Superlens transmission efficiency prediction method based on multi-mode decomposition

Abstract

The invention discloses a superlens transmission efficiency prediction method based on multi-mode decomposition, and the method comprises the steps: modeling a superatom structure, and constructing a diversified structure data set; obtaining transmission efficiency and phase curves of the superatoms at different incident angles, and constructing a data set comprising a plurality of groups of samples; decomposing the two low-rank adaptation LoRA into a superatom inherent structure characteristic and an optical response characteristic; performing two-stage prototype initialization, and performing feature-prototype matching; two-stage prototype scene embedding is calculated, and a two-way cross attention mechanism is used for enhancing features; performing mode matching alignment on the structural mode and the optical mode of each meta-structure atom; constructing a decomposition-fusion decoder, and carrying out dual decomposition and dual-path progressive fusion on the features; and finally, multi-angle prediction head design is carried out, and special prediction is provided for each incident angle. According to the invention, the problems of limited working bandwidth, difficulty in multi-angle response prediction and difficulty in consideration of high transmittance and high phase precision in the traditional super lens design are effectively solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical equipment, and particularly relates to a superlens transmission efficiency prediction method based on multi-modal decomposition. BACKGROUND

[0002] In recent years, the inverse design of superlenses (Metalens) has attracted widespread attention in the academic and industrial fields in the aspect of expanding the receiving field of view of ultrasonic laser radars. With the rapid development of optical communication and autonomous driving technologies, three-dimensional imaging laser radar systems are highly valued due to their ability to achieve precise environmental perception. Among them, MEMS has become a promising miniaturization solution due to its small size, low power consumption and high integration level. Compared with traditional mechanical rotating laser radars, the MEMS solution realizes higher system compactness by removing the heavy mechanical rotating components; compared with all-solid-state laser radars, the technology also shows higher maturity and feasibility, and the enhanced transmission power also endows it with long-distance detection capability. However, the practical application of such systems is restricted by the severe limitation of the receiving field of view, and the main reason is that the photosensitive area of the avalanche photodiode detector used in most MEMS systems is much smaller than that of the traditional area array CCD device. Although the avalanche photodiode has the advantages of high-speed response and internal signal gain, its effective photosensitive area is extremely small, and the large-area photodiode technology is still in the development stage, which has the bottleneck of complex process and high cost.

[0003] As an artificial optical device based on a two-dimensional superatom array, superlenses can achieve flexible wavefront control through phase modulation, and have shown great potential in many applications such as longitudinal light field control, wide-angle imaging, wideband chromatic aberration correction and all-solid-state laser radar. In the traditional superlens design process, establishing an accurate mapping between phase modulation and superatom parameters is a core link and also the most computationally intensive and iterative stage. Deep learning technology, especially deep neural networks, has been widely introduced as an efficient alternative to traditional numerical simulation, which can significantly alleviate the computational bottleneck and achieve compact parameterization of complex superatom geometry, but existing design methods still have four limitations: only suitable for simple nanostructures, limited working bandwidth, unable to predict multi-angle response, and difficult to achieve high transmittance and high phase accuracy at the same time. These deficiencies limit the performance of superlenses in real multi-angle receiving scenarios and restrict their effective application in systems requiring wide field of view detection.

[0004] Therefore, a superlens transmission efficiency prediction method based on multi-modal decomposition is provided. SUMMARY

[0005] To solve the above problems existing in the prior art, the application provides a multi-modal decomposition super-lens transmission efficiency prediction method, which effectively alleviates the problems of limited working bandwidth, difficult multi-angle response prediction, and difficult compatibility of high transmission rate and high phase accuracy in traditional super-lens design.

