Polymer composite material design method based on LLM knowledge database construction and reasoning optimization

Abstract

The invention belongs to the technical field of composite material design, and discloses a polymer composite material design method based on LLM knowledge database construction and reasoning optimization, and the method comprises the following steps: S1, target performance setting; s2, constructing a knowledge database; s3, constructing a reverse design model; s4, constructing a performance verification module; and S5, performing multi-round test-verification design loop optimization. According to the method, an intelligent design method integrating target setting, semantic knowledge database construction, a design assistant, a performance verifier, a prompt engineering module and a language model material knowledge system is provided, closed-loop reasoning and feedback of structure-mechanism-performance are achieved through multiple rounds of DA-VM collaborative optimization, and the design efficiency is improved. The method has the advantages of high design efficiency, strong structure generation rationality, accurate performance prediction, knowledge evolvable and the like, and is suitable for intelligent rapid development of the high thermal conductivity-wave absorbing-shielding-mechanical integrated functional polymer composite material.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of polymer composite material design and artificial intelligence optimization, in particular to a polymer composite material multi-objective design method based on LLM knowledge database construction and reasoning optimization, such as thermal-conductive-electromagnetic shielding-wave-absorbing-mechanical integrated polymer composite material design, which belongs to the cross technology of large language model knowledge database construction, intelligent material design, prompt engineering, and structure-function coupling optimization. BACKGROUND

[0002] With the rapid development of 5G communication, aerospace, intelligent terminal, flexible wearable and other technologies, the device miniaturization and integration degree are continuously improved, and the multi-physical field coupling problems such as thermal management, electromagnetic interference (EMI) shielding, microwave absorption, and mechanics are becoming increasingly serious. Although traditional polymer-based materials have the advantages of light weight, good flexibility, and excellent processability, their intrinsic thermal conductivity is low, the electrical conductivity is poor, and they lack a magnetic loss mechanism, making it difficult to meet the practical application requirements of integrated multi-functional materials such as "thermal-conductive-electromagnetic shielding-wave-absorbing".

[0003] To solve the above problems, researchers generally use functional fillers to enhance the performance of the polymer matrix, including the following main means: Thermal conduction: introducing high thermal conductivity fillers such as graphene, carbon nanotubes (CNT), boron nitride, etc., to build a continuous thermal channel network; Electromagnetic shielding: introducing conductive components such as MXene, carbon black, silver nanowires, carbon fibers, etc., to form a reflection-absorption-multiple scattering shielding mechanism; Wave absorption: adding magnetic nanoparticles (such as Fe3O4, Ni, Co), interface polarization sources (such as N-doped carbon, PDA, PPy) to build a composite loss path.

[0004] However, the above methods have the following core problems: Design process relies on experience, variables are complex and opaque: Most current research is based on the experience of "filler-matrix-structure-performance" construction, with many design parameters and mutual coupling, lacking a scientific modeling optimization system, low test efficiency and poor reproducibility.

[0005] The structure-performance mapping relationship is not clear: The contribution and weight of filler type, interface modification, and thermal / electromagnetic channel configuration to multi-objective performance are difficult to quantify systematically, and there is a lack of structure-mechanism-performance trinity feedback mechanism.

[0006] Multi-objective performance is mutually restrictive: improving thermal conductivity usually relies on high filler content and high density, while excellent wave absorption performance requires the introduction of loss channels and porous structures, and there is a reflection-absorption conflict between shielding and wave absorption, making the design trade-off difficult.

[0007] Data deficiency, traditional machine learning is difficult to generalize: the experimental data of complex multifunctional composites are scarce, the characteristic variables are highly unstructured, the traditional supervised learning model is difficult to train, the generalization ability is weak, and it is difficult to cope with the actual design scene.

[0008] Therefore, there is an urgent need for a new intelligent design method with prior knowledge guidance capability, which can process natural language material knowledge, support structure generation and performance prediction closed-loop feedback, and break through the bottleneck of traditional trial-and-error method in multi-objective material development.

[0009] In recent years, large language models (LLM) such as GPT series have shown strong capabilities in natural language understanding, knowledge reasoning, and complex task planning. Combined with Prompt Engineering (Prompt Engineering), Few-shot Learning (Few-shot Learning), Chain-of-Thought (Chain-of-Thought), etc. Mechanisms make the model can infer new structure schemes from natural language knowledge in the absence of large data support, while having strong logical consistency and semantic control ability. Existing research shows that large language models have shown structure generation and performance trend capture capabilities in building material (such as concrete) structure optimization, green formula design, etc. However, there is no specific method to apply such language-driven design systems to complex polymer-based heat-conducting-shielding-absorbing composite material design systems, and the existing framework of the prompt engineering mechanism (such as knowledge chain prompt CoK, automatic optimization APE, Tree-of-Thought, etc.) It is also difficult to combine the multi-scale structure mechanism model in the field of materials science to realize specific functions.

[0010] Therefore, constructing a multi-objective composite material design framework based on language model and knowledge-driven can realize the closed-loop control of target-generation-verification-feedback optimization, which has important theoretical innovation value and engineering application prospect. SUMMARY

[0011] The present application aims to solve the problems of complex multi-objective performance design process, variable redundancy, lack of structure-performance logic prediction and optimization capability in the prior art, and proposes a polymer composite material design method based on LLM knowledge database construction and reasoning optimization. An intelligent design method based on language model and prompt engineering is used to construct a multi-objective composite material design framework based on language model and knowledge-driven, which can realize the closed-loop control of target-generation-verification-feedback optimization, and can optimize the thermal conductivity, electromagnetic shielding efficiency and wave absorption performance of polymer composites under the premise of low sample, significantly improving the design efficiency and prediction accuracy.

[0012] To achieve the above purpose, the present application provides the following technical solutions: A polymer composite design method based on LLM knowledge database construction and reasoning optimization, suitable for formula generation and screening of heat-conducting-electromagnetic shielding-wave-absorbing-mechanical integrated materials, comprising the following steps: S1, target performance setting: set multiple target performance indicators including thermal conductivity, electromagnetic shielding efficiency, minimum reflection loss, effective absorption bandwidth, high heat resistance and strong mechanical performance, and express them in natural language form in prompt; S2, knowledge database construction: the knowledge database construction aims to create an embedded natural language material knowledge database as the cognitive basis of the system. The database is constructed by the following process: input professional literature, experimental data and patent description into large language model (LLM), use the semantic extraction ability of LLM to automatically extract "structure-mechanism-performance" related semantic fragments and causal chains, and store them in vector embedding form for prompting engineering to construct prompt context, generate structure design suggestions or verify target performance. The natural language knowledge database does not require traditional artificial labeling and structured format, has the advantages of fast construction, dynamic evolution, high semantic support, etc., and is the cognitive basis of supporting language model driven material design system. Its core role is: through the Retriever module, the relevant knowledge fragments are matched and injected into the prompt for the user's performance target; support for the design assistant (DA) to generate structure, ratio and interface strategy with literature basis; provide logical basis for the verifier (VM) to judge the scientificity and feasibility of the candidate scheme. In addition, the system can dynamically induce potential knowledge in high-performance schemes during the multi-round optimization process of DA-VM, feedback and supplement to the database, and realize self-evolution.

