A photonic crystal reverse design method and a valley photonic crystal topological waveguide device
By using a two-stage physical guided diffusion model in the design of valley photonic crystal optical devices, a high-fidelity layout is directly generated in the original pixel space, which solves the problems of high computational cost and blurred edges of dielectric apertures in existing technologies, and realizes high-precision photonic crystal manufacturing and device assembly.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies have high computational costs in the forward design of valley photonic crystal optical devices, making it difficult to meet the stringent tolerance requirements of high-precision physical manufacturing and device assembly. Furthermore, the generated models are prone to blurring the edges of dielectric apertures or compromising the accuracy of topological mapping.
A two-stage physical-guided diffusion model is adopted, including a diffusion backbone network and a low-rank adapter module. The global spatial symmetry features of the unit cell are extracted through pre-training, and a high-fidelity valley photonic crystal layout is generated by hard truncation binarization. The medium distribution without secondary smoothing is generated directly in the original pixel space.
It achieves end-to-end generation from target physical parameters to directly usable manufacturing, with clear edges of generated dielectric holes that meet physical manufacturing tolerance requirements, improving design efficiency and accuracy, robustly maintaining the C3 rotational symmetry of the cell, and avoiding global structural distortion of the local convolution generation model.
Smart Images
Figure CN122131482A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a reverse design method for photonic crystals and a valley photonic crystal topological waveguide device, belonging to the interdisciplinary field of machine learning and photonic crystal design. Background Technology
[0002] Valley photonic crystals, by breaking the spatial inversion symmetry of graphene-like lattices, unwind the Dirac cone degeneracy at the K-valley and K'-valley boundaries of the Brillouin zone, thereby exciting topological edge states with non-zero valley Chern numbers. These topological edge states possess anti-backscattering capabilities, allowing electromagnetic waves to pass through sharp bends and random defects without loss, making them crucial for achieving robust on-chip optical transmission. However, the topological phase transition and band structure of valley photonic crystals are highly dependent on the geometry of the dielectric pores within the unit cell and the global rotational symmetry; accurate geometric inverse reconstruction is a core prerequisite for device development.
[0003] Currently, most research on valley photonic crystal optical devices relies primarily on forward design, which depends on empirical scanning and trial and error of cell geometry parameters. This approach is computationally expensive when dealing with complex multi-degree-of-freedom configurations. In recent years, deep learning-based inverse design methods have been gradually introduced into the field of photonic crystals. However, existing solutions often suffer from the following shortcomings: generative models relying on variational autoencoder latent space compression often result in lossy space compression, leading to blurred edges of the generated dielectric apertures; and generative networks based on local convolution struggle to maintain lattice rotational symmetry globally. For valley photonic crystals, even extremely small distortions at geometric boundaries are sufficient to trigger bandgap closure or valley Chern number inversion, compromising the accuracy of topological mapping and making it difficult for the generated structures to meet the stringent tolerance requirements of high-precision physical manufacturing and device assembly. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a reverse design method for photonic crystals and a valley photonic crystal topological waveguide device. This method can directly generate a high-fidelity valley photonic crystal layout in the original pixel space, and generate a Z-shaped topological waveguide device that suppresses backscattering at sharp-angle bends. To achieve the above objective, this invention employs the following technical solution:
[0005] In a first aspect, the present invention provides a method for reverse design of photonic crystals, comprising:
[0006] Obtain the target physical condition vector;
[0007] A pre-trained two-stage physical guided diffusion model is used for reverse diffusion denoising to obtain a continuous grayscale image.
[0008] Hard truncation binarization is performed on the continuous grayscale image to obtain the binarized layout of the valley photonic crystal;
[0009] The two-stage physically guided diffusion model includes a diffusion backbone network and a low-rank adapter module, and the training steps include:
[0010] Obtain the prior dataset and high-quality sample set of valley photonic crystals;
[0011] Within the original pixel space, the prior dataset is input into the diffusion backbone network for pre-training, enabling the diffusion backbone network to extract the global spatial symmetry features of the valley photonic crystal cell.
[0012] Freeze the pre-trained parameters of the diffusion backbone network, fine-tune the low-rank adapter module using a high-quality sample set, update the parameters of the low-rank adapter module, and obtain the pre-trained two-stage physical guided diffusion model.
