On-chip diffractive optical neural network based on metasurface periodic nano-hole array

Abstract

An on-chip diffractive optical neural network based on a metasurface periodic nano-hole array, by introducing a periodic nano-hole array structure in the metasurface, combining a designed numerical model offline training algorithm, the precise controllable modulation of the phase of light waves is realized, has higher structural stability and good phase consistency, and is conducive to constructing an optical neural network with high-precision forward propagation capability.

Description

Technical Field

[0001] This invention relates to a technology in the field of optical neural networks, specifically an on-chip diffraction optical neural network based on a metasurface periodic nanopore array. Background Technology

[0002] Existing on-chip diffractive optical neural networks mostly employ metasurface structures as basic neuron units. These metasurface structures typically achieve phase modulation by constructing subwavelength-scale metal wire arrays. However, these structures are sensitive to geometric parameters, polarization direction, and incident angle, and are easily affected by fabrication errors and edge scattering, leading to decreased phase modulation accuracy. Under high-density integration, the enhanced scattering field between structures can also easily induce crosstalk between neurons, limiting the network's integration density and stability. Existing on-chip optical neural network technologies based on topological photonic crystal boundary state splicing rely on indirect phase modulation using thermo-optical materials in the boundary state path. This results in a limited modulation range and a lack of continuity, making high-precision forward propagation difficult and energy-intensive. Furthermore, these technologies are highly dependent on polarization state and incident direction, limiting the system's versatility and anti-interference capabilities. Summary of the Invention

[0003] This invention addresses the problems of unstable phase modulation, limited integration density, and low fabrication tolerance in existing on-chip optical neural networks. It proposes an on-chip diffraction optical neural network based on a metasurface periodic nanopore array. By introducing a periodic nanopore array structure into the metasurface, precise and controllable modulation of the light wave phase is achieved. This results in high structural stability and good phase consistency, which is beneficial for constructing optical neural networks with high-precision forward propagation capabilities.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to an on-chip diffractive optical neural network based on a metasurface periodic nanopore array, comprising: a laser input module, an on-chip beam splitter, a phase modulation component, a diffraction propagation structure, an output coupling component, and a signal receiving component connected sequentially in the optical path. The laser input module generates a stable and polarization-tunable single-mode optical signal. The on-chip beam splitter divides the single-path optical signal into multiple equal-power sub-signals. The phase modulation component maps the input vector data to the phase information of the corresponding channel and loads it onto the corresponding optical signal. The diffraction propagation structure performs precise phase delay modulation on the passing optical signal through the periodic nanopore array structure. The output coupling component extracts and spatially separates the modulated optical signal based on the light field interference results at different spatial positions on the chip, obtaining the light intensity distribution corresponding to the output layer of the optical neural network. The signal receiving component collects and performs photoelectric conversion on the light intensity of each output port, performs normalization processing, and outputs the final neural network classification or recognition result.

[0006] The diffraction propagation structure includes a multilayer periodic nanopore array structure integrated on a silicon-based metasurface. Each layer of the periodic nanopore structure consists of multiple circular nanopore units. Different units are provided with different pore diameters, pore spacings, or internal filling materials within the same period to regulate the local equivalent refractive index distribution, thereby achieving precise phase delay modulation of the light signal passing through the structure. Technical effect

[0007] This invention employs a periodic nanopore array structure with subwavelength scale dimensions. By adjusting the aperture, period length, and refractive index of the filling material, continuous phase control can be achieved within a limited area, increasing the integration density of neurons and meeting the requirements of highly integrated optical computing systems. Compared with existing technologies, this invention achieves high-precision and continuously tunable phase control of optical signals, while significantly suppressing edge scattering and crosstalk between neurons, improving the stability of phase modulation and the inference accuracy of the system. While ensuring phase control accuracy, it effectively reduces the size of neuron units, enhancing the integration density and scalability of the neural network, providing a novel device foundation and implementation path for constructing high-performance, low-power, and scalably integrated on-chip optical neural networks. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the structure of the present invention;

[0009] Figure 2 This is a schematic diagram of the diffraction propagation structure and a magnified schematic diagram of a portion of the nanopore.

[0010] Figure 3 For process flow diagram;

[0011] Figure 4 This is a schematic diagram of the main module structure in an embodiment;

[0012] Figure 5 The image shown is a scanning electron microscope image of the periodic nanopore array in the embodiment.