[0013] S3, reverse design model construction (DA): based on large language model, load the natural language knowledge base obtained in step S2 and the performance target set in step S1, generate candidate formula scheme through prompt engineering, including polymer matrix type, filler type and proportion, interface modification method and structure construction strategy; S4, performance verification module construction (Verifier, VM): based on the forward prediction ability of physical model, machine learning predictor or language model, build Verifier for scoring and sorting of candidate schemes in terms of thermal conductivity, electromagnetic shielding, wave absorption, mechanics and heat resistance.

[0014] S5, multi-round test-verification design cycle (TVDL) optimization: multi-round iteration optimization between reverse design module (DA) and performance verification module (VM), each round including: DA generates candidate design according to current target and knowledge; VM performs target performance prediction and multi-target scoring on all candidates; the top 10% of the scoring scheme is input into the next round of prompt as a high-performance solution. After at least 3 rounds of optimization, output the final candidate scheme.

[0015] The present application has the following beneficial effects: 1. The present application adopts a systematic optimization design path, constructs a multi-objective composite material design framework based on a language model and knowledge driving, and can realize closed-loop control from target generation, verification and feedback optimization. Compared with the traditional trial and error method, the present method effectively converges the multi-objective optimization process through language reasoning and knowledge embedding; 2. The prompt engineering of the present application enhances understanding ability, and through embedding CoT, ToT, CoK and other mechanisms, the large language model can output structure-performance causal chain instead of static templates; 3. The performance prediction of the present application is verifiable: through the construction of few-shot Verifier, rapid performance screening and scoring are realized, and the number of physical experiments is reduced; 4. The structure mechanism of the present application has high physical consistency: the generated multi-scale heterogeneous structure is highly consistent with the known material mechanism, and has good manufacturability and interpretability; 5. The present application is suitable for complex material systems: it can be extended to multiple thermal-electric-magnetic composite scenarios (such as 5G materials, electromagnetic interference protection, flexible device packaging, etc.); 6. The present application can be embedded in a material large model platform: the system can be part of a material science large model platform, and has universality and expandability.

[0016] The above is a summary of the technical solutions of the present application. The present application will be further described in conjunction with the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The main logic flowchart of the method designed for the embodiments of the present application is shown in the figure; Figure 2 The SEM picture of the C-PDA@PI@CNT carbon fiber prepared in the embodiments of the present application is shown in the figure; Figure 3 The enlarged SEM picture of the C-PDA@PI@CNT carbon fiber prepared in the embodiments of the present application is shown in the figure; Figure 4 The SEM picture of the C-PDA@PI@CNT@G prepared in the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0018] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined purpose, the specific embodiments of the present application are described in detail below in conjunction with the accompanying drawings, preferred embodiments and specific embodiments. Figure 1

[0019] Basic Embodiments Referring to the accompanying drawings, Figure 1 ​The polymer composite design method based on LLM knowledge database construction and reasoning optimization provided by the embodiment is suitable for formula generation and screening of heat-conducting-electromagnetic shielding-absorbing-mechanical integrated materials, and includes the following steps: S1, target performance setting: set multiple target performance indexes including thermal conductivity, electromagnetic shielding efficiency, minimum reflection loss, effective absorption bandwidth, high heat resistance, and strong mechanical properties, and express in natural language form in the prompt; Among them, the target performance at least includes the following constraints: the thermal conductivity is not less than 1.2 W / mK, the electromagnetic shielding efficiency is not less than 60 dB, the minimum reflection loss is less than -50 dB, the effective absorption bandwidth is not less than 4 GHz, and the matching mechanical and heat resistance performance is appropriate.

[0020] S2, knowledge database construction: the knowledge database construction aims to create an embedded natural language material knowledge database as the cognitive basis of the system. The database is constructed by the following process: input professional literature, experimental data and patent description to large language model (LLM), use the semantic extraction ability of LLM to automatically extract the semantic fragments and causal chains related to "structure-mechanism-performance", and store them in the form of vector embedding for prompting the construction of prompt context, generating structure design suggestions or verifying target performance. The natural language knowledge database does not need traditional artificial labels and structured format, has the advantages of fast construction, dynamic evolution, high semantic support degree and the like, and is the cognitive basis supporting the language model driven material design system. Its core role is: through the Retriever module, the relevant knowledge fragments are matched and injected into the prompt for the user's performance target; support is provided for the design assistant (DA) to generate structure, ratio and interface strategy with literature basis; the verifier (VM) provides logical basis for judging the scientificity and feasibility of the candidate scheme. In addition, in the multi-round optimization process of DA-VM, the system can dynamically induce the potential knowledge in the high-performance scheme, feed back and supplement to the database, and realize self-evolution.

[0021] The design knowledge database is automatically constructed by a language model based on inputted professional literature data, wherein the knowledge data input context is derived from pre-processed materials science literature, patent data, and experimental reports (at least ten thousand relevant contents), including but not limited to two-dimensional thermal conduction network, three-dimensional carbon heterostructure, electromagnetic shielding mechanism, wave absorption loss mechanism, high heat resistance, mechanical properties, etc. The database takes semantic fragments as the basic unit, records knowledge including but not limited to the following types: material structure characteristics and processing technology (such as two-dimensional graphene-three-dimensional carbon nanotube configuration, carbonization path, etc.); energy mechanism attribution (such as the influence of interface thermal resistance on thermal conduction, the contribution of magnetic filler to wave absorption); multi-objective conflict reconciliation strategy (such as the synergistic design logic of wave absorption-shielding-thermal conduction); specific structure-performance causal rules (such as "the improvement of graphitization degree can enhance SE but may reduce RL"). The entries are expressed in the form of natural language semantic fragments or structure-mechanism-performance chains and stored in the database in the form of vectors, supporting vectorized retrieval, prompt injection, and feedback induction update as a semantic support module for generating candidate structure schemes and performance evaluation.

[0022] S3, reverse design model construction (DA): based on a large language model, load the natural language knowledge base obtained in step S2 and the performance target set in step S1, and generate candidate formula schemes by prompting engineering, including polymer matrix type, filler type and proportion, interface modification method and structure construction strategy; In the candidate scheme output by the reverse design model DA in step S3, the filler content is not more than 10 wt%, and the filler includes but is not limited to carbon fiber, graphene, MXene, carbon nanotube, Fe3O4, silicon carbide, expanded graphite or a composite structure thereof.

