Multi-agent-based material performance prediction and synthesis method and system

Abstract

The invention relates to a multi-agent-based material performance prediction and synthesis system, and the system comprises a multi-agent data enhancement module which is configured to be used for firstly disassembling a complex problem into a plurality of subtasks, and then constructing a fine tuning data set comprising Sub-CoQ question and answer pairs by starting multi-source parallel retrieval; the multi-expert debate module is configured to be used for simulating decision conflicts of different roles in material engineering and generating a direct preference optimization DPO data set through debate; the training and verification module is configured to be used for training and verifying a large model MatMind in the field of materials by utilizing supervised fine tuning SFT and reinforcement learning RLHF based on the fine tuning data set and the DPO data set; and the material performance prediction and synthesis module is configured to be used for realizing intelligent recommendation of a material performance prediction and synthesis process by importing input parameters into the large model MatMind.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of material science and engineering, and particularly relates to a scheme for material performance prediction and synthesis based on multiple agents. BACKGROUND

[0002] With the progress of science and technology, large models are increasingly widely used in various fields. The current application of large models in the field of material science and engineering mainly faces two core challenges:

[0003] Firstly, the general large language model (GPT-4, DeepSeek and Qwen) lacks material professional knowledge and cannot effectively handle complex tasks such as crystal structure and synthesis path. Traditional incremental pre-training relies on limited labeled data, while the data generated by LLMs (Large Language Models) is mostly limited to qualitative description, lacking quantitative mapping of process-performance.

[0004] Secondly, the lack of engineering thinking, existing databases (such as Materials Project) mainly contain structured data in stable state, lacking the logic and conflict resolution process in actual material engineering decision-making. At the same time, non-structured data such as material engineering forum discussions are not effectively utilized, making it difficult for the model to handle complex requirements in real scenarios.

[0005] The root cause of this dilemma is that the material field model after incremental pre-training needs high-quality post-training data to learn material science tasks, but the current material engineering field lacks high-quality data, and the cost of manually building data is high. The information density of the generated data is insufficient, and the generated model lacks material physical constraints, which further limits the model's understanding of complex material engineering applications.

[0006] Due to the above reasons, AI large model technology has always been unable to be well used to solve various problems in the field of materials, especially in the aspects of material performance prediction and synthesis, and there is currently a lack of a scheme that can fully utilize large model means to assist technical personnel in material performance prediction, material optimization and screening, and material synthesis. SUMMARY

[0007] In the present application, a material performance prediction and synthesis method and system based on multiple agents are proposed, aiming to solve the problems of data limitation and lack of engineering thinking in the current large model training in the field of material science and engineering.

[0008] According to a first aspect of the present application, a material performance prediction and synthesis system based on multiple agents is provided, comprising:

[0009] a multi-agent data augmentation module configured to first decompose a complex problem into multiple sub-tasks, and then construct a fine-tuning dataset including Sub-CoQ question-answer pairs by starting multi-source parallel retrieval;

[0010] a multi-expert debate module configured to simulate decision conflicts of different roles in material engineering, and generate a direct preference optimization (DPO) dataset by debate;

[0011] a training and verification module configured to train and verify a large model MatMind in the field of materials based on the fine-tuning dataset and the DPO dataset, and utilize supervised fine-tuning (SFT) and reinforcement learning (RLHF); and

[0012] a material performance prediction and synthesis module configured to realize intelligent recommendation of material performance prediction and synthesis processes by importing input parameters into the large model MatMind.

[0013] According to a second aspect of the present application, a multi-agent based material performance prediction and synthesis method is provided, comprising:

[0014] decomposing a complex problem into multiple sub-tasks;

[0015] starting multi-source parallel retrieval according to characteristics of each sub-task to construct a fine-tuning dataset including Sub-CoQ question-answer pairs;

[0016] simulating decision conflicts of different roles in material engineering, and generating a direct preference optimization (DPO) dataset by debate;

[0017] training a large model MatMind based on the generated fine-tuning dataset and DPO dataset;

[0018] verifying the trained large model MatMind; and

[0019] after deploying the trained large model MatMind, importing input parameters into the large model MatMind to realize intelligent recommendation of material performance prediction and synthesis processes.

[0020] This summary is provided to introduce some concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to describe the manner in which the above-recited and other advantages and features of the application can be obtained, a more particular description of the application briefly described above will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. Understanding that these drawings depict only typical embodiments of the application and are not therefore to be considered to be limiting of its scope, the application will be described and explained with additional specificity and detail by the use of the accompanying drawings in which:

[0022] Figure 1 An example structural diagram of a multi-agent based material performance prediction and synthesis system according to an embodiment of the present application is shown.

[0023] Figure 2 An example schematic training method of a material engineering large model (Matmind) according to an embodiment of the present application is shown.

[0024] Figure 3 An example schematic flowchart of four-level preprocessing of collected raw text according to an embodiment of the present application is shown.

[0025] Figure 4 An example schematic diagram of model reward optimization dynamic analysis based on DPO training according to an embodiment of the present application is shown.

[0026] Figure 5 An example flowchart of a multi-agent based material performance prediction and synthesis method according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0027] In general, in the scheme of the present application, the core innovation is to propose a scheme of material dataset construction based on multi-agent external information enhancement and material performance prediction and synthesis using the constructed dataset. The scheme aims to solve the data limitation and lack of engineering thinking in the training of large models in the field of current material science and engineering, to provide a more intelligent and efficient means of material performance prediction and synthesis. The scheme cooperatively constructs a high-quality fine-tuning dataset through a multi-agent data enhancement module (including a sub-problem chain (Sub-CoQ) agent) and a multi-expert debate module, and uses training and verification modules such as supervised fine-tuning (SFT) and reinforcement learning (RLHF) to train and verify the material field large model MatMind, and finally realizes material performance prediction and synthesis process recommendation.

[0028] An example structural diagram of a multi-agent based material performance prediction and synthesis system according to an embodiment of the present application is described below based on Figure 1

[0029] As shown in the figure, the system mainly includes the following modules: ​

[0030] 1. Multi-agent data augmentation module 110

[0031] The core innovation of the present application is to construct a multi-agent collaborative data augmentation module, i.e., a multi-agent data augmentation module 110. The multi-agent data augmentation module 110 is configured to first decompose a complex problem into multiple sub-tasks, and then construct a fine-tuning data set including Sub-CoQ question and answer pairs by starting multi-source parallel retrieval.

[0032] In the present application, "agent" refers to an expert-type autonomous decision-making entity specially designed for material research and development. It integrates knowledge in a specific field (such as crystal structure rules, synthesis process constraints, or performance prediction models) and has the ability of task understanding, reasoning generation, and tool invocation.

