On-chip optical neural network and method based on topological photonic crystal boundary state splicing

Abstract

The present invention discloses an on-chip optical neural network and method based on the splicing of topological photonic crystal boundary states. The present invention uses the boundary states of topological photonic crystals as carriers for optical signal transmission, and proposes an on-chip integrated optical neural network with a certain tolerance to processing defects and accidental damage; it can reduce the impact of processing defects or accidental damage on the transmission of optical signals in the neural network, enhance the stability of the network, lay the foundation for the realization of large-scale photonic neural networks, and provide a wider application scenario for photonic neural networks; the collective effect produced by multiple periods of topological photonic crystals, this property makes it possible for damage to a unit in the structure to not directly cause a significant change in the properties of the boundary states; in addition, another property of the boundary states of topological photonic crystals is that the optical signal propagating thereon has two pseudospins, and the propagation directions of the two pseudospin optical signals are related to the splicing method of the topological photonic crystals.

Description

Technical Field

[0001] The present invention relates to optical signal processing technology, and in particular to an on-chip optical neural network based on topological photonic crystal boundary state splicing and an implementation method thereof. Background Art

[0002] The existing on-chip optical neural networks mainly use on-chip cascaded Mach-Zehnder interferometers (MZIs) and on-chip integrated waveguides to realize the transmission and processing of optical signals. Some designs use cascaded MZIs to achieve the regulation of optical signals by coupling optical signals between on-chip integrated waveguides. In this design, in order to realize the function of a linear layer, it is necessary to construct an MZI between every two adjacent waveguides in the input signal, with a long total waveguide length. Another part of the design makes the input end of each optical neuron only connected to the two adjacent neurons in the previous layer, and achieves a locally connected neural network configuration through waveguide splitting and combining. Such a design greatly reduces the length of the waveguide and the volume of the device, but each layer can only realize a tridiagonal matrix of optical signal changes. The challenge faced by the above two designs is that they are highly dependent on on-chip integrated waveguides as the medium for signal transmission, making the transmission of optical signals in these devices easily affected by waveguide processing defects or damage, and not stable at the hardware level. As the network scale increases, this lack of stability may have a serious impact on the function of on-chip optical neural networks.

[0003] When an optical signal is transmitted in a waveguide, if there are defects in the waveguide, additional scattering or reflection may be introduced in the signal transmission, thereby changing the phase of the optical signal or reducing the transmittance of the optical signal or even completely blocking the normal transmission of the optical signal. The design feature of the optical neural network based on on-chip integrated waveguide is to modulate the optical signal in layers, and rely on interference to adjust the signal intensity in each layer. If there is a waveguide defect in the previous layer or in the connection between layers, the phase and amplitude of the optical signal will change, which will cause the subsequent signal processing to deviate from the predetermined target. Such deviations may be amplified in the layer-by-layer propagation, causing the final output signal to completely deviate from the expected result. The current implementation of on-chip optical neural networks, whether using MZI linear layers or local connection layers, will inevitably encounter such problems. As the requirements for the functions of on-chip optical neural networks increase, the scale of the network becomes larger and larger, and the threat of devices using the above design being affected by waveguide processing defects or accidental damage will become higher. Summary of the invention

[0004] In order to solve the problem that optical signal transmission in optical hardware is affected by waveguide processing defects and damage, the present invention proposes an on-chip optical neural network based on topological photonic crystal boundary state splicing and an implementation method thereof.

[0005] One object of the present invention is to propose an on-chip optical neural network based on topological photonic crystal boundary state splicing.

