A dual-model cooperative lightweight material science intelligent research method and system

Abstract

The application relates to the technical field of computer systems, in particular to a lightweight material science intelligent research method and system based on double model cooperation, which comprises a user interaction interface, a double model cooperation intelligent agent core, a material science professional tool set and a bottom layer computing resource layer; the double model cooperation intelligent agent core comprises an execution coordination model and a knowledge core model; the execution coordination model and the knowledge core model are both self-recurrent language models with a parameter scale less than 100B; the execution coordination model analyzes a research task, understands a core intention and transmits a preliminary result to the knowledge core model; the knowledge core model performs deep analysis, interpretation and reasoning on the preliminary result, including physical rationality verification, and performs iterative decision-making. The application realizes decoupling and deep optimization of knowledge reasoning and complex tool chain operation in the field of material science, and aims to build an automatic and intelligent research platform with flexible architecture, strong performance and the ability to significantly accelerate the material discovery and development process.

Description

Technical Field

[0001] This application relates to the field of computer system technology, and in particular to a dual-model collaborative intelligent research method and system for lightweight materials science. Background Technology

[0002] Traditional materials science research relies heavily on experimental trial and error and expert experience, resulting in long development cycles and high costs. With the development of artificial intelligence (AI) technology, using AI to accelerate scientific discovery has become a significant trend. Particularly in materials science, AI can greatly improve research and development efficiency by integrating massive amounts of data, predicting material properties, and optimizing experimental procedures.

[0003] In recent years, large language models (LLMs) have offered the possibility of building more advanced automated scientific research platforms due to their powerful natural language understanding, reasoning, and planning capabilities. Researchers have begun to explore using LLMs as the "brain" to direct and schedule various software tools and hardware devices to simulate or even surpass the research processes of human experts. Currently, existing research has proposed an automated scientific research system based on a general-purpose large language model, called Coscientist. Coscientist is an AI agent system driven by a large-scale general-purpose language model (GPT-4). Its core architecture can be understood as a robotic scientist with a "brain" and "hands." The "brain," GPT-4, is responsible for high-level thinking, planning, and decision-making; the "hands" are a series of encapsulated external tools, including information tools (such as search engines), code tools (such as code interpreters), and hardware tools (such as automated chemical experiment platform APIs). Its workflow follows a "reasoning-action" cycle, autonomously completing a complete scientific research loop from literature review and experimental design to robot control through continuous thinking, tool usage, and result observation.

[0004] However, the scheme still has several key limitations in its architecture: First, it relies on a general large model, which is insufficient in the depth of domain knowledge when dealing with highly specialized materials science problems; second, its linear "reasoning-action" loop needs to be improved in terms of flexibility and efficiency when dealing with complex, nonlinear iterative research tasks; and third, its single-model architecture has inherent bottlenecks when simultaneously optimizing the two different capabilities of deep scientific reasoning and efficient tool execution. Summary of the Invention

[0005] This application provides a lightweight intelligent research method and system for materials science with dual-model collaboration, aiming to solve the fundamental problems of existing large language models when applied to complex materials science research, such as insufficient depth of domain knowledge, lack of professional tool calling ability, rigid system architecture, and difficulty in balancing knowledge reasoning and tool execution efficiency in single-model architecture.

[0006] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide a lightweight intelligent research system for materials science with dual-model collaboration, comprising: a user interaction interface, a dual-model collaborative intelligent agent core, a materials science professional toolset, and a bottom-level computing resource layer; the user interaction interface is used to receive research tasks input by the user in natural language; the dual-model collaborative intelligent agent core includes an execution coordination model and a knowledge core model; both the execution coordination model and the knowledge core model are autoregressive language models with a parameter scale of less than 100B, wherein the parameter scale of the knowledge core model is larger than that of the execution coordination model; the execution coordination model is used to parse the research task, understand its core intent, obtain preliminary results, and transmit the preliminary results to the knowledge core model; the knowledge core model is used to perform in-depth analysis, interpretation, and reasoning on the preliminary results, including physical rationality verification, and to make iterative decisions based on the results of in-depth analysis and physical rationality verification; the physical rationality verification includes: syntactic parsing verification of the crystal structure file CIF, rationality verification of crystal geometric parameters, and convex hull energy E of thermodynamic stability. hull Threshold verification; the materials science professional toolset is a software tool and database interface encapsulated in accordance with a standardized model context protocol; the materials science professional toolset includes multiple tools that are deployed as services that can be called via API; the underlying computing resource layer provides physical computing support for the research system.

[0007] In some exemplary embodiments, the underlying computing resource layer includes a central processing unit, a graphics processing unit, memory, and storage.

[0008] Secondly, this application provides a dual-model collaborative intelligent research method for lightweight materials science. This method is based on the dual-model collaborative intelligent research system for lightweight materials science described in the above embodiments. The method includes the following steps: Step S1: Receiving a research task and using an execution coordination model to parse the research task to obtain the task intent; Step S2: The execution coordination model autonomously plans and executes preliminary operations according to the task intent to obtain preliminary results; Step S3: The execution coordination model submits the obtained preliminary results and the research task to a knowledge core model; The knowledge core model performs in-depth analysis, interpretation, and reasoning on the preliminary results and the research task; The in-depth analysis includes physical rationality verification of the preliminary results, obtaining in-depth analysis results containing physical rationality verification results; The physical rationality verification includes: grammatical parsingability verification of the crystal structure file CIF, rationality verification of crystal geometric parameters, and convex hull energy E of thermodynamic stability. hullThreshold verification; Step S4: Based on the deep analysis results and the physical rationality verification results, the knowledge core model makes iterative decisions; If the knowledge core model determines that the current information is sufficient to generate the final answer, the knowledge core model will directly integrate all information to form a clear and scientifically insightful final report and output it to the user, and the process ends; If the knowledge core model determines that further calculations or information supplementation are needed to solve the problem, the knowledge core model will generate one or a series of new, more specific sub-task instructions; Step S5: The newly generated sub-task instructions are returned to the execution coordination model, and the system process returns to step S2. The execution coordination model parses the newly generated sub-task instructions, calls the corresponding tools to execute them, and submits the newly obtained results back to the knowledge core model for analysis. After closed-loop feedback and iterative execution, the process continues until the knowledge core model determines that all necessary information has been collected and a final conclusion can be formed.

[0009] In some exemplary embodiments, the execution coordination model autonomously plans and executes preliminary operations based on the task intent to obtain preliminary results, including: if the task intent is open exploration, the execution coordination model calls information retrieval tools to collect background knowledge; if the task intent is specific calculation, the execution coordination model selects and calls appropriate calculation tools to obtain structured execution results.

[0010] In some exemplary embodiments, the method for constructing the knowledge core model includes: integrating multi-source heterogeneous data through an automated data processing flow to construct a dataset; and performing full-parameter supervised fine-tuning of the knowledge core model based on a lightweight basic model that combines parametric inference capabilities and cost-effectiveness, as well as the dataset.

[0011] In some exemplary embodiments, the lightweight base model includes one or more of the Qwen2 series, Llama 2 series, Qwen3 series, Llama 3 series, and ChatGLM4.

[0012] In some exemplary embodiments, a dataset is constructed by integrating heterogeneous data from multiple sources through an automated data processing flow, including: obtaining crystal structure and computational performance data from public crystallography databases and associating them with academic literature using digital object identifiers; extracting unstructured text of synthesis methods and application scenarios from the literature using natural language processing technology; and constructing a high-quality instruction fine-tuning question-and-answer dataset containing full-link information on structure, performance, synthesis, and application using a "generate-distillation" strategy.

[0013] In some exemplary embodiments, the method for constructing the coordination model includes: constructing a standardized model context protocol platform, encapsulating all external tools into a unified API interface service, and explicitly defining the function, input parameters, and output format of each tool; selecting a computationally efficient lightweight base model and training it using a reinforcement learning algorithm; designing a composite reward function that includes the correctness of tool call format, syntactic executability, problem-solving efficiency, and the accuracy of the final answer, and training the model through multiple rounds of autonomous tool call training on a large number of tasks, enabling the model to learn how to plan complex tool sequences, handle tool errors, and efficiently complete specified tasks.