[0013] In conjunction with the first aspect, optionally, the prior dataset includes a pixel matrix and physical condition vectors of two-dimensional unit cell samples, wherein both the target physical condition vector and the physical condition vectors in the prior dataset include: upper bandgap frequency, lower bandgap frequency, valley number, and relative filling rate of internal dielectric pores.
[0014] In conjunction with the first aspect, optionally, the diffusion backbone network adopts a diffusion Transformer architecture; the step of inputting the prior dataset into the diffusion backbone network for pre-training includes:
[0015] The physical condition vector is mapped to a physical embedding vector, and the physical embedding vector is fused with the diffusion time step;
[0016] The fused features are then normalized through an adaptive layer and injected into each Transformer block of the diffusion backbone network.
[0017] By utilizing the multi-head self-attention mechanism of the Transformer block, long-range spatial correlations between pixels are captured in the original pixel space to extract the global spatial symmetry features of the valley photonic crystal cell and maintain the C3 rotational symmetry of the cell.
[0018] In conjunction with the first aspect, optionally, the fine-tuning of the low-rank adapter module using a high-quality sample set includes:
[0019] The low-rank adapter module is connected to the attention layer and feedforward fully connected layer of the diffusion backbone network in a bypass manner.
[0020] For the input feature map ,according to Forward propagation is performed, in which Forward propagation characteristics, The pre-trained weight matrix is frozen. For the dimension-reduced projection matrix, For the up-dimensional projection matrix, For the rank of the low-rank adapter module, This is the scaling factor;
[0021] During fine-tuning, only the dimension-reduced projection matrix is updated. and the upgraded projection matrix Parameters are used to achieve physical alignment in a low-dimensional parameter space.
[0022] In conjunction with the first aspect, optionally, the hard truncation binarization process for the continuous grayscale image includes:
[0023] Initialize the preset pixel grayscale threshold;
[0024] Pixels with grayscale values above a preset threshold are mapped to high-refractive-index medium regions, while pixels with grayscale values below or equal to the preset threshold are mapped to air regions, resulting in a binarized layout of the valley photonic crystal.
[0025] In conjunction with the first aspect, optionally, it also includes:
[0026] Based on the scale invariance of Maxwell's equations, the binarized layout characterized by normalized frequencies is scaled to a specified physical lattice constant, so that the relative bandgap is mapped to the absolute bandgap corresponding to the target physical condition vector.
[0027] In a second aspect, the present invention provides a valley photonic crystal topological waveguide device generated by the method described in the first aspect, comprising a background dielectric and two internal dielectric apertures arranged periodically.
[0028] The central transmission region of the valley photonic crystal topological waveguide device is a whisker-shaped interface formed by splicing together a first region and a second region with opposite valley numbers. The whisker-shaped interface forms a Z-shaped transmission path by two acute-angle bends in a two-dimensional plane.
[0029] In conjunction with the second aspect, optionally, the background medium is silicon, the internal medium pores are air pores, and the two internal medium pores have different shapes or sizes to break the spatial inversion symmetry and relieve the Dirac cone degeneracy at the momentum space K valley and K' valley.
[0030] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the photonic crystal reverse design method described in the first aspect.
[0031] Fourthly, the present invention provides a computer device, comprising:
[0032] Memory, used to store computer programs / instructions;
[0033] A processor for executing the computer program / instructions to implement the steps of the photonic crystal reverse design method described in the first aspect.
[0034] Compared with the prior art, the beneficial effects achieved by the photonic crystal reverse design method and valley photonic crystal topological waveguide device provided in this invention include:
[0035] This invention obtains the target physical condition vector; utilizes a pre-trained two-stage physical guided diffusion model for back-diffusion denoising to obtain a continuous grayscale image; performs hard-truncation binarization on the continuous grayscale image to obtain the binarized layout of the valley photonic crystal; this invention employs fixed-threshold hard-truncation binarization in the generation stage, directly converting the continuous grayscale output into a well-defined two-phase medium distribution, obtaining a clear-boundary, micro-jagged binarized layout without secondary smoothing or post-processing, meeting the stringent geometric tolerance requirements of photonic crystal physical manufacturing; this invention achieves end-to-end on-demand generation from target physical parameters to a layout directly usable for manufacturing, without the need for manual iterative adjustments, significantly improving design efficiency;
[0036] The two-stage physical-guided diffusion model of this invention includes a diffusion backbone network and a low-rank adapter module. The diffusion backbone network provided by this invention learns the general geometric priors and topological morphology rules of valley photonic crystal cells in the pre-training stage, while the low-rank adapter module adapts to specific physical targets with extremely low parameter counts in the fine-tuning stage. This division of labor and cooperation enables the model to maintain a wide range of geometric generation capabilities and accurately align specific bandgap targets, overcoming the problem that single-stage models cannot balance generalization and accuracy.