[0033] The multi-agent data augmentation module 110 further defines a sub-problem chain generation (Sub-CoQ) agent 112 and a data retrieval recall agent 114.

[0034] The Sub-CoQ agent 112 requires the agent to split the question into a sub-problem chain by using question splitting cues, and decomposes a complex problem into multiple sub-tasks. Specifically, the question chain (sub-coq) recursive decomposition algorithm refers to the decomposition of a complex problem (such as a complex problem of material design or performance optimization) into multiple sub-tasks that cannot be further split by adapting a large model. Its core process includes the following steps:

[0035] ‌ 1. Recursive decomposition of questions

[0036] When faced with a complex and comprehensive problem (for example: "Design a lightweight high-strength composite material suitable for high-temperature environments"), the large model first performs semantic understanding and recursively decomposes it into multiple sub-problems with clear logic and clear boundaries. For example:

[0037] Sub-problem 1: Which material system has good thermal stability at high temperatures?

[0038] Sub-problem 2: How to balance the density and mechanical strength of the material?

[0039] Sub-problem 3: What is the impact of available preparation processes on material performance?

[0040] The process continues recursively until each sub-problem reaches an "atomic level" - i.e., it can be directly solved by invoking a knowledge base, model, or tool. That is, the Sub-CoQ agent is further configured to recursively decompose sub-tasks until the generated sub-problems satisfy pre-set atomicity conditions, including that the problem can be answered by a single data source or that the problem corresponds to a standard operation unit in material R&D.

[0041] 2. Chain-of-Thought Reasoning on Sub-Problems

[0042] For each atomic sub-problem, a Chain-of-Thought (CoT) reasoning mechanism is enabled to guide the large model to generate a "thinking path" rather than directly output an answer.

[0043] By introducing CoT, the accuracy and explainability of each sub-problem answer are significantly improved.

[0044] 3. Parallel Solving & Result Aggregation

[0045] All sub-problems can be processed in parallel, supplemented with knowledge retrieval and recall agents 114, and combined with the concurrent reasoning capabilities of the large model. After the CoT reasoning of each sub-problem is completed, the system will structure and integrate the scattered answers to form a complete response to the original problem.

[0046] 4. Feedback & Refinement

[0047] The system supports a reverse verification mechanism: the generated answer is returned to the large model, and if the final answer does not meet the constraint conditions (such as high cost or unfeasible process), the key sub-problems are automatically identified and re-solved to achieve closed-loop optimization.

[0048] The data retrieval recall agent 114 initiates multi-source parallel retrieval according to the characteristics of each sub-task, including: obtaining crystallographic parameters through structured databases (such as Materials Project, ICSD), extracting material synthesis methods through academic literature databases (such as Sci-Hub, DOI), and collecting synthesis process parameters through patent databases (such as USPTO, CNIPA); then, low-relevance retrieval results are filtered out through semantic matching algorithms based on BERT models; and after the retrieval results are normalized, the Sub-CoQ answers in the form of question-evidence pairs are reorganized with the multiple sub-task problems; based on these Sub-CoQ question-answer pairs, the corresponding fine-tuning data set is finally constructed. The fine-tuning data set contains both material science structure and performance description and process synthesis parameters, and its data density is 5-8 times that of traditional manual annotation, and backtracking verification ensures 100% compliance with physical laws.

[0049] wherein the generated fine-tuning data set is a collection of question-evidence pairs in the form of Sub-CoQ question-answer pairs for fine-tuning.

[0050] 1. Input raw text

[0051] An unstructured natural language text (such as a scientific literature abstract, a Wikipedia entry, a news report, or a database description) is provided as input to the large model.

[0052] 2. Prompt Engineering

[0053] Through the designed prompt, the large model is guided to complete three tasks at the same time: identify key entity pairs, generate natural language questions based on the entity pairs, and provide answers. The prompt template is usually as follows:

[0054]

[0055] After obtaining the question, a more accurate answer is generated through the Sub-CoQ method to construct the fine-tuning data set. Specifically, it contains about 100,000 precisely paired Sub-CoQ question-answer pairs. It should be understood that as technology develops and databases are continuously updated, the number of Sub-CoQ question-answer pairs will also increase. Therefore, the multi-agent data augmentation module 110 needs to periodically perform the above processing to maintain the freshness and timeliness of the data set.

[0056] 2. Multi-expert debate module 120

[0057] Direct Preference Optimization (DPO) is an efficient fine-tuning method for aligning large language models with human preferences. Unlike traditional Reinforcement Learning-based Human Feedback Learning (RLHF), DPO does not require an explicit reward model. Instead, it directly utilizes human-labeled binary preference data (i.e., sample pairs each containing a preferred answer and a suboptimal answer) to optimize the language model's strategy end-to-end through a classification loss function based on the Bradley-Terry preference model.

[0058] The core idea is to maximize the log-likelihood ratio of the preferred answer over the suboptimal answer while keeping the model output distribution similar to the original reference model (constrained by KL divergence). This method implicitly treats the language model itself as a proxy for the reward function, simplifying the alignment process and improving training stability and computational efficiency.

[0059] In specialized fields such as materials science, DPO can effectively utilize expert preference judgments for different material design schemes, synthesis paths, or performance prediction results to guide large models to generate outputs that better align with field practices and multidisciplinary constraints.

[0060] However, the performance of DPO highly depends on the quality and domain relevance of the preference data. To address the problem of expert preference data scarcity in materials research and development, this application proposes a multi-expert debate mechanism that automatically generates preference pairs consistent with the logic of materials science by simulating decision-making conflicts between different role experts, providing high-quality training signals for DPO.

[0061] Overall, the basic construction process of the DPO dataset involves the following steps:

[0062] Predefine the role identities of multiple specialized agents, including "synthetic chemist," "process engineer," "performance test expert," "computational materials scientist," etc., each corresponding to a specific domain knowledge base and decision-making preference;

[0063] For a given materials research and development problem, each agent independently generates a candidate answer based on its role stance;

[0064] Through a multi-round interactive debate mechanism, each agent questions, refutes, or supplements the answers of other agents, using a sequence labeling model to extract key arguments from the debate content;

[0065] Reorganize the scattered arguments into a tree-like argument structure according to logical dependency relationships and verify the logical consistency between arguments using a materials knowledge graph;

[0066] Based on the preset multi-dimensional scoring system (including scientificity, implementability, interdisciplinary consistency, and data support), each candidate answer is quantitatively evaluated;

[0067] The multi-role argument enhancement strategy is adopted to guide the agent to supplement the missing arguments from different professional perspectives, and to optimize the completeness of the answer;

[0068] The optimal answer and the suboptimal answer are selected to form a pair of preferred answers.

[0069] Based on the optimal answer and the suboptimal answer, a pair of superior-inferior samples for direct preference optimization (DPO) training is constructed.