[0014] In some exemplary embodiments, the lightweight base model includes one or more of the Qwen2 series, Llama 2 series, Qwen3 series, Llama 3 series, and ChatGLM4.

[0015] In some exemplary embodiments, a reinforcement learning algorithm with generalized reward policy optimization is used for training. During training, the model autonomously generates tool call sequences for each problem, interacts with the tool server through a standardized tool call protocol, and continuously optimizes its policy network based on feedback from a composite reward function. The composite reward function is: R total = w1·R turns + w2·R think + w3·R format + w4·R syntax Among them, R turns R is a hierarchical step function based on the number of tool interaction rounds n; when round n reaches a preset depth threshold k, turns Take the maximum value, and the depth threshold k is not less than 4; R think Let R be the hyperbolic tangent function of the average token length L based on the content of the thought tag. think = tanh(L / λ), where λ is the preset scaling factor; R format To enforce a strict time constraint between thinking and action, meaning that at each intermediate step, the thought process must strictly precede the content invoked by the tool; R syntax The verification is determined through a two-layer check: the first layer is the existence check of the tool registry; the second layer is the parameter integrity and semantic resolvability check based on the Pydantic schema, and the semantic resolvability check includes the verification of the resolvability of the generated CIF crystal structure file.

[0016] Specifically, this application aims to solve the following technical problems: (1) Model capability bottleneck problem: Existing solutions either rely on a single general large model, which lacks depth and professionalism in domain knowledge; or adopt homogeneous model configurations, failing to carry out specialized and heterogeneous model design and optimization for the two very different task requirements of "deep knowledge reasoning" and "efficient tool execution", resulting in limited overall system performance.

[0017] (2) Incomplete technical solutions: Some technical solutions only focus on enhancing the ability of a single model to understand professional information, without providing a complete system-level solution that can autonomously plan and orchestrate complex multi-step toolchains.

[0018] (3) Rigid architecture problem: Existing multi-agent systems mostly adopt a linear "pipeline" architecture, with tasks being passed in a fixed order. They lack the flexibility to conduct dynamic feedback, iteration, and task replanning during the research process, and cannot adapt to the complexity and variability of real scientific research scenarios.

[0019] The technical solution provided in this application has at least the following advantages: This application provides a lightweight intelligent research method and system for materials science using a dual-model collaborative approach. The system includes: a user interface, a dual-model collaborative intelligent agent core, a materials science professional toolset, and a lower-level computing resource layer. The user interface receives research tasks input by the user in natural language. The dual-model collaborative intelligent agent core includes an execution coordination model and a knowledge core model. Both the execution coordination model and the knowledge core model are autoregressive language models with a parameter size less than 100 bytes, wherein the parameter size of the knowledge core model is larger than that of the execution coordination model. The execution coordination model parses the research task, understands its core intent, obtains preliminary results, and transmits the preliminary results to the knowledge core model. The knowledge core model performs in-depth analysis, interpretation, and reasoning on the preliminary results, including physical rationality verification, and makes iterative decisions based on the results of the in-depth analysis and physical rationality verification. The physical rationality verification includes: grammatical parsingability verification of the crystal structure file (CIF), rationality verification of crystal geometric parameters, and convex hull energy E for thermodynamic stability. hull Threshold verification; the materials science professional toolset is a software tool and database interface encapsulated in accordance with a standardized model context protocol; the materials science professional toolset includes multiple tools that are deployed as services that can be called via API; the underlying computing resource layer provides physical computing support for the research system.

[0020] This application provides a lightweight materials science intelligent research method and system with dual-model collaboration, achieving a flexible and dynamically iterative system architecture. One objective of this application is to construct a nonlinear closed-loop iterative workflow. Through this workflow, the system can autonomously perform feedback, adjustment, and task replanning based on intermediate results, thereby realistically simulating and accelerating the scientific exploration process to adapt to the complexity and variability of real scientific research.

[0021] On the other hand, it achieves decoupling and collaborative optimization of model capabilities. The core objective of this application is to decouple the capabilities of "deep knowledge reasoning" and "efficient tool execution" by constructing a functionally heterogeneous dual-model core. Through specially optimized "knowledge core model" and "execution coordination model" each performing their respective functions, it overcomes the bottlenecks of existing technologies in knowledge depth and tool orchestration capabilities, and enables the two to work collaboratively to achieve overall system performance of 1+1>2, while significantly reducing system deployment costs by reducing model parameters.

[0022] Furthermore, this application provides a complete, end-to-end automated research platform. The ultimate goal of this application is to provide a comprehensive automated research platform that seamlessly integrates the entire process from the formulation of scientific hypotheses, the autonomous invocation of complex toolchains, the comprehensive analysis of multi-source data, to the generation of final scientific conclusions. In this way, the professional barriers and time cycles of materials research and development are significantly reduced, accelerating the process of scientific discovery. Attached Figure Description

[0023] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0024] Figure 1 This is a schematic diagram of the structure of a lightweight materials science intelligent research system with dual-model collaboration, provided as an embodiment of this application.

[0025] Figure 2 This is a flowchart illustrating a method for constructing a knowledge core model according to an embodiment of this application.

[0026] Figure 3 This is a schematic diagram of a method for constructing an execution coordination model according to an embodiment of this application.

[0027] Figure 4 This is a flowchart illustrating a dual-model collaborative intelligent research method for lightweight materials science provided in an embodiment of this application.

[0028] Figure 5 This is a schematic diagram showing the system performance comparison and evaluation results with mainstream general-purpose large models provided in an embodiment of this application.

[0029] Figure 6 A comparison chart of deployment costs for different parameter models provided in one embodiment of this application.

[0030] Figure 7 The diagram shows the Cs2ErAgBr6 structure generated by the intelligent system provided in one embodiment of this application.

[0031] Figure 8 A ternary phase diagram of Li2ZrCl6 structure generated by an intelligent system provided in an embodiment of this application. Detailed Implementation

[0032] As can be seen from the background technology, existing large model machine agent technology still has several key limitations in its architecture: First, it relies on a general large model, which is insufficient in the depth of domain knowledge when dealing with highly specialized materials science problems; second, its linear "reasoning-action" loop needs to be improved in terms of flexibility and efficiency when dealing with complex, nonlinear iterative research tasks; and third, its single model architecture has inherent bottlenecks when simultaneously optimizing two different capabilities: deep scientific reasoning and efficient tool execution.

[0033] The related technology provides a multi-agent system composed of multiple agents for the development of water purification materials. The system includes multiple agents responsible for material design, evaluation and screening, synthesis and generation, characterization, operational condition matching, effect verification, and mechanism mining. These agents are all autoregressive large language models trained on a domain knowledge base. Its operating mode is a linear pipeline: the task starts with the "material design agent," and its output is sequentially passed to the next agent until the last agent outputs the final development report. This technology constructs a linear, unidirectional "pipeline" multi-agent system, which has shortcomings in terms of architectural flexibility and dynamic task adjustment.

[0034] Another related technology proposes a method for developing fiber masterbatch formulations based on a localized large model and a retrieval-enhanced generative framework. Its core technology is a "multimodal input embedding module," which encodes chemical structures and physical properties into vectors using techniques such as graph neural networks. These vectors, along with text, are then input into a large model, aiming to enable the single large model to directly understand the microstructure of molecules, thereby making formulation predictions more accurate. However, this technology focuses on optimizing the chemical information input capability of a single model and does not provide a complete, system-level solution for organizing and orchestrating complex multi-step toolchains.

[0035] Another related technology proposes a catalyst screening method that combines the Alpha-Beta pruning algorithm with a large language model. This method leverages the LLM's ability to represent material properties and the Alpha-Beta algorithm's search space pruning capabilities. Through the collaborative work of these two functional modules, it aims to efficiently screen promising catalytic materials from a vast pool of candidate combinations. However, this approach has a relatively specific application scenario, focusing on accelerating the specific task of material screening, and has not yet constituted an automated system covering the entire material research and development lifecycle.