[0013] Figure 6 This is a confusion matrix diagram showing the performance of an example in a classification task. Detailed Implementation

[0014] like Figure 1As shown in this embodiment, an on-chip diffractive optical neural network based on a metasurface periodic nanopore array is included. It comprises a laser input module, an on-chip beam splitter, a phase modulation component, a diffraction propagation structure, an output coupling component, and a signal receiving component, sequentially connected in the optical path. Specifically: the laser input module generates a stable and polarization-tunable single-mode optical signal; the on-chip beam splitter divides the single-path optical signal into multiple equal-power sub-light signals; the phase modulation component maps the input vector data to the phase information of the corresponding channel and loads it onto the corresponding sub-light signal; the diffraction propagation structure performs precise phase delay modulation on the passing sub-light signals through the periodic nanopore array structure; the output coupling component extracts and spatially separates the modulated optical signal based on the light field interference results at different spatial positions on the chip, obtaining the light intensity distribution corresponding to the output layer of the optical neural network; and the signal receiving component collects and photoelectrically converts the light intensity at each output port, performs normalization processing, and outputs the final neural network classification or recognition result.

[0015] The laser input module includes a laser and a polarization controller connected thereto for adjusting the polarization direction of the incident light. In practical applications, the laser wavelength can be set to 1550 nanometers in the C-band to be compatible with silicon photonics platforms.

[0016] The aforementioned beam splitting component is a multimode interference beam splitter. If 8-dimensional input data needs to be processed, the laser signal can be split into 8 channels of optical signal; if multiple input samples need to be loaded simultaneously, the optical signal can also be split into more channels.

[0017] The phase modulation component is implemented as a thermo-optic modulator, an electro-optic modulator, or other on-chip phase loading structure.

[0018] The input vector data is loaded onto the phase of the optical signal in each channel through a set linear mapping relationship, specifically: the phase value loaded into the i-th channel. ,in: This represents the normalized input value of the i-th input channel. and These are the lower and upper limits of the adjustable phase of the system, respectively, and are generally set between 0 and 2π.

[0019] The mapping is achieved by controlling the driving voltage of the thermo-optic modulator or the electro-optic modulator, preferably by adjusting the driving voltage to achieve continuous phase adjustment.

[0020] like Figure 2As shown, the diffraction propagation structure includes: a multilayer periodic nanopore array structure integrated on a silicon-based metasurface, i.e., a multilayer periodic nanopore modulation structure is stacked sequentially in the light propagation direction. Each layer of periodic nanopore structure consists of multiple sets of circular nanopore units, each with independent phase modulation function, used to carry weight information of different layers of the neural network. Different nanopore units are set with different apertures, aperture spacings or internal filling materials in the same period, thereby achieving local equivalent refractive index modulation in the subwavelength scale. Multiple periodic units are uniformly arranged in a plane perpendicular to the light propagation direction to form a two-dimensional planar array structure, thereby achieving precise phase delay modulation of the light signal passing through the structure.

[0021] like Figure 2 As shown, in the multilayer periodic nanopore array structure, a certain propagation spacing is provided between adjacent periodic circular nanopore structures to realize the interference and transmission of light field in space, simulate the propagation process between hidden layers in a neural network, and make the overall diffraction propagation structure have a stronger expressive ability.

[0022] The propagation process between hidden layers in the neural network follows the Huygens-Fresnel principle. Therefore, a numerical model can be constructed based on this principle during the design phase and trained offline to determine the specific size parameters of the corresponding periodic nanopore array structure and obtain the optimal phase parameters, specifically including:

[0023] Step A: Establish the optical field propagation model between the input layer and the output layer, specifically: Let the complex amplitude on the input surface be... The output light field at the receiving surface at a propagation distance of z Where: Fresnel diffraction kernel function λ is the wavelength of the light wave, the wave number k = 2π / λ, and j is the imaginary unit. By setting the propagation distance z and the phase modulation distribution between each layer, it can be regarded as the forward propagation process of an optical neural network, and the output light field can be numerically simulated.

[0024] Step B, based on the output light field simulated at the receiving surface Constructing a loss function corresponding to the target task and achieving end-to-end training of the phase array within the entire optical neural network structure, specifically including:

[0025] B1. Define the error function between the output light intensity and the target value, specifically: loss function. ,in: Let |Uᵢ|² be the complex amplitude of the optical field in the i-th output channel at the receiving surface, and |Uᵢ|² be the corresponding light intensity. Where N is the target output value, N is the number of output channels, and L is the mean square error loss function, used to measure the deviation between the analog output and the desired output under the current phase configuration.