[0037] This invention pre-trains the diffusion backbone network by inputting a prior dataset into the original pixel space, enabling the diffusion backbone network to extract the global spatial symmetry features of the valley photonic crystal cell. This invention operates directly in the original pixel space, avoiding the high-frequency geometric boundary blurring caused by the latent space compression of the variational autoencoder. The dielectric aperture edges generated by this invention can maintain pixel-level sharpness, meeting the tolerance requirements of physical manufacturing. This invention utilizes the global multi-head self-attention mechanism of the diffusion backbone network to accurately capture the long-range spatial correlation between discrete dielectric apertures, robustly maintain the C3 rotational symmetry of the cell, and avoid the global structural distortion commonly found in local convolutional generation models.
[0038] This invention freezes the pre-trained parameters of the diffusion backbone network, fine-tunes the low-rank adapter module using a high-quality sample set, updates the parameters of the low-rank adapter module, and obtains the pre-trained two-stage physical-guided diffusion model. The freezing of the pre-trained parameters in this invention can update only a very small number of network parameters in an extremely low-dimensional linear subspace, preventing the forgetting of topological priors under small sample conditions. The low-rank adapter module compresses the complex nonlinear physical constraints related to bandgap modulation into an extremely low-dimensional parameter space, suppressing overfitting and keeping the relative bandgap error stably at a low level, thereby improving the bandgap mapping accuracy. Attached Figure Description
[0039] Figure 1 This is a schematic diagram illustrating the working principle of the physical condition input and condition embedder in Embodiment 1 of the present invention;
[0040] Figure 2 This is a schematic diagram of the overall network architecture of the two-stage physical guided diffusion model in Embodiment 1 of the present invention;
[0041] Figure 3 This is a schematic diagram of the reverse diffusion denoising reasoning and closed-loop physical verification process of the reverse design of photonic crystal in Embodiment 1 of the present invention;
[0042] Figure 4 This is a dynamic curve of convergence of training loss and validation loss recorded in logarithmic coordinates during the pre-training stage of the diffusion backbone network in Embodiment 1 of the present invention.
[0043] Figure 5 This is a linear loss curve of the low-rank adapter module during the fine-tuning stage in Embodiment 1 of the present invention;
[0044] Figure 6 This is a quantitative comparison histogram of the relative bandgap error and relative bandgap ratio error between the two-stage generation model of the present invention and traditional generation models (CVAE, cDCGAN) in Embodiment 1 of the present invention.
[0045] Figure 7 This is a statistical analysis of the generation error and robustness of the two-stage network framework in Embodiment 1 of the present invention under different target bandgap constraints.
[0046] Figure 8 This is a schematic diagram of the projected band structure of the valley photonic crystal strip supercell generated in an embodiment of the present invention;
[0047] Figure 9 This is a broadband transmission spectrum analysis curve of the Z-type valley photonic crystal topological waveguide device assembled in Embodiment 2 of the present invention;
[0048] Figure 10This is a macroscopic splicing and assembly structure layout of the Z-type valley photonic crystal topological waveguide device generated in Embodiment 2 of the present invention;
[0049] Figure 11 This is a schematic diagram of the electromagnetic wave local transmission field distribution at a specific frequency in the Z-shaped topological waveguide generated in Embodiment 2 of the present invention;
[0050] Figure 12 This is a localized distribution diagram of electromagnetic energy in a Z-shaped topological waveguide generated according to Embodiment 2 of the present invention.