[0070] Specifically, the specific process of generating DPO data by the multi-expert debate module 120 is as follows:

[0071] 1. Data source and cleaning

[0072] The data comes from a materials engineering exchange forum, which contains diverse discussions among engineers, scholars and industry practitioners, covering practical problems such as material synthesis, performance optimization and failure analysis. In this process, various problems are first obtained from the discussions of each forum, and then irrelevant or low-quality problems are cleaned up through keyword filtering (such as "synthesis" and "failure analysis") and semantic analysis (BERT-based classifier) to ensure the domain relevance of the data set.

[0073] 2. Automatic answering and scoring

[0074] The enhanced high-quality question set is automatically answered by the supervised fine-tuned (SFT) material field large model MatMind. Based on the embedded material knowledge and reasoning ability of the large model, a structured and interpretable technical response is generated for each question. Then, the system quantitatively scores the quality of the answer based on multi-dimensional evaluation indicators, including the scientific accuracy of the answer (verified by comparing with the fact triples in the material knowledge graph), the logical completeness (measured by the argument coverage and reasoning chain depth), and the engineering implementability (judged according to the process constraints and cost feasibility rule base); further, the cognitive difficulty of the corresponding question is inversely deduced according to the breadth of knowledge required by the answer (such as the number of subfields involved), the complexity of reasoning steps and the depth of tool invocation, and the difficulty of the original question is classified and sorted according to the difficulty, forming a problem resource pool covering low, medium and high difficulty levels, which is used for subsequent multi-expert debate and differential scheduling of DPO data generation.

[0075] 3. Question expansion

[0076] After difficulty grading of enhanced high-quality problems, the system inputs each difficulty level (low, medium, high) problem as a seed to drive the field-adaptive problem generation model to expand differently: for low-difficulty seed problems, generate variant problems focusing on basic concepts and standard processes; for medium-difficulty seed problems, generate composite problems involving multi-factor coupling or cross-subfield correlation; for high-difficulty seed problems, generate exploratory problems that require the combination of cutting-edge theory, multi-scale simulation, or unexplored material systems. Through this hierarchical generation mechanism guided by seeds, the expanded problems ensure semantic relevance while accurately covering the full spectrum of cognitive needs in material research and development, from routine optimization to cutting-edge exploration, providing a balanced and controllable input problem set for subsequent multi-expert debates.

[0077] 4. Agent definition and construction

[0078] 1) Professionalized agent role definition

[0079] The system predefines multiple material field expert agents, including theoretical experts, process engineers, material doctoral students, and characterization analysts. Each agent is bound to a dedicated knowledge base and decision preferences: theoretical experts focus on thermodynamics and phase stability, process engineers focus on equipment parameters and mass production feasibility, doctoral students are good at cross-literature induction and method reflection, and characterization analysts can detect the feasibility of SEM / XRD-based logical verification schemes. The role identity remains consistent throughout the debate, ensuring the professionalism and perspective difference of the generated viewpoints.

[0080] 2) Multi-round adversarial debate and stance evolution trajectory modeling

[0081] For the same problem, each agent starts a multi-round debate, and the system records the entire stance evolution trajectory. In the first round, each agent independently generates an initial answer based on their own knowledge; from the second round, agents can selectively reference other agents' previous viewpoints and decide whether to modify their stance. If an agent recognizes that someone else's argument is more authoritative (such as a process engineer pointing out that the solvent proposed by a doctoral student is not available on the production line), or finds that their reasoning has a knowledge blind spot (such as a theoretical expert ignoring kinetic limitations), they will actively update their conclusions in subsequent rounds and explicitly label the modification reason (such as "original suggestion ignores the impact of sintering atmosphere, now adjusted according to ICSD#12345 data"). The system considers such self-correction behavior as a key signal of high-quality reasoning, recording the comparison of arguments before and after modification, changes in knowledge references, and logic reconstruction paths. All speeches are organized in chronological order and causal dependence as a directed trajectory graph, where nodes represent answers and edges represent stance inheritance, refutation, or modification relationships, forming a traceable and assessable collaborative cognitive evolution process.

[0082] 3) Whole-process answer evaluation and preference construction

[0083] The system scores each generated answer in the debate trajectory (including intermediate rounds and final output) in real time in multiple dimensions, including technical correctness, logical consistency, cross-role coordination, and practical feasibility. Based on the scores, the system constructs two types of preference pairs: one is a conclusion preference pair, with the final answer with the highest overall score as the chosen and the final answer with obvious knowledge conflicts as the rejected; the other is a process preference pair, with a high-quality intermediate answer in a certain round as the chosen and a low-quality answer that is effectively refuted by it in the same round or previous round as the rejected. This mechanism enables DPO training to not only learn "what conclusions are better" but also learn "how to reason step by step to better conclusions."

[0084] 4) Dynamic knowledge injection and understanding enhancement

[0085] During the debate process, if the agent determines that the current knowledge is insufficient to support effective argumentation (such as being unable to confirm the phase transition temperature of a certain alloy), it can actively trigger the tool invocation interface to query crystallography databases, perform CALPHAD calculations, or retrieve the latest literature in real time. The newly acquired knowledge is immediately injected into the debate context and used as the basis for subsequent statements. This mechanism enables the debate not only to be a clash of viewpoints but also a collaborative knowledge discovery process, often generating new understandings that are not covered by the initial knowledge base, such as identifying the "coupling inhibition mechanism of high-entropy effect and grain boundary segregation."

[0086] 5) Fully autonomous generation, no human intervention required.

[0087] The entire process - from problem input, multi-agent debate, trajectory recording, answer evaluation to preference pair output - is automatically completed by the system and does not rely on human annotation, forum crawling, or expert review. The generated DPO dataset is directly used for direct preference optimization training of the MatMind large model, enabling it to have a collaborative decision-making capability similar to an expert team in complex materials tasks.

[0088] 3. Training and verification module 130

[0089] The training and verification module 130 is configured to train and verify the large model MatMind in the field of materials based on the dataset and using supervised fine-tuning (SFT) and reinforcement learning (RLHF).

[0090] As previously described, using the above multi-agent data augmentation module 110 and multi-expert debate module 120, two types of training datasets are generated: one type of fine-tuning dataset for fine-tuning, and the other type of DPO dataset for reinforcement learning. Therefore, subsequently, the large model MatMind can be trained based on the two types of training datasets generated.

[0091] The specific process includes:

[0092] 1) In the Supervised Fine-Tuning (SFT) phase: Based on a fine-tuning dataset containing approximately 100,000 precisely paired Sub-CoQ question-answer pairs (samples), an evidence attention mechanism is used to optimize the pre-trained base model. The pre-trained base model can employ various mainstream models; this work uses the Qwen-2.5 7B model. Each question-answer sample in this fine-tuning dataset is accompanied by a multi-source evidence chain, including material calculation data, experimental characterization parameters, and process optimization records.