[0023] S4, performance verification module construction (Verifier, VM): based on the forward prediction ability of a physical model, a machine learning predictor or a language model, a Verifier is constructed for scoring and ranking the thermal conductivity, electromagnetic shielding, wave absorption, mechanical properties, and heat resistance of the candidate scheme.

[0024] The Verifier in step S4 is constructed based on at least one of the following models: Gaussian process regression (GPR), random forest (RF), Transformer structure language model, or integrated regressor thereof.

[0025] S5, multi-round test-verification design cycle (TVDL) optimization: multiple iterations are performed between the reverse design module (DA) and the performance verification module (VM), each round including: DA generates candidate designs according to the current target and knowledge; VM performs target performance prediction and multi-objective scoring on all candidates; the top 10% of the scoring scheme is input into the next round of prompt as a high-performance solution. After at least 3 rounds of optimization, the final candidate scheme is output.

[0026] The TVDL design loop is at least three rounds, each round including DA-VM-optimization feedback three stages; in each round, the DA inputs the optimal sample of the last round VM score as a new prompt context to improve the generation quality; the language interaction between the DA and the VM is based on the prompt engineering technology, which includes one or more of the following strategies: a. Chain-of-Thought (CoT): used to guide the causal reasoning chain between structure-process-performance of large language models, thereby enhancing the physical logic consistency of candidate designs; b. Tree-of-Thought (ToT): used to build a multi-path reasoning structure in the material formulation space, and systematically filter performance solutions through search and scoring mechanisms; c. Chain-of-Knowledge (CoK): connecting design tasks with embedded knowledge bases to guide language models to dynamically retrieve the most relevant material mechanisms, empirical laws, or verified literature facts for the current goal; d. Self-Consistency Prompting: in each round of candidate generation, multiple samples are output, and stable and consistent outputs are selected through clustering and structure matching; e. Active Prompting: based on target performance confidence or model uncertainty, preferentially select few-shot instances with the highest information gain to join the prompt; f. Automatic Prompt Engineer (APE): used to automatically construct or optimize prompt structure combinations, automatically construct or adjust the most suitable prompt structure combination for the current task through scoring feedback mechanisms, greedy search or strategy optimization algorithms.

[0027] The Tree-of-Thought prompt generates material design paths through depth-first search or breadth-first search strategies, and each "thought node" represents a candidate design. After multiple rounds of verification, the optimal performance branch is retained, and the performance degradation path is backtracked and pruned. The Self-Consistency prompting strategy requires at least three sets of formulation recommendations to be output during each generation process. Through structure similarity analysis and performance prediction aggregation, the most consistent solution is selected to enter the next round of feedback. The automatic prompt construction module is based on reinforcement learning, genetic algorithm or multi-armed bandit mechanism, and automatically optimizes the prompt structure and input segment combination based on the historical prompt effect score to improve the controllability and convergence efficiency of the DA module generation results.

[0028] The polymer composite comprises a multi-scale heterogeneous carbon structure network, and a structure design logic of the network is automatically generated by the DA module according to the following knowledge rules, specifically including: providing a three-dimensional porous scaffold by using a polyimide (PI) fiber felt to endow the composite with an initial macroscopic continuous path; constructing a high-polarity interface on the PI surface by introducing a dopamine (PDA) modification layer to form adsorption sites to promote the combination of two-dimensional fillers and metal ions, and meanwhile, the PDA modification layer is converted into N-doped carbon in subsequent carbonization to enhance dielectric loss; loading a graphene sheet layer to construct a two-dimensional thermal electron bridge to enhance a horizontal heat channel and an electromagnetic shielding path; realizing self-positioning of a catalytic site distribution by using Co²⁺ coordination with PDA, and triggering a Co reduction-CNT in-situ growth reaction in a co-carbonization process to form a one-dimensional vertical structure; and finally constructing CF-CNT, G-CNT, N-G-CNT and other heterogeneous interfaces to realize a multi-scale thermal-electric-magnetic synergistic network structure.

[0029] The structure design is generated by a language model through a prompt engineering mechanism, and the Prompt context comprises the following guiding strategies: explicitly combining target performance (such as “thermal conductivity needs to be > 1.25 W / m·K, EMI shielding > 65 dB, and RL < –60 dB”); providing a few-shot instance, such as “after adding PDA modification in the PI matrix, graphene can stably adsorb, and then Co²⁺ is introduced to in-situ grow CNT to improve the thermal conductivity and shielding performance”; embedding structure-performance knowledge rules, such as “CNT forms a vertical bridging structure between CF and G, which helps to form a 3D thermal conduction network and a microwave loss superposition mechanism”; guiding the model to generate a structure component-process path-structure morphology trinity; and reserving the structure with the best performance prediction output after multiple rounds of feedback by the VM verification module.

[0030] The performance prediction is realized by the Verifier module, and specifically comprises the following structure-performance mapping mechanism: if there is a CF-CNT-G bridging channel, the continuity of the system thermal conduction path is enhanced, and the thermal conductivity is expected to be improved by > 1 W / m·K; if the filler interface forms a coexisting structure of N-doped + Co particles + graphene, the dielectric polarization, magnetic loss and conductive shielding coupling are enhanced, the RL value is expected to decrease by 10-15 dB, and the shielding is expected to increase by 5-10 dB; the VM module introduces existing experimental performance samples in a few-shot supervised manner to perform regression evaluation, outputs the score of the candidate structure, and inputs the score as a TVDL feedback optimization iteration input.

[0031] Embodiment 1 The embodiment provided by the present application provides a polymer composite material design method based on LLM knowledge database construction and reasoning optimization. The embodiment is a specific application based on the basic embodiment and is based on the design and verification of knowledge-driven optimization of multi-scale heterogeneous carbon structure composites.

[0032] The embodiment is used to implement the design method and verify the feasibility of the design method in the development of actual polymer-based thermal-conductive-electromagnetic shielding-wave-absorbing integrated composites. The embodiment shows a language model driven formula generation process, a structure construction step, a performance verification method, and a structure-performance coupling logic.

[0033] Referring to the logic flowchart shown in FIG. 1, the design method specifically includes the following contents: Figure 1 I. Goal setting and prompt engineering construction The following multi-target performance indicators are set and expressed in natural language form in the Prompt: a. Thermal conductivity ≥ 1.25 W / m·K; b. Electromagnetic shielding efficiency ≥ 65 dB; c. Minimum reflection loss RLmin ≤ -60 dB; d. Absorption bandwidth EAB ≥ 5 GHz; e. The total content of fillers is not more than 10 wt%.