[0006] The on-chip optical neural network based on topological photonic crystal boundary state splicing of the present invention comprises: a topological photonic crystal unit, which is a topological non-trivial crystal or a topological trivial crystal; a plurality of topological photonic crystal units with the same edge shape are closely arranged to form a topological photonic crystal array, in which the topological non-trivial crystals and the topological trivial crystals are arranged alternately, and the outer edges are spliced ​​together, and boundary states are generated at the splicing, and photons are transmitted unidirectionally at the splicing, thereby forming a boundary state path at the splicing, and the boundary state path includes a transverse boundary state path and a longitudinal boundary state path; the pseudospin direction of the photons transmitted by the boundary state path is locked; for photons with the pseudospin direction upward, along the transmission direction, the left side is the topological non-trivial crystal, and the right side is the topological trivial crystal. body; for photons with downward pseudospin direction, along the transmission direction, the left side is a topological trivial crystal and the right side is a topological non-trivial crystal; photons with a specific pseudospin direction are transmitted from the transverse boundary state path on one side to the array, and then to the junction where the topological photonic crystal units are spliced, and then divided into two light signals, which are transmitted upward and downward along the longitudinal boundary state path respectively, and then to the junction where the topological photonic crystal units are spliced, and are transmitted along the transverse boundary state path respectively, and then to the junction where the topological photonic crystal units are spliced, and then the beams are split up and down again and transmitted along the longitudinal boundary state path respectively, and the light signals transmitted along the downward and upward longitudinal boundary state paths are combined, and after the beams are combined, they continue to be transmitted forward along the transverse boundary state path; the two light signals propagating along the two transverse boundary state paths , after being split up and down by the longitudinal boundary state pathway, interference occurs when the beams are combined and merged into a transverse boundary state pathway. This process constitutes a neuron. All such neurons are connected as a whole to form an optical neural network; photons with pseudospin direction upward are input into the optical neural network from the transverse boundary state pathway on the left and are finally output from the right side of the optical neural network, and share the same transmission path as photons with pseudospin direction downward that are input into the optical neural network from the transverse boundary state pathway on the right and are finally output from the left side of the optical neural network; similarly, photons with pseudospin direction downward are input into the optical neural network from the transverse boundary state pathway on the left and are finally output from the right side of the optical neural network, and are output from the transverse boundary state pathway on the right. The photons with pseudospin upward direction that are input into the optical neural network and finally output from the left side of the optical neural network share the same transmission path of the optical neural network; and the transmission paths of the photons with pseudospin upward direction and the photons with pseudospin downward direction that are input into the optical neural network along the transverse boundary state path on the same side are staggered; the way in which optical signals propagate in the optical neural network, including beam splitting and two beams of light signals phase-modulated through the transverse boundary state path and then interfering to obtain a combined light beam for transverse transmission, is completely consistent with the waveguide network, and the generation of boundary states is a collective effect produced by multiple periodically arranged topological photonic crystal units, so that damage to individual topological photonic crystal units in the optical neural network will not directly lead to significant changes in the properties of boundary states.

[0007] A vertical column in an optical neural network is called a layer, where the first layer is the input layer, the last layer is the output layer, and between the input layer are the first to nth hidden layers, where n is a natural number ≥1.

[0008] Furthermore, a thermo-optic modulation device is included to perform thermo-optic modulation on a lateral boundary state path, and each thermo-optic modulation device includes a thermo-optic material and two electrodes; in each lateral boundary state path of the optical neural network that needs to adjust the phase, the joint of the topological non-trivial crystal and the topological ordinary crystal is covered with the thermo-optic material, and electrodes are respectively arranged on the upper and lower sides of the thermo-optic material, and the two electrodes are respectively located on both sides of the joint; voltage is applied to the thermo-optic material through the electrodes, so that the thermo-optic material produces a thermal effect, causing the materials of the topological non-trivial crystal and the topological ordinary crystal to cause carrier migration, resulting in a change in the refractive index and a change in the phase of the optical signal; the optical signals on the two lateral boundary state paths of the previous layer after the phase is thermo-optically modulated are transmitted forward to the junction of the topological photonic crystal unit splicing and the combined beam, and the interference of the two optical signals with different phases changes the light intensity on this layer of lateral boundary state paths.