[0036] Based on the analysis of the above background technology and the closest similar solutions, the existing technology, when applied to complex materials science research, mainly suffers from the following three drawbacks that are directly related to this application and can be resolved by this application: (1) Rigid architectural design, unable to adapt to the dynamic and iterative nature of real scientific research. Scientific research is essentially a non-linear exploratory process, full of hypotheses, verification, failures, feedback, and adjustments. However, whether it is a "reasoning-action" chain or a "pipeline" system, its inherent task execution logic is linear and unidirectional. Such rigid architectures are difficult to dynamically and autonomously adjust subsequent research strategies or backtrack to previous steps based on the results of intermediate calculations or analyses. This greatly limits their practical value in dealing with complex exploratory research tasks that require repeated iterations.

[0037] (2) The model's capabilities have bottlenecks, making it difficult to balance knowledge depth and execution efficiency. Existing solutions involve fundamental compromises in model architecture. Reliance on a single, general-purpose model results in insufficient depth of professional knowledge in materials science, hindering profound scientific insights. While other solutions employ multiple models, these models are homogeneous. None of these solutions address a core contradiction: the optimal architecture and training methods for a model adept at creative, knowledge-intensive scientific reasoning and a model adept at precise, structured, multi-step tool invocation are drastically different. Allowing a single or homogeneous model to simultaneously undertake both tasks inevitably leads to performance trade-offs and bottlenecks.

[0038] (3) Fragmented technical solutions, lacking complete end-to-end system-level solutions. Although some existing technologies have made progress in specific aspects, such as optimizing the efficiency of material screening or enhancing the model's understanding of chemical information input, these are essentially "point" solutions. They optimize a certain isolated link in the entire scientific research process. They do not provide a "surface" platform system that can seamlessly integrate the entire process, including scientific hypotheses, multi-step tool calculations, data analysis, iterative optimization, and conclusion generation.

[0039] To address the shortcomings of existing technologies, such as rigid architecture, model capability bottlenecks, and fragmented solutions, this application provides a lightweight intelligent research method and system for materials science with dual-model collaboration. The system includes: a user interface, a dual-model collaborative intelligent agent core, a materials science professional toolset, and a lower-level computing resource layer. The user interface receives research tasks input by the user in natural language. The dual-model collaborative intelligent agent core includes an execution coordination model and a knowledge core model. Both the execution coordination model and the knowledge core model are autoregressive language models with a parameter size of less than 100 bytes, wherein the parameter size of the knowledge core model is larger than that of the execution coordination model. The execution coordination model parses the research task, understands its core intent, obtains preliminary results, and transmits the preliminary results to the knowledge core model. The knowledge core model performs in-depth analysis, interpretation, and reasoning on the preliminary results, including physical rationality verification, and makes iterative decisions based on the results of in-depth analysis and physical rationality verification. The physical rationality verification includes: grammatical parsingability verification of the crystal structure file (CIF), rationality verification of crystal geometric parameters, and convex hull energy E for thermodynamic stability. hull Threshold verification; the materials science professional toolset is a software tool and database interface encapsulated according to a standardized model context protocol; the materials science professional toolset includes multiple tools, which are deployed as services that can be called via API; the underlying computing resource layer provides physical computing support for the research system. This application proposes an innovative "dual-model collaborative" intelligent agent system architecture and method to achieve decoupling and deep optimization of knowledge reasoning and complex toolchain operations in the field of materials science, aiming to create an automated and intelligent research platform with flexible architecture, powerful performance, and the ability to significantly accelerate the materials discovery and development process.

[0040] The purpose of this application is to provide an innovative dual-model collaborative intelligent research system and method for materials science, in order to systematically overcome the limitations of existing solutions. The specific objectives are as follows: (1) To achieve a flexible and dynamically iterative system architecture. One of the objectives of this application is to construct a non-linear closed-loop iterative workflow. Through this workflow, the system can autonomously perform feedback, adjustment, and task replanning based on intermediate results, thereby realistically simulating and accelerating the scientific exploration process to adapt to the complexity and variability of real scientific research.

[0041] (2) Decoupling and Co-optimization of Model Capabilities. The core objective of this application is to decouple the capabilities of "deep knowledge reasoning" and "efficient tool execution" by constructing a dual-model core with heterogeneous functions. Through specially optimized "knowledge core model" and "execution coordination model" that each perform their respective functions, the bottlenecks of existing technologies in knowledge depth and tool orchestration capabilities are overcome, and the two work together to achieve overall system performance of 1+1>2, and the system deployment cost is significantly reduced by shrinking the model parameters.

[0042] (3) Provide a complete, end-to-end automated research platform. The ultimate goal of this application is to provide a complete automated research platform that seamlessly integrates the entire process from the formulation of scientific hypotheses, the autonomous invocation of complex toolchains, the comprehensive analysis of multi-source data, to the generation of final scientific conclusions. In this way, the professional threshold and time cycle of materials research and development are significantly reduced, and the process of scientific discovery is accelerated.

[0043] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0044] See Figure 1 This application provides a lightweight intelligent research system for materials science with dual-model collaboration, including: a user interface 101, a dual-model collaborative intelligent agent core 102, a materials science professional toolset 103, and a bottom-level computing resource layer (bottom-level resources) 104; the user interface 101 is used to receive research tasks input by the user in natural language; the dual-model collaborative intelligent agent core 102 includes a knowledge core model 102a and an execution coordination model 102b; both the knowledge core model 102a and the execution coordination model 102b are autoregressive language models with a parameter scale of less than 100B, wherein the knowledge core model... The parameter size of model 102a is larger than that of the execution coordination model 102b. The execution coordination model 102b is used to analyze the research task, understand its core intent, obtain preliminary results, and transmit these preliminary results to the knowledge core model 102a. The knowledge core model 102a is used to perform in-depth analysis, interpretation, and reasoning on the preliminary results, including physical rationality verification, and to make iterative decisions based on the results of in-depth analysis and physical rationality verification. Physical rationality verification includes: verification of the syntactic parsingability of the crystal structure file CIF, verification of the rationality of crystal geometric parameters, and verification of the convex hull energy E for thermodynamic stability. hullThreshold verification; Materials Science Professional Toolset 103 is a software tool and database interface encapsulated in accordance with a standardized model context protocol; Materials Science Professional Toolset 103 includes multiple tools that are deployed as services that can be called via API; The underlying computing resource layer 104 provides physical computing support for the research system.

[0045] Figure 1 This paper illustrates the four core components of the system provided in this application and their interrelationships. Users input natural language tasks through the user interface 101. The tasks are then passed to the dual-model collaborative agent core 102, which is the "brain" of the system. This core contains a knowledge core model 102a responsible for deep scientific reasoning and an execution coordination model 102b responsible for efficient tool orchestration. The agent core 102 invokes a materials science-specific toolset 103 to perform specific computational and retrieval tasks. The entire system operates on top of the underlying computing resource layer 104. Figure 1 The arrows clearly illustrate the core closed-loop iterative workflow of this invention.

[0046] Specifically, the user interaction interface 101 serves as the entry point for users to interact with the system. Technically, it can be a web-based graphical user interface or a command-line interface. Users input research tasks using natural language through this interface. For example, a user could input the text: "Please design a novel ternary metal sulfide catalyst for electrocatalytic nitrogen fixation and evaluate its stability." This interface receives this text string and passes it to the dual-model collaborative agent core 102 for processing.

[0047] The dual-model collaborative intelligent agent core 102 is the "brain" of this application, serving as the decision-making and control center of the entire system. It consists of two functionally heterogeneous, collaborative large language model intelligent agents: the knowledge core model 102a and the execution coordination model 102b.

[0048] The Materials Science Toolset 103, serving as the system's "hands," is a series of software tools and database interfaces encapsulated according to standardized model context protocols. These tools are deployed as services accessible via APIs, specifically including: Information retrieval tools, such as web search engines and professional literature database search engines, are used to obtain public information and academic literature.

[0049] Structural processing tools include tools for crystal structure generation, structural relaxation, symmetry analysis, and structural similarity comparison.

[0050] Performance prediction tool: Integrates multiple machine learning atomic potential models (such as CHGNet, M3GNet, MEGNet) for rapid prediction of material formation energy, band gap, mechanical properties, etc.

[0051] Analysis and visualization tools: including phase diagram construction tools, XRD diffraction pattern simulation tools, data analysis and plotting tools, etc.

[0052] The underlying computing resource layer 104 provides physical computing support for the entire system, including a central processing unit (CPU), a graphics processing unit (GPU), memory, and storage. High-performance GPU clusters are required for model training and inference; for tool execution, appropriate computing resources are configured according to the specific needs of the tool.

