A knowledge-guided multi-agent collaborative autonomous research framework and system for electromagnetic metasurfaces

By using a multi-agent collaborative framework, the problems of knowledge dispersion and fragmented automation processes in electromagnetic metasurface design are solved, realizing automated design from user requirements to executable simulation models, and ensuring the physical rationality and traceability of the design scheme.

CN123069868APending Publication Date: 2026-11-10SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611199343.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-09
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In the current electromagnetic metasurface design process, design knowledge is scattered and inconsistent, relying on human experience. The output of general large language models lacks physical reliability and lacks an end-to-end automated process from user requirements to executable simulation models.

Method used

A knowledge-guided multi-agent collaborative framework is adopted. By generating structured datasets and vector databases through multimodal literature analysis, combined with semantic and numerical retrieval, a physically reasonable design scheme is generated, and full-wave electromagnetic simulation is achieved by reproducing the agent.

Benefits of technology

It achieves end-to-end automation from literature knowledge to executable simulation models, ensuring the physical rationality and traceability of design schemes, reducing the need for manual intervention, and quickly identifying research hotspots and trends.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069868A_ABST
    Figure CN123069868A_ABST
Patent Text Reader

Abstract

This invention discloses a knowledge-guided multi-agent collaborative framework and system for autonomous research on electromagnetic metasurfaces. The method includes: constructing an offline knowledge base for electromagnetic metasurfaces; performing multimodal analysis on literature to generate structured datasets and vector databases; using a data analysis agent to perform autonomous literature mining and domain trend analysis on the structured datasets; upon receiving user design requirements, a retrieval and reasoning agent performs hybrid retrieval filtering to locate target literature and generates design schemes based on a retrieval-enhanced generation mechanism; and a reproduction agent converts the structural parameters in the design schemes into executable scripts for an electromagnetic solver, performing full-wave simulation. This invention, through a multi-agent collaborative mechanism, unifies literature knowledge mining, design requirement reasoning, and autonomous structural reproduction into a closed-loop automated research process, significantly reducing manual dependence in electromagnetic metasurface research and improving the conversion efficiency and physical reliability from literature knowledge to executable simulation models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic metasurface design and artificial intelligence-assisted scientific research automation technology, and in particular relates to a knowledge-guided multi-agent collaboration-based autonomous research framework method and system for electromagnetic metasurfaces. Background Technology

[0002] Electromagnetic metasurfaces are designable artificial electromagnetic materials composed of subwavelength-scale unit structures, enabling precise control of the amplitude, phase, and polarization of electromagnetic waves. By systematically designing the subwavelength units and their spatial arrangement, metasurfaces can achieve multifunctional wavefront manipulation such as beam deflection, polarization modulation, holographic imaging, stealth, and planar lenses, providing compact and efficient solutions for system-level applications in communication, sensing, imaging, and integrated photonics.

[0003] To meet specific wavefront manipulation requirements, metasurface design typically begins with finding suitable structural prototypes. This process relies on domain experts searching and screening a vast amount of literature, assessing applicability based on scattered structural parameters, electromagnetic responses, and design theories, and performing preliminary modeling and iterative full-wave simulation optimization under incomplete information. The entire process is time-consuming, heavily dependent on expert experience, and has limited reproducibility, hindering the rapid development and experimental verification of novel functional metasurfaces.

[0004] In recent years, large language models, with their knowledge synthesis and stepwise reasoning capabilities, have demonstrated the potential to drive the automation of scientific research in fields such as chemical synthesis, biology, and materials discovery. However, existing technologies face the following core problems in the field of electromagnetic metasurfaces: First, metasurface design knowledge is scattered throughout the text, structural diagrams, figures, tables, and supplementary materials of scientific literature. The information sources are discrete and the expression forms are inconsistent, making it difficult to reliably recover complete design information by relying solely on text extraction. Second, metasurface design requirements are reasoning tasks with strict physical constraints. Directly using general large language models to generate design schemes can easily produce a "design illusion" that appears coherent but lacks physical rationality. Third, existing automation methods typically handle literature analysis, parameter extraction, geometric modeling, and simulation verification in a fragmented manner, lacking a mechanism to unify knowledge retrieval, design reasoning, and structural reproduction into a closed-loop process, making it difficult to achieve end-to-end automation from user design requirements directly to an executable simulation model. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge-guided multi-agent collaborative autonomous research framework method and system for electromagnetic metasurfaces, in order to solve the technical problems in existing electromagnetic metasurface research, such as reliance on manual acquisition of literature knowledge, lack of physical reliability of general large language model output, and the disconnect between knowledge retrieval and structure reproduction processes.