[0009] The thermo-optic material uses a single layer or multi-layer graphene or a new phase change material Sb2Se3 film. The refractive index of the thermo-optic material is changed by applying voltage through the electrode, thereby modulating the phase by 0 to π; the phase change is positively correlated with the voltage change. The coverage length of a single layer of graphene is 50μm to 100μm, and the corresponding voltage of the phase π modulation is 1 to 2V. In addition, an isolation layer is set between the thermo-optic material and the topological photonic crystal array, and the isolation layer uses an insulating material SiO2.

[0010] Topological trivial crystals and topological non-trivial crystals are formed by punching holes in the background material. Topological trivial crystals and topological non-trivial crystals respectively include a number of basic structural unit cells closely arranged between lattices. The outer edge of each basic structural unit cell is a regular hexagon, and there are six rotationally symmetrically distributed regular triangular air holes inside. The overall structure has C 6v Symmetry: In topological trivial crystals, the distance between the centers of the six equilateral triangles and the center of the regular hexagon is less than 1 / 3 of the period of the two-dimensional topological photonic crystal, corresponding to a trivial topological state; in topological non-trivial crystals, the distance between the centers of the six equilateral triangles and the center of the regular hexagon is greater than 1 / 3 of the period of the two-dimensional topological photonic crystal, corresponding to a non-trivial topological state. When a topological trivial crystal and a topological non-trivial crystal are spliced ​​together, a boundary is formed at the splicing point, and two topological boundary states with opposite slopes appear in the dispersion curve of the boundary within the bandgap, corresponding to two different photon pseudospin directions, namely, pseudospin direction upward and pseudospin direction downward, and are localized near the boundary and transmitted unidirectionally along the boundary in opposite directions.

[0011] The topological photonic crystal unit includes at least dozens of periods of hexagonal basic structure cells in both the vertical and horizontal directions. The background material is silicon on insulator or silicon nitride.

[0012] Another object of the present invention is to propose a method for realizing an on-chip optical neural network based on topological photonic crystal boundary state splicing.

[0013] The method for realizing an on-chip optical neural network based on topological photonic crystal boundary state splicing of the present invention comprises the following steps:

[0014] 1) Providing a topological photonic crystal unit, wherein the topological photonic crystal unit is a topological non-trivial crystal or a topological trivial crystal;

[0015] 2) A plurality of topological photonic crystal units with the same edge shape are closely arranged to form a topological photonic crystal array, in which topological non-trivial crystals and topological trivial crystals are arranged alternately, and the outer edges are spliced ​​together, and boundary states are generated at the splicing. Photons are transmitted unidirectionally at the splicing, thereby forming a boundary state path at the splicing, and the boundary state path includes a transverse boundary state path and a longitudinal boundary state path;

[0016] 3) The pseudospin direction of photons transmitted through the boundary state channel is locked; for photons with pseudospins pointing upward, along the transmission direction, the left side is a topologically non-trivial crystal, and the right side is a topologically trivial crystal; for photons with pseudospins pointing downward, along the transmission direction, the left side is a topologically trivial crystal, and the right side is a topologically non-trivial crystal;

[0017] 4) Photons with a specific pseudospin direction are transmitted from the transverse boundary state path on one side to the array, and then to the junction where the topological photonic crystal units are spliced, and then divided into two light signals, which are respectively transmitted upward and downward along the longitudinal boundary state path, and then to the junction where the topological photonic crystal units are spliced, and then respectively transmitted along the transverse boundary state path, and then to the junction where the topological photonic crystal units are spliced, and then the light signals transmitted along the downward and upward longitudinal boundary state paths are combined, and then continue to be transmitted forward along the transverse boundary state path after being combined; the two light signals propagating along the two transverse boundary state paths are respectively divided up and down by the longitudinal boundary state path, and then interfered when combined and merged into one transverse boundary state path, and this process constitutes a neuron, and all such neurons are connected as a whole to constitute an optical neural network;

