Bionic design scheme intelligent generation method and system fused with dual-granularity knowledge base

Abstract

The invention discloses a bionic design scheme intelligent generation method and system fused with a dual-granularity knowledge base, and belongs to the technical field of bionic design and artificial intelligence. The method comprises the following steps: firstly, constructing a dual-granularity knowledge base containing fine-granularity knowledge and coarse-granularity knowledge; secondly, a dual-granularity knowledge retrieval method fusing semantic retrieval and sub-library scoring is adopted, and related biological strategies are accurately obtained; and finally, performing evaluation and iterative optimization on the generated bionic design scheme through a multi-language model expert consultation mechanism. The technical problems that in existing bionic design, the manual screening efficiency is low, and domain knowledge limitation exists in large language model application can be solved, a high-quality bionic design scheme can be intelligently generated, and the method has wide application prospects in the fields of materials, structures, system design and the like.

Description

Technical Field

[0001] This invention relates to the field of biomimetic design and artificial intelligence, specifically to a method and system for intelligently generating biomimetic design schemes that integrates a dual-granularity knowledge base. Background Technology

[0002] Bionics, serving as a bridge between biology and engineering, aims to solve engineering challenges by mimicking the ingenious structures, efficient energy mechanisms, and powerful adaptability of organisms in nature. However, traditional biomimetic design faces dual barriers in its widespread adoption: difficulty in retrieval and difficulty in translation. On the one hand, the vast amount of biological knowledge exists largely in unstructured text, making it difficult for engineers lacking specialized backgrounds to quickly identify effective strategies. On the other hand, the abstract transformation from biological morphology to engineering principles lacks standardized tools, relying heavily on personal experience, resulting in inefficient and unpredictable design processes.

[0003] To address these pain points, academia has developed biomimetic knowledge bases such as AskNature. While these systems improve information retrieval efficiency, they are limited by high manual annotation costs and slow updates. Furthermore, shallow keyword-based matching struggles to accurately interpret complex natural language descriptions from designers, resulting in unsatisfactory recall and precision rates.

[0004] In recent years, the rapid development of large language models (LLMs) has brought a turning point to the intelligent transformation of biomimetic design. General-purpose large models possess powerful semantic understanding and generation capabilities, quickly providing design inspiration. However, their limitations are equally significant: in the absence of domain-specific knowledge constraints, models are prone to "illusions" or ignoring engineering feasibility, and are difficult to evaluate systematically and multidimensionally. Therefore, the key challenge currently facing us is how to deeply integrate structured biomimetic knowledge with the reasoning capabilities of large models to construct an intelligent system that operates in a closed loop of "problem-knowledge-solution," thereby achieving automation and innovation in biomimetic design while ensuring the reliability of the solutions.

[0005] Chinese patent document CN 111310438A proposes a method and device for intelligent semantic matching of Chinese sentences based on a multi-granularity fusion model. This method improves the accuracy of sentence semantic encoding through the fusion modeling of character-level and word-level granularity, and is applicable to general text matching scenarios such as question-and-answer matching. However, the multi-granularity of this scheme is mainly reflected at the two linguistic unit levels of characters and words, remaining at the level of planar text semantic matching tasks. It does not build a structured, searchable knowledge base for a specific professional design field, nor does it distinguish between macro-level design strategies and micro-level implementation details, thus failing to support the dual-granularity organization and management of biological strategies in biomimetic design scenarios. This method only solves the problem of determining whether sentence A semantically matches sentence B, without designing a retrieval and ranking mechanism for the problem-oriented candidate biological strategy set surrounding complex engineering design problems. It struggles to address the technical problem of intelligently weighing macro-concepts and micro-details and selecting the most relevant biological strategies. This method does not introduce a multi-language model collaborative evaluation and iterative optimization mechanism, and cannot form a closed-loop process from knowledge retrieval to solution evaluation and optimization, thus failing to guarantee the reliability and feasibility of the generated design solution.

[0006] Chinese patent document CN 118966342B proposes an industrial knowledge-generating decision-making method based on multi-granularity semantics and large-model assistance. This method, targeting industrial decision-making scenarios, improves the matching ability between domain knowledge and problem semantics to some extent through graph embedding and large-model fine-tuning. However, this method mainly focuses on industrial business decisions (such as equipment operation and maintenance, process optimization, etc.) and does not consider the special needs of bio-engineering cross-domain mapping in biomimetic design. It also lacks a multi-language model expert consultation and evaluation framework for biomimetic design schemes, making it difficult to solve the technical problems of multi-dimensional evaluation and iterative optimization of biomimetic design schemes.

[0007] Chinese patent document CN 117057173B proposes a biomimetic design method that supports divergent thinking. This method constructs a semantic network with source, benefit, and application as nodes, and trains a retrieval model based on this semantic network. After receiving design-related keywords, the system can output a set of related triples, including shared biomimetic objects / applications and similar biomimetic objects / applications, thereby reducing information retrieval time and increasing the diversity of design results to some extent, supporting designers' divergent thinking. However, the knowledge representation of this method is mainly based on triples and semantic networks, which is essentially still a single-granularity, case-level knowledge organization method. It lacks hierarchical modeling of the same biomimetic strategy, using macro-level knowledge blocks and micro-level detailed knowledge blocks, making it difficult to simultaneously consider the macro-semantic framework and micro-implementation details of biological strategies. This method's retrieval process primarily expands based on nodes and their relationships, lacking a fusion scoring mechanism based on micro-level and macro-level knowledge entries. It cannot explicitly consider factors such as how many macro-level knowledge blocks a candidate biological strategy is covered in, or how the weight of each knowledge block is distributed at the micro-level, making it difficult to perform fine-grained comprehensive ranking of candidate strategies. Furthermore, this method mainly focuses on providing diverse inspiration and case studies, without utilizing large language models for end-to-end biomimetic design scheme generation and multi-expert collaborative evaluation. It also fails to construct an integrated closed-loop process from problem input and knowledge retrieval to scheme generation and iterative optimization, thus exhibiting shortcomings in scheme quality control and reliability.

[0008] As can be seen from the above, the existing technologies still cannot solve the following key technical problems at the same time: (1) How to construct a dual-granularity biomimetic knowledge base that simultaneously contains fine-grained biological strategy text blocks and coarse-grained strategy overview entries, and connect the macro-micro semantic relationships at the knowledge structure level; (2) How to design a retrieval algorithm that can simultaneously use Micro-level and Macro-level retrieval results for weighted fusion scoring, ensuring both the recall rate of retrieval results and the high relevance to specific design problems; (3) How to introduce a multi-language model expert consultation mechanism on the basis of knowledge retrieval, and conduct multi-dimensional collaborative review and iterative optimization of the generated biomimetic design scheme, thereby improving the professionalism, reliability and innovation of the scheme.

[0009] Therefore, in response to the aforementioned technical problems that have not yet been effectively solved, this invention proposes an intelligent generation method and system for biomimetic design schemes that integrates a dual-granularity knowledge base. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent generation method and system for biomimetic design schemes that integrates a dual-granularity knowledge base. This method and system construct a structured dual-granularity knowledge base containing macro-strategies and micro-details, combine it with a novel fusion retrieval mechanism, and introduce a multi-language model (LLM) consultation and evaluation framework that simulates the collaborative work of a human expert team. This enables end-to-end automated generation of high-quality, feasible biomimetic design schemes from user-input design problems.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows:

[0012] A method for intelligently generating biomimetic design schemes by integrating a dual-granularity knowledge base, the method comprising the following steps:

[0013] S1. Construct a dual-granularity knowledge base that includes both fine-grained and coarse-grained knowledge bases;

[0014] S2. Based on the dual-granularity knowledge base, perform a fusion search of fine-grained and coarse-grained retrieval to obtain a set of candidate biological strategies;

[0015] S3. Based on the candidate biological strategy set, the scheme is evaluated and iteratively optimized through a multi-language model expert consultation mechanism to generate the final biomimetic design scheme.

[0016] As a further improvement to the above technical solution, step S1 specifically includes:

[0017] S11. Construct a fine-grained knowledge base.

[0018] S111. Divide the biological strategy text into multiple text blocks according to the preset text block length;

[0019] S112. The multiple text blocks are vectorized using a preset first semantic embedding model to generate corresponding text block embedding vectors.

[0020] S113. Embed the text block into a vector and store it in a vector database to generate a fine-grained knowledge base.

[0021] S12. Construct a coarse-grained knowledge base.

[0022] S121. Extract the overview information of the biological strategy text as an independent text unit;

[0023] S122. The independent text units are vectorized using a preset first semantic embedding model to generate corresponding text unit embedding vectors.