[0006] The technical solution for achieving the above-mentioned purposes is: A multi-modal decomposition super-lens transmission efficiency prediction method, comprising: Step S1, setting super-atom structure parameters, performing super-atom structure modeling, and constructing a diversified structure data set through random generation, dense structure, sparse structure, and continuous block structure; Step S2, simulating by means of Lumerical FDTD software, obtaining the transmission efficiency and phase curve of the super-atom under different incident angles, and constructing a data set including multiple groups of samples; Step S3, representing the input by two low-rank adaptive LoRAs, respectively, and decomposing the input into super-atom inherent structure characteristics and optical response characteristics; Step S4, double-level prototype initialization and feature-prototype matching; Step S5, calculating double-level prototype scene embedding and using a bidirectional cross-attention mechanism to enhance the features; Step S6, modality matching alignment of the structure modality and optical modality of each super-constructed atom; Step S7, constructing a decomposition-fusion decoder, performing angle and frequency double decomposition on the features extracted by the encoder, and performing angle path progressive fusion and frequency path progressive fusion; Step S8, finally performing multi-angle prediction head design, non-shared parameter prediction head, first enhancing angle features, and then double-branch predicting phase and transmission efficiency to provide special prediction for each incident angle.

[0007] Preferably, in step S1, the super-atom structure modeling comprises: Firstly, the different design logics of the center and the edge of the super-lens are determined, the center super-lens size is set to be consistent with the input aperture as 500μm, and the edge super-lens is specially adapted to the focusing requirement of oblique incident light; Secondly, the super-atom phase distribution and structure parameters are determined, the height, period, and discrete period are set to be 700nm, 700nm, and 640nm respectively, and the super-atom structure modeling is a 32×32 binary discrete matrix, wherein “0” represents no structure, and “1” represents existing structure; Finally, the symmetry of the structure on the x-axis and the y-axis is utilized to compress the 32×32 binary discrete matrix to 16×16.

[0008] Preferably, in the step S2, the incident angle includes 0°, 10°, 20°, 30°, and the transmission efficiency and phase curve in the wavelength range of 1500-1600 nm; a data set including multiple groups of samples is constructed by diversifying the structure data set and the transmission efficiency and phase curve, and in the data set, 50% are random structures and 50% are directional generated structures.

[0009] Preferably, in the step S3, the decomposition technology for decomposing the input into the superatom intrinsic structure characteristics and the optical response characteristics is composed of prompt decomposition technology, column separation technology and block hierarchical technology, wherein, Prompt decomposition technology: In the typical cross-attention layer of the diffusion U-net model, the weight is used to embed the prompt into the key and value : ; Three independent prompts are used to calculate three groups of keys and values: One group uses the base model and the combined prompt , and the other two groups respectively use the LoRA of the optical response characteristics and the LoRA of the superatom intrinsic structure characteristics and their corresponding prompts, and then the obtained feature maps are added: ; In the formula, is the embedding vector of the original prompt, is the embedding vector of the optical response characteristics, is the embedding vector of the intrinsic structure characteristics, and are low-rank adaptation matrices for the two types of characteristics, Column separation technology: The concept of "column" is introduced to set the corresponding mask and for each LoRA, and the column mask controls the contribution of each element: ; Block hierarchical technology: Taking advantage of the fact that some modules are more critical in generating optical response characteristics, while others are more critical in generating superatom intrinsic structure characteristics, the column sparsity constraints of the superatom intrinsic structure characteristics and optical response characteristics LoRAs corresponding to the superatom intrinsic structure characteristics sensitive and optical response characteristics sensitive modules are respectively relaxed, so that more blocks are introduced to perform finer block-level subject and superatom intrinsic structure characteristic distribution.

[0010] Preferably, in the step S4, the two-level prototype initialization includes: Degenerate level prototype: considering 0°, 10°, 20°, 30° incident angles as different "degeneration types", each angle corresponds to a prototype vector: ; In the formula, is the degenerate level prototype matrix, is the real number space, is the 4 incident angles, is the channel dimension; The angle description text is encoded and normalized using pre-trained word embedding; Factor level prototype: under each incident angle, considering multiple physical factors affecting phase and transmission efficiency, each angle corresponds to factor prototype: ; In the formula, is the factor level prototype tensor; Wherein, an additional set of common prototypes is used to capture the shared features across angles; Then, feature-prototype matching is performed, the input is the discrete matrix structure of the superatom The Sinkhorn-Knopp iteration is used to calculate the matching matrix: ; ; In the formula, is the sensitivity of the superatom region to the incident angle, is the optimal transmission matching matrix between the two matrices calculated by Sinkhorn-Knopp iteration, is the regulation ability of the local structure to a specific physical factor, is the number of space grids.