[0018] 5) A photon with a pseudospin upward direction, which is input into the optical neural network from the left transverse boundary state path and finally output from the right side of the optical neural network, shares the same transmission path of the optical neural network with a photon with a pseudospin downward direction, which is input into the optical neural network from the right transverse boundary state path and finally output from the left side of the optical neural network; similarly, a photon with a pseudospin downward direction, which is input into the optical neural network from the left transverse boundary state path and finally output from the right side of the optical neural network, shares the same transmission path of the optical neural network with a photon with a pseudospin upward direction, which is input into the optical neural network from the right transverse boundary state path and finally output from the left side of the optical neural network; and the transmission paths of the pseudospin upward photons and the pseudospin downward photons input into the optical neural network along the same transverse boundary state path on the same side are staggered;

[0019] 6) The way in which light signals propagate in optical neural networks, including beam splitting and two light signals phase-modulated through the transverse boundary state path and then interfering to obtain a combined light beam for transverse transmission, is exactly the same as that in waveguide networks. The generation of boundary states is a collective effect produced by multiple periodically arranged topological photonic crystal units, so that damage to individual topological photonic crystal units in the optical neural network will not directly lead to significant changes in the properties of boundary states.

[0020] Advantages of the present invention:

[0021] The present invention utilizes the boundary states of topological photonic crystals as carriers for optical signal transmission, and proposes an on-chip integrated optical neural network with a certain tolerance to processing defects and accidental damage; it can reduce the impact of processing defects or accidental damage on the transmission of optical signals in the optical neural network to a certain extent, enhance the stability of the network, lay a certain foundation for the realization of large-scale photonic optical neural networks, and provide a broader application scenario for photonic optical neural networks, such as providing an on-chip optical research platform for more complex high-dimensional non-equilibrium physical problems; the collective effect produced by multiple periods of topological photonic crystals, this property makes it possible that damage to a unit in the structure will not directly lead to a significant change in the properties of the boundary states; in addition, another property of the boundary states of topological photonic crystals is that the optical signals propagating thereon have two pseudospins, and the propagation directions of the two pseudospin optical signals are related to the splicing method of the topological photonic crystals. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of the splicing of a topological non-trivial crystal and a topological trivial crystal of an on-chip optical neural network based on the splicing of topological photonic crystal boundary states of the present invention;

[0023] Figure 2A schematic diagram of a photon transmission process with the pseudospin direction upward and from left to right in an on-chip optical neural network based on topological photonic crystal boundary state splicing of the present invention;

[0024] Figure 3 It is a schematic diagram of the photon transmission process with the pseudospin direction downward and from left to right in the on-chip optical neural network based on topological photonic crystal boundary state splicing of the present invention;

[0025] Figure 4 A structural block diagram of an optical neural network extracted from an embodiment of an on-chip optical neural network based on topological photonic crystal boundary state splicing of the present invention;

[0026] Figure 5 A schematic diagram of setting up a thermo-optical modulation device for an embodiment of an on-chip optical neural network based on topological photonic crystal boundary state splicing of the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further described below through specific embodiments in conjunction with the accompanying drawings.

[0028] like Figure 1 As shown, the topological trivial crystal PhC2 and the topological non-trivial crystal PhC1 are formed by punching holes on the background material. The topological trivial crystal and the topological non-trivial crystal respectively include a plurality of basic structural unit cells closely arranged between the lattices. The outer edge of each basic structural unit cell is a regular hexagon, and there are six rotationally symmetrically distributed regular triangular air holes inside. The overall structure has C 6v Symmetry, the center of the triangle is located on the midline of one side of the hexagon, and its specific position is determined by the contraction-expansion ratio Q; in topological trivial crystals, the distance between the center of the six equilateral triangles and the center of the regular hexagon is less than 1 / 3 of the period of the two-dimensional topological photonic crystal, corresponding to a trivial topological state; in topological non-trivial crystals, the distance between the center of the six equilateral triangles and the center of the regular hexagon is greater than 1 / 3 of the period of the two-dimensional topological photonic crystal, corresponding to a non-trivial topological state. When a topological trivial crystal and a topological non-trivial crystal are spliced ​​together, a boundary is formed at the splicing point, and two topological boundary states with opposite slopes appear in the dispersion curve of the boundary within the bandgap, corresponding to two different photon pseudospin directions, namely, pseudospin direction upward and pseudospin direction downward, and are localized near the boundary and transmitted unidirectionally along the boundary in opposite directions.