[0024] S123. The text unit embedding vectors are processed using dimensionality reduction and clustering algorithms to achieve semantic grouping, form a structured knowledge organization, and generate a coarse-grained knowledge base.

[0025] As a further improvement to the above technical solution, step S2 specifically includes:

[0026] S21. Receive the input biological design problem, and perform parallel retrieval in the fine-grained knowledge base and the coarse-grained knowledge base according to the biological design problem to obtain an initial relevance score;

[0027] S22. According to the preset weight allocation mechanism, the retrieval results of the fine-grained knowledge base and the coarse-grained knowledge base are weighted and fused to calculate the fusion score.

[0028] The weight allocation mechanism uses a weight function. for:

[0029] ;

[0030] in, Indicates Macro-level entries The number of elements in the associated biological strategy set, For hyperparameters, defined as the contribution of macro-level knowledge blocks, if a macro-level knowledge block is composed of multiple micro-level sub-knowledge blocks, then its weight value is evenly distributed to the biological strategies associated with each sub-knowledge block that constitutes the macro-level knowledge block.

[0031] The fusion score is calculated using the following formula:

[0032] ;

[0033] in, Indicate candidate biological strategies, Represent each candidate biological strategy The fusion score.

[0034] Each candidate biological strategy The weights of all related knowledge items in the Micro-level and Macro-level granularity knowledge bases are summed into the final score. Based on the above definition, the method for integrating Micro-level (fine-grained) and Macro-level (coarse-grained) scoring proposed in this invention can be formally expressed as the above formula.

[0035] S23. Based on the preset relevance threshold, select samples with a fusion score not lower than the relevance threshold. Biological strategies are used to obtain a set of candidate biological strategies. .

[0036] The candidate biological strategy set As shown in the following formula:

[0037] .

[0038] As a further improvement to the above technical solution, step S3 specifically includes:

[0039] S31. Using the preset first language model, generate an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem.

[0040] S32. Using multiple preset second-largest language models as review experts, the initial biomimetic design scheme is collaboratively reviewed from multiple preset evaluation dimensions, and a multi-round consultation mechanism is adopted to obtain review opinions.

[0041] Let the first The i-th review expert in turn The solution selection result is Update the selection results using the following formula:

[0042] ;

[0043] in, This represents the rational adjustment function by which the expert made their choice after considering the opinions of two other experts. for The results of the solutions selected by the first, second, and third experts in the round.

[0044] If the solutions chosen by all review experts are in Consistency in the wheel, that is If the scheme is determined to be the Best Bio-Inspired Design (B-BID), no further discussion is needed.

[0045] S33. Using the third preset large language model as the authoritative expert E*, when there are disagreements in the review opinions of the experts, a final decision is made on the disagreements, and the initial biomimetic design scheme is iteratively optimized using the following formula to obtain the final biomimetic design scheme:

[0046] ;

[0047] in, For the first The review experts discussed the final round of audits. maxThe optimal solution for the wheel is i=1,2,3; B-BID is the best biomimetic design solution.

[0048] As a further improvement to the above technical solution, the preset text block length in step S111 is 200 tokens. The preset first semantic embedding model is the BGE-M3 embedding model.

[0049] As a further improvement to the above technical solution, the dimensionality reduction algorithm in step S123 is the UMAP algorithm, which optimizes the adjacency probability between samples in high-dimensional space. Probability corresponding to low-dimensional space Cross-entropy loss between for:

[0050] ;

[0051] The clustering algorithm in step S123 is the HDBSCAN algorithm; the clustering quality is optimized by calculating the silhouette coefficient, which is specifically defined as follows:

[0052] ;

[0053] in, For a non-noise sample index set, , Representative sample The average distance to other samples within its cluster. This represents the average distance to its nearest neighboring cluster sample. The average silhouette coefficient is a metric used to measure the overall quality of clustering.

[0054] As a further improvement to the above technical solution, the preset weight allocation mechanism includes: assigning a first preset weight to the retrieval results of the fine-grained knowledge base, and assigning a second preset weight to the retrieval results of the coarse-grained knowledge base, wherein the value of the second preset weight is greater than the value of the first preset weight; and the value of the second preset weight is twice the value of the first preset weight.

[0055] As a further improvement to the above technical solution, in step S32, multiple preset evaluation dimensions include: functionality and sustainability, function-based analogy matching, mechanism abstraction and cross-domain transformation, coherence, and relevance; the multi-round consultation mechanism is the Delphi method, and the Seminar function is used to show the opinions of other review experts to each review expert so that they can adjust their own choices.

[0056] As a further improvement to the above technical solution, the method also includes data preprocessing; the data preprocessing includes: obtaining raw data from a professional biomimetic database and performing quality control and standardization processing on the raw data.

[0057] The method also includes fine-tuning the large language model used for retrieval;

[0058] The fine-tuning employs the LoRA method shown in the following formula:

[0059] ;

[0060] in, This is the original weight matrix of the large language model; The equivalent weights of the large language model after incorporating the LoRA method; For the incremental update of the LoRA method, i.e., a low-rank matrix of rank r; A and B are two small matrices derived from the low-rank decomposition.

[0061] In a second aspect of the invention, a biomimetic design scheme intelligent generation system integrating a dual-granularity knowledge base is disclosed, the system comprising:

[0062] A dual-granularity knowledge base construction module is used to construct a dual-granularity knowledge base, which includes a fine-grained knowledge base and a coarse-grained knowledge base;

[0063] The dual-granularity fusion retrieval module is used to perform fusion retrieval in the dual-granularity knowledge base based on the input biological design problem, and filter to obtain a set of candidate biological strategies;

[0064] The multi-language model expert consultation, evaluation and optimization module is used to generate an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem, and to use multiple large language models for collaborative review and iterative optimization to generate the final biomimetic design scheme.

[0065] The construction of the dual-granularity knowledge base specifically includes:

[0066] The biological strategy text is segmented into multiple text blocks according to a preset text block length. A preset first semantic embedding model is used to vectorize the multiple text blocks, generating corresponding text block embedding vectors, which are stored in a vector database to construct the fine-grained knowledge base. The overview information of the biological strategy text is extracted as independent text units. The preset first semantic embedding model is used to vectorize the independent text units, generating corresponding text unit embedding vectors. Dimensionality reduction and clustering algorithms are used to process the text unit embedding vectors to achieve semantic grouping, thereby constructing the coarse-grained knowledge base.

[0067] The step involves performing a fusion search on the dual-granularity knowledge base based on the input biological design problem to obtain a set of candidate biological strategies, specifically including:

[0068] Parallel retrieval is performed in the fine-grained knowledge base and the coarse-grained knowledge base to obtain an initial relevance score; the retrieval results are weighted and fused according to a preset weight allocation mechanism to calculate a fusion score; biological strategies with fusion scores not lower than the preset relevance threshold are selected according to a preset relevance threshold to obtain the candidate biological strategy set.

[0069] The process of generating an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem, and then using multiple large language models for collaborative review and iterative optimization to generate the final biomimetic design scheme, specifically includes:

[0070] An initial biomimetic design scheme is generated using a first preset language model; multiple second preset language models are used as review experts to conduct a collaborative review of the initial biomimetic design scheme; a third preset language model is used as an authoritative expert to make a final decision and iterative optimization when there are disagreements in the review opinions of the review experts, thereby obtaining the final biomimetic design scheme.

[0071] The dual-granularity knowledge base construction module is also configured with a vector database for storing the fine-grained knowledge base and a structured database for storing the coarse-grained knowledge base;

[0072] The dual-granularity fusion retrieval module is equipped with a retrieval engine finely tuned using the LoRA method;

[0073] The multi-language model expert consultation, evaluation and optimization module includes a preset first language model, a preset second language model and a preset third language model deployed on the server, as well as a feedback mechanism for processing expert review opinions.

[0074] Compared with the prior art, the advantages of the present invention are:

[0075] This invention can solve the technical problems of low efficiency in manual screening and the limitation of domain knowledge in the application of large language models in existing biomimetic designs. It can intelligently generate high-quality biomimetic design solutions and has broad application prospects in the fields of materials, structure and system design. Attached Figure Description

[0076] Figure 1 This is a block diagram of the intelligent generation method for biomimetic design schemes that integrates a dual-granularity knowledge base in this invention;

[0077] Figure 2 This is a flowchart of the intelligent generation method for biomimetic design schemes that integrates a dual-granularity knowledge base in this invention.

[0078] Figure 3 Example diagram showing the impact of contribution weight value α;

[0079] Figure 4 This is a schematic diagram of the ablation experiment during the generation phase.