[0093] During training, a gradient accumulation strategy is adopted. For example, the batch size can be set to 64 and the learning rate to 5e-5. With the help of the cosine annealing scheduler, training is completed for 3 epochs (epoch: period / generation, refers to the process of the model completely traversing the entire training dataset once) on an 8×H100 GPU cluster.

[0094] It should be understood that the specific settings for batch size and learning rate, as well as the type of hardware support used, should be appropriately determined based on the application's requirements and available hardware resources. The values ​​and hardware examples above are merely illustrative.

[0095] 2) Reinforcement Learning Optimization Phase (RLHF): Based on the DPO data generated by the expert debate module 120, the Direct Preference Optimization (DPO) algorithm is used for deep training of the model. The Direct Preference Optimization (DPO) algorithm is an alignment algorithm for fine-tuning large language models (LLMs). It bypasses the traditional reinforcement learning process and directly optimizes the model strategy through human preference data, significantly simplifying the training process and improving efficiency.

[0096] By using the multi-expert debate module 120 to generate a DPO data model and constrain it with MatMind, it can not only provide more accurate solutions that conform to actual technical specifications, but also comprehensively consider the actual constraints of industrial implementation, realizing a complete knowledge transfer from theory to practice. The final output recommendation solution has both academic rigor and engineering practicality.

[0097] In the validation and iteration phase after model training is completed, this application uses the currently publicly available benchmark dataset Mat4llm-bench and a set of continuous training enhancement mechanisms to ensure continuous optimization of model performance:

[0098] Level 1: Validating the model's core capabilities through benchmark testing. Specifically, in the standard task of predicting material properties and synthesis processes based on CIF files, it surpasses large language models (Llama 2-7b, LLM-Prop) and is on par with or close to specialized models (CGCNN, MatBERT), demonstrating its efficiency and reliability in the field of materials science.

[0099] Level Two: Simultaneously, this application deploys an online continuous learning mechanism. This mechanism automatically acquires the latest published materials science papers and patent data daily via the arXiv API, and updates the model's knowledge base in real time after intelligent filtering and annotation. This not only addresses the pain point of traditional models' knowledge lag but also enables MatMind to continuously track cutting-edge advancements in materials science, maintaining the advanced nature and timeliness of its predictions and recommendations. This closed-loop optimization strategy of "evaluation-update" ensures stable model performance on classic tasks while endowing it with the ability to continuously evolve, providing dynamically updated intelligent support for materials research and development.

[0100] Thus, after completing the training and validation of the model, a large model MatMind based on multi-agent external information enhancement has been constructed.

[0101] The material property prediction and synthesis module 140 is configured to achieve intelligent recommendation of material property prediction and synthesis processes by importing input parameters into the large model MatMind based on multi-agent external information enhancement.

[0102] As mentioned earlier, by first generating two types of training datasets (including a fine-tuned dataset of Sub-CoQ question-answer pairs and a DPO dataset containing multi-expert debate) using the multi-agent data augmentation module 110 and the multi-expert debate module 120 respectively, and then using the two types of training datasets to train and validate the large model MatMind, a large model MatMind based on multi-agent external information augmentation can be constructed.

[0103] Subsequently, during the material performance prediction and synthesis process, by inputting the corresponding parameters into the model, the model combines the parameters to perform intelligent analysis, reasoning, and decision-making, ultimately achieving intelligent recommendations and suggestions for material performance prediction and synthesis.

[0104] These modules can communicate with each other through various wired or wireless communication methods, and the system can access various data websites, literature journals, materials forums and other Internet resources via local area networks, cellular networks, wide area networks, WIFI, WLAN and other means.

[0105] The following examples illustrate the specific application process of the above modules.

[0106] First, such as Figure 2 As shown in the example, this method provides a training approach for a large-scale materials engineering model, comprising three consecutive training phases. This method aims to address technical issues in existing large-scale materials engineering model training, such as incomplete knowledge coverage, insufficient reasoning ability, and inconsistent answer quality. Through a systematic three-stage training strategy, it significantly improves the model's professional capabilities in the field of materials engineering.

[0107] In the first stage, the basic large language model is pre-trained using text incrementally.

[0108] First, a multidisciplinary text dataset is constructed, encompassing text data from fields such as materials science, chemistry, physics, and biology. Specifically, materials science journal articles include, but are not limited to, full-text content from top journals such as *Materials Science and Engineering*, *Acta Materialia*, and *Materials Today*; chemistry database texts include, but are not limited to, structured text information such as compound descriptions, reaction mechanisms, and synthetic pathways from specialized databases such as PubChem, Reaxys, and SciFinder; physics journal articles include, but are not limited to, research papers related to the physical properties of materials from journals such as the *Physical Review* series, *Nature Physics*, and *Science*; biomaterials-related texts include, but are not limited to, professional literature on biocompatibility, biodegradability, and cell interactions from journals such as *Biomaterials*, *Acta Biomaterialia*, and *Journal of Biomedical Materials Research*; and materials engineering textbooks and handbooks include, but are not limited to, *Materials Science and Engineering: An Introduction*, *Introduction to Solid State Physics*, and *Polymer Science*. The materials are digitized from classic textbooks such as "Engineering and Engineering"; the material-related descriptions in the patent documents are mainly derived from material-related patent documents published by institutions such as the United States Patent and Trademark Office (USPTO), the European Patent Office (EPO), the China National Intellectual Property Administration (CNIPA), and the World Intellectual Property Organization (WIPO), especially patent texts involving new material synthesis, material modification, and material application.

[0109] The second stage involves four levels of preprocessing on the collected raw text. This preprocessing process consists of four sequentially executed steps. Figure 3 ).

[0110] First, OCR error correction is performed. This process uses LLM_correct from the Large Language Model (LLM) to correct the text D_ocr containing OCR recognition errors, resulting in the corrected text R_ocr, i.e., R_ocr = LLM_correct(D_ocr).

[0111] Because many historical documents and patent files exist in scanned form, OCR recognition often results in character errors and formatting issues, such as recognizing "TiO2" as "TiO2" or "TiOz", or recognizing "σ" as "6". Specially trained error-correcting language models, such as the aforementioned LLM_correct, can accurately identify and correct these errors, restoring the original text's correctness. This error-correction model is specially trained to understand the terminology and expression habits of materials science, thereby improving the accuracy of error correction.

[0112] The next step is noise removal, which involves cleaning the text after OCR correction to remove noise such as special symbols and garbled characters.