[0034] f. Consider high heat resistance (thermal decomposition temperature ≥ 500℃) and mechanical strength (tensile strength ≥ 50 MPa).

[0035] II. Knowledge database construction An embedded natural language material knowledge database is constructed, and the specific process is as follows: 1. Data source: Collect and preprocess at least ten thousand pieces of material science literature, patents and experimental reports, covering two-dimensional thermal conduction network, three-dimensional carbon heterogeneous structure, electromagnetic shielding mechanism, wave absorption loss mechanism, heat resistance and mechanical performance, etc.

[0036] 2. Knowledge extraction: Use large language models such as GPT-4 to perform semantic extraction on the corpus, automatically extract semantic fragments and causal chains related to “structure-mechanism-performance”, for example: “PDA modification can enhance Co²⁺ adsorption and guide in-situ growth of CNT”; “Graphene and CNT cooperatively construct three-dimensional thermal conduction network to improve TC and SE”; “N-doped carbon structure enhances dielectric polarization and improves wave absorption performance”.

[0037] ​3. Vectorized storage: Convert the extracted semantic fragments into vector embeddings and store them in a FAISS or similar vector database to support semantic retrieval and dynamic updates.

[0038] 4. Functional module integration: The database injects relevant knowledge into the Prompt through the Retriever module, supports the generation of structural suggestions for DA, and provides logical basis for VM.

[0039] III. Reverse design model (DA) construction and operation Based on GPT-4, the DA module is constructed, and the above knowledge base and performance targets are loaded. Through the following prompt engineering strategies, candidate solutions are generated: 1. Chain-of-Thought (CoT): Guide the model to output the "structure-process-performance" causal chain; 2. Knowledge Chain (CoK): Dynamically retrieve relevant rules from the knowledge base, such as "CNT bridging CF and G can enhance three-dimensional heat conduction"; 3. Self-Consistency: Generate 3 sets of solutions in each round, and select the most consistent one to enter the next round; 4. Automatic Prompt Engineering (APE): Automatically optimize the structure combination of the Prompt.

[0040] The TOP1 candidate solution output by DA is as follows: a. Matrix: Polyimide (PI) non-woven felt; b. Modifier: Dopamine (PDA); c. Two-dimensional filler: Graphene nanosheet (G); d. Metal ion: Co(NO3)2 (Co²⁺); e. Auxiliary carbon source: Melamine; f. Structure configuration: PI@PDA / G / Co²⁺ → co-carbonization → G–CNT–CF heterogeneous network structure IV. Performance verification module (VM) construction and operation (corresponding to claims 1, 5, and 10) VM is based on Few-shot supervised regression model (GPR + Transformer integration) to construct, and perform multi-objective performance prediction and scoring on candidate solutions: Input: Semantic description of candidate structure; Output: Prediction values of TC, SE, RLmin, EAB, and comprehensive score F; Scoring function: F = w1·TC + w2·SE − w3·|RLmin| + w4·EAB; The top 10% of the solutions in each round are fed back to the next round of Prompt.

[0041] The results show that the prediction and the actual measurement are in good agreement, proving that the Verifier module is feasible and the KDD optimization process is effective.

[0042] Five, multiple rounds of test-verification design loop (TVDL) optimization Three rounds of DA-VM iterative optimization are carried out: 1. First round: DA generates 10 groups of candidate solutions, and VM takes TOP1 after scoring (F=0.82) for feedback; 2. Second round: DA generates new solutions based on feedback, and VM takes TOP1 after scoring (F=0.91); 3. Third round: further optimize the CNT proportion, VM takes TOP1 after scoring (F=0.94), and output the final solution.

[0043] Six, experimental preparation and performance test Preparation process: 1. PI felt is self-polymerized with dopamine (2 mg / mL) in Tris-HCl buffer (pH 8.5) for 6 h to obtain PDA@PI; 2. Soak in graphene ethanol dispersion solution (3 mg / mL) for 30 min to obtain G@PDA@PI; 3. Soak in 0.1 mol / L Co(NO3)2 solution for 12 h to obtain G@PDA@PI@Co²⁺; 4. Mix with melamine 1:10, carbonize at 900°C for 2 h under N2 atmosphere to obtain CPPGCo composite material.

[0044] The prepared composite material is tested for performance, and the results are as shown in Table 1.

[0045] Table 1

[0046] Seven, structure-performance mechanism analysis CPPGCo has the following multi-scale structural characteristics through SEM, TEM and other characterization methods: a. Macro: PI-CF forms a continuous heat conduction backbone; b. Mesoscopic: Graphene sheets construct horizontal heat conduction and shielding paths; c. Micro: CNTs vertically bridge between CF and G, forming a three-dimensional heterogeneous network; d. Interface: PDA-derived N-doped carbon enhances dielectric polarization, and Co particles provide magnetic loss center.

[0047] The structure design is completely generated by the DA module under the knowledge driving, and has high physical consistency and manufacturability through VM prediction and experimental verification.

[0048] Figure 1 The main logic flow diagram of the embodiment of the application.

[0049] As Figure 1 shown, the overall logic flow of the method of the application follows a closed loop of "goal setting → knowledge construction → reverse design → performance verification → cycle optimization". This flow ensures that the material design process is driven by knowledge, and continuously approaches the optimal solution through multiple iterations. Specifically: after the multi-objective performance requirements set by the user (such as thermal conductivity ≥ 1.25 W / m·K, shielding ≥ 65 dB) are input into the system, the knowledge database (KDD) is guided to perform relevant semantic retrieval; the design assistant (DA) generates candidate formulations (such as "PI@PDA / G / Co²⁺") under the guidance of prompting engineering strategies (CoT, ToT, CoK, etc.); the performance verifier (VM) performs rapid prediction and scoring on the candidate scheme; the scheme with high score is fed back to the next round of optimization as new knowledge, until the final scheme is output. This flowchart intuitively shows how the application organically combines natural language understanding, knowledge reasoning, and multi-objective optimization, and fundamentally solves the problems of low efficiency and complex variable coupling of traditional trial-and-error methods.

[0050] Figure 2 SEM image of C-PDA@PI@CNT carbon fiber prepared in Example 1 of the application.

[0051] Figure 2 The SEM image clearly shows the macroscopic carbon fiber network skeleton formed after co-carbonization, corresponding to the design strategy of "using polyimide (PI) fiber felt to provide a three-dimensional porous scaffold" in Example 1. In the figure, the fibers are interwoven, forming a continuous porous channel structure. This macroscopic continuous structure is the primary basis for achieving high thermal conductivity and high electromagnetic shielding, as it provides the main channel for heat flow and current, directly meeting the design goals of high thermal conductivity (≥ 1.25 W / m·K) and high electromagnetic shielding efficiency (≥ 65 dB). Without this continuous structure, the performance will be as shown in Example 3, S1 sample, which is overall low.