[0029] like Figure 2 and 3As shown, the on-chip optical neural network based on topological photonic crystal boundary state splicing of this embodiment and its implementation method include: a topological photonic crystal unit, the topological photonic crystal unit is a topological non-trivial crystal or a topological trivial crystal; a plurality of topological photonic crystal units with the same edge shape are closely arranged to form a topological photonic crystal array, in which the topological non-trivial crystals and the topological trivial crystals are arranged alternately, and the outer edges are spliced ​​together, and boundary states are generated at the splicing, and photons are transmitted unidirectionally at the splicing, thereby forming a boundary state path at the splicing, and the boundary state path includes a transverse boundary state path and a longitudinal boundary state path; the pseudospin direction of the photons transmitted by the boundary state path is locked; for photons with pseudospin direction upward, along the transmission direction, the left side is a topological non-trivial crystal, and the right side is a topological trivial crystal; for photons with pseudospin direction downward, along the transmission direction, the left side is a topological trivial crystal, and the right side is a topological non-trivial crystal; as shown Figure 2 As shown, for photons with pseudospins in an upward direction, they are input into the optical neural network from the transverse boundary state pathway on the left. On the transverse boundary state pathway, the top is a topological non-trivial crystal and the bottom is a topological trivial crystal. When transmitting upward, on the longitudinal boundary state pathway, the left is a topological non-trivial crystal and the right is a topological trivial crystal. When transmitting downward, on the longitudinal boundary state pathway, the left is a topological trivial crystal and the right is a topological non-trivial crystal. For photons with pseudospins in a downward direction, they are input into the optical neural network from the transverse boundary state pathway on the right. On the transverse boundary state pathway, the top is a topological non-trivial crystal and the bottom is a topological trivial crystal. When transmitting upward, on the longitudinal boundary state pathway, the left is a topological trivial crystal and the right is a topological non-trivial crystal. When transmitting downward, on the longitudinal boundary state pathway, the left is a topological non-trivial crystal and the right is a topological trivial crystal. As shown Figure 3As shown, for photons with pseudospin downward direction, they are input into the optical neural network from the transverse boundary state path on the left. On the transverse boundary state path, the top is a topological trivial crystal and the bottom is a topological non-trivial crystal. When transmitting upward, on the longitudinal boundary state path, the left is a topological trivial crystal and the right is a topological non-trivial crystal. When transmitting downward, on the longitudinal boundary state path, the left is a topological non-trivial crystal and the right is a topological trivial crystal. For photons with pseudospin upward direction, they are input into the optical neural network from the transverse boundary state path on the right. On the transverse boundary state path, the top is a topological trivial crystal and the bottom is a topological non-trivial crystal. When transmitting upward, on the longitudinal boundary state path, the left is a topological non-trivial crystal and the right is a topological trivial crystal. When transmitting downward, on the longitudinal boundary state path, the left is a topological trivial crystal and the right is a topological non-trivial crystal. Ordinary crystal; photons with specific pseudospin directions are transmitted from the transverse boundary state path on one side to the array, transmitted to the junction of the topological photonic crystal unit splicing, and transmitted upward and downward along the longitudinal boundary state path respectively, and then to the junction of the topological photonic crystal unit splicing, respectively along the transverse boundary state path, and then to the junction of the topological photonic crystal unit splicing, and then split up and down again and transmitted along the longitudinal boundary state path respectively, and the light signals transmitted along the downward and upward longitudinal boundary state paths are combined, and after combining, they continue to be transmitted forward along the transverse boundary state path; two beams of light signals propagating along two transverse boundary state paths, after being split up and down by the longitudinal boundary state path respectively, interfere when combining and merge into a transverse boundary state path. This process constitutes a neuron, and all such neurons are connected as a whole to constitute an optical neural network, such as Figure 4 As shown; a photon with a pseudospin upward direction, which is input into the optical neural network from the left transverse boundary state channel and finally output from the right side of the optical neural network, shares the same transmission path with a photon with a pseudospin downward direction, which is input into the optical neural network from the right transverse boundary state channel and finally output from the left side of the optical neural network; similarly, a photon with a pseudospin downward direction, which is input into the optical neural network from the left transverse boundary state channel and finally output from the right side of the optical neural network, shares the same transmission path with a photon with a pseudospin upward direction, which is input into the optical neural network from the right transverse boundary state channel and finally output from the left side of the optical neural network. An optical neural network shares the same transmission path; and the transmission paths of photons with pseudospin upward and photons with pseudospin downward that are input into the optical neural network along the transverse boundary state path on the same side are staggered; the way in which light signals propagate in the optical neural network, including beam splitting and two light signals phase-modulated through the transverse boundary state path and then interfering to obtain a combined light beam for transverse transmission, is completely consistent with the waveguide network, and the generation of boundary states is a collective effect produced by multiple periodically arranged topological photonic crystal units, so that damage to individual topological photonic crystal units in the optical neural network will not directly lead to significant changes in the properties of boundary states.