[0113] In the process of document digitization, various noises are often introduced, such as scan spots, incomplete edge cropping, and header / footer information mixed into the text. These noises can affect the subsequent model training effect, thus requiring systematic noise removal. By designing specialized noise detection algorithms, these interfering information can be identified and removed while preserving the complete semantics of the text.

[0114] Next is the text deduplication process, which uses the MinHash-LSH algorithm, such as in the LLM model, to deduplicat the cleaned text D_clean. By calculating the MinHash signature and using the Locality Sensitive Hashing technique LSH_filter to filter duplicate text, the deduplicated text T_dedup is obtained, i.e., T_dedup = LSH_filter(MinHash_signatures(D_clean)).

[0115] Large-scale text datasets contain a significant amount of duplicate content, including different versions of the same document, repeated citations in review articles, and duplicate entries in databases. Directly using this duplicate content for training can cause the model to overemphasize certain specific elements, hindering balanced knowledge learning. The MinHash-LSH algorithm effectively identifies and removes semantically repetitive text paragraphs while maintaining computational efficiency, ensuring the diversity of training data.

[0116] Finally, there is the quality filtering process, which involves scoring the deduplicated text based on multiple dimensions of quality assessment indicators and setting reasonable thresholds to filter and retain high-quality text paragraphs. For example, the LLM model can be used with P_quality = Score_threshold (D_dedup, threshold).

[0117] The quality assessment includes multiple dimensions such as language fluency, professional accuracy, and logical consistency. A comprehensive score is used to select the most valuable training data. This quality filtering process ensures the overall quality of the training data and avoids the negative impact of low-quality text on model training.

[0118] After the above four-level preprocessing, word segmentation can be performed using algorithms such as BPE (Byte Pair Encoding) in the LLM model to construct a vocabulary. This preprocessing process also includes text cleaning, format standardization, and language detection steps to ensure data quality.

[0119] For example, when processing chemical formulas, formulas such as "C6H12O6" are converted into more readable text descriptions while retaining the original formatting information. For material property data, a unified numerical representation format is established, converting different unit representations such as "100 MPa," "1×10⁸ Pa," and "1000 bar" into the International System of Units (SI). This series of preprocessing steps ensures the quality and consistency of the training data, laying a solid foundation for subsequent model training.

[0120] The second phase also includes generating deep question chains based on entity information.

[0121] In this application, material-related entities are extracted from a pre-training corpus, including material entities (such as "Ti-6Al-4V alloy"), process entities (such as "hot isostatic pressing"), performance entities (such as "fatigue limit"), and application entities (such as "aircraft engine blade"). The entity extraction process combines domain dictionary rules with a sequence labeling model finely tuned on a materials science corpus to ensure high recall and high accuracy for technical terms.

[0122] Based on the extracted entities, the system first constructs an initial problem, such as "How to optimize the fatigue performance of Ti-6Al-4V alloy?", and initiates a dynamic sub-problem decomposition mechanism. This mechanism does not generate a problem chain of fixed depth, but rather adopts a context-aware recursive decision-making process: in each decomposition step, the original overall problem and the current sub-problem to be processed are input into the large model MatMind, which autonomously determines whether the sub-problem satisfies the atomicity condition—that is, whether it can be directly answered by a single knowledge source, or whether it corresponds to a standard operating unit in materials research and development; if not, the model generates a set of semantically clear, independently solvable lower-level sub-problems, ensuring that the newly generated sub-problems are consistent with the goal of the original overall problem.

[0123] The above decomposition process is executed recursively until all leaf node subproblems meet the atomicity requirement. Subsequently, the system schedules multiple specialized agents to solve each subproblem in parallel, and based on the logical dependencies between subproblems, the answers to each subproblem are merged from bottom to top, ultimately forming a structured, interpretable, and complete answer to the original overall problem.

[0124] This dynamic sub-problem decomposition mechanism has significant technical advantages. First, it simulates the cognitive process of materials science experts when facing complex engineering problems: experts do not mechanically apply problem templates, but dynamically judge "whether decomposition is necessary" and "how to decompose" based on the complexity of the problem. Second, the mechanism avoids over-decomposition or under-decomposition, refining only when necessary, thereby improving reasoning efficiency and logical compactness. Third, since each decomposition step uses the original overall problem as a contextual constraint, it effectively prevents sub-problems from deviating from the core objective. Finally, the mechanism can comprehensively cover the multi-scale knowledge system of materials engineering. For example, when dealing with the complex problem of "optimizing the fatigue performance of titanium alloys," the system can dynamically generate multi-level sub-problems such as "the influence of alloying elements on crack initiation?", "how to control the α / β phase ratio under heat treatment?", and "the correlation between surface shot peening parameters and residual stress?", and construct a closed-loop logical chain of "composition-process-microstructure-performance" during the answer integration stage, significantly improving the model's ability to solve real complex materials engineering problems.

[0125] In the third stage, DPO training data is constructed through multi-agent debate and used to train the MatMind model.

[0126] In this application, technical questions raised by users are first collected from professional exchange forums in the field of materials science and engineering. Participants in these forums include university researchers, corporate engineers, and field experts. To ensure question quality, the system semantically aligns the original questions with structured knowledge sources (including crystallography databases, materials patent databases, and academic literature databases), eliminating low-quality content that is vague, subjective, or lacks verifiable technical elements, and retaining high-quality questions with clear objectives, domain anchors, and solution orientations, thus forming an enhanced set of high-quality questions.

[0127] Subsequently, the enhanced high-quality question set is automatically answered using MatMind, a large-scale materials science model refined with Supervised Fine-tuning (SFT). Based on its embedded materials knowledge system and reasoning capabilities, MatMind generates structured and interpretable technical responses for each question. The system further quantifies and scores the quality of the responses based on multi-dimensional evaluation metrics, including: scientific accuracy (verified by comparison with fact triples in the materials knowledge graph), logical completeness (measured by argument coverage and reasoning chain depth), and engineering feasibility (judged based on process constraints and cost feasibility rule bases). Building upon this, the system reverse-engineers the cognitive difficulty of the original questions by combining the breadth of knowledge invoked during the answering process (e.g., the number of sub-domains involved), the complexity of the reasoning steps, and the depth of external tool invocation. Based on this, the questions are categorized into low, medium, and high difficulty levels, forming a structured question resource pool covering the entire spectrum of materials research and development.

[0128] After classifying the difficulty levels, the system uses questions at each difficulty level as seed inputs to drive a domain-adaptive question generation model for differentiated expansion: for low-difficulty seed questions, it generates variant questions focusing on basic concept understanding and standard process applications; for medium-difficulty seed questions, it generates composite questions involving multi-factor coupling (such as composition-process-performance correlation) or cross-subdomain collaboration (such as mechanics-thermal-chemical coupling); for high-difficulty seed questions, it generates exploratory questions requiring the integration of cutting-edge theories (such as high-entropy effects and topological defect manipulation), multi-scale simulations (such as from atomic scale to component scale), or those targeting unseen material systems. This seed-guided hierarchical generation mechanism ensures that the expanded questions, while maintaining semantic relevance, accurately cover diverse cognitive needs ranging from conventional process optimization to cutting-edge material design.