[0052] Figure 3 SEM image of C-PDA@PI@CNT carbon fiber prepared in Example 1 of the application.

[0053] Figure 3 At high magnification, it shows Figure 2The microstructure of the surface of a single fiber. A large number of carbon nanotubes (CNTs) catalytically grown in situ by Co²⁺ can be observed growing vertically from the fiber surface, forming a dense plush structure. This directly verifies the design logic generated by the DA module according to the knowledge rule "Co²⁺ coordinates with PDA to realize self-positioning of catalytic sites and trigger in-situ growth of CNTs". These CNTs greatly increase the specific surface area of the fiber, laying the foundation for the construction of heterogeneous interfaces, enhanced dielectric polarization, and multiple scattering, which is the key to achieving excellent wave absorption performance (RLmin ≤ -60 dB). At the same time, CNTs form nano-bridges between adjacent fibers or fillers, enhancing the continuity of the thermal and electrical networks, explaining why high performance can be achieved at low filler content.

[0054] Figure 4 SEM image of C-PDA@PI@CNT@G prepared for Example 1 of the present application.

[0055] Figure 4 The composite structure after the introduction of graphene is shown. In the figure, it can be seen that two-dimensional graphene sheets (G) are successfully attached to the CNT-modified carbon fiber skeleton, forming a "fiber-nanotube-sheet" multi-level heterogeneous structure. This perfectly achieves the goals of the DA module's design of "loading graphene sheets to construct two-dimensional thermal electron bridges, enhancing lateral heat channels and electromagnetic shielding paths" and "ultimately constructing a CF-CNT-G heterogeneous interface". The introduction of graphene sheets significantly enhances the in-plane thermal and electrical conductivity, combined with the vertical pathways constructed by CNTs, forming a robust 3D collaborative network. This microstructure is direct evidence of the synergistic optimization of thermal conductivity, shielding, and wave absorption performance, and its final measured performance (TC=1.25 W / m·K, SE=67.8dB, RLmin=-62.75 dB) fully meets and exceeds the preset design requirements, proving the effectiveness and accuracy of the method from design to preparation.

[0056] Example 2 The design method of the polymer composite material based on the LLM knowledge database construction and reasoning optimization provided by the present embodiment is based on the basic embodiment and Example 1, and further verifies the design method of the polymer composite material based on different prompt engineering strategies.

[0057] This embodiment is used to specifically illustrate the actual application and effect comparison of different prompt engineering strategies in the design method, and fully embodies the execution process of the five steps in claim 1.

[0058] I. Target performance setting The multi-target performance indicators are set as follows, and are expressed in natural language form in the system Prompt: Thermal conductivity (TC) ≥ 1.20 W / m·K Electromagnetic shielding efficiency (SE) ≥ 65 dB Minimum reflection loss (RLmin) ≤ -55 dB Effective absorption bandwidth (EAB) ≥ 5 GHz Total filler content ≤ 8 wt% Balancing mechanical strength (tensile strength ≥ 45 MPa) and heat resistance (heat distortion temperature ≥ 200°C) II. Knowledge database construction 1. Data input: The system calls the pre-constructed embedded natural language material knowledge database. This database is based on more than 10,000 pieces of material science literature, patents, and experimental reports, covering two-dimensional / three-dimensional thermal conduction networks, electromagnetic shielding mechanisms, wave absorption loss mechanisms, and interface modification strategies.

[0059] 2. Knowledge extraction and storage: Through the LLM semantic extraction capability, the "structure-mechanism-performance" related semantic fragments (such as "PDA modification layer can enhance Co²⁺ adsorption and guide in-situ growth of CNTs", "MXene and CNT complex can synergistically enhance the conductive network and multi-level scattering") are automatically extracted and stored in the form of vector embedding.

[0060] 3. Function realization: During the running process of this embodiment, the Retriever module dynamically retrieves matching knowledge fragments from the database according to the above performance targets and injects them into the subsequent Prompt, providing semantic support for DA generation and VM verification.

[0061] III. Reverse design model (DA) construction and operation Based on the large language model (GPT-4), the DA module is constructed, and the knowledge base obtained in step two and the performance targets set in step one are loaded. Through the following three prompt engineering strategies, candidate formula schemes (including polymer matrix type, filler type and proportion, interface modification method, and construction process) are generated: 1. Strategy A (baseline): The Prompt only contains performance targets and variable space, without additional reasoning guidance.

[0062] 2. Strategy B (CoT+CoK): Combined chain of thought (CoT) and knowledge chain (CoK) prompts. The Prompt embeds 3 rules (such as "magnetic fillers improve wave absorption but may reduce thermal conductivity") and 1 Few-shot reasoning example retrieved from the knowledge base, guiding the DA to output causal chains.

[0063] 3. Strategy C (ToT + CoK + Self-Consistency): Combine tree thinking (ToT), knowledge chain (CoK), and self-consistency prompts. The Prompt builds a three-layer design tree (filler combination - interface regulation - carbonization process), and injects relevant knowledge at each node through CoK; generate 3 groups of schemes for each path, and retain the optimal solution after consistency screening.

[0064] The core elements of the DA output candidate scheme include: Strategy A: PI + 8% CNT + KH550, hot pressing Strategy B: PI + 6% G / CNT (3:1) + PDA, melamine-assisted co-carbonization Strategy C: PI + 5% MXene@Fe3O4 core-shell structure + PDA + trace CNT, layer-by-layer assembly and co-carbonization.

[0065] Four, performance verification module (VM) construction and operation Based on the integration of Gaussian process regression (GPR) and random forest (RF), the Few-shot supervised model builds Verifier (VM) to predict and score the thermal conductivity, electromagnetic shielding, wave absorption, mechanical and heat resistance performance of all candidate schemes generated by DA. The scoring function is: F = 0.3*TC_norm + 0.3*SE_norm + 0.2*|RLmin|_norm + 0.2*EAB_norm.

[0066] Five, multi-round test-verification design cycle (TVDL) optimization For each prompt strategy, three rounds of DA-VM iterative optimization are performed: First round: DA generates 10 groups of candidate schemes, and VM predicts the score after scoring, and takes the top 10% (i.e. the highest F value scheme) to feedback to the next round of Prompt.

[0067] Second and third rounds: DA generates new candidates based on the context of high-performance schemes from the previous round of feedback; VM predicts the score again and feeds back.

[0068] After three rounds, output the final optimal scheme of each strategy.

[0069] Six, results and discussion After three rounds of TVDL optimization, the performance prediction results of the final schemes of each strategy are as follows in Table 2. Note: In Table 2, the F value is a multi-objective weighted function, the higher the better; the smaller the standard deviation, the higher the stability of multiple generations.

