is a semiconductor material with low transmission loss of communication wave band compatible with Complementary Metal Oxide Semiconductor (CMOS) process, and the transmission loss is less than 3 dB/cm.

The invention has the advantages that:

according to the invention, the MZI waveguide configuration and graphene heteroenhancement Bi are adopted₂Te₃Nonlinear material combination, and on-chip waveguide structure design is utilized to further amplify graphene heteroenhancement Bi₂Te₃The nonlinear response of the material completes the design of the on-chip integrated nonlinear activation layer, realizes the optical nonlinear calculation in the on-chip integrated waveguide, solves the problem of weak nonlinear degree of the optical nonlinear material, expands the functions of the optical neural network and provides possibility for the use of a multilayer pure optical neural network; the optical nonlinear activation layer provided by the invention can be used for an on-chip integrated optical neural network and can also be used in scenes needing nonlinear calculation in other integrated optical signal processing platforms, the response speed is very high, and the requirements of low energy consumption and high speed calculation can be met.

Drawings

FIG. 1 shows a Bi-based composition of the present invention₂Te₃A schematic diagram of one embodiment of a cell structure of an all-optical nonlinear active layer of an optical neural network of materials;

FIG. 2 shows a Bi-based composition of the present invention₂Te₃A comparative graph of input and output power relations of bismuth telluride before and after using MZI structure modification of one embodiment of the all-optical nonlinear active layer of the optical neural network of materials.

Detailed Description

The invention will be further elucidated by means of specific embodiments in the following with reference to the drawing.

As shown in FIG. 1, the Bi-based material of the present embodiment₂Te₃The optical neural network all-optical nonlinear activation layer of the material comprises a plurality of unit structures, and each unit structure comprises: the device comprises a waveguide structure, a first electric control phase shifter PS1, a graphene heterogeneous enhanced nonlinear material BiTe and a second electric control phase shifter PS 2; the waveguide structure comprises a waveguide beam splitter BS, a first upper branch waveguide, a first lower branch waveguide, a Directional Coupler (DC), a second upper branch waveguide and a second lower branch waveguide, wherein the part of the waveguide beam splitter, which is connected to the first upper branch waveguide and the first lower branch waveguide, is spliced by two arcs respectively, in order to reduce bending loss introduced during turning, the curvature radius of each arc is designed to be 20 mu m, the part of the waveguide beam splitter, which is connected to the first upper branch waveguide and the first lower branch waveguide, is divided into two branch waveguides with equal light intensity through the output end of the waveguide beam splitter, which are the first upper branch waveguide and the first lower branch waveguide respectively, and the tail ends of the first upper branch waveguide and the first lower branch waveguide are coupled and enter the Directional Coupler; the output end of the directional coupler is divided into two branches of waveguides, namely a second upper branch waveguide and a second lower branch waveguide, the thickness of the waveguides is 220nm, and the width of the waveguides is 500 nm; arranging a first electrically controlled phase shifter on the first upper branch waveguide; covering the second upper branch waveguide with single-layer graphene, and depositing a layer of bismuth telluride (Bi) on the single-layer graphene₂Te₃The thickness of bismuth telluride is 20nm, and the area is 2 Mum multiplied by 2 Mum, thus realizing the heterostructure of single-layer graphene-bismuth telluride, therebyObtaining graphene heteroenhancement Bi₂Te₃A non-linear material; graphene heteroenhancement of Bi₂Te₃The carrier relaxation in the heterostructure of the nonlinear material becomes faster than that in bismuth telluride, and the optical absorption intensity can be obviously improved, thereby being beneficial to enhancing the nonlinearity of light and improving the modulation depth; arranging a second electrically controlled phase shifter on the second lower branch waveguide; the ends of the second upper arm waveguide and the second lower arm waveguide are combined. The whole unit structure forms a Mach-Zehnder interferometer structure; and connecting the input ends of the waveguide beam splitters of the Mach-Zehnder interferometer structures with the output ends of the corresponding linear computing unit waveguides respectively to form an all-optical nonlinear active layer of the optical neural network, wherein the number of the Mach-Zehnder interferometer structures is consistent with that of the linear computing unit waveguides.

An optical signal input from the input end of the waveguide beam splitter is divided into two paths with equal light intensity through the output end of the waveguide beam splitter, one path of the optical signal passes through a first upper branch waveguide provided with a first electric control phase shifter, the other path of the optical signal passes through a first lower branch waveguide, a phase difference is introduced into the first upper branch waveguide by adjusting the voltage of the first electric control phase shifter, and the two paths of the optical signal are coupled through a directional coupler, so that the proportion of the two paths after coupling is 1.79: 1 light intensity splitting ratio; bi with two paths of optical signals respectively enhanced through graphene heterogeneously arranged₂Te₃The nonlinear material second upper branch waveguide and the second lower branch waveguide are provided with a second electric control phase shifter, the voltage of the second electric control phase shifter is adjusted, so that the light of the second upper branch waveguide and the light signal of the second lower branch waveguide have a phase difference pi, the linear absorption part of bismuth telluride on the light signal is offset, the nonlinear absorption part of bismuth telluride on the light signal is amplified, and finally the light is combined into one path, so that the nonlinear transmission part of the light signal is obtained.

The combined optical signal is output from the rightmost end output end, and the relation curve of the input and output light intensity shown by the solid line in fig. 2 is obtained, so that the nonlinear effect of the bismuth telluride is greatly enhanced compared with the nonlinear (before correction) input and output power response which is not further amplified by utilizing the on-chip MZI waveguide structure design.

And a plurality of Mach-Zehnder interferometer structures longitudinally arranged in a one-dimensional array are arranged at the tail end of the linear calculation unit and used as a nonlinear activation layer, so that optical nonlinear calculation is introduced into the on-chip optical neural network.

Finally, it is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.