[0006] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows: A knowledge-guided multi-agent collaboration-based autonomous research framework for electromagnetic metasurfaces includes the following steps: Literature in the field of electromagnetic metasurfaces is acquired, multimodal analysis is performed on the literature, textual and visual information is extracted and fused into a unified corpus, and a structured dataset and vector database are generated based on the unified corpus to build an offline knowledge base. By using a data analysis intelligent agent to perform autonomous literature mining and domain trend analysis on structured datasets, a visual analysis result reflecting the research status of electromagnetic metasurfaces is generated. After receiving the user's design requirements, the retrieval reasoning agent performs a hybrid retrieval. First, semantic retrieval is used to retrieve the context of candidate documents from the vector database, and then numerical retrieval is used to filter the context of candidate documents to locate the target documents that match the user's design requirements. Finally, a design scheme containing structural parameters is generated based on the retrieval enhancement generation mechanism. By reproducing the intelligent agent, the structural parameters in the design scheme are converted into an executable script for an electromagnetic solver, and full-wave electromagnetic simulation is performed to complete the autonomous research process from literature knowledge to an executable simulation model in a closed loop.

[0007] Furthermore, a document parsing tool is used to convert the PDF files of the documents into a machine-readable format; a visual language model is used to parse the figures and tables in the documents and extract the physical parameters and topological information embedded in the figures and tables; the visual information extracted by the visual language model is merged with the main text of the documents to form a unified corpus containing graphic and textual information.

[0008] Furthermore, a structured dataset and a vector database are generated, including: extracting structural parameters and constructing a structured dataset based on a unified corpus using a large language model; converting the unified corpus into vector embeddings and storing them in a vector database, so that physical concepts are mapped to a consistent embedding space.

[0009] Furthermore, hybrid retrieval includes both semantic and numerical retrieval. Semantic retrieval matches relevant document context from a vector database based on the semantic description in the user's design requirements. Numerical retrieval precisely filters the candidate document context returned by semantic retrieval based on the numerical conditions in the user's design requirements, in order to locate the single target document that best matches the requirements.

[0010] Furthermore, the retrieval enhancement generation mechanism uses the corpus content of the target document as contextual enhancement information input into the large language model. The large language model then generates a design scheme containing structural parameters based on the contextual enhancement information, thereby eliminating the physical unreliability of direct generation.

[0011] Furthermore, the data analysis agent performs autonomous document mining and domain trend analysis, including: autonomously extracting metadata features of documents from structured datasets, including metasurface type, target frequency, electromagnetic function, and application system; automatically synthesizing data visualization code to generate visualization charts that reflect the multidimensional mapping relationship between metadata features.

[0012] Furthermore, the replicating agent converts structural parameters into executable scripts for the electromagnetic solver, including converting the design scheme into parametric modeling instructions, material assignment instructions, boundary setting instructions, excitation setting instructions, and solver control instructions supported by the electromagnetic solver; after performing full-wave electromagnetic simulation, the replicating agent compares and verifies the simulation response results with the reference results in the target literature.