[0070] Table 2

[0071] The conclusions are as follows: a. After introducing the Tree-of-Thought structure and knowledge chain embedding in Scheme C, the DA output path has a significantly higher performance prediction value; b. The self-consistency sampling mechanism (Self-Consistency) improves the convergence of multiple rounds of generation and avoids unstable structures; c. The CoK strategy allows the language model to retrieve relevant rules, such as "metal oxide + two-dimensional sheet coupling structure can balance magnetic loss and impedance matching", further enhancing the rationality of the structure.

[0072] Example 3 The polymer composite design method based on LLM knowledge database construction and reasoning optimization provided by the embodiment of the present embodiment is further optimized for filler ratio and carbonization temperature multi-objective design based on KDD and TVDL on the basis of the basic embodiment and Examples 1 and 2.

[0073] This embodiment is used to illustrate how the language model driven multi-objective optimization system in the present invention selects the structure-performance optimal path through the reasoning loop (TVDL) under complex variable space (filler ratio, carbonization temperature), further verifying its actual prediction ability and experimental consistency.

[0074] I. Target performance setting The multi-objective performance indicators are set as follows, and are expressed in natural language form in the system Prompt: a. Thermal conductivity (TC) ≥ 1.20 W / m·K b. Electromagnetic shielding efficiency (SE) ≥ 65 dB c. Minimum reflection loss (RLmin) ≤ –60 dB d. Effective absorption bandwidth (EAB) ≥ 5 GHz e. Explore the best combination of filler ratio (3 wt%, 5 wt%, 7 wt%) and carbonization temperature (800°C, 900°C, 1000°C).

[0075] II. Knowledge database construction Data input: The system calls a pre-constructed embedded natural language material knowledge database. This database is constructed based on professional literature and experimental data, and contains "structure-process-performance" semantic knowledge such as "filler content improves thermal conductivity but may compromise wave absorption", "carbonization near 900°C can balance graphitization and interface integrity", etc.

[0076] Knowledge retrieval and application: The Retriever module retrieves matching knowledge pieces from the database based on the target performance and variable space, and injects prompts such as "Medium filler content (5-7 wt%) combined with medium-high carbonization temperature (900°C) may achieve performance balance."

[0077] III. Reverse design model (DA) construction and operation The DA module is constructed based on a large language model, loads the knowledge base and performance targets, and generates candidate formulation schemes through prompt engineering. The prompt explicitly guides the DA to consider variable interaction effects: a. "High filler ratio improves thermal conductivity but may cause excessive shielding" b. "Carbonization temperature affects graphitization degree and interface integrity" c. "Melamine-assisted carbon source is beneficial to co-carbonization uniformity" The DA outputs candidate schemes represented by 5 wt% filler and 900°C carbonization based on knowledge reasoning: PI matrix + 5wt% G / CNT composite filler (G:CNT=3:2) + PDA surface modification + Co²⁺ adsorption + melamine-assisted carbon source.

[0078] IV. Performance verification module (VM) construction and operation The Verifier (VM) is constructed based on Gaussian Process Regression (GPR), which performs performance prediction and multi-objective scoring for all candidate schemes generated by DA, including thermal conductivity, electromagnetic shielding, and wave absorption. The scoring function is: F = w1·TC + w2·SE - w3·|RLmin| + w4·EAB.

[0079] V. Multi-round test-verification design loop (TVDL) optimization Three rounds of DA-VM iterative optimization are performed, and the process of each round is as follows: 1. DA generates 10 groups of candidate designs based on the current target and knowledge 2. VM performs target performance prediction and multi-objective scoring on all candidates 3. The top 10% of the scoring schemes are input into the next round of prompt as high-performance solutions The three-round optimization path is shown in Table 3: Table 3

[0080] The results show that the model can automatically exclude weak performance paths and converge to the optimal parameter combination (filler 5 wt%, 900°C, G:CNT=3:2).

[0081] VI. Experimental verification and performance testing Three typical formula composite samples were prepared: S1 (control): 3 wt% filler, carbonized at 800 °C S2 (recommended): 5 wt% filler, carbonized at 900 °C S3: 7 wt% filler, carbonized at 1000 °C The performance test results are shown in Table 4 below: Table 4

[0082] Anastomosis analysis: a. S2 is the model recommended solution, which is highly consistent with the predicted value; b. S3, although TC and SE are improved, RLmin and EAB are degraded, verifying the "over-shielding" mechanism; c. S1 performance is low, confirming the problems of insufficient filler and insufficient carbonization.

[0083] Seven, structure-performance mechanism analysis Through SEM, TEM and other characterization means to confirm: a. S2 sample forms an ideal multi-scale heterogeneous structure: CF-G-CNT three-channel network, complete interface, N-doped carbon enhanced dielectric polarization b. S1 sample is discontinuous due to insufficient filler and insufficient carbonization c. S3 sample, although the thermal conductivity and shielding are improved, the over-graphitization leads to structure densification, which damages the wave absorption performance Eight, Conclusion This embodiment fully demonstrates the implementation process of the five-step design method, and proves: 1. KDD system can effectively capture the "carbonization temperature-structure integrity-performance balance" complex relationship 2. TVDL multi-round feedback ensures that the DA model is optimized towards a better physical mechanism path 3. Verifier can accurately predict the effect of the structure scheme, saving the experimental trial and error cost 4. The final recommended scheme (5 wt%, 900°C) achieves the best balance among multiple target performances Embodiment 4

[0084] The polymer composite design method based on LLM knowledge database construction and reasoning optimization provided by the embodiment of the present embodiment is further based on the basic embodiment and embodiments 1-3, and further to the structure guiding design and performance prediction based on the differences of carbon source types and catalyst types.

[0085] This example aims to illustrate how the knowledge-driven design method described in the present invention influences the heterogeneous structure construction path of composites by combining different carbon sources (such as PDA and melamine) and catalyst types (such as Co²⁺ and Fe³⁺), and realizes structure difference perception and performance prediction optimization through language model-driven prompting strategies.

[0086] I. Target performance setting The following multi-target performance indicators are set, and are expressed in natural language form in the system Prompt: a. TC ≥ 1.20 W / m·K; b. SE ≥ 65 dB; c. RLmin ≤ –55 dB; d. EAB ≥ 5 GHz; e. Explore the influence of different carbon sources (PDA vs melamine) and catalysts (Co²⁺ vs Fe³⁺) on structure / performance.

[0087] II. Knowledge database construction 1. Data input: The system calls a pre-constructed embedded natural language material knowledge database. This database is based on professional literature and experimental data, and contains more than ten thousand relevant semantic knowledge.