[0051] In the picture:
[0052] 100. Valley photonic crystal topological waveguide device; 101. First valley photonic crystal region; 102. Second valley photonic crystal region; 103. Beard-like interface; 104. Acute angle bend. Detailed Implementation
[0053] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0054] Example 1:
[0055] This embodiment provides a method for reverse design of photonic crystals, such as... Figure 3 As shown, the entire lifecycle process from target condition input, discrete noise reduction, hard truncation thresholding to final MPB software calculation of band structure is illustrated, including the following specific steps.
[0056] Step 1: Obtain the target physical condition vector.
[0057] The target physical condition vector is a four-dimensional feature vector, such as Figure 1 As shown, a four-dimensional physical condition vector containing bandgap characteristics and silicon fill rate is presented, specifically including: upper bandgap frequency. lower bandgap frequency Gu Chenshu Relative fill rate of internal dielectric pores (silicon fill rate) ).
[0058] The target physical condition vector is specified by the user based on the target performance parameters of the photonic crystal device to be designed.
[0059] like Figure 1 As shown, it also includes the four-dimensional physical condition vector and the diffusion time step. Blend and merge, then input the conditional embedder.
[0060] Step 2: Use a pre-trained two-stage physical guided diffusion model to perform reverse diffusion denoising to obtain a continuous grayscale image.
[0061] like Figure 2 As shown, the two-stage physically guided diffusion model (DiT+LoRA) includes a diffusion backbone network and a low-rank adapter module. The diffusion backbone network adopts a diffusion Transformer architecture, and the low-rank adapter module (LoRA) is connected to the attention layer and feedforward fully connected layer of the diffusion backbone network in a bypass manner.
[0062] The reverse diffusion denoising process is as follows: driven by the target physical condition vector, a set multi-step discrete-time inference is performed to gradually denoise the pure Gaussian random noise into a continuous grayscale image of 128×128 pixels.
[0063] Specifically, such as Figure 3 As shown, the input to the backdiffusion denoising process is pure Gaussian random noise that conforms to a standard normal distribution. Driven by the target physical condition vector, discrete-time inference of a set number of steps is performed (250 inference steps in this embodiment). As the inference time steps... from Gradually decreasing towards 0, the diffusion backbone network predicts and removes the current noise component at each time step, so that the potential geometric features are transformed from pure Gaussian random noise in a completely chaotic state. It gradually evolves into an intermediate state with a preliminary outline. And finally in Decoding continuous grayscale images with clear boundaries and high-fidelity physical morphological features at every moment is represented as... .
[0064] Step 3: Perform hard truncation binarization on the continuous grayscale image to obtain the binarized layout of the valley photonic crystal;
[0065] Step 3.1: Initialize the preset pixel grayscale threshold.
[0066] In this embodiment, a pixel grayscale value greater than 0.2 is used as the preset pixel grayscale threshold.
[0067] Step 3.2: Map pixels with grayscale values above the preset threshold to high-refractive-index medium regions, and pixels with grayscale values below or equal to the preset threshold to air regions, to obtain the binarized layout of the valley photonic crystal.
[0068] In this embodiment, the background medium is silicon with a refractive index of approximately 2.83; internal dielectric holes A and B are air holes with a refractive index of approximately 1.0. The geometric areas of internal dielectric holes A and B differ precisely controlled by an algorithm, breaking the spatial inversion symmetry at the physical level.
[0069] Step 4: Based on the scale invariance of Maxwell's equations, the binarized layout characterized by normalized frequencies is scaled to a specified physical lattice constant, so that the relative bandgap is mapped to the absolute bandgap corresponding to the target physical condition vector.
[0070] Specifically, the absolute bandgap corresponding to the target physical condition vector is the absolute macroscopic physical bandgap of the target terahertz or near-infrared frequency band.
[0071] Step 5: As Figure 3 As shown, in order to further ensure that the physical performance of the generated structure meets the preset indicators, this embodiment also includes a closed-loop physical verification process.
[0072] The binarized layout and physical dimension parameters of the valley photonic crystal (VPC) generated in step 3 are imported into a physical simulator based on the plane wave expansion method (PWEM), such as MPB simulation software. By solving Maxwell's characteristic equations, the eigenfrequency distribution of the generated structure in momentum space is calculated, and its actual band structure is obtained. The actual band structure obtained from the simulation is compared with the target physical condition vector input in step 1 in a closed-loop manner to verify the accuracy of the generated valley photonic crystal structure in terms of the target bandgap position, bandwidth, and topological properties.