[0053] This application also provides a dual-model collaborative intelligent research method for lightweight materials science. This method is based on the dual-model collaborative intelligent research system for lightweight materials science described in the above embodiments, and is characterized by including the following steps: Step S1: Receive the research task and use the execution coordination model to parse the research task to obtain the task intent.

[0054] Step S2: The execution coordination model autonomously plans and executes preliminary operations based on the task intent to obtain preliminary results.

[0055] Step S3: The execution coordination model submits the obtained preliminary results and the research task to the knowledge core model; the knowledge core model performs in-depth analysis, interpretation, and reasoning on the preliminary results and the research task; the in-depth analysis includes physical rationality verification of the preliminary results, obtaining in-depth analysis results containing physical rationality verification results; the physical rationality verification includes: syntactic parsingability verification of the crystal structure file CIF, rationality verification of crystal geometric parameters, and convex hull energy E of thermodynamic stability. hull Threshold verification.

[0056] Step S4: Based on the deep analysis results and the physical rationality verification results, the knowledge core model makes iterative decisions; if the knowledge core model determines that the current information is sufficient to generate the final answer, the knowledge core model will directly integrate all information to form a clear and scientifically insightful final report and output it to the user, and the process ends; if the knowledge core model determines that further calculations or information supplementation are needed to solve the problem, the knowledge core model will generate one or a series of new and more specific sub-task instructions. Step S5: The newly generated subtask instruction is returned to the execution coordination model, and the system flow returns to step S2. The execution coordination model parses the newly generated subtask instruction, calls the corresponding tool to execute it, and submits the newly obtained results to the knowledge core model for analysis. After closed-loop feedback and iterative execution, the process continues until the knowledge core model determines that all necessary information has been collected and a final conclusion can be formed.

[0057] In some embodiments, in step S2, the execution coordination model autonomously plans and executes preliminary operations according to the task intent to obtain preliminary results, including: if the task intent is open exploration, the execution coordination model calls information retrieval tools to collect background knowledge; if the task intent is specific calculation, the execution coordination model selects and calls the corresponding calculation tools to obtain structured execution results.

[0058] In some embodiments, the method for constructing a knowledge core model includes: integrating multi-source heterogeneous data through an automated data processing flow to construct a dataset; and performing full-parameter supervised fine-tuning of the knowledge core model based on a lightweight basic model that combines parametric reasoning capabilities and cost-effectiveness, as well as the dataset.

[0059] In some embodiments, the lightweight base model is an autoregressive language model with a parameter size of less than 100B, including one or more of the Qwen2 series, Llama 2 series, Qwen3 series, Llama 3 series, and ChatGLM4; preferably, the lightweight base model adopts Qwen3-30B.

[0060] In some embodiments, a dataset is constructed by integrating heterogeneous data from multiple sources through an automated data processing flow, including: obtaining crystal structure and computational performance data from public crystallography databases and associating them with academic literature using digital object identifiers; extracting unstructured text of synthesis methods and application scenarios from the literature using natural language processing technology; and constructing a high-quality instruction fine-tuning question-and-answer dataset containing full-link information on structure, performance, synthesis, and application using a "generate-distillation" strategy.

[0061] In some embodiments, the in-depth analysis in step S3 includes a physical plausibility check of the preliminary results. The physical plausibility check includes at least one of the following: (1) a syntactic parsing check of the crystal structure file (CIF), i.e., determining whether the CIF file contained in the preliminary results can be correctly parsed by a crystallographic analysis tool (such as Pymatgen); (2) a plausibility check of the crystal geometric parameters, i.e., determining whether the bond lengths, bond angles, and space group symmetry conform to crystal chemistry rules; (3) the convex hull energy E of thermodynamic stability. hull Threshold verification involves determining whether the Ehull values ​​of the candidate materials given in the preliminary results meet a preset stability threshold. The result of the physical rationality verification serves as one of the bases for the iterative decision-making of the knowledge core model. When the physical rationality verification fails, the knowledge core model generates sub-task instructions for the corresponding defects and returns them to the execution coordination model, triggering a new round of iteration.

[0062] In some embodiments, the method for constructing the coordination model includes: constructing a standardized model context protocol platform, encapsulating all external tools into a unified API interface service, and explicitly defining the function, input parameters, and output format of each tool; selecting a computationally efficient lightweight base model and training it using a reinforcement learning algorithm; designing a composite reward function that includes the correctness of tool call format, syntactic executability, problem-solving efficiency (round penalty), and the accuracy of the final answer, and training the model through multiple rounds of autonomous tool call on a large number of tasks, enabling the model to learn how to plan complex tool sequences, handle tool errors, and efficiently complete specified tasks.

[0063] In some embodiments, a reinforcement learning algorithm optimized by a generalized reward policy is used for training. During training, the model autonomously generates tool call sequences for the problem, interacts with the tool server through a standardized tool call protocol, and continuously optimizes its policy network based on feedback from the composite reward function.

[0064] In some embodiments, the composite reward function R total The reward is expressed as a weighted linear combination of four reward components: R total = w1·R turns + w2·R think + w3·R format + w4·R syntax The preferred ratio of weights w1, w2, w3, and w4 is 0.1:0.3:0.25:0.35.

[0065] The R turns A hierarchical ladder function based on the number of interaction rounds n is adopted: when n=0, R turns =0; when n=1, R turns =0.5; when n=2, R turns =0.7; when n≥4, R turns =1.0, where the preset depth threshold is preferably k=4. The purpose of this reward is to encourage the model to iterate through complex tasks multiple times, rather than giving conclusions too early.

[0066] The R think Using the hyperbolic tangent function R think = tanh(L / λ), where L is the thought label. <think>The average token length of the content, λ is a preset scaling factor, preferably λ=500. This reward increases smoothly with the inference length, encouraging the model to generate detailed and logically rigorous reasoning chains rather than short, intuitive judgments. When L≈500, R think It is approximately 0.76 and converges asymptotically to 1.0 as L increases.

[0067] The R format Employing a segmented weighted mechanism: intermediate steps require <think>and <tool call >existing simultaneously and <think>Strict preorder <tool call The constraint weight α = 0.4; the termination step requires... <think>and <answer>Co-occurrence, with a constraint weight β=0.6. This mechanism prevents the model from prematurely generating the final answer in intermediate stages, ensuring the structural integrity of the agent's workflow. It effectively suppresses reward hacking.

[0068] The R syntax Each tool call must undergo a two-layer validation: the first layer verifies whether the called function exists in the Mat-MCP namespace; the second layer validates parameter integrity and type consistency based on the Pydantic schema. Specifically, the generated CIF string must be correctly parsed by Pymatgen. R syntax Take the average success rate of all tool calls.

[0069] The reinforcement learning training preferably employs the DAPO (Decoupled Clip and Dynamic Sampling Policy Optimization) algorithm.

[0070] The core of this application lies in constructing a dual-model intelligent agent architecture with decoupled functions and collaborative operation, and realizing the automated and intelligent processing of complex materials science research tasks through a closed-loop iterative workflow.

[0071] The intelligent research system for materials science proposed in this application has the following system architecture: Figure 1 As shown, it mainly includes: a user interaction interface, a dual-model collaborative intelligent agent core, a materials science professional toolset, and a bottom-level computing resource layer.

[0072] The core of the dual-model collaborative intelligent agent: As the "brain" of the system, it consists of two heterogeneous language model intelligent agents that work together, responsible for decision-making, reasoning, and tool orchestration.

[0073] The core knowledge model is a large-parameter-scale language model (e.g., at the 30-B parameter level) that has been deeply fine-tuned using literature and data in the field of materials science. Its primary responsibilities are to perform deep scientific reasoning, analyze and interpret computational results, generate scientific hypotheses, and guide future research directions in complex situations. This model optimizes the understanding of the complex physicochemical relationships between crystal structures, material properties, and synthetic pathways.

[0074] The execution coordination model is a lightweight language model (e.g., with 14-byte parameters) specifically trained for efficient and accurate execution of tool calls. Its primary responsibility is to parse user or knowledge core model instructions, autonomously select tools, configure parameters, orchestrate multi-step tool call sequences, and return structured tool execution results. This model is optimized for complex workflows through reinforcement learning training targeting standardized tool protocols.