[0088] 2. Knowledge retrieval and application: The Retriever module retrieves and injects the following key knowledge fragments from the database according to the target performance and variable space: a. "Melamine releases nitrogen source at high temperature, forming sheet graphitization area and N doping" b. "PDA modification layer can enhance metal ion adsorption and guide interface carbonization to form carbon shell structure" c. "Co²⁺ is more easily reduced to metal particles than Fe³⁺ and triggers CNT growth reaction" d. "Fe³⁺ forms stronger magnetic nanoparticles, but has weaker influence on thermal conduction channels" III. Construction and operation of reverse design model (DA) Based on a large language model, the DA module is constructed, the knowledge base and performance targets are loaded, and four groups of candidate schemes for different carbon source-catalyst combinations are generated through prompting engineering techniques, as shown in Table 5: Table 5

[0089] IV. Construction and operation of performance verification module (VM) Based on the few-shot supervised regression model, a Verifier (VM) is constructed to predict and score the performance of the four candidate schemes. The scoring function is: F = 0.3*TC_norm + 0.3*SE_norm + 0.2*|RLmin|_norm + 0.2*EAB_norm.

[0090] V. Multi-round test-verification design loop (TVDL) optimization Three rounds of DA-VM iterative optimization are performed: 1. First round: DA generates four sets of schemes, and VM predicts the scores after feedback to take the TOP scheme (P1) 2. Second and third rounds: DA generates new schemes based on feedback optimization, and VM predicts the scores again 3. The final output is the performance prediction result of each optimized scheme VI. Results and discussion The VM performance prediction results are shown in Table 6: Table 6

[0091] Experimental verification (supplementary preparation and testing): The above schemes are prepared and tested, and the key results are as follows in Table 7: Table 7

[0092] VII. Structure-performance mechanism analysis Through characterization analysis, it is confirmed that: 1. P1 scheme: PDA provides excellent interface modification, Co²⁺ effectively catalyzes CNT growth, forming ideal multi-scale heterostructures, realizing heat conduction-shielding-absorption synergy 2. P2 scheme: melamine is not combined with PDA modification, the structure is loose, and CNT connection is insufficient 3. P3 scheme: Fe³⁺ forms an oxide core, enhancing magnetic loss but difficult to form a heat conduction network 4. P4 scheme: combination of Fe³⁺ and melamine leads to excessive agglomeration, and the structure is discontinuous VIII. Conclusion This example fully demonstrates the implementation process of the five-step method of claim 1, and proves that: 1. The KDD system can effectively understand the differential influence mechanism of carbon source-catalyst combination 2. The DA module can intelligently select the optimal structure path according to the material mechanism knowledge 3. The VM module accurately predicts the performance trend of different combinations, which is highly consistent with the experimental results 4. PDA+Co²⁺ combination is the best choice, achieving optimal balance among multiple target properties The design method provided by each of the above embodiments of the present application significantly reduces the experimental trial and error cost of multivariate combination, and provides an effective intelligent solution for complex material system design.

[0093] Each of the above embodiments of the present application proposes a polymer composite material design method based on LLM knowledge database construction and reasoning optimization through the fusion of polymer material and artificial intelligence cross technology. The method constructs an intelligent design road system integrating target setting, semantic knowledge database construction, design assistant (DA), performance verifier (VM), prompt engineering module and language model material knowledge system. The system inputs professional literature, experimental data and other material corpus, uses large language model (LLM) for natural language understanding and knowledge extraction, constructs embedded semantic knowledge database, and generates composite material structure scheme meeting the multi-target performance constraints of heat conduction, electromagnetic shielding, wave absorption and mechanics under the guidance of prompt engineering. The prompt engineering combines chain reasoning (CoT), tree thinking (ToT), knowledge chain retrieval (CoK), self-consistency sampling and automatic prompt optimization strategies to strengthen the design generation and explainability of the language model. Through multiple rounds of DA-VM collaborative optimization, the structure-mechanism-performance closed-loop reasoning and feedback are realized. The system relies on the literature knowledge corpus of high-performance polymer-based composite materials and can output composite configurations such as multi-scale carbon heterogeneous network structure, realize the multi-performance collaborative optimization of thermal conductivity ≥1.25 W / m·K, electromagnetic shielding efficiency ≥65 dB, minimum reflection loss ≤-60 dB under low filler conditions, and simultaneously consider high heat resistance and mechanical strength. The method has the advantages of high design efficiency, strong structure generation rationality, accurate performance prediction, and evolvable knowledge, and is suitable for intelligent rapid development of high-thermal-conductivity-wave-absorption-shielding-mechanical-integrated functional polymer composites.

[0094] It should be noted that within the scope of the above description of the present application, other technical solutions using different models, algorithms, components, proportions and preparation processes can also achieve the technical effects described in the present application, and therefore will not be listed one by one.

Claims

1. A polymer composite material design method based on LLM knowledge database construction and reasoning optimization, used for the formulation generation and screening of thermally conductive, electromagnetically shielding, wave-absorbing, and mechanically integrated materials, characterized in that... Includes the following steps: S1. Target performance setting: Set multiple target performance indicators, including at least thermal conductivity, electromagnetic shielding efficiency, minimum reflection loss, effective absorption bandwidth, high heat resistance, and strong mechanical properties, and describe them in natural language in the prompt. S2. Knowledge Database Construction: Create an embedded natural language material knowledge database as a cognitive foundation; This natural language knowledge database does not require traditional manual labeling and structured formats, and has advantages such as rapid construction, dynamic evolution, and high semantic support. It is the cognitive foundation that supports language model-driven material design. S3. Reverse Design Model (DA): Based on the large language model, load the natural language knowledge base obtained in step S2 and the performance target set in step S1, and generate candidate formulation schemes through prompting engineering, including polymer matrix type, filler type and ratio, interface modification method and structure construction strategy. S4. The performance verification module builds a Verifier, VM: Based on the positive prediction capabilities of physical models, machine learning predictors, or language models, a Verifier is built to score and rank candidate solutions in terms of thermal conductivity, electromagnetic shielding, wave absorption, mechanical properties, and heat resistance. S5. Multi-round test-verification design loop TVDL optimization: Multi-round iterative optimization is performed between the reverse design module DA and the performance verification module VM. Each round includes: DA generates candidate designs based on the current goal and knowledge; VM performs target performance prediction and multi-objective scoring on all candidates; the top 10% of the solutions are used as high-performance solutions as inputs to the next round prompt; and the final candidate solution is output after at least 3 rounds of optimization.

2. The design method according to claim 1, characterized in that, In step S1, the target performance includes at least the following constraints: thermal conductivity not less than 1.2 W / mK, electromagnetic shielding efficiency not less than 60 dB, minimum reflection loss less than -50 dB, effective absorption bandwidth not less than 4 GHz, and matching appropriate mechanical and heat resistance properties.