[0080] Figure 5 This is a schematic diagram illustrating an example of the application of the present invention in designing a contaminated soil remediation system using the principles of wetland ecosystems. Detailed Implementation

[0081] To provide a better understanding of the structural features and effects achieved by the present invention, the following detailed description is provided in conjunction with preferred embodiments and accompanying drawings:

[0082] This invention relates to the interdisciplinary field of biomimetic design and artificial intelligence, specifically to a method, system, and application for intelligently and automatically generating biomimetic design solutions using large language models (LLMs) and a dual-granularity knowledge base. This invention aims to address the problems of low knowledge acquisition efficiency and insufficient information utilization in traditional biomimetic design processes, as well as the limitations of existing AI-based solutions in terms of domain knowledge depth, logical reasoning, and solution innovation. It is particularly suitable for complex scenarios that require drawing inspiration from natural biological systems to solve engineering problems, such as new material development, functional structure design, robotic system construction, and sustainable technology innovation.

[0083] The technical problem to be solved by this invention is:

[0084] (1) The problem of granularity imbalance in knowledge representation: How to construct a knowledge base that can both grasp the overall picture of biological strategies from a macro perspective and form a structured knowledge system, and delve into the specific implementation details to meet the retrieval needs at different levels.

[0085] (2) The issue of accuracy and comprehensiveness of knowledge retrieval: How to design a retrieval algorithm that can accurately understand the user's complex design intentions and make intelligent trade-offs between macro concepts and micro details to recommend the most relevant biological inspirations while ensuring high recall and high precision.

[0086] (3) Quality and reliability of the generated scheme: How to ensure that the biomimetic design scheme generated by the large language model is not only relevant in content, but also rigorous in technical logic, complete in structure, feasible in practice, and innovative.

[0087] (4) Automation and closed-loop design process: How to seamlessly integrate knowledge retrieval, solution generation, solution evaluation and solution optimization to form an automated design closed loop that can improve itself and iterate continuously.

[0088] Example 1

[0089] This invention provides an intelligent generation method for biomimetic design schemes that integrates a dual-granularity knowledge base. The core idea of ​​this method is to simulate the cognitive process of human biomimetic design experts, combining structured representation of knowledge, multi-dimensional retrieval and collective verification, and creative generation to construct an automated, high-quality biomimetic design process.

[0090] The complete process of this method will be explained in detail below, such as Figure 1 and Figure 2 As shown, the process mainly includes three core stages: building a dual-granularity knowledge base, dual-granularity fusion retrieval, and evaluation and optimization based on expert consultation using a multi-language model (LLM).

[0091] like Figure 1 and Figure 2 As shown, the intelligent generation method for biomimetic design schemes that integrates a dual-granularity knowledge base includes the following steps:

[0092] S1. Construction of a Two-Granularity Knowledge Base

[0093] The goal of step S1 is to construct a high-quality, structured external knowledge source as the knowledge foundation for the entire method. Unlike the single, flat knowledge bases in existing technologies, this invention innovatively constructs a knowledge base with both coarse-grained and fine-grained levels, aiming to simultaneously capture the macroscopic semantic framework and microscopic implementation details of biological strategies. In this embodiment, the knowledge source primarily comes from the internationally renowned biomimetic database AskNature, which contains a large number of biological strategies, biomimetic innovation cases, and related biological questions that have been compiled and reviewed by experts.

[0094] Before starting to build a knowledge base, a data preprocessing step is usually required. This step mainly includes:

[0095] (1) Data acquisition: Download raw data from target data sources such as AskNature through web crawlers or API interfaces. This data is usually unstructured HTML pages or JSON objects, containing rich information such as the name, overview, detailed description, related organisms, functions, and application scenarios of biological strategies.

[0096] (2) Data Cleaning and Structuring: The acquired raw data is parsed to extract key fields, such as strategy, biological problem, and bio-inspired design. Noise such as HTML tags, advertisements, and irrelevant links is removed. The extracted information is organized into a structured format, such as JSON or a database table, for easier subsequent processing. In this embodiment, 185 core biological strategies, along with 237 related biological problems and 237 bio-inspired design cases, were identified.

[0097] (3) Quality control: Perform manual or semi-automatic quality checks on the structured data to ensure the accuracy and completeness of key information. For example, check whether the description of the biological strategy is clear and whether there are any factual errors, and ensure that its association with the corresponding biological problem and biomimetic design is correct.

[0098] After preprocessing is complete, the formal construction of the two-granularity knowledge base begins:

[0099] S11. Construct a micro-level knowledge base, which aims to store the detailed implementation mechanisms and specific knowledge points of biological strategies.

[0100] S111, Text Segmentation.

[0101] Detailed descriptions of biological strategies are selected as the processing objects. To achieve accurate matching in subsequent vector retrieval, long texts need to be segmented into semantically complete and appropriately sized text chunks. In this embodiment, a preset text chunk length of 200 tokens is set. Tokens are the basic units for language models to process text. The length of 200 tokens roughly corresponds to 100-150 Chinese characters or English words. This length is usually sufficient to contain a complete and independent knowledge point (e.g., a functional description of a specific biological structure), while also being short enough to avoid semantic ambiguity due to including too many topics during vectorization. Segmentation employs sliding windows or semantic boundary-based strategies (such as periods or paragraphs) to ensure the coherence of text chunks. For example, a long text describing the biological strategies of kingfishers diving to catch fish will be segmented into multiple independent text chunks, such as the streamlined design and drag reduction principle of the beak, the nictitating membrane protection mechanism of the eye, and body posture control and entry angle into the water.

[0102] S112, Vectorization processing.

[0103] For each segmented text block, a pre-trained, powerful semantic embedding model is used for vectorization. This embodiment preferably employs the BGE-M3 multilingual embedding model as the semantic embedding model. This model is one of the industry's leading text embedding models; it not only supports multiple languages ​​and can handle mixed Chinese and English corpora, but more importantly, it can capture deep semantic information of the text, not just surface keywords. The model maps each text block (a string) to a high-dimensional (e.g., 1024-dimensional) floating-point vector, i.e., the text block embedding vector. In the vector space, semantically similar text blocks have corresponding vectors that are spatially close to each other (e.g., measured by cosine similarity or Euclidean distance).

[0104] S113, Data storage.

[0105] The original text of all text blocks and their corresponding embedding vectors are stored in Milvus, a vector database specifically designed for efficient vector retrieval. By constructing a special index structure (such as IVF-FLAT), efficient approximate nearest neighbor (ANN) searches can be performed on massive amounts of vectors, reaching billions of data points, thus providing performance guarantees for subsequent real-time semantic retrieval. At this point, the fine-grained knowledge base is complete.

[0106] S12. Construct a coarse-grained knowledge base. The coarse-grained knowledge base aims to reveal the macro-themes and semantic relationships between biological strategies, forming a structured knowledge map.

[0107] S121. Extract core text units. For each biological strategy, instead of using its lengthy detailed description, extract its most core and concise summary information. This part is usually a high-level summary of the entire biological strategy and best represents its core ideas. This summary information is treated as an independent text unit.

[0108] S122. Vectorization. Similar to building a fine-grained knowledge base, the BGE-M3 semantic embedding model is used to transform each independent text unit (i.e., policy summary) into a high-dimensional text unit embedding vector.

[0109] S123, Semantic Clustering and Structuring. This is the most crucial step in building a coarse-grained knowledge base. The goal is to organize semantically similar biological strategies together to form higher-level knowledge clusters. This embodiment employs a dimensionality reduction + clustering technology pipeline:

[0110] First, the UMAP (Uniform Manifold Approximation and Projection) algorithm is used to reduce the dimensionality of all text unit embedding vectors. UMAP is an advanced nonlinear manifold learning technique that preserves the global topological structure of the data to the greatest extent while reducing dimensionality. This means that the clustering and separation relationships between vectors in the original high-dimensional space are well preserved in the reduced-dimensional space. This is crucial for the effectiveness of subsequent clustering algorithms.

[0111] Then, in the dimensionality-reduced vector space, the HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) algorithm is used for clustering. HDBSCAN was chosen because it has two main advantages: first, it can automatically determine the optimal number of clusters, unlike algorithms such as K-Means, which require pre-specification; second, it can identify and handle noise points, i.e., outlier biological strategies that do not belong to any cluster. The clustering result is that each biological strategy is assigned a cluster ID or labeled as noise. Semantically similar biological strategies, for example, all strategies related to self-cleaning surfaces, regardless of their biological prototype (lily leaf, sharkskin, or pitcher plant), are grouped into the same cluster.