[0013] This invention also discloses a knowledge-guided multi-agent collaborative autonomous research framework system for electromagnetic metasurfaces, comprising: The knowledge base construction module is used to acquire multiple documents in the field of electromagnetic metasurfaces, perform multimodal analysis on the documents to extract textual and visual information and fuse them into a unified corpus, and generate a structured dataset and vector database based on the unified corpus to build an offline knowledge base. The data analysis intelligent agent module is used to perform autonomous literature mining and domain trend analysis on structured datasets, and generate visual analysis results that reflect the research status of electromagnetic metasurfaces. The retrieval and reasoning agent module is used to perform hybrid retrieval after receiving the user's design requirements. First, semantic retrieval is used to retrieve the context of candidate documents from the vector database. Then, numerical retrieval is used to filter the context of candidate documents to locate the target document. Finally, a design scheme containing detailed structural parameters is generated based on the retrieval enhancement generation mechanism. The intelligent agent module is used to convert the structural parameters in the design scheme into an executable script for an electromagnetic solver, and to perform full-wave electromagnetic simulation. The data analysis agent module, the retrieval and reasoning agent module, and the reproduction agent module work together in the workflow sequence of offline knowledge base construction, data analysis, retrieval and reasoning, and structure reproduction.

[0014] Furthermore, the knowledge base construction module includes: a document parsing unit, used to convert PDF files of documents into machine-readable formats; a visual language modeling unit, used to parse the charts in the documents to extract the physical parameters and topological information embedded in the charts; a corpus fusion unit, used to merge the visual information extracted by the visual language modeling unit with the main text of the documents into a unified corpus; a data structuring unit, used to extract structural parameters based on the unified corpus to construct a structured dataset; and a vectorization unit, used to convert the unified corpus into vector embeddings and store them in a vector database.

[0015] The present invention provides a knowledge-guided multi-agent collaborative autonomous research framework method and system for electromagnetic metasurfaces, which has the following advantages: 1. By constructing an offline knowledge base and performing multimodal text and image analysis on the documents, the difficulty of designing knowledge scattered in text and charts can be overcome, providing a complete information foundation for subsequent retrieval and reasoning; 2. By introducing a hybrid retrieval mechanism that combines semantic retrieval with numerical retrieval, it is possible to accurately match the most relevant documents corresponding to the user's design requirements from a large-scale document database, thus avoiding the semantic drift problem of pure semantic retrieval. 3. By using a retrieval-enhanced generation mechanism, the output of the large language model is anchored to the content of peer-reviewed scientific literature, effectively eliminating the "design illusion" of the general large language model in metasurface design tasks, and ensuring that the generated design scheme has traceability and physical rationality; 4. By linking and collaborating the retrieval and reasoning agent with the reproduction agent, end-to-end automation from literature knowledge to an executable electromagnetic simulation model is achieved, significantly reducing the need for manual intervention. 5. Through the large-scale literature statistics and trend visualization capabilities provided by the data analysis intelligent agent, researchers can quickly identify the current research hotspots and evolution of a specific field; 6. Through a closed-loop architecture of multi-agent collaboration, literature knowledge mining, design requirement reasoning, and autonomous structural reproduction are unified into a coherent automated research process, opening up a new paradigm of autonomous scientific research based on AI agents in the field of electromagnetic metasurfaces. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall method and multi-agent cooperation framework of the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the working principle of the data analysis intelligent agent of the present invention.

[0018] Figure 3 This is a schematic diagram illustrating a Sankey diagram of the data analysis agent of this invention.

[0019] Figure 4 This is a schematic diagram of the time-application heatmap of the data analysis agent of the present invention.

[0020] Figure 5 This is a schematic diagram of the frequency-application heatmap of the data analysis intelligent agent of the present invention.

[0021] Figure 6 This is a schematic diagram illustrating the working principle of the retrieval and reasoning intelligent agent of this invention.

[0022] Figure 7 This diagram illustrates the performance comparison between the retrieval and reasoning agent of this invention and other large models.

[0023] Figure 8 This is a schematic diagram illustrating the working principle of the intelligent agent in this invention.

[0024] Figure 9 This is a schematic diagram illustrating the full-wave electromagnetic simulation results of the target literature for reproducing the intelligent agent case in this invention.

[0025] Figure 10 This is a schematic diagram of the full-wave electromagnetic simulation results of the design structure for reproducing the intelligent agent case of this invention.