[0030] like Figure 4 As shown, a vertical column in the optical neural network is called a layer, where the first layer is the input layer, the last layer is the output layer, and between the input layer are the first to fourth hidden layers.

[0031] like Figure 5 As shown, each thermo-optical modulation device includes a thermo-optical material H and two electrodes E; in the lateral boundary state paths of all optical neural networks, the joints of the topological non-trivial crystal and the topological trivial crystal are covered with the thermo-optical material, and electrodes are respectively arranged on the upper and lower sides of the thermo-optical material, and the two electrodes are respectively located on both sides of the joints. Voltage is applied to the thermo-optical material through the electrodes, so that the thermo-optical material produces a thermal effect, causing the materials of the topological non-trivial crystal and the topological trivial crystal to cause carrier migration, resulting in a change in the refractive index and a phase change of the optical signal. The optical signals on the two lateral boundary state paths of the previous layer after the phase is thermo-optically modulated are transmitted forward to the junction of the topological photonic crystal unit splicing and converge into a beam. The interference generated when the two optical signals with different phases are combined changes the light intensity on this layer of the lateral boundary state paths.

[0032] In this embodiment, the distance between the centers of two adjacent basic structure unit cells is 798nm, and the side length of the six rotationally symmetrically distributed equilateral triangular air holes inside is 276nm; the contraction and expansion ratios Q of the topological trivial crystal PhC2 and the topological non-trivial crystal PhC1 are 0.95 and 1.05 respectively; the transverse and longitudinal period numbers of the topological photonic crystal array are 160 and 80 respectively; the thermo-optical material adopts a single-layer graphene with a length of 50μm and a width of 20μm, and the corresponding phase realizes a π-modulated voltage of 1.8V, and a SiO2 with a thickness of 150nm is arranged between the topological photonic crystal array and the single-layer graphene as an isolation layer; the background material adopts silicon on insulator with a thickness of 220nm; the background material is arranged on a SiO2 substrate.