[0026] B2. Perform gradient backpropagation on the adjustable phase delay parameter according to the loss function, and update the phase delay parameter through an optimization algorithm, specifically: ,in: Let be the trainable phase delay parameter of a certain periodic nanopore array. For learning rate, This is the partial derivative of the loss function with respect to the phase delay unit, which can be obtained by combining the chain rule with the propagation model and is used to guide the update direction of the phase delay array in each training iteration.

[0027] Step C involves mapping the optimal phase distribution parameters obtained during the offline training phase to specific phase delay structures, determining the pore size, arrangement, and internal filling material type of the circular pores in each periodic nanopore array, and realizing the phase-to-structure mapping relationship. This specifically includes:

[0028] C1. Based on the relationship between the phase delay value obtained during training and the equivalent refractive index of the periodic nanopore array, the effective refractive index is deduced, specifically: Where: λ is the wavelength of the incident light, ρ is the effective refractive index of the periodic nanopore region, n0 is the background refractive index without the structure, ρ is the periodic spacing between the centers of the two nanopores, and m·ρ is the total length of the structure in the direction of light propagation.

[0029] C2, when the refractive index n of the filling material fill and the refractive index n of the substrate material sub When determining the effective refractive index, the phase delay φ obtained from training is used to inversely calculate the corresponding effective refractive index. Then, by combining the fill factor and the mapping relationship between the effective refractive index and the nanopore structure, the pore size r and period length ρ of the nanopore are deduced.

[0030] The fill factor, i.e., the area ratio of nanopores per unit period, is specifically as follows: .

[0031] The mapping relationship between the effective refractive index and the nanoporous structure is as follows: ,in: The refractive index of the filling material, denoted as the refractive index of the substrate material.

[0032] Step D involves establishing a structural parameter lookup table and combining it with a minimum error matching algorithm to achieve the mapping process between the phase delay and the periodic nanopore structure parameters described in Step C. Specifically, this involves: pre-modeling and calculating the equivalent refractive index under different pore sizes, period lengths, and filling material combinations using simulation software to generate a database of structural parameters and phase responses; after offline training, for the optimal phase value at each spatial location, using the minimum error matching algorithm to retrieve the corresponding pore size and period length from the database based on the refractive indices of the filling material and the base material, as a manufacturable graphic output, ultimately obtaining a periodic nanopore array pattern that meets the design accuracy requirements, thus realizing the entire process of conversion from the target phase distribution to structural design parameters.

[0033] like Figure 3 As shown, this embodiment relates to a method for preparing the above-mentioned diffraction propagation structure, including:

[0034] Step 1: Prepare a substrate material consisting of a 1-micron thick buried oxide layer of silica and a 220-nanometer thick single-crystal silicon layer.

[0035] Step 2: After determining the specific size parameters, i.e. the optimal phase parameters, of the corresponding periodic nanopore array structure through offline training, the pattern of these size parameters is realized on the surface of a single-crystal silicon layer by electron beam exposure.

[0036] Step 3: Transfer the pattern obtained in Step 2 to the single-crystal silicon layer using reactive ion etching.

[0037] Step 4: Fill the etched nanopores with silica material by chemical vapor deposition to form a functional structure array with different pore sizes and filling ratios.

[0038] Step 5: Deposit a silicon dioxide cladding layer on the surface of the single-crystal silicon layer.

[0039] like Figure 4 As shown in the figure, through specific experiments, a single-mode laser source with an input power of 30dBm and a center wavelength of 1550nm, a multi-channel voltage source to power the phase modulation unit, and the above-mentioned diffraction optical neural network chip with a periodic circular nanopore array structure were used. The light intensity distribution at the output coupling port was recorded by a photodetector, and the output signal was normalized and used as the neural network output for classification tasks. The dataset used was the standard IRIS (iris) dataset, and the function of the neural network was to classify and identify different types of iris samples.

[0040] The diffractive optical neural network chips described have two overall dimensions: 0.5mm × 0.15mm and 1.0mm × 0.15mm, corresponding to single-layer and three-layer periodic nanopore array structures. For example... Figure 5The image shown is a scanning electron microscope (SEM) image of the periodic nanopore array structure. The designed nanopores have a pore size range of 170 nm to 270 nm, and the fabrication error is controlled within ±10 nm, meeting the subwavelength scale structural accuracy requirements and verifying the device's manufacturability and experimental repeatability.