[0011] Preferably, in the step S5, the two-level prototype scene embedding is calculated, including: Degenerate level scene embedding: ; In the formula, is the matching weight vector of the th spatial position of the superatom and the incident angle prototype; Factor level scene embedding: ; In the formula, is the quantization of the a spatial position matching weight for all physical factor prototypes; image features , scene embeddings , enhanced features using bidirectional cross-attention mechanism: ; wherein, is the output enhanced features, is the degradation-level scene embedding provided by , is the factor-level scene embedding provided by , is the attention weight, is the weighted aggregation, is the temperature coefficient.

[0012] Preferably, in the step S6, the structural modalities: features obtained after feature extraction of the discrete matrix, are denoted as ; optical modalities: features obtained by an encoder from the phase and transmission efficiency curves, are denoted as ; Two-level prototypes are used to match the two modalities, obtaining the matching matrices of the two modalities and ; The alignment loss between the matching distribution of the two modalities at the degradation level and the factor level is calculated: ; wherein, is the sensitivity of the structural features to the incident angle, is the dependence of the optical response on the incident angle, is the association of the structural features with the microscopic physical factors, is the mapping of the optical response to the physical factors; The degradation-level scene embedding of the structural modalities and the degradation-level scene embedding of the optical modalities are taken as positive pairs, while the optical modalities embedding of other samples are taken as negative pairs, and the same processing is performed for the factor-level scene embedding; For the degradation-level scene embedding, let and denote the degradation-level scene embedding of the structural modalities and the optical modalities of a sample, respectively, then the InfoNCE loss is: ; wherein, is the exponential cosine similarity, which amplifies the difference between positive and negative samples, is the sample index within the batch, is the batch size, is the temperature parameter; Similarly, for factor-level scene embedding computation ; The total alignment loss is: ; The total loss function becomes: ; where, is the mean square error of the target phase distribution and the predicted phase of the optical modality, is the difference between the transmission efficiency of the optical modality and the target value, is the balance between optical performance and structural feasibility, is the manufacturability constraint of the structural modality, is the weight of the alignment loss.

[0013] Preferably, in the step S7, the features extracted by the encoder are subjected to angle and frequency double decomposition, and angle path progressive fusion and frequency path progressive fusion, including: Angle decomposition: For the input feature map , four angle features are obtained through four convolutional layers: ; where, is the convolutional layer; Frequency decomposition: Four convolutional layers are used, corresponding to low frequency, horizontal high frequency, vertical high frequency, and diagonal high frequency, to decompose the input feature map: ; where, is the low-pass filter convolutional layer; Angle path progressive fusion: Starting from 0°, gradually fuse to 30°: ; The fusion step uses a double attention mechanism, where, Spatial attention weight: ; where, is the Sigmoid activation function, which compresses the weight to the interval, is the large kernel convolutional fusion pooling result, generating a spatial sensitive weight map, is the average pooling along the channel dimension, is the maximum pooling along the channel dimension; Channel attention weight: ; where, and are both fully connected layer weight matrices, is a nonlinear activation, is a global average pooling, and are both bias terms; Combining spatial attention weights with channel attention weights to realize feature fusion : ; In the formula, is the current angle feature, is the next angle feature, is an element-wise multiplication; Frequency path fusion: Low-frequency features guide high-frequency features to enhance: ; ; ; In the formula, is the enhanced horizontal high-frequency feature, is a frequency band fusion module, which guides high-frequency features to enhance through low-frequency features, is a horizontal high-frequency feature, is a low-frequency feature, is an enhanced vertical high-frequency feature, is a vertical high-frequency feature, is an enhanced diagonal high-frequency feature, is a diagonal high-frequency feature; High-frequency feature fusion: ; In the formula, is a tensor splicing; Final frequency fusion, low-frequency features further interact with fused high-frequency features: .

[0014] Preferably, in the step S8, angle-specific feature enhancement: ; In the formula, is the exclusive feature of angle , is the fused global feature, is the attention weight of angle ; Double-branch prediction, including phase prediction branch and transmission efficiency prediction branch; wherein, Phase prediction branch: ; In the formula, is the phase prediction result, is compressed to , multiplied by After covering Phase range, is a multilayer perception machine; Transmission efficiency prediction branch: ; In the formula, is the transmission efficiency prediction result, is limited to Interval.