[0075] It should be noted that the lightweight materials science intelligent research system with dual-model collaboration provided in this application can be physically implemented as one or more servers, or a distributed computing cluster. Logically, the system consists of multiple functional modules.

[0076] The method provided in this application is based on a closed-loop iterative research process using a dual-model collaborative approach. This process enables the system to dynamically adjust its research strategy based on intermediate results, simulating the exploration process of real scientific research. The specific method steps are as follows (refer to...). Figure 1 Workflow shown: Step 1: Task Reception and Parsing. The system receives research tasks input by users in natural language form (e.g., "Please design a novel ternary metal sulfide catalyst for electrocatalytic nitrogen fixation"). The execution coordination model first parses the task to understand its core intent.

[0077] Step 2: Preliminary Planning and Tool Execution. Based on the task intent, the execution coordination model autonomously plans and executes the first step. If the task is an open-ended exploration, it will invoke information retrieval tools to gather background knowledge; if the task involves specific calculations, it will select and invoke the appropriate computational tools. The execution coordination model sends tool invocation requests to a materials science-specific toolset and obtains structured execution results (such as literature abstracts, database entries, computational data, etc.).

[0078] Step 3: Knowledge Reasoning and Deep Analysis. The execution coordination model submits the collected preliminary results, along with the original task, to the knowledge core model. The knowledge core model leverages its deep domain knowledge to perform in-depth analysis, interpretation, and reasoning on these results. For example, it analyzes technical gaps in the literature, explains the physical meaning behind the computational data, or assesses whether the current results meet the task requirements.

[0079] Step 4: Iterative Decision Making and Task Replanning. Based on deep analysis, the knowledge core model makes iterative decisions: (1) Case A: Sufficient Information. If the knowledge core model determines that the current information is sufficient to generate the final answer, it will directly integrate all the information to form a well-organized final report containing scientific insights and output it to the user. The process ends. (2) Case B: Insufficient Information. If the knowledge core model determines that further calculations or information supplementation are needed to solve the problem, it will generate one or a series of new, more specific sub-task instructions.

[0080] Step 5: Closed-Loop Feedback and Iterative Execution. The newly generated subtask instructions are returned to the execution coordination model. The system flow returns to Step 2, where the execution coordination model parses the new instructions, calls the corresponding tools to execute them, and submits the newly acquired results back to the knowledge core model for analysis. This closed-loop process of "execution-analysis-decision" will continue to iterate until the knowledge core model determines that all necessary information has been gathered and a final conclusion can be reached.

[0081] The method for constructing the knowledge core model 102a provided in this application is as follows: (Refer to...) Figure 2 The method includes: a) Dataset construction: such as Figure 2 As shown, multi-source heterogeneous data is integrated through an automated data processing workflow. Crystal structure and computational performance data are obtained from public crystallography databases (such as the Materials Project), and DOIs are linked to academic literature (PDFs). Unstructured text, including synthesis methods and application scenarios, is extracted from the literature using natural language processing techniques. Finally, a high-quality instruction fine-tuning question-answer dataset containing full-chain information on structure, performance, synthesis, and application is constructed using a "generate-distillate" strategy.

[0082] b) Model Fine-tuning: Select a lightweight base model that combines parametric inference capabilities with cost-effectiveness, and fine-tune it using the domain-specific datasets and general scientific inference datasets constructed above, with full parameter supervision. This process aims to inject deep materials science expertise and complex physicochemical relationships into the model while maintaining its general scientific inference capabilities.

[0083] Specifically, the technical implementation scheme of Knowledge Core Model 102a is as follows: Functional positioning: to conduct in-depth scientific reasoning, generate scientific hypotheses, interpret calculation results, evaluate research progress, and guide the next research direction.

[0084] Technical Implementation: The model construction process is as follows Figure 2 As shown in a) to b). Figure 2 Figure a) illustrates the data set construction module, showcasing the process of building a domain-specific dataset. This process begins with acquiring heterogeneous data from multiple sources (such as public databases and academic literature). Through data association, unstructured information extraction, and particularly by employing a "generate-distillation" strategy, this raw data is transformed into a high-quality instruction-fine-tuned question-answer pair dataset. This figure highlights the key steps in integrating disparate materials science knowledge into a structured format that the model can learn. Figure 2 Figure b) illustrates the supervised fine-tuning framework for the Knowledge Core Model (Mat-R1). This figure shows that by selecting a base model with a large number of parameters and using… Figure 2 The domain-specific dataset and general scientific reasoning dataset constructed in section a) are fine-tuned with full parameter supervision on a distributed computing cluster. This process aims to inject deep materials science expertise into the model, thereby building a knowledge core focused on "what to think about".

[0085] Step 1: Construct a domain-specific instruction fine-tuning dataset. Please refer to... Figure 2 As shown in Figure a), the process of constructing the dataset is illustrated.

[0086] Data acquisition: Crystal structure data and corresponding computational performance data are acquired in batches from public crystallography databases.

[0087] Data association: Using the numeric object identifiers provided in the data entries, associate and download the corresponding full text of academic literature.

[0088] Unstructured information extraction: Natural language processing tools are used to convert documents into structured Markdown text. Then, through multi-model voting classification and algorithms, key paragraphs related to the material's "synthesis method" and "application scenario" are automatically extracted from the text.

[0089] The "Generation-Distillation" Strategy: To transform the aforementioned multi-source heterogeneous data into high-quality instruction-question-answer pairs, a "generation-distillation" strategy is adopted. First, using retrieval-enhanced generation technology, well-related information such as structure, properties, synthesis, and applications is input into a generation model to generate specialized questions related to materials science. Then, the generated questions and the original information paragraphs are input together into a stronger reasoning model to "distill" accurate and professional answers.

[0090] Quality control: Low-quality question-answer pairs are filtered out through manual sampling and automated model self-checking (e.g., scoring the consistency and factuality of answers), ultimately forming an instruction fine-tuning dataset containing high-quality question-answer pairs.

[0091] Step Two: Full Parameter Supervision and Fine-Tuning of the Model. Please refer to... Figure 2 As shown in b), it illustrates the fine-tuning framework of the model.

[0092] Base model selection: Choose a large-parameter base model with strong general reasoning capabilities and deep prior knowledge of physics and chemistry (e.g., Qwen3-30B with 30 parameters). The selection criteria can be its performance on standard scientific benchmarks (such as GPQA) and its initial training loss value on domain datasets.

[0093] Fine-tuning execution: Using the domain-specific dataset and general scientific reasoning dataset built in step one, perform fully parameter-supervised fine-tuning of the base model on a distributed computing cluster.

[0094] 2. Methods for constructing the coordination model: Refer to... Figure 3 The method includes: a) Tool protocol standardization: Build a standardized model context protocol platform, encapsulate all external tools into a unified API interface service, and clearly define the function, input parameters and output format of each tool.

[0095] b) Reinforcement Learning Training: Select a computationally efficient, lightweight base model. Train the model using reinforcement learning algorithms. Design a composite reward function that includes correctness of tool invocation format, syntactic feasibility, problem-solving efficiency (round penalty), and accuracy of the final answer. Through multiple rounds of autonomous tool invocation training on a large number of tasks, the model learns how to plan complex tool sequences, handle tool errors, and efficiently complete the specified task.

[0096] The technical implementation scheme for implementing coordination model 102b is as follows.

[0097] Functionality: Parse instructions, select tools, configure parameters, arrange multi-step tool call sequences, and return execution results in a structured manner.

[0098] Technical Implementation: The model construction process is as follows Figure 3 As shown in a) to b). Figure 3 Figure a) illustrates the standardized encapsulation of a materials science toolset. This diagram shows that tools for information retrieval, structure processing, performance prediction, and analysis visualization are all encapsulated according to a unified model context protocol and deployed as services accessible via APIs. This standardization is a prerequisite for ensuring reliable and efficient tool calls from models. Figure 3 Figure b) illustrates the reinforcement learning training module, showcasing the reinforcement learning training framework for the execution coordination model. The figure shows that a lightweight base model is selected and trained using a reinforcement learning algorithm. The core of the training lies in a composite reward function that comprehensively evaluates the correctness of the tool call format, syntactic executableness, problem-solving efficiency (rounds), and the accuracy of the final answer. This is achieved through... Figure 3 In section a), the MCP server interacts with the model, which is trained on a large number of tasks and eventually masters the ability to autonomously plan and orchestrate complex toolchains, focusing on "how to execute".