[0026] Figure 11 This is a schematic diagram of the target document scattering diagram for the intelligent agent case of this invention.

[0027] Figure 12 This is a schematic diagram of the scattering pattern of the design structure for reproducing the intelligent agent case of this invention.

[0028] Figure 13 This is a schematic diagram of the experimental testing environment for this invention.

[0029] Figure 14 This is a schematic diagram showing the comparison between the simulated and measured radar cross-section patterns of this invention. Detailed Implementation

[0030] To better understand the purpose, structure, and function of this invention, the following detailed description, in conjunction with the accompanying drawings, provides a framework method and system for autonomous research of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration.

[0031] A knowledge-guided multi-agent collaboration-based autonomous research framework for electromagnetic metasurfaces includes the following steps: Literature in the field of electromagnetic metasurfaces is acquired, multimodal analysis is performed on the literature, textual and visual information is extracted and fused into a unified corpus, and a structured dataset and vector database are generated based on the unified corpus to build an offline knowledge base. By using a data analysis intelligent agent to perform autonomous literature mining and domain trend analysis on structured datasets, a visual analysis result reflecting the research status of electromagnetic metasurfaces is generated. After receiving the user's design requirements, the retrieval reasoning agent performs a hybrid retrieval. First, semantic retrieval is used to retrieve the context of candidate documents from the vector database, and then numerical retrieval is used to filter the context of candidate documents to locate the target documents that match the user's design requirements. Based on the corpus content of the target documents, a design scheme containing detailed structural parameters is generated through a retrieval-enhanced generation mechanism. By reproducing the intelligent agent, the structural parameters in the design scheme are converted into an executable script for an electromagnetic solver, and full-wave electromagnetic simulation is performed to complete the autonomous research process from literature knowledge to an executable simulation model in a closed loop.

[0032] The document parsing tool was used to convert the PDF files of the documents into a machine-readable format; a visual language model was used to parse the figures and tables in the documents and extract the physical parameters and topological information embedded in the figures and tables; the visual information extracted by the visual language model was merged with the main text of the documents to form a unified corpus containing text and image information.

[0033] Generate structured datasets and vector databases, including: extracting structural parameters and constructing structured datasets based on a unified corpus using a large language model; converting the unified corpus into vector embeddings and storing them in a vector database, so that physical concepts are mapped to a consistent embedding space.

[0034] In hybrid retrieval, semantic retrieval matches relevant document context from a vector database based on the semantic description in the user's design requirements; numerical retrieval precisely filters the candidate document context returned by semantic retrieval based on the numerical conditions in the user's design requirements, in order to locate the single target document that best matches the requirements.

[0035] The retrieval enhancement generation mechanism uses the corpus content of the target document as contextual enhancement information input into a large language model. The large language model then generates a design scheme containing structural parameters based on the contextual enhancement information, thereby eliminating the physical unreliability of direct generation.

[0036] The data analysis agent performs autonomous document mining and domain trend analysis, including: autonomously extracting metadata features from structured datasets, including metasurface type, target frequency, electromagnetic function, and application system; automatically synthesizing data visualization code to generate visualization charts that reflect the multidimensional mapping relationship between metadata features.

[0037] The replicating agent converts structural parameters into executable scripts for the electromagnetic solver, including converting the design scheme into parametric modeling instructions, material assignment instructions, boundary setting instructions, excitation setting instructions, and solver control instructions supported by the electromagnetic solver. After performing full-wave electromagnetic simulation, the replicating agent compares and verifies the simulation response results with the reference results in the target literature.