[0015] Compared with the prior art, the beneficial effects of the present application are: The present application decouples the super-atomic structure (optical response characteristics) and the optical response (super-atomic inherent structure characteristics) features by prompting decomposition, column separation and block grading technology, avoids cross interference, and lays a foundation for subsequent accurate modeling; The double-level prototype feature extraction module constructs the degradation level (corresponding to multiple angles) and the factor level (corresponding to physical factors) prototype, realizes feature-prototype matching combined with Sinkhorn-Knopp iteration, and efficiently captures the optical property correlation of super-atoms under multiple angles; Further, the modal matching alignment constrains the matching distribution of the structure mode and the optical mode with InfoNCE loss, ensures the consistency of the double-modal information, and improves the prediction accuracy; The angle-frequency decomposition fusion decoder first decomposes the features according to angle and frequency, and then gradually fuses through double paths (angle gradually fuses from 0° to 30°, and frequency guides high frequency with low frequency), realizes the deep integration of multi-angle and multi-frequency features, and adapts to the wide field of view detection demand; In summary, the present application effectively alleviates the problems of limited working bandwidth, difficult multi-angle response prediction, difficult high transmittance and high phase accuracy in traditional super-lens design, significantly expands the effective receiving field of view of the laser radar system, improves the light field reconstruction accuracy, and provides technical support for miniaturized, wide field of view laser radar application. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 is a flowchart of a multi-modal decomposition super-lens transmission efficiency prediction method based on the present application; Figure 2This is another flowchart of a method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 , 2 As shown, a method for predicting the transmission efficiency of a superlens based on multimodal decomposition includes: Step S1: Set the superatomic structure parameters, perform superatomic structure modeling, and construct diverse structure datasets through random generation, dense structure, sparse structure, and continuous block structure to ensure that the model has good generalization ability for superatoms with different geometric shapes.

[0019] In this embodiment, superatomic structure modeling includes: First, the different design logics of the center and edge of the superlens are clarified. The size of the center superlens is set to 500μm, which is consistent with the input aperture, while the edge superlens is specially adapted to the focusing requirements of oblique incident light. Secondly, the phase distribution and structural parameters of the superatoms were determined. The height, period, and discrete period were set to 700nm, 700nm, and 640nm, respectively. The superatom structure was modeled as a 32×32 binary discrete matrix, where "0" represents no structure and "1" represents the presence of structure. Finally, by taking advantage of the symmetry of the structure on the x-axis and y-axis, the 32×32 binary discrete matrix is ​​compressed to 16×16. The working principle of the superatom structure is as follows: When light shines perpendicularly onto the front surface, it passes through the input aperture and directly enters the center of the superlens, focusing onto the imaging plane. When light is incident at a certain angle, it is blocked by the aperture, allowing only the central beam to reach the edge of the superlens. The design of the edge superlens differs from that of the center; it is specifically configured to focus the tilted light back to the center of the image plane. By applying this basic concept, the input position on the superlens at different incident angles can be calculated, and corresponding features can be designed to effectively focus light from various angles near the center of the imaging plane. For the superlens at the center, its overall size is consistent with the input aperture, set to 500 μm.

[0020] Step S2, simulation is carried out by means of Lumerical FDTD software, transmission efficiency and phase curve of superatom under different incident angles are obtained, and a data set including multiple groups of samples is constructed.

[0021] In the embodiment, the incident angles include 0°, 10°, 20° and 30°, and the transmission efficiency and phase curve in the wavelength range of 1500-1600 nm; the data set including multiple groups of samples is constructed by diversifying the structure data set and the transmission efficiency and phase curve, in which 50% are random structures and 50% are directional generated structures, and the data set is divided into a training set, a verification set and a test set at a ratio of 8:1:1.

[0022] Step S3, the input is represented by two low-rank adaptive LoRAs respectively, and the input is decomposed into superatom intrinsic structure characteristics and optical response characteristics.