[0112] To ensure effective clustering, this embodiment also introduces the Silhouette Score as an indicator of clustering quality. The Silhouette Score comprehensively considers the density within clusters and the spatiality between clusters; a score closer to 1 indicates better clustering. The clustering quality can be further optimized by adjusting the hyperparameters of UMAP and HDBSCAN (such as the number of neighbors in UMAP and the minimum cluster size in HDBSCAN), with the goal of maximizing the Silhouette Score.

[0113] Through the above steps, a structured, coarse-grained knowledge base is constructed. It not only contains an overview of each biological strategy, but more importantly, it reveals the inherent, semantic-based relationships between these strategies at the class, order, family, and genus levels through clustering. This information can be stored in traditional relational databases or graph databases.

[0114] S2. Based on the dual-granularity knowledge base, perform dual-granularity fusion retrieval.

[0115] Once the knowledge base is built, the system will be able to respond to user queries. The goal of this stage is to accurately and comprehensively retrieve relevant biological strategies from the two-granularity knowledge base based on the biological design questions input by the user.

[0116] S21, Parallel Semantic Retrieval.

[0117] When a user inputs a BioP design problem described in natural language, such as "How to design a strong adhesive that can be repeatedly used in humid environments?", the system first uses the same BGE-M3 model as when building the knowledge base to transform the problem into a query vector. Then, the system simultaneously sends this query vector to both the fine-grained knowledge base (vector database) and the coarse-grained knowledge base (clustering result database) to perform the retrieval.

[0118] In the fine-grained knowledge base, a vector similarity search is performed, returning a batch of text blocks most similar to the query vector. In the coarse-grained knowledge base, also a vector similarity search is performed, returning a batch of biological strategy summaries most similar to the query vector. In this way, the system obtains two independent initial search result lists from different granularities, along with their respective initial relevance scores (usually cosine similarity) to the query question.

[0119] S22, Weighted Fusion Score.

[0120] To integrate information at these two levels of granularity and fully leverage the advantages of a two-layer knowledge structure, this invention proposes an innovative weighted fusion scoring mechanism. The core idea of ​​this mechanism is that macro-level matching (i.e., the overall concept of the biological strategy being relevant to the problem) is more important than micro-level matching of a specific detail. Therefore, this invention assigns different weights to retrieval results from different knowledge bases.

[0121] In this embodiment, a first preset weight, set to 1, is assigned to the retrieval results from the fine-grained knowledge base. A second preset weight, α, is assigned to the retrieval results from the coarse-grained knowledge base, set to a hyperparameter greater than 1. Experiments have shown that the overall retrieval performance (measured by the MAP index) of the system reaches its optimal level when α=2. This means that the contribution of the coarse-grained knowledge base is considered twice that of the fine-grained knowledge base when calculating the final score.

[0122] For each candidate biological strategy, its final fusion score consists of two parts: one is its weighted score in the coarse-grained knowledge base (i.e., its summary's similarity to the question multiplied by a weight α), and the other is the sum of the scores of all relevant text blocks retrieved in the fine-grained knowledge base. During the retrieval process, for each candidate biological strategy b... j The system will score the data based on its relevance to the query. The score will be updated using the following formula:

[0123]

[0124] in, Representing biological strategies Cumulative ratings This indicates evidence or text blocks related to the strategy. Indicates the weight of the evidence.

[0125] By setting a relevance threshold τ, a set of candidate biological strategies B* that meets the criteria is selected:

[0126] .

[0127] S23, Candidate Strategy Screening.

[0128] After calculating the fusion score of all candidate biological strategies, a preset relevance threshold τ is set. This threshold is an empirical value used to filter out results that are retrieved but have low relevance. Only biological strategies with a fusion score of not less than τ are considered truly relevant and are included in the final candidate biological strategy set (B) for submission to the next stage of processing.

[0129] S3, Multi-Language Model (LLM) Expert Consultation, Evaluation and Optimization

[0130] This is the most innovative part of the invention. It does not stop at simple Retrieval Enhanced Generation (RAG), but designs a complex socialized collaborative creation and verification process that simulates the work of a team of human experts to ensure the quality of the final generated solution.

[0131] S31. Generate the initial biomimetic design scheme (BID).

[0132] First, the system invokes a large language model (the default first LLM) designated as the "Design Draft Expert." This LLM can be a powerful, general-purpose model like GPT-4. The system provides the LLM with the *candidate biological strategy set (B) selected in the previous stage and the user's original biological design problem (BioP), along with a carefully designed prompt template. The task of this LLM is: The generation process of the initial biomimetic design scheme BID can be represented as:

[0133]

[0134] Where F represents the function that integrates and knowledge-based sub-outlines O_i and BioP to finally generate the complete BID.

[0135] S32. Multi-expert collaborative review.

[0136] The generated initial BID is not directly accepted but undergoes a rigorous peer review process. This invention establishes a review committee comprised of multiple pre-defined second LLMs. These LLMs are models from different companies (such as Qwen, LLaMA, and ChatGLM) to introduce diversity of perspectives.

[0137] Each reviewer receives an initial BID and is asked to score it and provide written comments based on a set of pre-defined evaluation dimensions that are critical to the design. In this embodiment, these dimensions include: Functionality and Sustainability (F&S), Function-Based Analogy Matching (FuncMatch), and Mechanism Abstraction and Cross-Domain Transformation (M&C).

[0139] To facilitate effective interaction among experts, this embodiment introduces a multi-round consultation mechanism based on the Delphi Method. In the first round, each expert conducts their review independently and anonymously. Starting from the second round, the system uses a Seminar function to summarize the anonymous scores and comments from all experts in the previous round and display them to each expert. This allows each expert to understand the perspectives of others and has the opportunity to reflect on and adjust their own judgments. This process can be repeated multiple times (e.g., with a preset tmax=3 rounds), aiming to converge expert opinions or discover new and valuable insights.

[0140] S33, expert adjudication and iterative optimization.

[0141] After multiple rounds of consultation, two scenarios may occur: first, experts reach a consensus, unanimously agreeing that a certain solution is optimal or that a modified solution is acceptable; second, significant disagreements remain among the experts. At this point, the authoritative expert role (the third pre-defined LLM) intervenes. This LLM is typically the most capable and reliable model (e.g., the Qwen model with a larger number of parameters), with the highest decision weight. Final Ruling: The authoritative expert receives the final review results (scores and comments) from all review experts, along with the initial BID. It comprehensively analyzes this information, paying particular attention to points of contention, and makes a final, binding ruling based on its superior reasoning ability and knowledge base. For example, between two solutions, A and B, each with its own advantages and disadvantages, it might rule that solution A is more innovative in its core functionality and thus serves as the foundation, while also indicating the need to incorporate the cost control advantages of solution B. Iterative Optimization: After the ruling, the authoritative expert also assumes the crucial role of optimizer. Based on the constructive feedback (including positive affirmations and negative criticisms) collected throughout the review process, it conducts a comprehensive and in-depth modification and refinement of the selected solution. This process is not merely simple text editing, but a deep technical reconstruction, which may include: supplementing missing technical details, correcting logical loopholes, enhancing the feasibility analysis of the solution, and improving the overall structure and expression. The solution optimized by authoritative experts becomes the final biomimetic design solution (B-BID). Through the close cooperation of the above three stages, the embodiments of the present invention can systematically and automatically transform a vague, high-level biomimetic design problem into a concrete, high-quality, and multi-verified engineering solution.

[0142] This embodiment will demonstrate the execution process of the method of the present invention in a specific application scenario, "Design of a Contaminated Soil Remediation System Based on Wetland Ecological Principles". Suppose an environmental technology company faces a technical challenge: to design a system capable of in-situ remediation of industrial sites contaminated by heavy metals and hydrocarbons.

[0143] User input (BioP design problem): "Design an automated system capable of effectively selecting, deploying, and monitoring specific fungal and microbial communities for in-situ bioremediation of contaminated industrial soils. The system needs to ensure the effective degradation of the target pollutants while maintaining and restoring soil health and ecosystem integrity."

[0144] Phase 1: Knowledge Base Construction

[0145] The system first constructs a two-granularity knowledge base based on the pre-processed AskNature database. It assumes the database contains an article on the biological strategies of a wetland ecosystem, detailing how wetlands purify water through complex interactions between plants, microorganisms, and the physical environment.

[0146] 1. Building a fine-grained knowledge base: The detailed description text of this strategy is segmented into multiple 200-token text blocks, for example:

[0147] Chunk 1: "Plants such as reeds and cattails in wetlands provide a huge surface area for microorganisms to attach to and transport oxygen to the rhizosphere, forming a microenvironment that alternates between aerobic and anaerobic processes..."