[0041] This experiment performed image classification using the above settings, and the results are shown in Table 1. This table shows the actual reasoning process of different input samples in the three-layer diffraction neural network, verifying that the structure has clear classification and judgment capabilities and good readability in the integrated state.

[0042] Table 1

[0043] Table 2 shows the performance changes corresponding to different numbers of structural layers.

[0044] Table 2

[0045] like Figure 6 As shown, the confusion matrices of single-layer and three-layer diffraction neural networks on the IRIS dataset reflect the discriminative power and stability of multi-layer structures in practical classification tasks.

[0046] In summary, this invention achieves spatial modulation and forward propagation of optical signals by integrating a beam splitter, a phase modulation structure, and a multilayer periodic nanopore array on a silicon-based chip. Combined with the designed numerical model offline training algorithm, the phase modulation strategy can be optimized at the system-level design stage, further improving overall inference accuracy and functional reliability. The periodic nanopore array structure used in this chip offers advantages such as compactness, high phase modulation accuracy, and strong manufacturing process compatibility, which is beneficial for realizing high-density integrated on-chip optical neural networks.

[0047] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. An on-chip diffraction optical neural network based on a metasurface periodic nanopore array, characterized in that, include: The optical path is sequentially connected to a laser input module, an on-chip beam splitter, a phase modulation module, a diffraction propagation structure, an output coupling module, and a signal receiving module. Specifically: the laser input module generates a stable, polarization-tunable single-mode optical signal; the on-chip beam splitter divides the single-path optical signal into multiple equal-power sub-signals; the phase modulation module maps the input vector data to the phase information of the corresponding channel and loads it onto the corresponding optical signal; the diffraction propagation structure uses a periodic nanopore array structure to precisely delay the phase of the passing optical signal; the output coupling module extracts and spatially separates the modulated optical signal based on the light field interference results at different spatial positions on the chip, obtaining the light intensity distribution corresponding to the output layer of the optical neural network; and the signal receiving module collects and photoelectrically converts the light intensity at each output port, performs normalization processing, and outputs the final neural network classification or recognition result. The diffraction propagation structure includes a multilayer periodic nanopore array structure integrated on a silicon-based metasurface. Each layer of the periodic nanopore structure consists of multiple circular nanopore units. Different units are provided with different pore diameters, pore spacings, or internal filling materials within the same period to regulate the local equivalent refractive index distribution, thereby achieving precise phase delay modulation of the light signal passing through the structure.

2. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 1, characterized in that, The input vector data is loaded onto the phase of the optical signal in each channel through a set linear mapping relationship, specifically: the phase value loaded into the i-th channel. ,in: This represents the normalized input value of the i-th input channel. and These are the lower and upper limits of the adjustable phase of the system, respectively. The mapping is achieved by controlling the driving voltage of the thermo-optic modulator or the electro-optic modulator.

3. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 1, characterized in that, The diffraction propagation structure includes: a multilayer periodic nanopore array structure integrated on a silicon-based metasurface, i.e., a multilayer periodic nanopore modulation structure is stacked sequentially in the light propagation direction. Each layer of periodic nanopore structure consists of multiple sets of circular nanopore units, each with independent phase modulation function, used to carry weight information of different layers of the neural network. Different nanopore units are set with different apertures, spacings or internal filling materials in the same period, thereby achieving local equivalent refractive index modulation in the subwavelength scale. Multiple periodic units are uniformly arranged in a plane perpendicular to the light propagation direction to form a two-dimensional planar array structure, thereby achieving precise phase delay modulation of the light signal passing through the structure.

4. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 3, characterized in that, In the aforementioned multi-layer periodic nanopore array structure, a propagation spacing is provided between adjacent periodic circular nanopore structures to realize the interference and transmission of light field in space, simulating the propagation process between hidden layers in a neural network, thereby enabling the overall diffraction propagation structure to have a stronger expressive power.

5. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 3, characterized in that, The propagation process between hidden layers in the neural network follows the Huygens-Fresnel principle. Therefore, a numerical model can be constructed based on this principle during the design phase and offline training can be performed to determine the specific size parameters of the corresponding periodic nanopore array structure and obtain the optimal phase parameters.

6. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 5, characterized in that, The optimal phase parameters are obtained in the following way: Step A: Establish the optical field propagation model between the input layer and the output layer, specifically: Let the complex amplitude on the input surface be... The output light field at the receiving surface at a propagation distance of z Where: Fresnel diffraction kernel function λ is the wavelength of light, the wave number k = 2π / λ, and j is the imaginary unit. By setting the propagation distance z and the phase modulation distribution between each layer, it can be used as the forward propagation process of the optical neural network to numerically simulate the output light field. Step B, based on the simulated output light field at the receiving surface Construct a loss function corresponding to the target task to achieve end-to-end training of the phase array in the entire optical neural network structure; Step C involves mapping the optimal phase distribution parameters obtained during the offline training phase to specific phase delay structures, determining the pore size, arrangement, and internal filling material type of the circular holes in each periodic nanopore array, and realizing the mapping relationship from phase to structure. Step D involves establishing a structural parameter lookup table and combining it with a minimum error matching algorithm to realize the mapping process between the phase delay and the periodic nanopore structure parameters described in Step C. Specifically, this involves: using simulation software to model and calculate the equivalent refractive index under different pore sizes, period lengths, and filling material combinations in advance, generating a database of structural parameters and phase responses; after offline training, for the optimal phase value at each spatial location, using the minimum error matching algorithm to retrieve the pore size and period length corresponding to the refractive index of the filling material and the refractive index of the base material from the database, as a manufacturable graphic output, ultimately obtaining a periodic nanopore array pattern that meets the design accuracy requirements, thus realizing the full-process conversion from the target phase distribution to the structural design parameters.

7. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 6, characterized in that, Step B specifically includes: B1. Define the error function between the output light intensity and the target value, specifically: loss function. ,in: Let be the complex amplitude of the optical field of the i-th output channel at the receiving surface. ² represents the corresponding light intensity. The target output value is N, the number of output channels is N, and L is the mean square error loss function, which is used to measure the deviation between the analog output and the desired output under the current phase configuration. B2. Perform gradient backpropagation on the adjustable phase delay parameter according to the loss function, and update the phase delay parameter through an optimization algorithm, specifically: ,in: Let be the trainable phase delay parameter of a certain periodic nanopore array. For learning rate, This is the partial derivative of the loss function with respect to the phase delay unit, which can be obtained by combining the chain rule with the propagation model and is used to guide the update direction of the phase delay array in each training iteration.

8. The on-chip diffraction optical neural network based on a metasurface periodic nanopore array according to claim 6, characterized in that, Step C specifically includes: C1. Based on the relationship between the phase delay value obtained during training and the equivalent refractive index of the periodic nanopore array, the effective refractive index is deduced, specifically: Where: λ is the wavelength of the incident light, The effective refractive index of the periodic nanoporous region, ρ is the background refractive index without the structure loaded, ρ is the periodic spacing between the centers of the two nanopores, and m·ρ is the total length of the structure in the direction of light propagation. C2, when the refractive index n of the filling material fill and the refractive index n of the substrate material sub When determined, based on the phase delay obtained during training. Inversely calculate the corresponding effective refractive index Then, by combining the fill factor and the mapping relationship between the effective refractive index and the nanopore structure, the pore size r and period length ρ of the nanopore are deduced. The fill factor, which is the area ratio of nanopores per unit period, is specifically: ; The mapping relationship is as follows: ,in: The refractive index of the filling material, denoted as the refractive index of the substrate material.

9. A method for preparing an on-chip diffractive optical neural network according to any one of claims 1-8, characterized in that, include: Step 1: Prepare a substrate material consisting of a 1-micron thick buried oxide layer of silica and a 220-nanometer thick single-crystal silicon layer; Step 2: After determining the specific size parameters, i.e. the optimal phase parameters, of the corresponding periodic nanopore array structure through offline training, the pattern of these size parameters is realized on the surface of a single-crystal silicon layer by electron beam exposure. Step 3: Transfer the pattern obtained in Step 2 to the single-crystal silicon layer using reactive ion etching (RIE) process; Step 4: Fill the etched nanopores with silica material by chemical vapor deposition to form a functional structure array with different pore sizes and filling ratios. Step 5: Deposit a silicon dioxide cladding layer on the surface of the single-crystal silicon layer.

Images