[0023] In the embodiment, the decomposition technology for decomposing the input into superatom intrinsic structure characteristics and optical response characteristics is composed of prompt decomposition technology, column separation technology and block hierarchical technology, wherein, Prompt decomposition technology: In the typical cross-attention layer of the diffusion U-net model, the weight is used. The prompt is embedded into the key and value : ; Three independent prompts are used to calculate three groups of keys and values: One group uses the base model and the combined prompt , and the other two groups respectively use the LoRA of the optical response characteristics and the LoRA of the superatom intrinsic structure characteristics and their corresponding prompts, and then the obtained feature maps are added: ; In the formula, is the embedding vector of the original prompt, is the embedding vector of the optical response characteristics, is the embedding vector of the intrinsic structure characteristics, and are the low-rank adaptive matrices for the two types of characteristics, Column separation technology: The concept of "column" is introduced to set the corresponding mask and for each LoRA, and the column mask controls the contribution of each element: Block-wise hierarchical technique: Taking advantage of the fact that certain modules are more critical in generating optical response characteristics while others are more critical in generating intrinsic superatom structural characteristics, the column sparsity constraints on intrinsic superatom structural characteristics and optical response characteristics LoRs of modules sensitive to intrinsic superatom structural characteristics and optical response characteristics are relaxed respectively, thus, by introducing more blocks, finer block-wise assignment of modules to intrinsic superatom structural characteristics is performed, and the method is extended to further improve accuracy.

[0024] Step S4, two-level prototype initialization and feature-prototype matching.

[0025] In the embodiment, two-level prototype initialization includes: Degenerate level prototype: 0°, 10°, 20°, 30° incident angles are regarded as different “degeneration types”, and each angle corresponds to a prototype vector: ; In the formula, is the degenerate level prototype matrix, is the real number space, is the 4 incident angles, is the channel dimension; The angle description text (such as “0° incident angle”) is encoded and normalized using pre-trained word embeddings; Factor level prototype: under each incident angle, multiple physical factors (such as wavelength, structure size, material properties, etc.) that affect phase and transmission efficiency are considered, and each angle corresponds to factor prototypes: ; In the formula, is the factor level prototype tensor; where an additional set of common prototypes is used to capture shared features across angles; Then, feature-prototype matching is performed, and the input is the discrete matrix structure of the superatom The Sinkhorn-Knopp iteration is used to calculate the matching matrix: ; ; In the formula, is the sensitivity of the superatom region to the incident angle, is the optimal transport matching matrix between two matrices calculated by Sinkhorn-Knopp iteration, is the regulation ability of the local structure to a specific physical factor, is the number of space grids.

[0026] Step S5: Calculate the two-level prototype scene embedding and enhance the features using a bidirectional cross-attention mechanism.

[0027] In this embodiment, calculating the two-level prototype scene embedding includes: Degradation-level scene embedding: ; In the formula, For the first superatomic atom Spatial location and Matching weight vectors for each incident angle prototype; Factor-level scene embedding: ; In the formula, To quantify the first The matching weight of each spatial location to all physical factor prototypes; Image features Scene embedding Enhance features using a bidirectional cross-attention mechanism: ; In the formula, To output the enhanced features, Embedded for degradation-level scenes, by supply, For factor-level scene embedding, by supply, For attention weights, For weighted aggregation, This is the temperature coefficient.

[0028] Step S6: Perform mode matching alignment on the structural and optical modes of each metaatom.

[0029] In the embodiment, structural mode: the features obtained after feature extraction from the discrete matrix, denoted as ; Optical mode: The characteristic obtained by passing the phase and transmission efficiency curves through an encoder, denoted as... ; The two modes are matched using a bi-level prototype to obtain a matching matrix for the two modes. and ; Calculate the alignment loss between the matching distributions of the two modes at the degradation level and the factor level: ; In the formula, Assigning sensitivity of structural features to incident angle. This refers to the dependence of the optical response on the incident angle. to associate the structural features with the microscopic physical factors, to map the optical response with the physical factors; embed the structural modal degradation level scenario and the optical modal degradation level scenario as positive pairs, and embed the optical modal scenario of other samples as negative pairs, and process the factor level scenario embedding in the same way; For the degradation level scenario embedding, let and denote the degradation level scenario embedding of the structural modal and the optical modal of a sample respectively, then the InfoNCE loss is: ; In the formula, is the exponential cosine similarity, which amplifies the difference between positive and negative samples, is the sample index within the batch, is the batch size, is the temperature parameter; Similarly, the factor level scenario embedding is calculated ; The total alignment loss is: ; The total loss function becomes: ; In the formula, is the mean square error of the target phase distribution and the predicted phase of the optical modal, is the difference between the transmission efficiency of the optical modal and the target value, is the balance between optical performance and structural feasibility, is the manufacturability constraint of the structural modal, is the weight of the alignment loss.