[0073] In this embodiment, a fixed threshold hard-truncation binarization is used in the generation stage to directly convert the continuous grayscale output into a clear two-phase medium distribution. Without secondary smoothing or post-processing, a binarized layout with clear boundaries and no microscopic jagged edges can be obtained, which meets the strict requirements of geometric tolerance in photonic crystal physical manufacturing.
[0074] This embodiment enables end-to-end on-demand generation of layouts from target physical parameters to designs that can be directly used for manufacturing, without the need for manual iterative adjustments, thus significantly improving design efficiency.
[0075] The training steps for the two-stage physical guided diffusion model used in step 2 include the following specific steps.
[0076] Step A: Obtain the prior dataset and high-quality sample set of valley photonic crystals.
[0077] The prior dataset includes discrete pixel matrices of two-dimensional valley photonic crystal cell samples with different geometries, and physical condition feature vectors that correspond one-to-one with the discrete pixel matrices.
[0078] In this embodiment, the resolution of the discrete pixel matrix is 128×128 pixels.
[0079] In this embodiment, the physical condition feature vector is a four-dimensional vector, specifically including the upper bandgap frequency value, the lower bandgap frequency value, the valley number, and the relative filling rate of the dielectric pores.
[0080] The high-quality sample set contains valley photonic crystal samples of specific absolute bandgap targets.
[0081] Step B: In the original pixel space, the prior dataset is input into the diffusion backbone network for pre-training, so that the diffusion backbone network can extract the global spatial symmetry features of the valley photonic crystal cell.
[0082] Step B01: Map the physical condition vector to a physical embedding vector, and fuse the physical embedding vector with the diffusion time step.
[0083] Specifically, the target physical condition vector is fed into the physical condition embedder, which uses a multilayer perceptron to map the four-dimensional physical condition vector into a high-dimensional physical embedding vector.
[0084] Step B02: Inject the fused features into each Transformer block of the diffusion backbone network through adaptive layer normalization.
[0085] Step B03: Utilize the multi-head self-attention mechanism of the Transformer block to capture long-range spatial correlations between pixels in the original pixel space, thereby extracting the global spatial symmetry features of the valley photonic crystal cell and maintaining the C3 rotational symmetry of the cell.
[0086] Specifically, by utilizing the multi-head self-attention mechanism of the Transformer block, the model can simultaneously perform dot product calculations across the entire pixel space, calculating the long-range spatial relationships between pixels in the discrete pixel matrix. This allows it to directly capture the global C3 rotational symmetry of periodic lattices within the original 128×128 pixel space without relying on lossy compression from local convolutional kernels and variational autoencoders. This captures the global spatial symmetry features of periodic lattices within the original 128×128 pixel space without relying on lossy compression from variational autoencoders.
[0087] This embodiment operates directly in the original pixel space, avoiding the high-frequency geometric boundary blurring caused by the latent space compression of the variational autoencoder. The dielectric aperture edges generated by this embodiment can maintain pixel-level sharpness, meeting the tolerance requirements of physical manufacturing. By utilizing the global multi-head self-attention mechanism of the diffusion backbone network, it can accurately capture the long-range spatial correlation between discrete dielectric apertures, robustly maintain the C3 rotational symmetry of the cell, and avoid the global structural distortion commonly found in local convolutional generation models.
[0088] Step C: Freeze the pre-trained parameters of the diffusion backbone network, fine-tune the low-rank adapter module using a high-quality sample set, update the parameters of the low-rank adapter module, and obtain the pre-trained two-stage physical guided diffusion model.
[0089] During fine-tuning, all pre-trained Transformer weight matrices in the diffusion backbone network are completely frozen to protect the robust topology generation prior established by the base model.
[0090] The low-rank adapter module is bypassed into the attention layer and feedforward fully connected layer of the Transformer block. The attention layer includes a query matrix, a key matrix, a value matrix, and an output mapping matrix.