[0099] Step 1: Build a standardized tool platform. Please refer to... Figure 3 a) Tool protocol standardization, which shows a schematic diagram of a materials science professional toolset.

[0100] Tool encapsulation: All external tools, whether database query APIs, local computation scripts, or machine learning models, are encapsulated according to a unified model context protocol. Each encapsulated tool has its functional description, input parameters, and output format clearly defined.

[0101] Service-oriented deployment: Deploying the packaged tools as one or more independent services that can be called via API. For example, using Docker to containerize and deploy an MCP server ensures the stability, isolation, and scalability of tool calls.

[0102] Step Two: Model Reinforcement Learning Training. Please refer to... Figure 3 b) Reinforcement learning training, which illustrates the reinforcement learning training framework for the model.

[0103] Base model selection: Select a computationally efficient base model with a small number of parameters to ensure low latency during inference and tool invocation decisions.

[0104] Training data preparation: Sample a portion of the dataset that is relevant to the specific computational task to form a reinforcement learning training set.

[0105] Design a composite reward function: To guide the model in learning efficient and accurate tool calling strategies, a composite reward function with multi-dimensional indicators is designed.

[0106] Through the above technical solution, this application constructs an intelligent agent system. Its core knowledge model focuses on "what to think about," while its execution coordination model focuses on "how to execute." The two collaborate closely through a closed-loop iterative workflow. This architecture not only achieves decoupling and collaborative optimization between domain knowledge depth and tool execution efficiency, but also endows the system with the ability of dynamic planning and iterative exploration. Thus, it systematically solves the limitations of existing technical solutions and provides a complete and efficient platform for realizing automated and intelligent materials science research.

[0107] In summary, this application provides a lightweight intelligent research method and system for materials science with dual-model collaboration. The core technical solution of this system and method lies in: constructing a functionally decoupled dual-model intelligent agent core, in which the "knowledge core model" is responsible for deep scientific reasoning and the "execution coordination model" is responsible for efficient tool orchestration; and through a closed-loop iterative workflow, the two are closely coordinated to achieve automated and intelligent processing of complex materials science research tasks, thereby solving the problems of rigid architecture, model capability bottlenecks and fragmented solutions in existing technologies.

[0108] The method provided in this application is a closed-loop iterative research method based on the aforementioned system device and employing a dual-model approach. Please refer to... Figure 4 The diagram shows a detailed flowchart of the method described in this application. Figure 4 The flowchart details the closed-loop iterative process of dual-model collaboration step by step. The process begins in step S1 (task reception and parsing), where the execution coordination model (Mat-T1) processes user input. In step S2 (preliminary planning and tool execution), Mat-T1 calls the Mat-MCP toolset to obtain preliminary data. In step S3 (knowledge reasoning and deep analysis), the knowledge core model (Mat-R1) performs deep analysis on the data submitted by Mat-T1. In the crucial step S4 (iterative decision-making and task replanning), Mat-R1 determines whether the information is sufficient. If sufficient, the process ends and a report is output; if insufficient, a new subtask instruction is generated. If a new instruction is generated, the process proceeds to step S5 (closed-loop feedback and iterative execution), returning the new instruction to Mat-T1, and the process jumps back to step S2, forming a dynamic, iterative closed loop until the problem is solved. This diagram visually demonstrates the dynamic and intelligent characteristics of this invention, distinguishing it from linear workflows.

[0109] The method specifically includes the following steps: Step S1: Task Reception and Parsing. The system receives research tasks input by the user in natural language through the user interaction interface 101. The execution coordination model (Mat-T1) 102b first parses the task text, identifying the user's core intent, key entities (such as material system, target performance), and constraints. For example, for the task "Find a stable Cs-Pb-I perovskite structure," Mat-T1 will identify key information such as "find structure," "Cs-Pb-I system," "perovskite," and "stable."

[0110] Step S2: Preliminary Planning and Tool Execution. Mat-T1 autonomously plans and executes the first step operation based on the parsed intent.

[0111] Technical Implementation: Mat-T1 generates one or more tool invocation instructions in its policy network. For example, it might first decide to invoke the search_crystal_structures_from_materials_project tool and set the parameters elements=['Cs', 'Pb', 'I'] to query the database to see if the relevant structure already exists.

[0112] Mat-T1 sends the generated tool call request (a JSON object conforming to the MCP format) to Materials Science Professional Toolset (Mat-MCP) 103 via API.

[0113] The corresponding tool service in Mat-MCP 103 executes the request and returns the execution result (e.g., a list of queried database entries or "not found" information) to Mat-T1 in a structured format (such as JSON).

[0114] Step S3: Knowledge Reasoning and Deep Analysis. Mat-T1 integrates the original task description, its own thought process, and the structured execution results obtained from the toolset into a complete context and submits it to the knowledge core model (Mat-R1) 102a.

[0115] Technical Implementation: After receiving the context, Mat-R1 utilizes its powerful reasoning capabilities, refined with domain knowledge, to perform in-depth analysis of the information. For example, if the result of step S2 is "no relevant structure found," Mat-R1 will infer that the next step may require generating the structure from scratch; if multiple structures are found, Mat-R1 will analyze the characteristics of these structures and determine which structure is most worthy of further study.

[0116] Step S4: Iterative Decision Making and Task Replanning. Based on the in-depth analysis in Step S3, Mat-R1 performs iterative decision-making to determine whether the current information is sufficient to answer the user's initial question.

[0117] Scenario A: Sufficient Information. If Mat-R1 determines that enough information has been collected to reach a final conclusion (e.g., the target structure has been found and its stability calculated, with results meeting requirements), it will directly integrate all analysis processes and results to generate a final research report containing scientific insights and rich graphics, which will be output to the user through the user interface 101. At this point, the entire workflow is complete.

[0118] Scenario B: Insufficient Information. If Mat-R1 determines that further computation or additional information is needed, it will generate one or more new, more specific subtask instructions. These instructions are highly structured and explicit, serving as "commands" to guide the next steps. For example, it might generate the instruction: "Based on the above analysis, no structure in the existing database meets the conditions. Please call the generate_crystal_structures_crystallm tool, using the chemical formula 'CsPbI3' and space group 'Pm-3m' as input, to generate a new perovskite structure and perform structural relaxation on it." Step S5: Closed-loop feedback and iterative execution. The new subtask instruction generated by Mat-R1 is returned to the execution coordination model (Mat-T1). The system flow automatically returns to step S2. Mat-T1 receives and parses this new, internally generated instruction, selects and calls the appropriate tools (e.g., generate_crystal_structures_crystallm and relax_crystal_structure_MatGL), and submits the newly acquired results (e.g., the newly generated crystal structure CIF file and the relaxed energy) back to Mat-R1 for analysis in step S3.

[0119] This closed-loop process of "execution (Mat-T1) - analysis (Mat-R1) - decision (Mat-R1) - replanning (Mat-R1) - re-execution (Mat-T1)" will continue to iterate until, in step S4 of a certain iteration, Mat-R1 determines that the task is completed and exits the loop.

[0120] Through the aforementioned system apparatus and working method, this application realizes an intelligent system capable of simulating the research process of human scientists. By decoupling the functions of the two models, it allows the depth of knowledge reasoning and the efficiency of tool execution to be optimized to their respective extremes. Through a closed-loop iterative workflow, it endows the system with the ability to dynamically adjust research strategies and handle complex nonlinear tasks, thus systematically solving the shortcomings of existing technologies and providing a complete, efficient, and powerful technical solution for automated materials science research.

[0121] In summary, this application provides a lightweight dual-model cooperative intelligent agent system architecture with functional decoupling. The core of this application is a system architecture comprising two functionally heterogeneous, cooperative language model intelligent agents: Knowledge Core Model (Mat-R1): A large-parameter model finely tuned with data from the field of materials science, focusing on deep scientific reasoning, outcome analysis, and research direction decisions ("what to think about").