[0148] Chunk 2: "The rhizosphere microbial community, especially certain strains of Pseudomonas, has been shown to be highly efficient at degrading stubborn organic pollutants such as polycyclic aromatic hydrocarbons..."

[0149] Chunk 3: "Some fungi, such as Trichoderma, can produce a variety of extracellular enzymes that decompose complex hydrocarbons and can accumulate heavy metals such as cadmium and lead in the soil..."

[0150] Chunk 4: "The physicochemical conditions of wetland soils, such as adequate moisture content, neutral pH, and abundant organic matter, are key to maintaining microbial activity and pollutant degradation efficiency..."

[0151] These text blocks are vectorized using the BGE-M3 model and then stored in a vector database.

[0152] 2. Constructing a coarse-grained knowledge base: The wetland ecosystem strategy is summarized as follows: "Wetlands, through their unique plant-microbe-soil complex system, exhibit strong self-purification capabilities, efficiently removing various pollutants from water and soil, and achieving a stable ecosystem balance." This summary text was vectorized using the BGE-M3 model and, along with summaries of 184 other biological strategies, underwent UMAP dimensionality reduction and HDBSCAN clustering. It is assumed that this strategy, along with strategies such as forest litter decomposition and coral reef symbiotic systems, is clustered into the same cluster representing a multi-species synergistic ecosystem.

[0153] Phase Two: Dual-Granularity Fusion Retrieval

[0154] 1. Parallel retrieval: After the user's BioP is vectorized, it is retrieved simultaneously in two knowledge bases.

[0155] Fine-grained searches might return Chunk 2, Chunk 3, and Chunk 4, as they directly match the deep semantics of keywords such as microorganisms, fungi, pollutant degradation, and heavy metals. Coarse-grained searches, on the other hand, might reveal that the overview of wetland ecosystems highly matches the core concepts of BioP, such as "ecosystem restoration" and "microbial communities."

[0156] 2. Fusion Score and Selection: Calculate the fusion score for the wetland ecosystem strategy. Due to its high matching at the coarse-grained level (assuming a similarity of 0.9) and the support of multiple strongly related text blocks at the fine-grained level, its fusion score will be very high. This strategy will be successfully selected into the candidate biological strategy set B.

[0157] Phase 3: Multi-LLM Expert Consultation, Evaluation, and Optimization

[0158] 1. Initial BID Generation: After receiving the BioP and "Wetland Ecosystem" strategies, the "Design Draft Expert" (GPT-4) generates the following initial BID:

[0159] Initial biomimetic design scheme (BID-0)

[0160] Name: MIMIC-Wetland System (In-situ Bioremediation System that Mimics Wetland Ecological Interactions)

[0161] Design Inspiration: This design mimics the complex synergy between plants, microorganisms and the physical environment in natural wetland ecosystems to achieve efficient and sustainable remediation of polluted soil.

[0162] Core components:

[0163] (1) Microbial and fungal compound preparation (M&F Inoculum): contains screened strains of Pseudomonas and Trichoderma that have the ability to efficiently degrade or enrich specific heavy metals (such as lead and cadmium) and hydrocarbons (such as polycyclic aromatic hydrocarbons).

[0164] (2) Root-Mimic Matrix: A biodegradable porous polymer scaffold used to simulate plant roots, providing an attachment surface for microorganisms, and with an internal microtube network for delivering oxygen and nutrient solution.

[0165] (3) Intelligent monitoring and control unit (SMCU): integrates multiple sensors (pH, humidity, temperature, pollutant concentration) and actuators (micro pumps, valves) to monitor the soil environment in real time and dynamically regulate the supply of oxygen and nutrient solution according to preset algorithms.

[0166] Workflow:

[0167] (1) Deployment: The root system simulation matrix array is implanted into the contaminated soil area.

[0168] (2) Inoculation: The M&F compound preparation is injected into the matrix through SMCU.

[0169] (3) Operation: Based on sensor data, SMCU simulates the alternating aerobic-anaerobic environment of wetland rhizosphere, intermittently pumping air and solutions containing specific nutrients to maintain the optimal activity of the microbial community.

[0170] (4) Monitoring: SMCU continuously monitors changes in pollutant concentration and wirelessly transmits the data to the cloud platform for operators to assess the remediation effect.

[0171] 2. Multi-expert collaborative review: Three "review experts" (LLM-A (Qwen), LLM-B (LLaMA), LLM-C (ChatGLM)) review BID-0.

[0172] LLM-A (Rating: 3 / 5): "The concept is clear and innovative. However, there may be a contradiction between the degradability and long-term stability of the 'root simulation matrix.' In addition, the description of how to select and cultivate the compound formulation is too general and lacks specific standards."

[0173] LLM-B (Rating: 4 / 5): "This is a good starting point. The SMCU design is comprehensive, but energy consumption may be an issue. Solar power should be considered. Additionally, the proposal does not address how to prevent the introduced microorganisms from causing unforeseen negative impacts on the native soil ecosystem."

[0174] LLM-C (Rating: 4 / 5): "The solution is logically sound. However, for assessing the remediation effectiveness, monitoring pollutant concentrations alone may not be sufficient; monitoring of biological indicators of soil microbial diversity and activity should be added."

[0175] 3. Expert Adjudication and Iterative Optimization: After receiving the above review comments, the "authoritative expert" (GPT-4.5-Turbo) will adjudicate and optimize the system.

[0176] Ruling: The authoritative experts believe that the core idea of ​​BID-0 is correct and feasible, but the issues raised by the review experts are very critical and must be addressed in the final solution.

[0177] Optimization: Leading experts have made significant modifications and expansions to BID-0, generating the final B-BID.

[0178] Final biomimetic design scheme (B-BID)

[0179] Name: Adaptive MIMIC-Wetland System (Mimicking Wetland Ecological Interactions)

[0180] ...(The design inspiration is the same as BID-0)...

[0181] Core components:

[0182] (1) Functional Microbial Screening and Compounding Module (FMSM):

[0183] Screening criteria: The quantitative indicators for strain screening are clearly defined, including: target pollutant degradation rate > 90% invitro, heavy metal enrichment coefficient > 1000, adaptability to local soil environment (pH 5-9, temperature 10-40°C), and mutualism with other strains (no antagonistic effect).

[0184] The principle of compounding is to adopt a compounding strategy of "core functional bacteria + auxiliary functional bacteria". For example, Pseudomonas aeruginosa, which is highly efficient at degrading polycyclic aromatic hydrocarbons, is used as the core, supplemented by Bacillus, which can produce biosurfactants to improve the bioavailability of pollutants.

[0185] (2) Hierarchical Root-Mimic Matrix:

[0186] Materials: The core framework is made of biocompatible materials such as polylactic acid (PLA) and manufactured by 3D printing to ensure long-term structural stability; the surface is coated with a layer of chitosan hydrogel to provide a hydrophilic environment and initial nutrition, which can slowly degrade in the later stages of repair.

[0187] Structure: The internal microtube network is designed with a fractal structure to mimic the transport network of real plant roots, achieving maximum coverage with minimal energy consumption.

[0188] (3) Low-power intelligent monitoring and control unit (LP-SMCU):

[0189] Energy: Integrating flexible solar thin films and energy storage batteries, prioritizing the use of renewable energy.

[0190] Biosafety controls: A "biobarrier" monitoring function has been added. eDNA (environmental DNA) sensors are deployed at the boundary of the remediation area to monitor in real time for any leakage of functional bacterial strains. Once detected, the system automatically initiates boundary UV disinfection or phage release procedures.

[0191] Comprehensive benefit assessment: The monitoring indicators have been expanded to include online monitoring of soil respiration rate, microbial biomass carbon (MBC), and key enzyme activities (such as dehydrogenase and polyphenol oxidase) to comprehensively assess the degree of soil health restoration.

[0192] (The workflow section has been updated to include closed-loop feedback for biosafety monitoring and comprehensive benefit assessment.)

[0193] As can be seen from this embodiment, the method of the present invention is not simply about finding a biological inspiration, but rather about using a systematic and progressive process to gradually concretize, refine, and optimize a vague idea, ultimately forming a technically mature and well-considered engineering solution.

[0194] To verify the rationality of the key hyperparameter settings in the method of this invention and to find their optimal values, a series of parameter optimization experiments were conducted. Taking the contribution weight α of the coarse-grained knowledge base as an example, this parameter determines the importance of macro-concept matching in the fusion retrieval stage. This invention uses the MAP (mean accuracy) and P@1 (first-recommendation accuracy) of the final retrieval results as evaluation indicators and conducted a set of comparative experiments. The experimental settings are as follows: keeping all other parameters and components unchanged, the value of the weight α is gradually adjusted from 1 to 5, with a step size of 0.5. α=1 means that the contributions of the coarse-grained and fine-grained knowledge bases are treated equally. On a dataset containing 237 biological questions, for each α value, the retrieval process is run completely once, and its MAP and P@1 scores are calculated.