[0033] Finally, it should be noted that the purpose of publishing the embodiments is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. An on-chip optical neural network based on topological photonic crystal boundary state splicing, characterized in that: The optical neural network includes: a topological photonic crystal unit, which is a topological non-trivial crystal or a topological trivial crystal; a plurality of topological photonic crystal units with the same edge shape are closely arranged to form a topological photonic crystal array, in which topological non-trivial crystals and topological trivial crystals are arranged alternately, and the outer edges are spliced ​​together, and boundary states are generated at the splicing points, and photons are transmitted unidirectionally at the splicing points, thereby forming boundary state pathways at the splicing points, and the boundary state pathways include transverse boundary state pathways and longitudinal boundary state pathways; the pseudospin direction of photons transmitted by the boundary state pathway is locked; for photons with pseudospin directions upward, along the transmission direction, the left side is a topological non-trivial crystal, and the right side is a topological trivial crystal; for photons with pseudospin directions downward, along the transmission direction, the left side is a topological trivial crystal, and the right side is a topological non-trivial crystal; photons with specific pseudospin directions are transmitted from the transverse boundary state pathway on one side to the array, The light signals are transmitted to the junction where the topological photonic crystal units are spliced, divided into two beams of light signals, which are transmitted upward and downward along the longitudinal boundary state path respectively, and then to the junction where the topological photonic crystal units are spliced, and are transmitted along the transverse boundary state path respectively, and then to the junction where the topological photonic crystal units are spliced, the beams are split up and down again and are transmitted along the longitudinal boundary state path respectively, and the light signals transmitted along the downward and upward longitudinal boundary state paths are combined, and after the beams are combined, they continue to be transmitted forward along the transverse boundary state path; the two beams of light signals propagating along the two transverse boundary state paths, after being respectively split up and down through the longitudinal boundary state path, interfere when they are combined and merge into a transverse boundary state path, this process constitutes a neuron, and all such neurons are connected as a whole to constitute an optical neural network; and the transmission paths of the photons with the pseudospin direction upward and the photons with the pseudospin direction downward input into the optical neural network along the transverse boundary state path on the same side are staggered.

2. The optical neural network according to claim 1, wherein: Photons with pseudospin upward direction are input into the optical neural network from the left transverse boundary state channel and finally output from the right side of the optical neural network, and share the same transmission path of the optical neural network with photons with pseudospin downward direction which are input into the optical neural network from the right transverse boundary state channel and finally output from the left side of the optical neural network; similarly, photons with pseudospin downward direction are input into the optical neural network from the left transverse boundary state channel and finally output from the right side of the optical neural network, and share the same transmission path of the optical neural network with photons with pseudospin upward direction which are input into the optical neural network from the right transverse boundary state channel and finally output from the left side of the optical neural network; the way in which optical signals propagate in the optical neural network, including beam splitting and two beams of light signals phase-modulated through the transverse boundary state channel and then interfering to obtain a combined light beam for transverse transmission, is completely consistent with the waveguide network, and the generation of boundary states is a collective effect generated by multiple periodically arranged topological photonic crystal units, so that damage to individual topological photonic crystal units in the optical neural network will not directly lead to significant changes in the properties of boundary states.

3. The optical neural network according to claim 1, wherein: The invention comprises a thermo-optic modulation device for thermo-optically modulating a lateral boundary state path, each of which comprises a thermo-optic material and two electrodes. In each lateral boundary state path of the optical neural network that needs to adjust the phase, the joint of the topological non-trivial crystal and the topological ordinary crystal is covered with the thermo-optic material, and electrodes are respectively arranged on the upper and lower sides of the thermo-optic material, and the two electrodes are respectively located on both sides of the joint. Voltage is applied to the thermo-optic material through the electrodes, so that the thermo-optic material produces a thermal effect, causing the materials of the topological non-trivial crystal and the topological ordinary crystal to cause carrier migration, resulting in a change in the refractive index and a change in the phase of the optical signal. The optical signals on the two lateral boundary state paths of the previous layer after the phase is thermo-optically modulated are transmitted forward to the junction of the topological photonic crystal unit splicing, and the interference of the two optical signals with different phases changes the light intensity on this layer of lateral boundary state paths.

4. The optical neural network according to claim 3, wherein: The thermo-optical material adopts a single-layer or multi-layer graphene or a novel phase-change material Sb2Se3 film.

5. The optical neural network according to claim 4, wherein: The single-layer graphene coverage length is 50 μm to 100 μm, and the corresponding phase The modulated voltage is 1~2V.

6. The optical neural network according to claim 3, wherein: It also includes an isolation layer, which is arranged between the thermo-optical material and the topological photonic crystal array.