3. The design method according to claim 1, characterized in that, The database in step S2 is constructed through the following process: By inputting professional literature, experimental data, and patent descriptions into a Large Language Model (LLM), and leveraging the semantic extraction capabilities of the LLM, semantic fragments and causal chains related to "structure-mechanism-performance" are automatically extracted and converted into vector embeddings for storage. This data is then used in engineering to construct contextual suggestions, generate structural design recommendations, or verify target performance. Its core function is: The Retriever module matches and injects relevant knowledge fragments into the Prompt to match the user's performance goals; it supports the Design Assistant (DA) in generating literature-based structures, proportions, and interface strategies; it provides a logical basis for the Verifier (VM) to judge the scientific validity and feasibility of candidate solutions; and during the multi-round optimization process of DA-VM, it can dynamically summarize the potential knowledge in high-performance solutions, feed it back and supplement it to the database, and achieve self-evolution. The database is constructed automatically by a language model based on input professional literature data. The input context of this knowledge data comes from preprocessed materials science literature, patent data, and experimental reports, including but not limited to content on two-dimensional thermal conductive networks, three-dimensional carbon heterostructures, electromagnetic shielding mechanisms, wave absorption loss mechanisms, high heat resistance, and mechanical properties. The database uses semantic fragments as basic units and records knowledge including but not limited to the following types: material structural characteristics and processing technology; energy mechanism attribution; multi-objective conflict reconciliation strategies; and specific structure-performance causal rules. Entries are expressed in the form of natural language semantic fragments or structure-mechanism-performance chains and stored in the database in vector form. This supports vectorized retrieval, prompt injection, and feedback inductive updates, serving as a semantic support module for generating candidate structural schemes and performance evaluations.

4. The design method according to claim 1, characterized in that, In step S3, during the construction of the reverse design model (DA), the filler content in the candidate solutions output by the reverse design model (DA) does not exceed 10 wt%. The filler includes, but is not limited to, carbon fiber, graphene, MXene, carbon nanotubes, Fe3O4, silicon carbide, expanded graphite, or composite structures thereof.

5. The design method according to claim 1, characterized in that, The Verifier in step S4 is constructed based on at least one of the following models: Gaussian Process Regression (GPR), Random Forest (RF), Transformer Structural Language Model, or their ensemble regressors.

6. The method according to claim 1, characterized in that, In step S5, the TVDL design loop is performed at least three times, with each round including three stages: DA, VM, and optimization feedback. In each round, the DA input is the best sample from the previous round's VM score as a new prompt context to improve the generation quality.

7. The design method according to claim 6, characterized in that, The language interaction between the DA and the VM is based on a prompting project, which includes one or more of the following strategies: a. Chain-of-Thought (CoT) prompts: used to guide the output of large language models to establish causal reasoning chains between structure, process, and performance, thereby enhancing the physical-logical consistency of candidate designs; b. Tree-like thinking prompts ToT: used to construct multi-path reasoning structures in the material formulation space, and systematically filter optimal solutions through search and scoring mechanisms; c. Knowledge Chain Hints (CoK): Connects the design task with an embedded knowledge base to guide the language model to dynamically retrieve the material mechanisms, empirical rules, or verified literature facts most relevant to the current goal; d. Consistency Hint: In each round of candidate generation, multiple samples are taken, and multiple structures or schemes are output. Stable and consistent outputs are filtered through clustering and structure matching. e. Proactive example prompts: Based on the target performance confidence or model uncertainty, prioritize adding the few-shot instances with the highest information gain to the prompt; f. Automatic Prompt Construction Module (APE): Used to automatically build or optimize prompt structure combinations. Through scoring feedback mechanisms, greedy search, or strategy optimization algorithms, it automatically builds or adjusts the prompt structure combination that best suits the current task. The Tree-of-Thought prompt guides the generation of material design paths through depth-first search or breadth-first search strategies. Each "thinking node" represents a candidate design. After multiple rounds of verification, the branch with the best performance is retained, and paths with deteriorating performance are backtracked and pruned. The Self-Consistency prompt strategy requires at least three sets of formulation suggestions to be output in each round of generation. Through structural similarity analysis and performance prediction aggregation, the most consistent solution is selected to enter the next round of feedback. The automatic prompt construction module is based on reinforcement learning, genetic algorithms, or multi-armed gambling machine mechanisms. It automatically optimizes the combination of prompt structure and input fragments based on historical prompt effect scores to improve the controllability and convergence efficiency of the DA module's generated results.

8. The design method according to claim 1, characterized in that, The polymer composite material contains a multi-scale heterogeneous carbon structure network. Its structural design logic is automatically generated by the DA module according to the following knowledge rules, specifically including: using polyimide (PI) fiber felt to provide a three-dimensional porous scaffold, giving the composite material an initial macroscopic continuous pathway; introducing a dopamine PDA modified layer to construct a highly polar interface on the PI surface, forming adsorption sites to promote the combination of two-dimensional filler and metal ions, while converting it into N-doped carbon in subsequent carbonization to enhance dielectric loss; loading graphene sheets to construct two-dimensional hot electron bridges to enhance lateral thermal channels and electromagnetic shielding paths; using Co²⁺ to coordinate with PDA to achieve self-positioning catalytic site distribution, and triggering the Co reduction-CNT in-situ growth reaction during co-carbonization to form a one-dimensional vertical structure; finally constructing heterogeneous interfaces such as CF-CNT, G-CNT, and N-G-CNT to realize a multi-scale thermo-electric-magnetic synergistic network structure.

9. The design method according to claim 8, characterized in that, The structural design is generated by the language model through a prompting engineering mechanism. The Prompt context includes the following guidance strategies: clearly defining the target performance combination; providing Few-shot instances; embedding structure-performance knowledge rules; and guiding the model to generate a three-in-one structure composition, process path, and structural morphology. After multiple rounds of feedback from the VM verification module, the retained structure has the best performance prediction output.

10. The design method according to claim 9, characterized in that, The performance prediction is implemented by the Verifier module, specifically including the following structure-performance mapping mechanism: if a CF-CNT-G bridging channel exists, the continuity of the system's thermal conductivity path is enhanced, and the expected thermal conductivity improvement is > 1 W / m·K; if an N-doped + Co-particle + graphene coexisting structure is formed at the filler interface, the dielectric polarization, magnetic loss, and conductive shielding coupling are enhanced, and the predicted RL value decreases by 10–15 dB, while the shielding is improved by 5–10 dB; the VM module introduces existing experimental performance samples for regression evaluation through few-shot supervision, outputs the score of the candidate structure, and uses it as the input for TVDL feedback optimization iteration.

Images