[0195] The experimental results of the above parameter optimization experiment are as follows: Figure 3 As shown. From Figure 3 The data clearly shows that as α increases from 1.0, both MAP and P@1 indicators exhibit a significant upward trend. This indicates that giving higher weight to the coarse-grained knowledge base (macro-concept matching) can indeed more effectively help the system grasp user intent as a whole, thereby improving retrieval accuracy. When α reaches 2.0, both indicators reach their peak simultaneously. This shows that setting the importance of macro-concept matching to twice that of micro-detail matching is an optimal balance point in the system framework and dataset of this invention. At this point, the system can quickly locate the correct semantic category while making full use of detailed information for precise matching. When α exceeds 2.0 and continues to increase, both indicators begin to decline. This may be because overemphasizing macro-concept matching may cause the system to tend to recommend biological strategies that "look very similar" but do not actually match the details perfectly, leading to "overgeneralization" errors and reducing accuracy. Therefore, this embodiment determines α=2 as the optimal default parameter for the method of this invention. This process also demonstrates the optimizability and scientific nature of the method of this invention; its key parameters are not set arbitrarily but are supported by experimental data.

[0196] To further verify the necessity and effectiveness of the various innovative components of this invention, a series of ablation studies were designed. That is, a key component was "removed" from the complete system, and the performance degradation was observed. The degree of performance degradation reflects the importance of that component. This invention uses the quality of the final generated solution (blindly evaluated and scored by domain experts on a 5-point scale) and retrieval performance (MAP) as evaluation indicators.

[0197] Figure 4 The results of the ablation experiment are presented, and the performance changes of the complete IBID method and the removal of different components are compared by bar chart. Figure 4 The ablation study includes five sets of comparative experiments: The complete IBID (blue bars) represents the complete approach including all components, serving as the baseline; "without review experts" (orange bars) indicates the removal of the review expert committee, retaining only the design draft experts and authoritative experts; "without authoritative experts" (green bars) indicates the removal of authoritative experts, retaining only the design draft experts and the review expert committee; "without outline" (red bars) indicates the removal of the structured outline generation stage; and "without recommendation table" (purple bars) indicates the removal of the expert recommendation table. The ablation study results show that removing the structured outline has the most significant negative impact on relevance and coherence; removing the review expert committee reduces FuncMatch, M&C, and F&S metrics; removing authoritative experts significantly reduces professional quality metrics; and removing the recommendation table also significantly reduces metrics in the final stage. This figure demonstrates the significant contributions of each component in the IBID design and verifies the rationality of the system architecture.

[0198] The experimental setup is as follows:

[0199] (1) Full System: That is, the complete method of the present invention.

[0200] (2) No coarse-grained knowledge base (- Macro KB): Remove the coarse-grained knowledge base and use only the fine-grained knowledge base for traditional vector retrieval. This is equivalent to an optimized version of Naive RAG.

[0201] (3) No Multi-Expert Consultation: The third-stage multi-LLM expert consultation module is removed. After the search is completed, a single LLM (Design Draft Expert) is used to generate the solution and serve as the final result. This is equivalent to a RAG with dual-granularity search.

[0202] (4) No Authority Expert: The "Authority Expert" role is removed in the multi-expert consultation module. When the opinions of the review experts differ, a simple voting or average score mechanism is used to decide.

[0203] The experimental results are shown in the table below:

[0204]

[0205] Analysis of the above experimental results leads to the following conclusions: The necessity of a coarse-grained knowledge base: Removing the coarse-grained knowledge base caused the retrieval performance (MAP) to plummet from 0.578 to 0.449, a decrease of approximately 22.3%. This directly resulted in a drop in the quality of the final generated solution from 4.78 to 4.15. This strongly demonstrates that the dual-granularity knowledge structure proposed in this invention is key to improving retrieval accuracy, and a macroscopic knowledge map is crucial for navigating complex biological knowledge spaces. The necessity of multi-expert consultation: While retaining high-quality retrieval, removing the multi-expert consultation module caused the quality of the generated solution to drop drastically from 4.78 to 3.52, a decrease of approximately 26.4%. This indicates that even with high-quality knowledge input, a single LLM approach, working in isolation, cannot guarantee high-quality output. Introducing a consultation mechanism involving peer review, critical thinking, and iterative optimization is the core driving force for achieving the creative leap from "relevant knowledge" to "high-quality solutions." The Necessity of Authoritative Experts: In the consultation module, simply removing authoritative experts significantly reduced the quality of the proposed solutions (from 4.78 to 4.41). Actual testing revealed that without the adjudication and final optimization by authoritative experts, the system, when faced with differing opinions from review experts, often resorted to simple voting to select the "less controversial" solution rather than the "highest quality" one. Furthermore, it missed the opportunity for a final step of incorporating all opinions for comprehensive refinement. This demonstrates the crucial role of authoritative experts in ensuring decision-making quality and the completeness of the proposed solutions. In conclusion, the ablation experiment results clearly demonstrate that the dual-granularity knowledge base, multi-expert consultation mechanism, and authoritative expert role proposed in this invention are all indispensable key components. Their collaborative efforts enable this invention to achieve performance far exceeding existing technologies.

[0206] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0207] Figure 2 The end-to-end workflow of this invention is demonstrated, mainly divided into three stages. The workflow begins with the "User Input BioP Problem (BioP)" in the upper left corner. The first stage is the construction of a two-granularity knowledge base. Figure 2The diagram illustrates the process of constructing a "fine-grained knowledge base (Micro-level KB)" and a "coarse-grained knowledge base (Macro-level KB)" starting from raw data (such as AskNature). The second stage is "dual-grained fusion retrieval," where arrows indicate that the user's BioP query is simultaneously sent to both knowledge bases for retrieval. The results are processed by a box labeled "fusion scoring module," outputting "candidate biological strategy set B." The third stage is "multi-LLM expert consultation," where a loop structure is used to visually represent the "design draft expert" generating the initial BID, followed by a review process composed of "review experts A / B / C." After the review opinions are summarized, the "authoritative expert" makes the final decision and optimization, finally outputting the "final biomimetic design scheme (B-BID)" in the lower right corner. The entire flowchart clearly reveals the entire process of information flow and value-added in this invention. The upper part of the diagram, the "Fine-grained Knowledge Base Construction Process," illustrates how a long text on biological strategies is processed through "text segmentation (200-token chunks)" to obtain multiple text blocks. Each text block then undergoes "semantic embedding (BGE-M3)" to generate vectors, which are finally stored in a "vector database." The lower part, the "Coarse-grained Knowledge Base Construction Process," demonstrates how "overview information" is extracted from the long text on biological strategies, then "semantic embedding (BGE-M3)" is used to generate vectors, followed by "UMAP dimensionality reduction" and "HDBSCAN clustering" to ultimately form "structured knowledge clusters (clustering results)." This diagram visually compares the similarities and differences between the two granularities of knowledge processing.

[0208] Figure 5 The consultation process is illustrated in sequence, starting with the "Initial BID". First, the BID is distributed to three parallel "Review Experts LLM-A / B / C", each conducting independent "evaluation and scoring". Second, their review opinions are fed into a "Seminar function / opinion summary" module. Third, a decision loop is initiated: "Has a consensus been reached?" If "yes", the process proceeds directly to the "Final Optimization by Authoritative Experts" stage. If "no", the summarized opinions are fed back to the review experts, entering the "Next Round of Review", forming a loop until a consensus is reached or the maximum number of rounds is reached. Outside of this loop, the path for the "Final Ruling" by the "Authoritative Experts" is also shown when irreconcilable disputes arise. The entire diagram clearly demonstrates a socialized collaborative mechanism based on the Delphi method and authoritative decision-making.

[0209] Compared with the prior art, the innovation of this invention is as follows:

[0210] (1) This invention significantly improves the accuracy and comprehensiveness of biomimetic knowledge retrieval, effectively solving the bottleneck problem of knowledge acquisition. This invention fundamentally optimizes the organization and representation of biomimetic knowledge by constructing an innovative dual-granularity knowledge base. The coarse-grained (Macro-level) knowledge base forms a structured knowledge outline by clustering the overview of biological strategies, enabling the system to quickly grasp the overall concept and semantic category of biological strategies at the macro level, greatly improving the precision of retrieval. When a user inputs a design problem, the system can first locate the most relevant biological strategy category at the macro level, avoiding invalid searches in a large number of irrelevant details. At the same time, the fine-grained (Micro-level) knowledge base divides the detailed description of biological strategies into fine text blocks, ensuring comprehensive coverage of micro-information such as specific implementation mechanisms and structural features, thereby improving the recall of retrieval.