[0122] The execution coordination model (Mat-T1) is a lightweight model trained through reinforcement learning, specifically designed for efficient tool invocation. It is responsible for parsing instructions and autonomously planning and orchestrating multi-step tool invocation sequences ("how to execute"). This architecture decouples and coordinates "deep knowledge reasoning" with "efficient tool execution" capabilities.

[0123] Furthermore, this application provides a closed-loop iterative working method. This method defines a collaborative workflow with two core models: the execution coordination model (Mat-T1) performs tool operations to acquire data, and then the knowledge core model (Mat-R1) conducts in-depth analysis of the data and makes iterative decisions. If information is insufficient, the knowledge core model generates new sub-task instructions and returns them to the execution coordination model, forming a closed loop of "execution-analysis-decision-replanning". This non-linear, dynamically adjustable automated research process is a key innovation that distinguishes it from existing linear "pipeline" or "reasoning-action" chains.

[0124] Furthermore, this application also provides methods for constructing the knowledge core model (Mat-R1) and the execution coordination model (Mat-T1); specifically, the method for constructing the knowledge core model (Mat-R1) includes: a method for constructing a full-link instruction fine-tuning dataset (such as Mat-252K-SFT) from multi-source heterogeneous data (crystal structure, computational performance, academic literature), particularly the "generate-distillation" strategy employed therein; and a technical solution for using domain-specific datasets to perform fully parameter-supervised fine-tuning (SFT) of a large-parameter basic model.

[0125] The construction method of the execution coordination model (Mat-T1) includes: encapsulating professional toolsets according to the standardization protocol (MCP); and using reinforcement learning (such as the GRPO algorithm) and a composite reward function that includes dimensions such as format, syntax, efficiency, and accuracy to train a lightweight model to master the ability to orchestrate complex toolchains.

[0126] The system provided in this application and its wide range of applications in materials science. The entire system, including the user interface, the dual-model collaborative intelligent agent core, and the materials science professional toolset (Mat-MCP) packaged according to the standardized protocol (MCP), together constitute a complete automated research platform. Specific applications of this platform in materials science research, such as catalyst discovery, structure prediction, stability analysis, performance prediction, and synthetic route design, and their complete end-to-end solutions, should also be protected.

[0127] Compared with existing technologies, such as the Coscientist automated scientific research system built on a single general large model, and linear multi-agent "pipeline" systems (such as CN120524711A), this application has the following significant advantages: (1) The model capabilities are decoupled and collaboratively optimized, breaking through the bottleneck of single-model capabilities and resulting in stronger overall system performance. Coscientist relies on a single, general-purpose large model (such as GPT-4), making it difficult to balance the depth of knowledge in the materials field with efficient tool execution capabilities. This application decouples these two capabilities through a dual-model collaborative architecture: The Knowledge Core Model (Mat-R1), a lightweight model with 30 parameters, acquires deep professional knowledge through full parameter fine-tuning on the domain dataset (Mat-252K-SFT).

[0128] The execution coordination model (Mat-T1), a lightweight model with 14 parameters, utilizes reinforcement learning to optimize tool calls, resulting in high execution efficiency and low latency. This "expert thinking" + "assistant execution" model achieves a synergistic effect greater than the sum of its parts.

[0129] It should be noted that this application uses a 30B lightweight model for Mat-R1 and a 14B lightweight model for Mat-T1, totaling 44B, which is much smaller than general solutions (such as DeepSeek-R1's 671B).

[0130] Based on specialized training with Mat-R1 and Mat-T1, Mat-R1 acts as the right brain for knowledge, while Mat-T1 acts as the left brain for execution. This application constructs a dual-brain collaborative system—MatBrain. This system follows the biomimetic concept of the human "left and right brains." Mat-T1 (14B), with fewer parameters but trained through reinforcement learning (RL), acts as the "execution left brain," responsible for transforming abstract scientific questions into specific Mat-MCP tool call sequences and ensuring the syntactic correctness of the execution. Meanwhile, Mat-R1 (30B), with relatively more parameters and fine-tuned through extensive literature review, acts as the "knowledge right brain," responsible for scientifically analyzing the user's initial intent and verifying the physical rationality and providing in-depth interpretation of the tool execution results returned by Mat-T1.

[0131] (2) The system architecture is more flexible and can simulate the iterative process of real scientific research, solving the rigidity problem of the existing architecture. The linear "reasoning-action" loop and "pipeline" system of Coscientist has a unidirectional task logic, making it difficult to dynamically adjust strategies or backtrack based on intermediate results. This application realizes the "execution-analysis-decision-replanning" loop through the closed-loop iterative workflow of the knowledge core model (Mat-R1) and the execution coordination model (Mat-T1). When Mat-R1 determines that the information is insufficient, it can autonomously generate new sub-task instructions, enabling the system to explore, try and fail, and iteratively optimize like a human scientist. This design not only significantly reduces the reasoning cost of the entire system, but also effectively avoids the illusion problem common in general large models in the scientific field through the "generation-verification-feedback" mechanism.

[0132] (3) It provides a complete end-to-end solution, overcoming the fragmentation problem of existing technologies. Some existing technologies only optimize specific steps (such as screening or input understanding) without providing a complete automated research platform. This application integrates the entire process from task reception, autonomous multi-step tool invocation (through the Mat-MCP standardized toolset), data analysis, iterative optimization to the generation of scientific conclusions.

[0133] The feasibility of the system and method provided in this application was demonstrated through simulation evaluation, and the results are as follows: (1) Simulation Evaluation: To comprehensively verify the effectiveness of the intelligent agent architecture in solving practical materials science problems, this application designed a rigorous evaluation benchmark based on the cleaned Mat-252K test set, covering two core dimensions: crystal structure generation (structural similarity, CIF syntax accuracy) and material performance prediction (stability, band gap type, metallicity, magnetism, and magnetic order). Performance evaluations were conducted on the system of this application (MatBrain) and other mainstream large models (such as GPT-5 and Gemini-2.5-Pro). The results show that the knowledge core model (Mat-R1) of this application outperforms mainstream models in all test dimensions.

[0134] Figure 5 The results of the system performance comparison evaluation with mainstream general-purpose models are shown. The figure, presented as a multi-dimensional radar chart, comprehensively and comprehensively illustrates the differences in capabilities between the proposed model and existing top-tier general-purpose models in materials science tasks. The evaluation covers seven key dimensions, including stability, structural similarity, and CIF syntax accuracy. The comparison begins with basic indicators (such as metallicity and magnetism), with each model exhibiting its own strengths and weaknesses. The key to the data presentation lies in the two core dimensions of "structural similarity" and "CIF syntax accuracy." The figure clearly shows that the scores of general-purpose models such as Deepseek-R1 and GPT-5 have significantly declined, while the curve (solid line) of the proposed MatBrain system exhibits a full, enveloping shape, with all indicators approaching a perfect score of 1.0. This figure intuitively demonstrates the comprehensive leading advantage achieved by the proposed model through domain-specific training, addressing the pain point of general-purpose models "not understanding material structures."

[0135] A complete MatBrain system also demonstrates a significant advantage in deployment costs, reducing hardware expenses by 95%. Specific cost comparisons are as follows: Figure 6 As shown.

[0136] Figure 6 A comparison chart of deployment costs for different parameter models is shown. Figure 6 The present invention uses a four-grid bar chart to detail the significant differences in resource consumption between the Mat series model of this invention and the baseline model based on Qwen3-235B. The comparative analysis begins with the "Minimum VRAM (GB)" sub-chart, showing that the baseline model requires nearly 500GB, while the present invention's model requires only a fraction of that. The subsequent comparison of hardware quantity (number of 4090s and H100s) demonstrates that the present invention's model can reduce the massive cluster requirements to single-card or low-card requirements. In the crucial "Hardware Cost" sub-chart, the bar heights of the baseline model and the extremely low bar heights of the present invention stand in stark contrast. This chart visually illustrates the lightweight, low-cost, and easily localized deployment characteristics of the present invention's solution, distinguishing it from traditional large models.