[0038] This invention also discloses a knowledge-guided multi-agent collaborative autonomous research framework system for electromagnetic metasurfaces, comprising: The knowledge base construction module is used to acquire multiple documents in the field of electromagnetic metasurfaces, perform multimodal analysis on the documents to extract textual and visual information and fuse them into a unified corpus, and generate a structured dataset and vector database based on the unified corpus to build an offline knowledge base. The data analysis intelligent agent module is used to perform autonomous literature mining and domain trend analysis on structured datasets, and generate visual analysis results that reflect the research status of electromagnetic metasurfaces. The retrieval and reasoning agent module is used to perform hybrid retrieval after receiving the user's design requirements. First, semantic retrieval is used to retrieve the context of candidate documents from the vector database. Then, numerical retrieval is used to filter the context of candidate documents to locate the target document. Finally, a design scheme containing detailed structural parameters is generated based on the retrieval enhancement generation mechanism. The intelligent agent module is used to convert the structural parameters in the design scheme into an executable script for an electromagnetic solver, and to perform full-wave electromagnetic simulation. The data analysis agent module, the retrieval and reasoning agent module, and the reproduction agent module work together in the workflow sequence of offline knowledge base construction, data analysis, retrieval and reasoning, and structure reproduction.

[0039] The knowledge base construction module includes: a document parsing unit, used to convert PDF files of documents into machine-readable format; a visual language modeling unit, used to parse the charts in the documents to extract the physical parameters and topological information embedded in the charts; a corpus fusion unit, used to merge the visual information extracted by the visual language modeling unit with the main text of the documents into a unified corpus; a data structuring unit, used to extract structural parameters based on the unified corpus to construct a structured dataset; and a vectorization unit, used to convert the unified corpus into vector embeddings and store them in a vector database.

[0040] like Figure 1 As shown, the overall framework of this invention includes an offline knowledge base construction phase and an online multi-agent collaboration phase. In the offline knowledge base construction phase, the knowledge base construction module performs multimodal analysis on a large-scale literature database in the field of electromagnetic metasurfaces, generating a structured dataset and a vector database. In the online multi-agent collaboration phase, the data analysis agent, the retrieval and reasoning agent, and the reproduction agent collaborate according to the workflow sequence, respectively performing literature mining and trend analysis, hybrid retrieval and design scheme reasoning, and structural reproduction and simulation verification, forming a closed-loop automated process from domain knowledge acquisition to design requirement response and structural verification.

[0041] In one specific implementation, the offline knowledge base construction phase receives a corpus containing approximately 2000 highly cited peer-reviewed papers in the field of electromagnetic metasurfaces. First, an open-source document parsing tool is used to convert the original PDF documents into a machine-readable format, removing typesetting artifacts. Then, addressing the issue that key physical parameters in the electromagnetic metasurface field—such as the metasurface's geometric dimensions, periodicity, and scattering parameters—are often embedded in complex scientific diagrams, the Qianwen open-source visual language model is deployed to parse the diagrams, ensuring that the physical parameters in the diagrams are indexed along with the main text. Visual information is directly incorporated into the unified corpus to compensate for the insufficient information capacity of pure text. The DeepSeek large language model is used to extract structural parameters from the unified corpus and construct a structured dataset; simultaneously, the unified corpus is converted into vector embeddings and stored in a vector database, ensuring that physical concepts are mapped to a consistent embedding space. The structured dataset and the vector database together form the information foundation for subsequent hybrid retrieval and autonomous analysis.

[0042] like Figure 2 As shown, the data analysis agent performs autonomous literature mining and domain trend analysis on a structured dataset. The agent autonomously executes the entire workflow from initial semantic understanding and structured data extraction to the synthesis of executable data visualization code.

[0043] In one specific implementation, the data analysis agent autonomously extracts metadata features from a structured dataset, including metasurface type (e.g., reflective, transmissive, programmable), target frequency (e.g., IEEE standard microwave band), electromagnetic functions (e.g., beam deflection, wavefront shaping, polarization modulation, stealth / RCS reduction, holography, focusing), and corresponding application systems (e.g., wireless communication, smart reflective surfaces). Subsequently, the agent automatically synthesizes data visualization code to generate visual charts reflecting the multidimensional mapping relationships between metadata features.