[0211] This invention uses P@K and MAP as the main retrieval performance evaluation indicators. The formula for calculating P@K is:

[0212] ;

[0213] Among them, rel i Let K be the binary relevance label for the i-th search result, and K be the set of search requests. The formula for calculating MAP (Mean Average Precision) is:

[0214] ;

[0215] Where Q is the set of retrieval requests, and AP(q) represents the average precision of query q.

[0216] More importantly, this invention features a unique dual-granularity fusion retrieval mechanism. By assigning different weights to retrieval results at different granularities and performing fusion scoring, it achieves a perfect combination of macro-level positioning and micro-level mining. Experimental results show that the method of this invention has achieved industry-leading levels in several key retrieval performance indicators. For example, in the P@1 (first-recommendation accuracy) metric, the method of this invention reaches 0.785, a performance improvement of up to 31.4% compared to the traditional Naive Retrieval Augmentation (Naive RAG) method that only uses a single-granularity text block. This means that the first biological strategy recommended by this invention is highly likely to be a highly relevant strategy, greatly saving users' filtering time. In the MAP (mean mean accuracy), a comprehensive measure of retrieval quality, this invention also reaches 0.578, significantly better than the baseline model. This highly accurate and efficient knowledge retrieval capability effectively solves the core pain point of traditional biomimetic design: the time-consuming, labor-intensive, and inefficient manual information filtering.

[0217] (2) This invention greatly enhances the professionalism, logic, and feasibility of the generated solutions, ensuring high-quality design output. Existing technologies often directly use large language models (LLMs) to generate design solutions, which frequently result in inconsistent solution quality and even factual errors (illusions) due to a lack of professional knowledge and evaluation mechanisms. This invention introduces a multi-LLM expert consultation, evaluation, and optimization module that simulates the collaborative work of a human expert team, thus ensuring the quality of the generated solutions from a mechanistic perspective. This module includes three key roles: the "Design Draft Expert" is responsible for transforming highly relevant retrieved knowledge into a structured initial design solution, ensuring the reliability of the knowledge source and the relevance of the content. The "Review Expert Committee" introduces a peer review mechanism. Multiple LLMs critically examine and evaluate the draft from different professional dimensions (such as functionality, sustainability, cost, manufacturability, etc.), exposing potential problems and defects that a single model might overlook. The "Authoritative Expert" plays the role of the final decision-maker and optimizer, weighing and deciding in case of disputes, and is responsible for absorbing all review opinions, iteratively refining and improving the solution, ensuring the logical rigor, structural integrity, and technical feasibility of the final solution.

[0218] Experiments have demonstrated that the biomimetic design schemes generated using the method of this invention perform excellently across multiple dimensions. On a series of evaluation indicators developed by domain experts, including Functionality and Sustainability (F&S), Function-Based Analogy Matching (FuncMatch), Mechanism Abstraction and Cross-Domain Transformation (M&C), Coherence, and Relevance, the overall score of this invention far surpasses that of methods using only LLM or naive RAG. For example, on Mechanism Abstraction and Cross-Domain Transformation, a key indicator of innovativeness, this invention scores as high as 4.72 (out of 5), indicating its ability to effectively translate biological principles into engineering-usable solutions. This high-quality output means that the automatically generated biomimetic design schemes are no longer abstract concepts, but truly possess the potential to enter the subsequent detailed design and engineering practice stages.

[0219] (3) This invention achieves end-to-end automation from problem input to solution output, significantly improving the efficiency and accessibility of biomimetic design. This invention seamlessly integrates a series of complex activities, such as knowledge base construction, intelligent retrieval, solution generation, multi-party evaluation, and iterative optimization, into a unified, automated process. Users only need to describe the engineering problem they encounter in natural language, and the system can automatically complete all intermediate steps and ultimately deliver a high-quality, multi-verified biomimetic design solution. This greatly lowers the barrier to biomimetic design, enabling engineers and designers without a strong biological background to easily utilize biomimetic principles for innovative design. The entire process requires no cumbersome manual intervention, shortening the design cycle from weeks or even months under traditional methods to minutes or hours, achieving an order-of-magnitude improvement in design efficiency.

[0220] (4) This invention has good scalability and adaptability, and shows broad application prospects in multiple technical fields. The framework of this invention is highly modular and flexible. Its dual-granularity knowledge base can not only accommodate knowledge from different sources and fields, but also continuously optimize its internal structure by updating the embedding model and clustering algorithm. The multi-LLM expert consultation module can also flexibly configure the number, role and evaluation dimensions of experts according to the needs of specific application scenarios. This good scalability makes the method of this invention easy to apply to various innovative activities that need to draw inspiration from nature, including but not limited to: new material development: for example, designing new functional materials with self-healing, self-cleaning or superhydrophobic properties. Bionic structural design: for example, optimizing the load-bearing structure of buildings, the aerodynamic shape of aircraft or the flexible joints of robots. Bionic system design: for example, developing efficient energy harvesting and conversion systems, sustainable pollution control systems or intelligent swarm robot collaborative systems. Through the above case of designing a polluted soil remediation system using the principles of wetland ecosystems, the powerful ability of this invention to solve complex system design problems is clearly demonstrated. This proves that the present invention is not only a theoretical framework, but also an engineering tool with great practical application value, which can provide a continuous source of inspiration and practical solutions for technological innovation in various industries.

[0221] Example 2

[0222] A biomimetic design scheme intelligent generation system integrating a dual-granularity knowledge base, the system comprising:

[0223] A dual-granularity knowledge base construction module is used to construct a dual-granularity knowledge base, which includes a fine-grained knowledge base and a coarse-grained knowledge base;

[0224] The dual-granularity fusion retrieval module is used to perform fusion retrieval in the dual-granularity knowledge base based on the input biological design problem, and filter to obtain a set of candidate biological strategies;

[0225] The multi-language model expert consultation, evaluation, and optimization module is used to generate an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem, and to use multiple large language models for collaborative review and iterative optimization to generate the final biomimetic design scheme.

[0226] The construction of the dual-granularity knowledge base specifically includes:

[0227] The biological strategy text is segmented according to a preset text block length to obtain multiple text blocks; the multiple text blocks are vectorized using a preset first semantic embedding model to generate corresponding text block embedding vectors, which are stored in a vector database to construct the fine-grained knowledge base; the overview information of the biological strategy text is extracted as an independent text unit; the independent text units are vectorized using the preset first semantic embedding model to generate corresponding text unit embedding vectors; the text unit embedding vectors are processed using dimensionality reduction and clustering algorithms to achieve semantic grouping, thereby constructing the coarse-grained knowledge base.

[0228] The step involves performing a fusion search on the dual-granularity knowledge base based on the input biological design problem to obtain a set of candidate biological strategies, specifically including:

[0229] Parallel retrieval is performed in the fine-grained knowledge base and the coarse-grained knowledge base to obtain an initial relevance score; the retrieval results are weighted and fused according to a preset weight allocation mechanism to calculate a fusion score; biological strategies with fusion scores not lower than the preset relevance threshold are selected according to a preset relevance threshold to obtain the candidate biological strategy set.

[0230] The process of generating an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem, and then using multiple large language models for collaborative review and iterative optimization to generate the final biomimetic design scheme, specifically includes:

[0231] An initial biomimetic design scheme is generated using a first preset language model; multiple second preset language models are used as review experts to conduct collaborative reviews of the initial biomimetic design scheme; a third preset language model is used as an authoritative expert to make a final decision and iterative optimization when there are disagreements in the review opinions of the review experts, thereby obtaining the final biomimetic design scheme.

[0232] The dual-granularity knowledge base construction module is also configured with a vector database for storing the fine-grained knowledge base and a structured database for storing the coarse-grained knowledge base.

[0233] The dual-granularity fusion retrieval module is equipped with a retrieval engine finely tuned using the LoRA method.

[0234] The multi-language model expert consultation, evaluation and optimization module includes a preset first language model, a preset second language model and a preset third language model deployed on the server, as well as a feedback mechanism for processing expert review opinions.

[0235] Example 3

[0236] An electronic device includes: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the above-described intelligent generation method for biomimetic design schemes incorporating a dual-granularity knowledge base.

[0237] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable computing device, consumer electronic device, etc.

[0238] Example 4

[0239] A machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the aforementioned intelligent generation method for biomimetic design schemes that integrates a dual-granularity knowledge base.