[0129] The expanded question set serves as input, which is then used by multiple agents to generate answers. This application deploys three specialized agents: a theoretical expert agent, an engineering expert agent, and an experimental expert agent. These agents differ not only in their specialized areas but are also bound to their respective domain knowledge bases and tool invocation capabilities: the theoretical expert agent integrates first-principles calculation rules, phase diagram databases, and thermodynamic models, and can invoke CALPHAD or DFT calculation interfaces; the engineering expert agent has a built-in process constraint library, equipment parameter manual, and cost model, supporting manufacturability assessment; and the experimental expert agent is associated with material characterization standards (such as ASTM and ISO), test method libraries, and error analysis rules, and can parse data logic from SEM, XRD, etc.

[0130] For each deep problem generated through dynamic sub-problem decomposition, three agents generate initial answers in parallel and enter a multi-round debate phase. During the debate, each agent can proactively revise its own position based on the viewpoints of other agents. For example, when an engineering expert points out that a theoretically recommended sintering temperature cannot be achieved in an industrial furnace, the theoretical expert can recalculate the phase stability based on new constraints and update its conclusions; if an experimental expert finds that a performance prediction lacks measurable indicators, it can request supplementary characterization schemes. The system fully records the answers, position changes, and knowledge sources cited by each agent in each round, forming a traceable debate trajectory.

[0131] After the debate, the system automatically scores all responses (including intermediate rounds and the final output) throughout the process using a unified evaluation module. The scoring is based on preset multi-dimensional indicators, including technical correctness (consistency with the material's knowledge graph), logical completeness (unbroken chains of arguments), cross-role coordination (effective response to other perspectives), and practical feasibility (whether executable parameters are included). A weighted summation method is used to calculate the overall score. Where Si represents the score for each dimension, and wi represents the corresponding weight. This evaluation process requires no additional "judgment agent" and does not rely on human intervention.

[0132] Based on the scoring results, the system constructs two types of DPO preference pairs: For the final answer, the one with the highest overall score is selected as the preferred sample (chosen), and those with obvious knowledge conflicts or logical flaws are selected as the second-best sample (rejected); for intermediate answers, if an answer in a certain round effectively corrects a previous error or introduces a key new argument, it is selected as the chosen answer, and the previous answer refuted by it is rejected. All preference pairs together constitute the DPO training dataset.

[0133] Subsequently, the MatMind large model was trained using Direct Preference Optimization (DPO) on this dataset. DPO directly aligns the model output with the preferences of multi-expert collaborative decision-making by maximizing the log-likelihood ratio of the best answer to the second-best answer, without requiring explicit reward modeling or reinforcement learning (RLHF). The training process does not involve reward function design, reward shift, or policy gradient updates; model parameters are optimized solely through standard classification loss.

[0134] The MatMind model trained as described above possesses strong interdisciplinary collaborative reasoning capabilities: when faced with complex problems such as "optimizing the oxidation resistance of high-temperature alloys," it can automatically integrate oxidation thermodynamic mechanisms (theory), coating process feasibility (engineering), and oxide layer characterization and verification schemes (experiment) to generate professional recommendations that are structurally complete, logically consistent, and feasible. This model can be directly deployed in material performance prediction and synthesis process recommendation scenarios, providing users with high-precision and highly reliable intelligent decision support.

[0135] Below, in conjunction with Figure 5 Here is an example flowchart describing a multi-agent-based material property prediction and synthesis method according to an embodiment of this application.

[0136] As shown in the figure, the method mainly includes three stages.

[0137] In the data set generation phase :

[0138] First, in step 502, for example, a recursive decomposition algorithm based on a materials knowledge graph is used to break down the complex problem into multiple subtasks. Each subtask is labeled with key entities and relationships using a predefined domain ontology. The recursive decomposition algorithm uses a structured knowledge base to break down complex problems (such as materials design or performance optimization) into manageable subtasks.

[0139] Subsequently, in step 504, a multi-source parallel retrieval is initiated based on the characteristics of each sub-task to construct a fine-tuned dataset including Sub-CoQ question-answer pairs. The retrieval includes: obtaining crystallographic parameters through structured databases (such as the Materials Project and ICSD); extracting material synthesis methods through academic literature databases (such as Sci-Hub and DOI); and collecting synthesis process parameters through patent databases (such as USPTO and CNIPA). Then, low-relevance search results are filtered out using semantic matching algorithms such as those based on the BERT model. After normalization, the search results are recombined with the multiple sub-task questions into "Sub-CoQ question-answer pairs" in the form of question-evidence pairs. In practical applications, the retrieval process needs to be performed periodically (e.g., once a week) to ensure the timeliness of the fine-tuned dataset.

[0140] Then, in step 506, decision-making conflicts among different roles in materials engineering are simulated, generating a high-quality Direct Preference Optimization (DPO) dataset through debate. Specifically, the DPO technique utilizes binary preference data (i.e., sample pairs labeled "good" or "bad") to directly optimize language model parameters through a classification loss function, implicitly transforming the language model itself into a reward function (see [link to relevant documentation]). Figure 4 (For example, KL divergence constraints are used to prevent the fine-tuned model from deviating from the original model.)

[0141] The specific generation flow of the DPO dataset is as follows:

[0142] In step 506-1, the data from the Materials Engineering Exchange Forum is cleaned. This cleaning can be achieved using a BERT-based classifier.

[0143] In step 506-2, questions are classified according to their complexity and the requirements for answering them. For example, questions can be classified as complete answer type and high-precision type.

[0144] In step 506-3, for the debate threads in the forum, the debate content is extracted and the arguments are constructed into a logical chain (e.g., reorganizing scattered arguments into a tree structure). Different weights can be assigned to different roles in the debate.

[0145] In step 506-4, a multi-dimensional scoring system is designed to quantitatively evaluate the quality of the answers. Specifically, for each dimension, a weighted summation method can be used to calculate the comprehensive score of the answer, i.e.:

[0146]

[0147] Among them, S total This represents the overall score of the answer, where the overall score indicates the quality of the answer. i w represents the scoring indicators for each dimension. i This represents the corresponding weighting coefficient. Finally, all answers are sorted according to their overall score, and the top 10-20% of answers are retained as candidate positive examples.