[0044] Visual charts include at least one of the following: Figure 3 The Sankey diagram shown is used to map the hierarchical relationships between metasurface types, target frequencies, electromagnetic functions, and their corresponding application systems, quantitatively revealing the progressive relationships between different physical mechanisms and engineering applications; such as Figure 4 The time-application distribution heatmap shown illustrates the annual publication trends of various functional directions in the field of metasurfaces from 2015 to 2025, intuitively quantifying the changing trajectory of various research trends. For example... Figure 5 The frequency-application distribution heatmap shown illustrates the density of cross-research between different IEEE standard microwave frequency bands and application fields. These visualizations allow researchers to quickly identify the current research status and cutting-edge trends in specific fields without manually reading and analyzing large amounts of literature.

[0045] like Figure 6As shown, the retrieval and reasoning agent receives user design requirements and generates a design scheme containing structural parameters through a hybrid retrieval and retrieval enhancement generation mechanism.

[0046] In one specific implementation, the hybrid retrieval first employs semantic retrieval, searching for relevant literature context in a vector database based on semantic descriptions in the user's design requirements (e.g., "reflective coded metasurface," "X-band beam deflection," etc.). After semantic retrieval and positioning, precise numerical retrieval is then performed on the candidate literature context using numerical constraints in the user's design requirements (e.g., target frequency range, unit period range, etc.), matching them one by one to locate the single target literature that best matches the user's requirements. By combining semantic retrieval and numerical retrieval in a hierarchical manner, the semantic drift problem that may occur when using semantic retrieval alone is effectively avoided, and the limitation of pure numerical retrieval in capturing conceptual relevance is overcome.

[0047] After locating the target literature, the retrieval and reasoning agent inputs the corpus content of the target literature as contextual enhancement information into the large language model. The large language model then generates a design scheme based on this contextual enhancement information, incorporating elements such as metasurface geometric parameters, material parameters, operating frequency, boundary conditions, and excitation conditions. Unlike general large language models that directly generate design schemes, the retrieval-enhanced generation mechanism employed in this invention strictly anchors the model output to peer-reviewed scientific literature content, fundamentally suppressing the "design illusion"—i.e., outputs that appear reasonable on the surface but lack physical basis—that arises from the model's free interpretation. Every generated parameter can be traced back to its source literature, ensuring the factual consistency and physical rationality of the design scheme.

[0048] like Figure 7 As shown, to verify the effectiveness of the retrieval and inference agent, in one specific implementation, a test benchmark covering 236 cross-scenario metasurface design requirements was constructed. An evaluation process using a large language model as the evaluator was employed, comparing the retrieval and inference agent with four baseline models: DeepSeek, Qwen, Gemini, and ChatGPT, across three orthogonal dimensions: "user requirement alignment," "literature knowledge reference," and "no hypothesis gap." The retrieval and inference agent achieved near-perfect ranking results in the "no hypothesis gap" dimension, being rated first in 233 out of 236 test samples, verifying the effectiveness of anchoring generated content to retrieved literature. Since the retrieval and inference agent is built on the DeepSeek large language model, the performance comparison between the two constitutes a natural ablation experiment. The retrieval and inference agent achieved consistent improvements across all three evaluation dimensions, indicating that the introduction of a structured knowledge base and retrieval pipeline in metasurface reverse design brings key gains to physical accuracy, fully demonstrating the significant advantages of the retrieval enhancement architecture compared to using a base model alone.

[0049] like Figure 8 As shown, the replicating agent converts the structural parameters in the design scheme generated by the retrieval reasoning agent into an electromagnetic solver executable script.

[0050] In one specific implementation, the input received by the replication agent includes metasurface geometric parameters, material parameters, operating frequency, boundary conditions, and excitation conditions extracted from the target literature and verified by the retrieval and inference agent. The replication agent converts these parameters into parametric modeling instructions, material assignment instructions, boundary setting instructions, excitation setting instructions, and solver control instructions supported by the electromagnetic solver, generating an editable and reusable executable script for the electromagnetic solver (e.g., a VBA script). After script execution, full-wave electromagnetic simulation results are obtained, including physical responses such as scattering parameters, reflection / transmission amplitude and phase, near-field distribution, far-field distribution, and radar cross-section pattern.

