CN-121995626-A - Reverse design method of superlens

CN121995626ACN 121995626 ACN121995626 ACN 121995626ACN-121995626-A

Abstract

The application discloses a reverse design method of a superlens, and belongs to the technical field of optical equipment. The method comprises the steps of S1, constructing a simulation data set containing paired structural images and spectrum data, S2, adopting a double-channel variation self-encoder to encode the structural images and the spectrum data into a low-dimensional latent space respectively, S3, constructing a reverse design single-step diffusion model, generating the structural latent characterization and decoding the structural latent characterization into the structural images through single-step denoising under the condition of a target spectral latent characterization, S4, constructing a forward prediction single-step diffusion model, generating the spectral latent characterization and decoding the spectral latent characterization into the spectrum data through single-step denoising under the condition of the structural latent characterization, S5, introducing a period consistency training mechanism, and carrying out joint optimization on the two models through two closed loops of a structural period and a spectral period to force the models to keep physical consistency in round-trip mapping. The model stability and the data utilization rate are improved through the two-channel subspace coding-two-way single-step diffusion-period consistency combined training closed-loop design.

Inventors

Ju Fayin
LI NING

Assignees

浙江优众新材料科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A method of reverse design of a superlens comprising the steps of: S1, constructing a simulation data set containing paired structural images and spectrum data; S2, constructing a double-channel latent space characterization framework, and respectively encoding the structural image and the spectrum data into independent latent spaces; s3, constructing a reverse design single-step diffusion model, generating a corresponding structural latent representation through single-step denoising under the condition of the latent representation of the target spectrum, and decoding the structural latent representation into a structural image; S4, constructing a forward prediction single-step diffusion model, generating a corresponding spectrum latent representation through single-step denoising under the condition of the latent representation of the structural image, and decoding the spectrum latent representation into spectrum data; S5, introducing a periodic consistency training mechanism, and performing joint optimization on the reverse design single-step diffusion model and the forward prediction single-step diffusion model by constructing two closed loops of a structural period and a spectrum period so as to force the models to keep consistency in round trip mapping.
2. The superlens retrospective design method according to claim 1, wherein in step S2, the two-channel latent space characterization framework is implemented by two independent variational automatic encoders, comprising: A structural variation automatic encoder for encoding the structural image into a first latent vector and reconstructing; A spectral variation automatic encoder for encoding spectral data into a second latent vector and reconstructing; The denoising processes in steps S3 and S4 are performed in the latent spaces where the first latent vector and the second latent vector are located, respectively.
3. The method of claim 1, wherein in steps S3 and S4, the reverse-engineered single-step diffusion model and the forward-predicted single-step diffusion model are both constructed based on a diffusion transducer architecture.
4. The method of claim 3, wherein the reverse-designed single-step diffusion model and the forward-predicted single-step diffusion model both perform single-step denoising operations at a fixed time step that is the most noisy during diffusion when reasoning.
5. The superlens reverse design method according to claim 1, wherein the loss function employed in training the reverse design single step diffusion model and the forward predicted single step diffusion model separately comprises: A loss function for structural image reconstruction that combines pixel-level mean square error and perceived quality loss; a loss function for spectral data reconstruction that uses a mean square error.
6. The method according to claim 1, wherein in step S5, the structural period is that a real structural image is reconstructed into a first reconstruction structure through forward prediction and reverse design, and a first period loss between the real structural image and the first reconstruction structure is calculated; the spectrum period is that real spectrum data is reversely designed and then is reconstructed into a first reconstructed spectrum through forward prediction, and a second period loss between the real spectrum data and the first reconstructed spectrum is calculated; The jointly optimized total loss function includes the first periodic loss and the second periodic loss.
7. The superlens reverse design method according to claim 6, wherein the first period loss and the second period loss are both calculated using L2 norms.
8. The superlens reverse design method according to claim 1 or 6, wherein the period consistency training mechanism supports self-supervised training using unlabeled data, wherein unlabeled data including only structural images participates in training of the structural period and unlabeled data including only spectral data participates in training of the spectral period.
9. The superlens reverse design method according to claim 1, further comprising a one-to-many reverse design step: Fixing a latent characterization of the target spectrum; Sampling different Gaussian noise for multiple times, and respectively inputting the Gaussian noise to the trained reverse design single-step diffusion model; and respectively denoising and decoding the single-step diffusion model through the reverse design to obtain a plurality of different metamaterial structure images.
10. The method of claim 9, wherein the plurality of different metamaterial structure images are validated via the forward predictive single step diffusion model with predicted spectral data that are all consistent with the target spectrum within a predetermined tolerance.