[0148] In step 506-5, a multi-role argument enhancement strategy is adopted to enhance the role-based text. That is, the existing arguments are first supplemented by simulating the viewpoints of different roles, and then the "Materials PhD Student" role-based intelligent agent reviews and corrects the generated role-based text to ensure the professionalism and accuracy of the content, thereby forming a more comprehensive argument system.

[0149] In steps 506-6, the best answer is generated according to the question's objective and the answer's objective. First, the system analyzes the user's question to identify their intent and context, then retrieves relevant knowledge bases or external data sources to obtain information. During the generation phase, a pre-trained language model is used to generate multiple candidate answers through conditional generation and diversity generation techniques. Subsequently, these candidate answers are evaluated for quality, including logical consistency and relevance to the question, and a combination of automatic and manual evaluation is used to select the answer with the highest score. Finally, the generation model is iteratively optimized based on user feedback to improve the accuracy and satisfaction of subsequent answer generation.

[0150] In steps 506-7, a superior-inferiority pair is constructed by evaluating the quality of the arguments and answers, as well as the commonality of the questions and answers.

[0151] First, the system will evaluate the quality of each argument, taking into account its logical rigor, relevance, and the authority of its source, thereby determining the quality of the argument.

[0152] Meanwhile, the quality of the answer is evaluated by its relevance to the question, clarity, and accuracy.

[0153] Next, the system will compare these evaluation results with common question-and-answer patterns to identify high-frequency effective question-and-answer structures.

[0154] Finally, by comprehensively analyzing these factors, a system of optimal and unoptimized pairs is constructed, forming a standardized evaluation system.

[0155] The generated superior-inferiority pairs serve as a direct source of the DPO dataset, which can be directly converted into the DPO dataset required by DPO.

[0156] It should be understood that in practical applications, the multi-agent debate process also needs to be executed periodically (e.g., once a week) to ensure the timeliness of the DPO dataset.

[0157] After completing the construction of the fine-tuning dataset and the DPO dataset, the method proceeds to the training and validation phase.

[0158] In the training and validation phase :

[0159] First, in step 508, the MatMind large-scale model for materials science and engineering is trained based on the generated dataset. The training includes two phases: (1) Supervised Fine-tuning (SFT); and (2) Reinforcement Learning Optimization (RLHF).

[0160] During the supervised fine-tuning phase, the pre-trained base model is optimized using an evidence attention mechanism based on a fine-tuning dataset containing, for example, 100,000 precisely paired Sub-CoQ question-answer pairs.

[0161] In the reinforcement learning optimization phase (RLHF), the model is deeply trained using the Direct Preference Optimization (DPO) algorithm, based on a dataset such as the DPO dataset generated from 100,000 multi-expert debates. During this deep training, a multi-objective reward function is designed, material thermodynamic constraints are embedded in the policy gradient update process, and the Elastic Weight Consolidation (EWC) algorithm is employed. This enables the trained MatMind model to not only provide technically feasible solutions but also comprehensively consider the practical constraints of industrial application, achieving a complete knowledge transfer from theory to practice. The final recommendation solution output combines academic rigor with engineering practicality.

[0162] Subsequently, in step 510, a two-level evaluation system is used to validate the trained model MatMind. The two-level evaluation system is as follows:

[0163] Level 1: Verify the core capabilities of the model through benchmark testing.

[0164] Level 2: At the same time, an online continuous learning mechanism is deployed in this application.

[0165] Having completed the training and validation of the model, a large-scale model, MatMind, based on multi-agent external information augmentation, has been constructed. The method then proceeds to the model application phase.

[0166] In the model application phase :

[0167] In step 512, after deploying the trained large model MatMind, the input parameters are imported into the large model MatMind based on multi-agent external information enhancement to realize intelligent recommendation of material property prediction and synthesis process.

[0168] To further illustrate the technical advantages of this invention, the following specific application examples demonstrate how the model can be used to solve various problems in the field of materials:

[0169] Example 1: Handling Complex Materials Problems

[0170] For the problem "Design a new titanium alloy for aerospace applications, requiring excellent high-temperature strength and good machinability", traditional single-agent models may only provide answers from a single perspective. However, the model trained by this application can comprehensively consider multiple aspects such as alloy design theory, processing requirements, and performance testing methods, and provide a complete solution including alloy composition design, heat treatment process, and performance verification scheme.

[0171] Example 2: Integration of interdisciplinary problems

[0172] Regarding the question "explain the application mechanism of graphene in the biomedical field", the model can integrate knowledge from multiple disciplines such as materials science, biology, and medicine, and conduct in-depth analysis from multiple levels such as the physicochemical properties of graphene, biocompatibility mechanism, cell interaction mechanism, and clinical application prospects.

[0173] Example 3: Engineering Practice Guidance

[0174] Regarding the question "How to optimize the surface treatment process of 3D printed titanium alloy implants", the model can provide practical and guiding process optimization suggestions by combining the characteristics of 3D printing technology, the properties of titanium alloy materials, the requirements of biomedical applications, surface modification technology, and other aspects.

[0175] Example 4: Performance Prediction Task

[0176] When predicting the bandgap value of semiconductor materials under the generalized gradient approximation method, MatMind's MAD / MAE ratio is 2.47, indicating that the model is relatively reliable and superior to other models.

[0177] Although the techniques have been described using language specific to structural features and / or methodological actions, it should be understood that the appended claims are not necessarily limited to the described features or actions. Rather, these features and actions are described as exemplary forms of implementing these techniques.

[0178] The operations of the example processes are shown in separate boxes and are summarized with reference to these boxes. These processes are shown as a flow of logical boxes, each of which may represent one or more operations that can be implemented using hardware, software, or a combination thereof. In the context of software, these operations represent computer-executable instructions stored on one or more computer-readable media that, when executed by one or more processors, cause one or more processors to perform a given operation. Generally, computer-executable instructions include routines, programs, objects, modules, components, data structures, etc., that perform a particular function or implement a particular abstract data type. The order in which the operations are described is not intended to be construed as limiting, and any number of the operations may be executed in any order, combined in any order, subdivided into multiple sub-operations, and / or executed in parallel to implement the described process. The described process may be executed by resources associated with one or more computing devices, such as one or more internal or external CPUs or GPUs, and / or one or more pieces of hardware logic, such as FPGAs, DSPs, or other types of accelerators.

[0179] All of the methods and processes described above can be embodied in software code modules executed by one or more general-purpose computers or processors, and can be fully automated via these software code modules. These code modules can be stored on any type of computer-executable storage medium or other computer storage device. This code can also be packaged into corresponding computer program products. Some or all of these methods can alternatively be embodied in dedicated computer hardware.