[0051] The simulated intelligent agent is further compared and verified with the reference results reported in the target literature. Taking a polarization-insensitive reflective coded metasurface as an example, such as... Figure 9 As shown, the target literature reports a metasurface that achieves "0" and "1" encoded states by changing the width of the metal patch, and satisfies a phase difference requirement of approximately 180° near 8.7 GHz. Figure 10 As shown, the reproduced intelligent agent reconstructs the corresponding two metasurfaces based on the extracted parameters and performs full-wave electromagnetic simulation. The reconstruction diagram shows that the phase response curve obtained from the simulation is highly consistent with the reference results in the literature, confirming that the reproduced unit-level electromagnetic behavior is faithful to the original design.

[0052] To verify the physical reliability of the reproduced structure, the metasurface array generated by the reproduced intelligent agent was fabricated and experimentally characterized. The scattering pattern reported in the target literature is shown below. Figure 11 As shown, the full-wave electromagnetic simulation scattering pattern of the structure designed in this invention is as follows. Figure 12 As shown. The experimental testing environment is as follows. Figure 13 As shown: In a microwave anechoic chamber, the metasurface sample was mounted on a rotating stage. A rectangular horn antenna approximately 1 meter away illuminated the sample under normal incidence conditions. A common-polarized rectangular horn antenna approximately 10 meters away served as the receiver. The radar cross-section pattern of the metasurface was recorded via a horizontally rotating platform. The comparison results between the simulated and measured radar cross-section patterns are shown below. Figure 14As shown, the two approaches closely match near the designed anomalous reflection angle—according to phase gradient design and the generalized Snell's law, the anomalous reflection angle under the used coding period is approximately ±29.5°, and both simulation and experimental results exhibit significant scattering lobes near this direction. Slight differences can be attributed to manufacturing tolerances, assembly errors, finite size effects, and measurement uncertainties. The results verify that the multi-agent collaborative framework proposed in this invention can autonomously transform design knowledge from the literature into executable electromagnetic structures and verify them through physical experiments, achieving a complete closed loop from knowledge retrieval to structure reproduction and experimental verification.

[0053] It should be understood that the document parsing tool, visual language model, and large language model in this invention can be implemented in different ways, and the field design of the structured dataset, physical verification rules, and electromagnetic solver interface can also be adjusted according to the application scenario. The core of this invention lies in introducing a knowledge-guided multi-agent collaborative mechanism into the electromagnetic metasurface research process. It lays the information foundation through offline knowledge base construction and multimodal graph and text parsing, provides large-scale trend insights through data analysis agents, ensures the physical reliability of design reasoning through hybrid retrieval and retrieval enhancement generation, and achieves end-to-end automation from literature parameters to executable modeling and simulation through reproducibility agents. All equivalent substitutions or modifications made under the concept of this invention should fall within the protection scope of this invention.

Claims

1. A knowledge-guided multi-agent collaborative autonomous research framework for electromagnetic metasurfaces, characterized in that, Includes the following steps: Literature in the field of electromagnetic metasurfaces is acquired, multimodal analysis is performed on the literature, textual and visual information is extracted and fused into a unified corpus, and a structured dataset and vector database are generated based on the unified corpus to construct an offline knowledge base. The data analysis agent performs autonomous literature mining and domain trend analysis on the structured dataset to generate visual analysis results that reflect the research status of electromagnetic metasurfaces. After receiving the user's design requirements, a hybrid retrieval is performed by a retrieval reasoning agent. First, semantic retrieval is used to retrieve the context of candidate documents from the vector database. Then, numerical retrieval is used to filter the context of candidate documents, locate the target documents that match the user's design requirements, and generate a design scheme containing structural parameters based on the retrieval enhancement generation mechanism. By reproducing the intelligent agent, the design scheme is converted into an executable script for an electromagnetic solver, and full-wave electromagnetic simulation is performed to complete the autonomous research process from literature knowledge to simulation model in a closed loop.

2. The method for autonomous research framework of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration according to claim 1, characterized in that, A document parsing tool is used to convert the PDF files of the documents into a machine-readable format; a visual language model is used to parse the figures and tables in the documents and extract the physical parameters and topological information embedded in the figures and tables; the visual information extracted by the visual language model is merged with the main text of the documents to form a unified corpus containing graphic and textual information.