[0091] Step C01: For the input feature map ,according to Forward propagation is performed, in which Forward propagation characteristics, The pre-trained weight matrix is frozen. For the dimension-reduced projection matrix, For the up-dimensional projection matrix, For the rank of the low-rank adapter module, This is the scaling factor.
[0092] In this embodiment, the rank of the low-rank adapter module scaling factor .
[0093] Step C02: During fine-tuning, only the dimension-reduced projection matrix is updated. and the upgraded projection matrix Parameters are used to achieve physical alignment in a low-dimensional parameter space.
[0094] A grouped learning rate strategy is adopted, allocating a relatively high learning rate to the low-rank adapter module, while the learning rate of the physical condition embedder is moderately suppressed. During fine-tuning, the actual number of updated parameters is far less than 1% of the total number of parameters, which can effectively compress the nonlinear physical constraints controlling the bandgap evolution into a low-dimensional linear subspace and prevent catastrophic forgetting of topological priors.
[0095] The freezing of pre-trained parameters in this embodiment can update only a very small number of network parameters in an extremely low-dimensional linear subspace, preventing the forgetting of prior topological features under small sample conditions. The low-rank adapter module compresses the complex nonlinear physical constraints related to bandgap modulation into an extremely low-dimensional parameter space, suppressing overfitting and keeping the relative bandgap error stably at a low level, thereby improving the bandgap mapping accuracy.
[0096] The diffusion backbone network provided in this embodiment learns the general geometric priors and topological morphology of valley photonic crystal cells during the pre-training stage. The low-rank adapter module adapts to specific physical targets with extremely low parameter counts during the fine-tuning stage. This division of labor and cooperation enables the model to maintain a wide range of geometric generation capabilities and accurately align specific bandgap targets, overcoming the problem that single-stage models cannot balance generalization and accuracy.
[0097] like Figure 4 As shown, the loss convergence of the diffusion backbone network (basic DiT model) in this embodiment during the pre-training phase is illustrated. In logarithmic coordinates, the training loss and validation loss curves exhibit a smooth and consistent downward trend after 100 epochs of iteration, demonstrating that the model has fully learned the global spatial symmetry and geometric priors of the valley photonic crystal cell.
[0098] like Figure 5 As shown, during the fine-tuning phase after introducing the low-rank adapter (LoRA), the linear loss curve exhibits fine-grained optimization dynamics, and the model achieves precise alignment of physical features within an extremely short training period.
[0099] like Figure 6 As shown, in order to verify the performance advantages of the present invention, the two-stage physical guided diffusion model (DiT+LoRA) provided in this embodiment is compared with two mainstream deep learning generative models.
[0100] The two main deep learning generative models are: Conditional Variational Autoencoder (CVAE) and Conditional Deep Convolutional Generative Adversarial Network (cDCGAN).
[0101] The quantitative results show that:
[0102] The two-stage physical guided diffusion model (DiT+LoRA) provided in this embodiment has a relative bandgap error of only 3.96% and a relative bandgap ratio error of only 3.63%.
[0103] CVAE model: relative error of band gap is 37.33%, relative error of band gap ratio is 42.88%;
[0104] cDCGAN model: The relative error of band gap is as high as 126.99%, and the relative error of band gap ratio is 122.53%.
[0105] Comparative data shows that this invention overcomes the boundary blurring problem caused by CVAE latent space compression and the limitations of cDCGAN in maintaining global rotational symmetry by utilizing a global self-attention mechanism and physical guided fine-tuning in the original pixel space, achieving extremely high-precision physical inverse mapping.
[0106] like Figure 7As shown, the robustness of the framework in this embodiment under different target bandgap constraints was further tested. Under multiple target bandgap gradients ranging from 1% to 20%, the 10 randomly generated structures all exhibited extremely narrow error bands, and the average relative error remained stably at a low level. This demonstrates that the present invention has extremely strong generalization ability within the full parameter search space and can robustly generate valley photonic crystal structures that meet stringent bandgap requirements.
[0107] Example 2:
[0108] To verify the engineering potential of the geometry generated in Example 1 in a practical topological optoelectronic device, this example provides a model that is directly generated and assembled using the above-described reverse design method. Figure 10 The valley photonic crystal topological waveguide device 100 shown is illustrated.