Description

Reverse design method of superlens Technical Field The application relates to the technical field of optical equipment and nano photonics, in particular to a reverse design method of a superlens. Background Reverse engineering of superlenses and metamaterials is a central challenge in the nanophotonics field. The traditional method relies on experience and electromagnetic simulation (such as FDTD or COMSOL) based on the first principle, and is high in calculation cost, long in time consumption and difficult to cover a wide design space. In recent years, deep learning, and in particular generative models, have been introduced to learn complex mappings between structure and spectra. However, the existing method is generally used for independently training a forward model and a reverse model, and has the problems of inconsistent design results and target spectrum due to period consistency deficiency, unstable training, low reasoning speed and high data acquisition cost due to dependence on a large amount of paired annotation data. In view of the above problems, a reverse design framework capable of simultaneously guaranteeing efficient reasoning and physical consistency and supporting unlabeled data training has not been proposed in the prior art. Therefore, a novel cross-modal generation model is needed to realize rapid and diversified superlens reverse design on the premise of keeping physical rationality. Disclosure of Invention In order to solve the problems, the application provides a superlens reverse design method to solve the problems of low reasoning efficiency, poor physical consistency, strong data dependence and the like in the prior art. The first technical scheme adopted by the application is that the application provides a reverse design method of a superlens, which comprises the following steps: S1, constructing a simulation data set containing paired structural images and spectrum data; S2, constructing a double-channel latent space characterization framework, and respectively encoding the structural image and the spectrum data into independent latent spaces; s3, constructing a reverse design single-step diffusion model, generating a corresponding structural latent representation through single-step denoising under the condition of the latent representation of the target spectrum, and decoding the structural latent representation into a structural image; S4, constructing a forward prediction single-step diffusion model, generating a corresponding spectrum latent representation through single-step denoising under the condition of the latent representation of the structural image, and decoding the spectrum latent representation into spectrum data; S5, introducing a periodic consistency training mechanism, and performing joint optimization on the reverse design single-step diffusion model and the forward prediction single-step diffusion model by constructing two closed loops of a structural period and a spectrum period so as to force the models to keep consistency in round trip mapping. In an alternative embodiment, in step S2, the dual-channel latent space characterization framework is implemented by two independent variational automatic encoders, including: A structural variation automatic encoder for encoding the structural image into a first latent vector and reconstructing; A spectral variation automatic encoder for encoding spectral data into a second latent vector and reconstructing; The denoising processes in steps S3 and S4 are performed in the latent spaces where the first latent vector and the second latent vector are located, respectively. In an alternative embodiment, in steps S3 and S4, the reverse-engineered single-step diffusion model and the forward-predicted single-step diffusion model are both constructed based on a diffusion transducer architecture. In an alternative embodiment, the reverse-engineered single-step diffusion model and the forward-predicted single-step diffusion model both perform a single-step denoising operation at a fixed time step that is the noisiest in the diffusion process when reasoning. In an alternative embodiment, the loss function employed in training the reverse-engineered single-step diffusion model and the forward-predicted single-step diffusion model separately includes: A loss function for structural image reconstruction that combines pixel-level mean square error and perceived quality loss; a loss function for spectral data reconstruction that uses a mean square error. In an optional embodiment, in step S5, the structure period is that the real structure image is reconstructed into a first reconstructed structure through forward prediction and reverse design, and a first period loss between the real structure image and the first reconstructed structure is calculated; the spectrum period is that real spectrum data is reversely designed and then is reconstructed into a first reconstructed spectrum through forward prediction, and a second period lo