[0180] Any routine description, element, or box in the flowcharts described herein and / or in the accompanying drawings should be understood as potentially representing a module, segment, or portion of code comprising one or more executable instructions for implementing a specific logical function or element in that routine. Alternative implementations are included within the scope of the examples described herein, wherein elements or functions may be removed or performed inconsistently with the order shown or discussed, including substantially synchronous or reverse order execution, depending on the functionality involved, as will be understood by those skilled in the art.

[0181] While different embodiments have been described above, it should be understood that they are merely examples and not limitations. Those skilled in the art will appreciate that various modifications in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. A multi-agent-based material property prediction and synthesis system, comprising: The multi-agent data augmentation module is configured to first break down complex problems into multiple sub-tasks, and then construct a fine-tuned dataset including Sub-CoQ question-answer pairs by initiating multi-source parallel retrieval. The multi-expert debate module is configured to simulate decision-making conflicts among different roles in materials engineering, and generates direct preference optimization DPO datasets through debate; The training and validation module is configured to train and validate large-scale models in the materials domain, MatMind, based on the fine-tuning dataset and the DPO dataset, and utilizing supervised fine-tuning SFT and reinforcement learning RLHF. as well as The material property prediction and synthesis module is configured to intelligently recommend material property prediction and synthesis processes by importing input parameters into the large model MatMind.

2. The material property prediction and synthesis system as described in claim 1, characterized in that, The multi-agent data augmentation module includes: Sub-CoQ agents are configured to use a model-knowledge-based recursive decomposition algorithm to break down complex problems into multiple subtasks. The data retrieval recall agent is configured to initiate multi-source parallel retrieval based on the features of each subtask to build a fine-tuned dataset including Sub-CoQ question-answer pairs; The step of initiating multi-source parallel retrieval based on the characteristics of each sub-task to construct a fine-tuned dataset including Sub-CoQ question-answer pairs includes: obtaining crystallographic parameters through a structured database, extracting material synthesis methods through an academic literature database, and collecting synthesis process parameters through a patent database; filtering out low-relevance search results using a semantic matching algorithm based on a dual-tower model, and recombining them with multiple sub-task questions into corresponding question-evidence pairs; and finally constructing the corresponding fine-tuned dataset based on the Sub-CoQ question-answer pairs.

3. The material property prediction and synthesis system as described in claim 2, characterized in that, The Sub-CoQ agent is configured to recursively decompose subtasks until the generated sub-problems satisfy preset atomicity conditions. These atomicity conditions include: the problem can be answered by a single data source, or the problem corresponds to a standard operating unit in material development.

4. The material property prediction and synthesis system as described in claim 1, characterized in that, The simulated decision-making conflicts among different roles in materials engineering, which generate a Direct Preference Optimization (DPO) dataset through debate, include: Predefined roles are assigned to multiple specialized intelligent agents, including "synthetic chemist", "process engineer", "performance testing expert", "computational materials scientist", etc., with each role corresponding to a specific domain knowledge base and decision preferences; Given a material development problem, each agent independently generates candidate answers based on its role and position. Through a multi-round interactive debate mechanism, each agent questions, refutes, or supplements the answers of other agents, and uses a sequence labeling model to extract key arguments from the debate content. The scattered arguments are reorganized into a tree-like argument structure according to their logical dependencies, and the logical consistency between the arguments is verified by combining the material knowledge graph. Based on a pre-set multi-dimensional scoring system (including scientific validity, feasibility, interdisciplinary consistency, and data support), each candidate answer is quantitatively evaluated. A multi-role argument enhancement strategy is adopted to guide the agent to supplement missing arguments from different professional perspectives, thereby optimizing the completeness of the response. By combining the scoring results with the question objectives, the optimal answer is determined, and the second-best answer is selected to form a preference pair; Based on the optimal and suboptimal answers, good-bad sample pairs are constructed for Direct Preference Optimization (DPO) training.

5. The material property prediction and synthesis system as described in claim 4, characterized in that, The multi-expert debate module performs processing periodically to maintain the timeliness of the DPO dataset.

6. The material property prediction and synthesis system as described in claim 1, characterized in that, The training process of the training and validation module is as follows: In the supervised fine-tuning SFT stage: based on the fine-tuning dataset, the pre-trained base model is optimized using instruction fine-tuning; In the reinforcement learning optimization RLHF stage, the large model MatMind is deeply trained using the direct preference optimization algorithm based on the DPO dataset. The verification process for the training and verification module adopts the following two-level evaluation system: Level 1: Verify the model's core capabilities through benchmark testing; Level 2: Deploy an online continuous learning mechanism to automatically acquire the latest published materials science papers and patent data daily, and update the model knowledge base in real time after intelligent filtering and annotation.

7. The material property prediction and synthesis system as described in claim 1, characterized in that, The pre-trained base model is Qwen-2.5 7B.

8. A method for predicting and synthesizing material properties based on multi-agent systems, comprising: Break down complex problems into multiple sub-tasks; Multi-source parallel retrieval is initiated based on the characteristics of each subtask to construct a fine-tuned dataset including Sub-CoQ question-answer pairs; Simulate decision-making conflicts among different roles in materials engineering, and generate a Direct Preference Optimization (DPO) dataset through debate; The large model MatMind was trained based on the generated fine-tuning dataset and the DPO dataset; Validate the trained large model MatMind; as well as After deploying the trained large model MatMind, the input parameters are imported into the large model MatMind to realize material property prediction and intelligent recommendation of synthesis process.

9. The material property prediction and synthesis method as described in claim 8, characterized in that, The simulated decision-making conflicts among different roles in materials engineering, which generate a Direct Preference Optimization (DPO) dataset through debate, include: Predefined roles are assigned to multiple specialized intelligent agents, including "synthetic chemist", "process engineer", "performance testing expert", "computational materials scientist", etc., with each role corresponding to a specific domain knowledge base and decision preferences; Given a material development problem, each agent independently generates candidate answers based on its role and position. Through a multi-round interactive debate mechanism, each agent questions, refutes, or supplements the answers of other agents, and uses a sequence labeling model to extract key arguments from the debate content. The scattered arguments are reorganized into a tree-like argument structure according to their logical dependencies, and the logical consistency between the arguments is verified by combining the material knowledge graph. Based on a pre-set multi-dimensional scoring system (including scientific validity, feasibility, interdisciplinary consistency, and data support), each candidate answer is quantitatively evaluated. A multi-role argument enhancement strategy is adopted to guide the agent to supplement missing arguments from different professional perspectives, thereby optimizing the completeness of the response. By combining the scoring results with the question objectives, the optimal answer is determined, and the second-best answer is selected to form a preference pair; Based on the optimal and suboptimal answers, good-bad sample pairs are constructed for Direct Preference Optimization (DPO) training.

10. The material property prediction and synthesis method as described in claim 8, characterized in that, Each subtask annotates key entities and relationships using a predefined domain ontology.

Images