[0109] The device includes a silicon background dielectric and two internal air holes with independent shape features arranged periodically in the silicon background dielectric.
[0110] The central transmission region of the valley photonic crystal topological waveguide device 100 includes a whisker-like interface 103 (the junction between the blue and red areas in the figure) formed by splicing together a first valley photonic crystal region 101 (blue area in the figure) and a second valley photonic crystal region 102 (red area in the figure). The unit cells used by the first valley photonic crystal region 101 and the second valley photonic crystal region 102 have opposite valley Chern numbers (i.e., single valley Chern numbers). In the diagram, the crystals are circled with boxes in the blue and red areas to highlight that the unit cells within these areas have opposite valley Chern numbers, resulting in a difference in valley Chern numbers on both sides of the interface. Based on the bulk-boundary correspondence, a gapless topological valley-locked edge state is excited in the bulk bandgap of the whisker-shaped interface 103. Based on the whisker-shaped interface 103, the valley photonic crystal topological waveguide device 100 is assembled on a two-dimensional plane into a Z-shaped topological waveguide transmission path containing two sharp bends, including acute-angle bends 104, as shown in... Figure 10 The upper right and lower left acute angles of the Z-shaped waveguide are shown. The Z-shaped topological waveguide transmission path is used to guide terahertz or near-infrared electromagnetic waves of a specific frequency band for defect-free transmission, and suppresses backscattering of electromagnetic waves at the acute angle bend 104.
[0111] like Figure 8As shown, to verify the topological physical properties of the generated structure, this embodiment performed projected band structure analysis on the reverse-designed valley photonic crystal stripe supercell. In the figure, the gray area represents bulk bands, and the white area represents the topological bandgap formed by splicing the first valley photonic crystal region 101 and the second valley photonic crystal region 102. Within the bandgap range of normalized frequencies from 0.22 to 0.28, a clear red solid line spanning the bandgap is observed, representing the excited topological valley-locked edge state. This curve has a large slope near the center of the Brillouin zone, indicating high group velocity and good optical transmission capability.
[0112] like Figure 9 As shown, through finite-difference time-domain (FDTD) full-wave electromagnetic simulation, the Z-shaped topological waveguide exhibits extremely high flat transmittance, approaching unity, within the topological bandgap frequency range. Simulation results show that the transmittance remains stably above 0.6 within a wide bandwidth range of 0.26 to 0.28 normalized frequencies, demonstrating that the structure generated by this invention possesses excellent broadband transmission characteristics and can meet the bandwidth requirements of high-performance photonic integrated circuits.
[0113] like Figure 11 As shown, the real-time magnetic field distribution inside the waveguide is demonstrated through finite-difference time-domain (FDTD) full-wave electromagnetic simulation. Re ( H z )).Depend on Figure 11 It was observed that when the electromagnetic wave propagates along the Z-shaped path, the magnetic field phase maintains a high degree of continuity even after passing through the two acute-angle bends at 104, without any drastic phase distortion or wavefront collapse. This physically demonstrates that the structure generated by this invention can perfectly maintain the phase-locked characteristics of the topological edge states.
[0114] like Figure 12 As shown, the corresponding electromagnetic energy field distribution is illustrated. The electromagnetic energy is highly localized near the whisker-shaped interface 103 and is strictly constrained by the band gap of the bulk energy bands on both sides of the interface. Even when encountering the extreme geometric path abrupt change at the acute-angle bend 104, the energy flow still diffracts smoothly, and no standing wave interference fringes caused by backscattering are observed in the transmission path, with no energy oscillation phenomenon.
[0115] Combination Figure 11 and Figure 12 The simulation results fully verify that the valley photonic crystal device generated by this invention has extremely strong topological protection characteristics and anti-backscattering capability under complex geometric configurations, and achieves robust optical transmission with near-unit efficiency.
[0116] Example 3:
[0117] This embodiment provides a computer-readable storage medium storing a computer program / instruction thereon. When the computer program / instruction is executed by a processor, it implements the steps of the photonic crystal reverse design method described in Embodiment 1.
[0118] Example 4:
[0119] This embodiment provides an electronic device, including:
[0120] Memory, used to store computer programs / instructions;
[0121] A processor is used to execute the computer program / instructions to implement the steps of the photonic crystal reverse design method described in Embodiment 1.