7. The optical neural network according to claim 1, wherein: The topological trivial crystal and the topological non-trivial crystal are formed by punching holes on the background material. The topological trivial crystal and the topological non-trivial crystal respectively include a plurality of basic structural unit cells closely arranged between lattices. The outer edge of each basic structural unit cell is a regular hexagon, and there are six rotationally symmetrically distributed regular triangular air holes inside. The overall structure has C 6v Symmetry; in topologically trivial crystals, the distance between the centers of the six equilateral triangles and the center of the regular hexagon is less than 1 / 3 of the period of the two-dimensional topological photonic crystal, corresponding to a trivial topological state; in topologically non-trivial crystals, the distance between the centers of the six equilateral triangles and the center of the regular hexagon is greater than 1 / 3 of the period of the two-dimensional topological photonic crystal, corresponding to a non-trivial topological state.

8. The optical neural network according to claim 7, wherein: The topological photonic crystal unit includes at least ten periods of hexagonal basic structure cells in both the longitudinal and lateral directions.

9. The optical neural network according to claim 7, wherein: The background material uses silicon on insulator or silicon nitride.

10. A method for realizing an on-chip optical neural network based on topological photonic crystal boundary state splicing as claimed in claim 1, characterized in that: The implementation method comprises the following steps: Providing a topological photonic crystal unit, the topological photonic crystal unit is a topological non-trivial crystal or a topological trivial crystal; A plurality of topological photonic crystal units with the same edge shape are closely arranged to form a topological photonic crystal array, in which topological non-trivial crystals and topological trivial crystals are arranged alternately, and the outer edges are spliced ​​together, and a boundary state is generated at the splicing. Photons are transmitted unidirectionally at the splicing, thereby forming a boundary state path at the splicing, and the boundary state path includes a transverse boundary state path and a longitudinal boundary state path; The pseudospin direction of photons transmitted through the boundary state channel is locked; for photons with pseudospins pointing upward, along the transmission direction, the left side is a topologically non-trivial crystal, and the right side is a topologically trivial crystal; for photons with pseudospins pointing downward, along the transmission direction, the left side is a topologically trivial crystal, and the right side is a topologically non-trivial crystal; Photons with a specific pseudospin direction are transmitted from the transverse boundary state path on one side to the array, and then to the junction where the topological photonic crystal units are spliced, and then divided into two light signals, which are respectively transmitted upward and downward along the longitudinal boundary state path, and then to the junction where the topological photonic crystal units are spliced, and then respectively transmitted along the transverse boundary state path, and then to the junction where the topological photonic crystal units are spliced, and then split up and down again and respectively transmitted along the longitudinal boundary state path, and the light signals transmitted along the downward and upward longitudinal boundary state paths are combined, and then continue to be transmitted forward along the transverse boundary state path after being combined; the two light signals propagating along the two transverse boundary state paths are respectively split up and down by the longitudinal boundary state path, and then interfere when combined and merge into one transverse boundary state path, and this process constitutes a neuron, and all such neurons are connected as a whole to constitute an optical neural network; Photons with pseudospin upward direction, which are input into the optical neural network from the left transverse boundary state pathway and finally output from the right side of the optical neural network, share the same transmission path of the optical neural network with photons with pseudospin downward direction, which are input into the optical neural network from the right transverse boundary state pathway and finally output from the left side of the optical neural network; similarly, photons with pseudospin downward direction, which are input into the optical neural network from the left transverse boundary state pathway and finally output from the right side of the optical neural network, share the same transmission path of the optical neural network with photons with pseudospin upward direction, which are input into the optical neural network from the right transverse boundary state pathway and finally output from the left side of the optical neural network; and the transmission paths of photons with pseudospin upward direction and photons with pseudospin downward direction, which are input into the optical neural network along the same transverse boundary state pathway on the same side, are staggered; The way in which light signals propagate in optical neural networks, including beam splitting and two light signals phase-modulated through the transverse boundary state path and then interfering to obtain a combined light beam for transverse transmission, is exactly the same as that in waveguide networks. The generation of boundary states is a collective effect produced by multiple periodically arranged topological photonic crystal units, so that damage to individual topological photonic crystal units in the optical neural network will not directly lead to significant changes in the properties of boundary states.

Images