[0240] Specifically, a system or apparatus equipped with a readable storage medium storing software program code that implements the functions of any of the embodiments described above, and enabling the computer or processor of the system or apparatus to read and execute instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the embodiments described above; therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification. Embodiments of the readable storage medium include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code can be downloaded from a server computer or the cloud via a communication network. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0241] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0242] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for intelligently generating biomimetic design schemes by integrating a dual-granularity knowledge base, characterized in that, The method includes the following steps: S1. Construct a dual-granularity knowledge base that includes both fine-grained and coarse-grained knowledge bases; S2. Based on the dual-granularity knowledge base, perform a fusion search of fine-grained and coarse-grained retrieval to obtain a set of candidate biological strategies; S3. Based on the candidate biological strategy set, the scheme is evaluated and iteratively optimized through a multi-language model expert consultation mechanism to generate the final biomimetic design scheme.

2. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base as described in claim 1, characterized in that, Step S1 specifically includes: S11. Construct a fine-grained knowledge base; S111. Divide the biological strategy text into multiple text blocks according to the preset text block length; S112. The multiple text blocks are vectorized using a preset first semantic embedding model to generate corresponding text block embedding vectors. S113. Embed the text block into a vector and store it in a vector database to generate a fine-grained knowledge base; S12. Construct a coarse-grained knowledge base; S121. Extract the overview information of the biological strategy text as an independent text unit; S122. The independent text units are vectorized using a preset first semantic embedding model to generate corresponding text unit embedding vectors. S123. The text unit embedding vectors are processed using dimensionality reduction and clustering algorithms to achieve semantic grouping, form a structured knowledge organization, and generate a coarse-grained knowledge base.

3. The intelligent generation method for biomimetic design schemes integrating a dual-granularity knowledge base according to claim 2, characterized in that, Step S2 specifically includes: S21. Receive the input biological design problem, and perform parallel retrieval in the fine-grained knowledge base and the coarse-grained knowledge base according to the biological design problem to obtain an initial relevance score; S22. According to the preset weight allocation mechanism, the retrieval results of the fine-grained knowledge base and the coarse-grained knowledge base are weighted and fused to calculate the fusion score; The weight allocation mechanism uses a weight function. for: ； in, Indicates Macro-level entries The number of elements in the associated biological strategy set, For hyperparameters, defined as the contribution of macro-level knowledge blocks, if a macro-level knowledge block is composed of multiple micro-level sub-knowledge blocks, then its weight value is evenly distributed to the biological strategies associated with each sub-knowledge block that constitutes the macro-level knowledge block. The fusion score is calculated using the following formula: ； in, Indicate candidate biological strategies, Represent each candidate biological strategy The fusion score; S23. Based on the preset relevance threshold, select samples with a fusion score not lower than the relevance threshold. Biological strategies are used to obtain a set of candidate biological strategies. ; The candidate biological strategy set As shown in the following formula: 。 4. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base according to claim 2, characterized in that, Step S3 specifically includes: S31. Using the preset first language model, generate an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem; S32. Using multiple preset second-largest language models as review experts, the initial biomimetic design scheme is collaboratively reviewed from multiple preset evaluation dimensions, and a multi-round consultation mechanism is adopted to obtain review opinions. Let the first The i-th review expert in turn The solution selection result is Update the selection results using the following formula: ； in, This represents the rational adjustment function by which the expert made their choice after considering the opinions of two other experts. for The selection results of the 1st, 2nd, and 3rd experts in each round; If the solutions chosen by all review experts are in Consistency in the wheel, that is If the scheme is determined to be the Best Bio-Inspired Design (B-BID), no further discussion is needed. S33. Using the third preset large language model as the authoritative expert E*, when there are disagreements in the review opinions of the experts, a final decision is made on the disagreements, and the initial biomimetic design scheme is iteratively optimized using the following formula to obtain the final biomimetic design scheme: ； in, For the first The review experts held the final discussion round t max The optimal solution for the wheel is i=1,2,3; B-BID is the best biomimetic design solution.

5. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base according to claim 2, characterized in that, The preset text block length in step S111 is 200 tokens; The preset first semantic embedding model is the BGE-M3 embedding model.

6. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base according to claim 2, characterized in that, The dimensionality reduction algorithm in step S123 is the UMAP algorithm, which optimizes the adjacency probability between samples in high-dimensional space. Probability corresponding to low-dimensional space Cross-entropy loss between for: ； The clustering algorithm in step S123 is the HDBSCAN algorithm; the clustering quality is optimized by calculating the silhouette coefficient, which is specifically defined as follows: ； in, For a non-noise sample index set, , Representative sample The average distance to other samples within its cluster. This represents the average distance to its nearest neighboring cluster sample. The average silhouette coefficient is a metric used to measure the overall quality of clustering.

7. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base according to claim 3, characterized in that, The preset weight allocation mechanism includes: assigning a first preset weight to the retrieval results of the fine-grained knowledge base, and assigning a second preset weight to the retrieval results of the coarse-grained knowledge base, wherein the value of the second preset weight is greater than the value of the first preset weight; the value of the second preset weight is twice the value of the first preset weight.

8. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base according to claim 2, characterized in that, In step S32, multiple preset evaluation dimensions include: functionality and sustainability, function-based analogy matching, mechanism abstraction and cross-domain transformation, coherence, and relevance; the multi-round consultation mechanism is the Delphi method, and the Seminar function is used to show the opinions of other review experts to each review expert so that they can adjust their own choices.

9. The intelligent generation method for biomimetic design schemes based on a dual-granularity knowledge base according to claim 1, characterized in that, This method also includes data preprocessing; The data preprocessing includes: obtaining raw data from a professional biomimetic database and performing quality control and standardization on the raw data; The method also includes fine-tuning the large language model used for retrieval; The fine-tuning employs the LoRA method shown in the following formula: ； in, This is the original weight matrix of the large language model; The equivalent weights of the large language model after incorporating the LoRA method; For the incremental update of the LoRA method, i.e., a low-rank matrix of rank r; A and B are two small matrices derived from the low-rank decomposition.

10. A biomimetic design scheme intelligent generation system integrating a dual-granularity knowledge base, characterized in that, The system includes: A dual-granularity knowledge base construction module is used to construct a dual-granularity knowledge base, which includes a fine-grained knowledge base and a coarse-grained knowledge base; The dual-granularity fusion retrieval module is used to perform fusion retrieval in the dual-granularity knowledge base based on the input biological design problem, and filter to obtain a set of candidate biological strategies; The multi-language model expert consultation, evaluation and optimization module is used to generate an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem, and to use multiple large language models for collaborative review and iterative optimization to generate the final biomimetic design scheme. The construction of the dual-granularity knowledge base specifically includes: The biological strategy text is segmented into multiple text blocks according to a preset text block length. A preset first semantic embedding model is used to vectorize the multiple text blocks, generating corresponding text block embedding vectors, which are stored in a vector database to construct the fine-grained knowledge base. The overview information of the biological strategy text is extracted as independent text units. The preset first semantic embedding model is used to vectorize the independent text units, generating corresponding text unit embedding vectors. Dimensionality reduction and clustering algorithms are used to process the text unit embedding vectors to achieve semantic grouping, thereby constructing the coarse-grained knowledge base. The step involves performing a fusion search on the dual-granularity knowledge base based on the input biological design problem to obtain a set of candidate biological strategies, specifically including: Parallel retrieval is performed in the fine-grained knowledge base and the coarse-grained knowledge base to obtain an initial relevance score; the retrieval results are weighted and fused according to a preset weight allocation mechanism to calculate a fusion score; biological strategies with fusion scores not lower than the preset relevance threshold are selected according to a preset relevance threshold to obtain the candidate biological strategy set. The process of generating an initial biomimetic design scheme based on the candidate biological strategy set and the biological design problem, and then using multiple large language models for collaborative review and iterative optimization to generate the final biomimetic design scheme, specifically includes: An initial biomimetic design scheme is generated using a first preset language model; multiple second preset language models are used as review experts to conduct a collaborative review of the initial biomimetic design scheme; a third preset language model is used as an authoritative expert to make a final decision and iterative optimization when there are disagreements in the review opinions of the review experts, thereby obtaining the final biomimetic design scheme. The dual-granularity knowledge base construction module is also configured with a vector database for storing the fine-grained knowledge base and a structured database for storing the coarse-grained knowledge base; The dual-granularity fusion retrieval module is equipped with a retrieval engine finely tuned using the LoRA method; The multi-language model expert consultation, evaluation and optimization module includes a preset first language model, a preset second language model and a preset third language model deployed on the server, as well as a feedback mechanism for processing expert review opinions.

Images