[0122] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0123] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0126] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A method for reverse design of photonic crystals, characterized in that, include: Obtain the target physical condition vector; A pre-trained two-stage physical guided diffusion model is used for reverse diffusion denoising to obtain a continuous grayscale image. Hard truncation binarization is performed on the continuous grayscale image to obtain the binarized layout of the valley photonic crystal; The two-stage physically guided diffusion model includes a diffusion backbone network and a low-rank adapter module, and the training steps include: Obtain the prior dataset and high-quality sample set of valley photonic crystals; Within the original pixel space, the prior dataset is input into the diffusion backbone network for pre-training, enabling the diffusion backbone network to extract the global spatial symmetry features of the valley photonic crystal cell. Freeze the pre-trained parameters of the diffusion backbone network, fine-tune the low-rank adapter module using a high-quality sample set, update the parameters of the low-rank adapter module, and obtain the pre-trained two-stage physical guided diffusion model.
2. The reverse design method for photonic crystals according to claim 1, characterized in that, The prior dataset includes the pixel matrix and physical condition vector of the two-dimensional unit cell sample. Both the target physical condition vector and the physical condition vector in the prior dataset include: upper bandgap frequency, lower bandgap frequency, valley number, and relative filling rate of internal dielectric pores.
3. The reverse design method for photonic crystals according to claim 1, characterized in that, The diffusion backbone network adopts a diffusion Transformer architecture; The step of inputting the prior dataset into the diffusion backbone network for pre-training includes: The physical condition vector is mapped to a physical embedding vector, and the physical embedding vector is fused with the diffusion time step; The fused features are then normalized through an adaptive layer and injected into each Transformer block of the diffusion backbone network. By utilizing the multi-head self-attention mechanism of the Transformer block, long-range spatial correlations between pixels are captured in the original pixel space to extract the global spatial symmetry features of the valley photonic crystal cell and maintain the C3 rotational symmetry of the cell.
4. The reverse design method for photonic crystals according to claim 1, characterized in that, The fine-tuning of the low-rank adapter module using a high-quality sample set includes: The low-rank adapter module is connected to the attention layer and feedforward fully connected layer of the diffusion backbone network in a bypass manner. For the input feature map ,according to Forward propagation is performed, in which Forward propagation characteristics, The pre-trained weight matrix is frozen. For the dimension-reduced projection matrix, For the up-dimensional projection matrix, For the rank of the low-rank adapter module, This is the scaling factor; During fine-tuning, only the dimension-reduced projection matrix is updated. and the upgraded projection matrix Parameters are used to achieve physical alignment in a low-dimensional parameter space.
5. The reverse design method for photonic crystals according to claim 1, characterized in that, The hard truncation binarization process for continuous grayscale images includes: Initialize the preset pixel grayscale threshold; Pixels with grayscale values above a preset threshold are mapped to high-refractive-index medium regions, while pixels with grayscale values below or equal to the preset threshold are mapped to air regions, resulting in a binarized layout of the valley photonic crystal.
6. The reverse design method for photonic crystals according to claim 1, characterized in that, Also includes: Based on the scale invariance of Maxwell's equations, the binarized layout characterized by normalized frequencies is scaled to a specified physical lattice constant, so that the relative bandgap is mapped to the absolute bandgap corresponding to the target physical condition vector.
7. A valley photonic crystal topological waveguide device generated using the method described in any one of claims 1 to 6, characterized in that, Includes a background medium and two internal medium pores arranged periodically; The central transmission region of the valley photonic crystal topological waveguide device is a whisker-shaped interface formed by splicing together a first region and a second region with opposite valley numbers. The whisker-shaped interface forms a Z-shaped transmission path by two acute-angle bends in a two-dimensional plane.
8. The valley photonic crystal topological waveguide device according to claim 7, characterized in that, The background medium is silicon, and the internal medium pores are air pores. The two internal medium pores are not equal in shape or size to break the spatial inversion symmetry and to resolve the Dirac cone degeneracy at the K-valley and K'-valley in momentum space.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the reverse design method for photonic crystals as described in any one of claims 1-6.
10. An electronic device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the photonic crystal reverse design method according to any one of claims 1-6.