3. The method for autonomous research framework of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration according to claim 1, characterized in that, Based on a unified corpus, a large language model is used to extract structural parameters and construct a structured dataset; The unified corpus is converted into vector embeddings and stored in a vector database, so that physical concepts are mapped to a consistent embedding space.

4. The method for autonomous research framework of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration according to claim 1, characterized in that, Hybrid retrieval involves both semantic and numerical retrieval. Semantic retrieval matches relevant document context from a vector database based on the semantic description in the user's design requirements. Numerical retrieval, on the other hand, precisely filters the candidate document context returned by semantic retrieval based on the numerical conditions in the user's design requirements, in order to locate the single target document that best matches the user's design requirements.

5. The method for autonomous research framework of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration according to claim 1, characterized in that, The retrieval enhancement generation mechanism uses the corpus content of the target document as contextual enhancement information input into the large language model. The large language model then generates a design scheme containing structural parameters based on the contextual enhancement information to eliminate the physical unreliability in direct generation.

6. The method for an autonomous research framework of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration according to claim 1, characterized in that, The data analysis agent performs autonomous document mining and domain trend analysis, including the following steps: autonomously extracting metadata features of documents from the structured dataset, the metadata features including metasurface type, frequency, electromagnetic function and application system; automatically synthesizing data visualization code to generate visualization charts reflecting the multidimensional mapping relationship between the metadata features.

7. The method for an autonomous research framework of electromagnetic metasurfaces based on knowledge-guided multi-agent collaboration according to claim 1, characterized in that, The replication agent converts the design scheme into an executable script for the electromagnetic solver, including the following steps: converting the design scheme into parametric modeling instructions, material assignment instructions, boundary setting instructions, excitation setting instructions, and solver control instructions supported by the electromagnetic solver; after performing full-wave simulation, the replication agent compares and verifies the simulation response results with the reference results in the target literature.

8. A knowledge-guided multi-agent collaboration-based autonomous research framework system for electromagnetic metasurfaces, used in the knowledge-guided multi-agent collaboration-based autonomous research framework method for electromagnetic metasurfaces as described in any one of claims 1-7, characterized in that, include: The knowledge base construction module is used to acquire multiple documents in the field of electromagnetic metasurfaces, perform multimodal analysis on the documents to extract textual and visual information and fuse them into a unified corpus, and generate a structured dataset and a vector database based on the unified corpus to construct an offline knowledge base. The data analysis intelligent agent module is used to perform autonomous literature mining and domain trend analysis on the structured dataset, and generate visual analysis results that reflect the research status of electromagnetic metasurfaces. The retrieval and reasoning agent module is used to perform hybrid retrieval after receiving user design requirements. First, it uses semantic retrieval to retrieve candidate document context from the vector database, then uses numerical retrieval to filter the candidate document context to locate the target document, and finally generates a design scheme containing detailed structural parameters based on the retrieval enhancement generation mechanism. The intelligent agent module is used to convert the structural parameters in the design scheme into an executable script for an electromagnetic solver, and to perform full-wave electromagnetic simulation. The data analysis agent module, the retrieval and reasoning agent module, and the reproduction agent module work together in the workflow sequence of offline knowledge base construction, data analysis, retrieval and reasoning, and structure reproduction.

9. The knowledge-guided multi-agent collaborative autonomous research framework system for electromagnetic metasurfaces according to claim 8, characterized in that, The knowledge base construction module includes: a document parsing unit for converting PDF files of documents into a machine-readable format; a visual language modeling unit for parsing charts in documents to extract physical parameters and topological information embedded in the charts; a corpus fusion unit for merging the visual information extracted by the visual language modeling unit with the main text of the documents into a unified corpus; a data structuring unit for extracting structural parameters based on the unified corpus to construct a structured dataset; and a vectorization unit for converting the unified corpus into vector embeddings and storing them in a vector database.