[0030] Step S7, construct a decomposition-fusion decoder, perform angle and frequency double decomposition on the features extracted by the encoder, and perform angle path progressive fusion and frequency path progressive fusion.

[0031] In the embodiment, the angle and frequency double decomposition on the features extracted by the encoder, and the angle path progressive fusion and the frequency path progressive fusion, include: Angle decomposition: For the input feature map , four angle features are obtained through four convolution layers: ; In the formula, is a convolution layer; Frequency decomposition: Use four convolution layers, respectively corresponding to low frequency, horizontal high frequency, vertical high frequency and diagonal high frequency, to decompose the input feature map: ; wherein, is a low-pass filter convolution layer; Angle path progressive fusion: Start from 0°, gradually fuse to 30°: ; The fusion step uses a double attention mechanism, wherein, Spatial attention weight: ; wherein, is a Sigmoid activation function, compressing the weight to interval, is a large kernel convolution fusion pooling result, generating a spatial sensitive weight map, is an average pooling along the channel dimension, is a maximum pooling along the channel dimension; Channel attention weight: ; wherein, and are full connection layer weight matrices, is a nonlinear activation, is a global average pooling, and are bias terms; Combine spatial attention weight and channel attention weight to realize feature fusion : ; wherein, is the current angle feature, is the next angle feature, is an element-wise multiplication; Frequency path fusion: Low-frequency features guide high-frequency features to enhance: ; ; ; wherein, is the enhanced horizontal high-frequency feature, is a frequency band fusion module, which guides high-frequency features to enhance through low-frequency features, is a horizontal high-frequency feature, is a low-frequency feature, is an enhanced vertical high-frequency feature, is a vertical high-frequency feature, To enhance the diagonal high-frequency features, It is a diagonal high-frequency feature; High-frequency feature fusion: ; In the formula, For tensor splicing; The final frequency fusion yields low-frequency characteristics. Further interaction with fused high-frequency features: .

[0032] Step S8: Finally, a multi-angle prediction head is designed. The non-shared parameter prediction head first enhances the angle features, and then predicts the phase and transmission efficiency results in a dual-branch manner, providing a dedicated prediction for each incident angle, thereby achieving accurate prediction of key optical parameters of the superatomic spectrum under multiple angles.

[0033] In the embodiment, angle-specific features are enhanced: ; In the formula, For angle Its unique characteristics For the fused global features, For angle Attention weights; Two-branch prediction includes a phase prediction branch and a transmission efficiency prediction branch; in, Phase prediction branch: ; In the formula, For phase prediction results, To compress the predicted values ​​to multiplied by Back Cover Phase range, It is a multilayer perceptron; Transmission efficiency prediction branch: ; In the formula, For transmission efficiency prediction results, To limit efficiency to Interval.

[0034] Finally, it should be noted that the above is only the preferred embodiment of the present application, and is not intended to limit the present application, although the foregoing embodiments of the present application are described in detail, for those skilled in the art, it still can be modified to the technical solution recorded in the foregoing embodiments, or equivalent replacement of some technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the transmission efficiency of a superlens based on multimodal decomposition, characterized in that, include: Step S1: Set the superatomic structure parameters, perform superatomic structure modeling, and construct diverse structure datasets through random generation, dense structure, sparse structure, and continuous block structure. Step S2: Use Lumerical FDTD software to perform simulations, obtain the transmission efficiency and phase curves of superatoms at different incident angles, and construct a dataset including multiple sets of samples. Step S3: The input is represented by two low-rank adapted LoRAs, and its input is decomposed into the intrinsic structural characteristics of the superatoms and the optical response characteristics. Step S4: Two-level prototype initialization and feature-prototype matching; Step S5: Calculate the two-level prototype scene embedding and enhance the features using a bidirectional cross-attention mechanism; Step S6: Perform mode matching alignment on the structural and optical modes of each metaatom; Step S7: Construct a decomposition-fusion decoder to perform dual decomposition of the features extracted by the encoder in terms of angle and frequency, as well as progressive fusion of the angle path and the frequency path. Step S8: Finally, a multi-angle prediction head is designed. The non-shared parameter prediction head first enhances the angle features, then predicts the phase and transmission efficiency in a dual-branch manner, providing a dedicated prediction for each incident angle.

2. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 1, characterized in that, In step S1, the superatomic structure modeling includes: First, the different design logics of the center and edge of the superlens are clarified. The size of the center superlens is set to 500μm, which is consistent with the input aperture, while the edge superlens is specially adapted to the focusing requirements of oblique incident light. Secondly, the phase distribution and structural parameters of the superatoms were determined. The height, period, and discrete period were set to 700nm, 700nm, and 640nm, respectively. The superatom structure was modeled as a 32×32 binary discrete matrix, where "0" represents no structure and "1" represents the presence of structure. Finally, by utilizing the symmetry of the structure on the x and y axes, the 32×32 binary discrete matrix is ​​compressed to 16×16.

3. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 2, characterized in that, In step S2, the incident angles include 0°, 10°, 20°, and 30°, and the transmission efficiency and phase curves are in the wavelength range of 1500–1600 nm. A dataset including multiple sets of samples is constructed by using a diverse set of structural data and the transmission efficiency and phase curves. In the dataset, 50% are random structures and 50% are directional generated structures.

4. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 3, characterized in that, In step S3, the decomposition technique for the input into superatomic intrinsic structural characteristics and optical response characteristics consists of cue decomposition technique, column separation technique, and block hierarchical technique, wherein... Tip: Decomposition technique In the typical cross-interest layer of the diffusion U-net model, weights are used. Embed the prompt Keywords mapped to the converter Sum : ； Calculate three sets of keywords and values ​​using three independent hints: A set using the base model Combination tips The other two groups use optical response characteristics respectively. LoRA and the intrinsic structural properties of superatoms The LoRA and its corresponding hints are then used to sum the resulting feature maps: ； In the formula, The embedding vector of the original prompt. This is the embedding vector of the optical response characteristics. An embedding vector representing inherent structural characteristics. and For low-rank fitting matrices for two types of properties, Column separation technology: The concept of "columns" is introduced, and a corresponding mask is set for each LoRA. and The column mask controls the contribution of each element: ； Block hierarchical technology: Taking advantage of the fact that some modules are more critical in generating optical response characteristics while others are more critical in generating superatomic intrinsic structure characteristics, the column sparsity constraints of superatomic intrinsic structure characteristics and optical response characteristics LoRAs corresponding to modules sensitive to superatomic intrinsic structure characteristics and optical response characteristics are relaxed respectively. Therefore, by introducing more blocks, a more refined block-level allocation of the main body and superatomic intrinsic structure characteristics can be achieved.

5. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 4, characterized in that, In step S4, the two-level prototype initialization includes: Degradation level prototype: 0°, 10°, 20°, and 30° incident angles are considered different "degradation types," with each angle corresponding to a prototype vector. ； In the formula, For the degenerate level prototype matrix, For the real number space, There are 4 incident angles. For channel dimensions; The angle description text is encoded and normalized using pre-trained word embeddings; Factor-level prototype: For each incident angle, multiple physical factors affecting phase and transmission efficiency are considered, with each angle corresponding to... One factor prototype: ； In the formula, It is a factorial prototype tensor; An additional set of common prototypes is used to capture shared features across angles; Then feature-prototype matching is performed, with the input being the discrete matrix structure of the superatoms. The matching matrix is ​​calculated using the Sinkhorn-Knopp iterative method: ； ； In the formula, The sensitivity of the metaatomic region to the incident angle. To compute the optimal transmission matching matrix between two matrices using the Sinkhorn-Knopp iteration, This refers to the ability of a local structure to regulate specific physical factors. This represents the number of spatial grid cells.

6. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 5, characterized in that, In step S5, calculating the two-level prototype scene embedding includes: Degradation-level scene embedding: ； In the formula, For the first superatomic atom Spatial location and Matching weight vectors for each incident angle prototype; Factor-level scene embedding: ； In the formula, To quantify the first The matching weight of each spatial location to all physical factor prototypes; Image features Scene embedding Enhance features using a bidirectional cross-attention mechanism: ； In the formula, To output the enhanced features, Embedded for degradation-level scenes, by supply, For factor-level scene embedding, by supply, For attention weights, For weighted aggregation, This is the temperature coefficient.

7. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 6, characterized in that, In step S6, structural mode: the feature obtained after feature extraction of the discrete matrix, denoted as... ; Optical mode: The characteristic obtained by passing the phase and transmission efficiency curves through an encoder, denoted as... ; The two modes are matched using a bi-level prototype to obtain a matching matrix for the two modes. and ; Calculate the alignment loss between the matching distributions of the two modes at the degradation level and the factor level: ； In the formula, Assigning sensitivity of structural features to incident angle. This refers to the dependence of the optical response on the incident angle. To establish the relationship between structural features and microscopic physical factors, This represents the mapping between optical response and physical factors. Degenerate-level scene embeddings of structural modes and optical modes are treated as positive pairs, while optical mode embeddings of other samples are treated as negative pairs. Factor-level scene embeddings are treated in the same way. For embedding in degradation-level scenarios, set and Let the degraded-level scene embeddings of the structural and optical modes of a sample be represented respectively. Then the InfoNCE loss is: ； In the formula, To achieve exponential cosine similarity, the differences between positive and negative samples are amplified. For sample index within a batch, For batch size, For temperature parameters; Similarly, factor-level scene embedding computation ; The total alignment loss is: ； The total loss function becomes: ； In the formula, The mean square error between the target phase distribution and the predicted phase of the optical mode is given. The difference between the transmission efficiency of the optical mode and the target value, To balance optical performance and structural feasibility, For the manufacturability constraints of structural modes, Weights for alignment loss.

8. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 7, characterized in that, In step S7, the features extracted by the encoder are subjected to dual decomposition of angle and frequency, as well as progressive fusion of angle path and progressive fusion of frequency path, including: Angle breakdown: For the input feature map Four angular features are obtained through four convolutional layers: ； In the formula, It is a convolutional layer; Frequency decomposition: Four convolutional layers are used to decompose the input feature map, corresponding to low frequency, horizontal high frequency, vertical high frequency, and diagonal high frequency, respectively: ； In the formula, It is a low-pass filtered convolutional layer; Gradual fusion of angles and paths: Starting from 0°, gradually blending up to 30°: ； The fusion step uses a dual attention mechanism, in which, Spatial attention weights: ； In the formula, The Sigmoid activation function compresses the weights to... interval, A spatially sensitive weight map is generated from the large kernel convolution fusion pooling results. For average pooling along the channel dimension, Max pooling along the channel dimension; Channel attention weights: ； In the formula, and All are weight matrices of fully connected layers. It is a non-linear activation. For global average pooling, and All are bias terms; Feature fusion is achieved by combining spatial attention weights and channel attention weights. : ； In the formula, For the current angle features, For the next angular feature, This is element-wise multiplication; Frequency path fusion: Low-frequency features guide high-frequency feature enhancement: ； ； ； In the formula, To enhance the horizontal high-frequency characteristics, As a frequency band fusion module, it enhances high-frequency features by guiding low-frequency features. It is a horizontal high-frequency feature. It is a low-frequency characteristic. For the enhanced vertical high-frequency features, It is a vertical high-frequency feature. To enhance the diagonal high-frequency features, It is a diagonal high-frequency feature; High-frequency feature fusion: ； In the formula, For tensor splicing; The final frequency fusion yields low-frequency characteristics. Further interaction with fused high-frequency features: 。 9. The method for predicting the transmission efficiency of a superlens based on multimodal decomposition according to claim 8, characterized in that, In step S8, angle-specific features are enhanced: ； In the formula, For angle Its unique characteristics For the fused global features, For angle Attention weights; Two-branch prediction includes a phase prediction branch and a transmission efficiency prediction branch; in, Phase prediction branch: ； In the formula, For phase prediction results, To compress the predicted values ​​to multiplied by Back Cover Phase range, It is a multilayer perceptron; Transmission efficiency prediction branch: ； In the formula, For transmission efficiency prediction results, To limit efficiency to Interval.

Images