[0137] Case Study Usage: To comprehensively demonstrate the practical application value of MatBrain in materials science research, this application selects a series of representative cases covering the core aspects of materials science research. Crystal structure prediction is particularly important for materials that have not yet been synthesized or whose precise structures are difficult to obtain experimentally. Figure 7 As shown, when MatBrain was handling the task of generating the rare-earth double perovskite Cs₂ErAgBr₆ structure, the system first used Mat-T1 to intelligently query the Materials Project and OQMD databases. After confirming that no existing records existed, it automatically invoked the CrystalLM structure generation tool to generate the corresponding crystal structure based on the space group and chemical formula as constraints. The generated CIF file contained complete unit cell parameters and atomic coordinates. Subsequently, Mat-R1 analyzed the structure and confirmed that it conformed to the typical characteristics of double perovskites. This workflow significantly simplified the process from requirement description to structure generation, reduced the workload of researchers in tool selection and parameter configuration, and provided theoretical guidance for subsequent experimental synthesis and performance testing.

[0138] Material stability is a key indicator for evaluating the practical value of materials. To achieve this complex analysis, this application developed and integrated a dedicated phase diagram analysis tool, which integrates three key functions: Materials Project database query, CHGNet energy calculation, and Pymatgen phase diagram construction.

[0139] Figure 7 A schematic diagram of the Cs2ErAgBr6 structure generated by the MatBrain system is shown. Figure 7 The microstructure of complex double perovskite materials is described in detail using a three-dimensional crystal structure model. The generated Cs2ErAgBr6 perovskite unit cell structure is shown, clearly presenting the complete space lattice structure. Figure 7 The figure clearly demonstrates a coordinated octahedral structure centered on Er (erbium) and Ag (silver), with Cs (cesium) atoms filling the interstices. A key spatial feature is that the bond lengths and bond angles between the atoms conform to the rules of crystal chemistry. This figure visually illustrates the invention's ability to directly output high-fidelity three-dimensional crystal structures that conform to objective physical and chemical laws.

[0140] Figure 8 This is a ternary phase diagram of the Li₂ZrCl₆ structure generated by the MatBrain system. Mat-T1 first invokes this integrated tool, querying the structure and thermodynamic data of known phases in the Cs-Pb-I system through the Materials Project interface, including the formation energy of the Li-Zr-Cl ternary system. Subsequently, the tool utilizes Pymatgen's phase diagram analysis module to construct a complete ternary phase diagram based on the collected thermodynamic data, clearly showing the stable phase regions and phase boundaries under different compositions. Mat-R1 conducts in-depth analysis based on the phase diagram, indicating that Li₂ZrCl₆ exhibits a metastable state (E0) in the Li–Zr–Cl ternary system. hull = 0.028 eV / atom), which can exist stably at room temperature and under controlled conditions, but may decompose into adjacent stable phases (such as the combination of LiCl + ZrCl4) under thermodynamic equilibrium or high temperature conditions. Therefore, attention should be paid to heat treatment and phase environment control when applying it.

[0141] Figure 8 The thermodynamic stability verification results of the generated material are described in detail using a ternary equilibrium phase diagram. The diagram begins with a triangular frame, with the three vertices representing the three elements Li, Zr, and Cl, respectively, and the internal lines representing stable phase equilibrium paths. The key marker is the "pentagram" in the diagram, which precisely locates the current compositional coordinates of the Li₂ZrCl₆ compound generated by this system. The determination is based on the E marked at the bottom. hull =0.028 eV / atom, this extremely low convex hull energy value indicates that the material is in a thermodynamically stable (or metastable) range. This figure visually demonstrates that the material generated by this invention does not merely exist in theoretical calculations, but is a real substance with extremely high experimental synthetic feasibility.

[0142] Materials synthesis is a crucial step connecting theoretical design with practical applications. Using MatBrain as an example, CsPb(Cl) synthesis... 0.2 Br 0.4 I 0.4 Taking the room-temperature synthesis scheme of perovskite as an example, the system first retrieves relevant literature information through the Mat-T1 search tool. Then, Mat-R1, based on its professional knowledge, analyzes the synthesis characteristics of mixed-halogen perovskites and designs a detailed synthesis process, including precise precursor ratios, solvent selection, and reaction conditions. This precise synthesis scheme design, down to the operational details, significantly lowers the barrier between theoretical design and experimental realization, providing researchers with directly executable experimental guidance. This case demonstrates the application value of MatBrain in the field of materials synthesis, especially for planning the synthesis pathways of materials with complex components. By combining literature data and professional knowledge, the system can provide reasonable synthesis suggestions for newly designed materials, accelerating the transformation process from theoretical design to practical application.

[0143] Based on the above technical solutions, this application provides a lightweight intelligent research method and system for materials science with dual-model collaboration. The system includes: a user interface, a dual-model collaborative intelligent agent core, a materials science professional toolset, and a bottom-level computing resource layer. The user interface receives research tasks input by the user in natural language. The dual-model collaborative intelligent agent core includes an execution coordination model and a knowledge core model. Both the execution coordination model and the knowledge core model are autoregressive language models with a parameter size of less than 100 bytes, wherein the parameter size of the knowledge core model is larger than that of the execution coordination model. The execution coordination model parses the research task, understands its core intent, obtains preliminary results, and transmits the preliminary results to the knowledge core model. The knowledge core model performs in-depth analysis, interpretation, and reasoning on the preliminary results, including physical rationality verification, and makes iterative decisions based on the in-depth analysis and physical rationality verification results. The physical rationality verification includes: grammatical parsingability verification of the crystal structure file (CIF), rationality verification of crystal geometric parameters, and convex hull energy E for thermodynamic stability. hull Threshold verification; the materials science professional toolset is a software tool and database interface encapsulated in accordance with a standardized model context protocol; the materials science professional toolset includes multiple tools that are deployed as services that can be called via API; the underlying computing resource layer provides physical computing support for the research system.

[0144] This application provides a lightweight materials science intelligent research method and system with dual-model collaboration, achieving a flexible and dynamically iterative system architecture. One objective of this application is to construct a nonlinear closed-loop iterative workflow. Through this workflow, the system can autonomously perform feedback, adjustment, and task replanning based on intermediate results, thereby realistically simulating and accelerating the scientific exploration process to adapt to the complexity and variability of real scientific research.

[0145] On the other hand, it achieves decoupling and collaborative optimization of model capabilities. The core objective of this application is to decouple the capabilities of "deep knowledge reasoning" and "efficient tool execution" by constructing a functionally heterogeneous dual-model core. Through specially optimized "knowledge core model" and "execution coordination model" each performing their respective functions, it overcomes the bottlenecks of existing technologies in knowledge depth and tool orchestration capabilities, and enables the two to work collaboratively to achieve overall system performance of 1+1>2, while significantly reducing system deployment costs by reducing model parameters.

[0146] Furthermore, this application provides a complete, end-to-end automated research platform. The ultimate goal of this application is to provide a comprehensive automated research platform that seamlessly integrates the entire process from the formulation of scientific hypotheses, the autonomous invocation of complex toolchains, the comprehensive analysis of multi-source data, to the generation of final scientific conclusions. In this way, the professional barriers and time cycles of materials research and development are significantly reduced, accelerating the process of scientific discovery.

[0147] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.< / answer> < / think> < / think> < / think> < / think>

Claims

1. A lightweight intelligent research system for materials science with dual-model collaboration, characterized in that, include: User interaction interface, dual-model collaborative intelligent agent core, materials science professional toolset and underlying computing resource layer; The user interaction interface is used to receive research tasks input by users in natural language. The dual-model collaborative agent core includes an execution coordination model and a knowledge core model; both the execution coordination model and the knowledge core model are autoregressive language models with a parameter size of less than 100B, wherein the parameter size of the knowledge core model is larger than that of the execution coordination model. The execution coordination model is used to analyze research tasks, understand their core intent, obtain preliminary results, and transmit the preliminary results to the knowledge core model. When constructing the execution coordination model, a computationally efficient lightweight base model is selected and trained using a reinforcement learning algorithm optimized by a generalized reward policy. During training, the lightweight base model autonomously generates tool call sequences for the problem, interacts with the tool server through a standardized tool call protocol, and continuously optimizes its policy network based on the feedback from the composite reward function. The composite reward function is: R total = w1·R turns + w2·R think + w3·R format + w4·R syntax Among them, w1, w2, w3, and w4 are four reward components R. turns R think R format R syntax The weights of R; turns R is a hierarchical ladder function based on the number of tool interaction rounds n; when round n reaches a preset depth threshold k, turns Take the maximum value, and the depth threshold k is not less than 4; R think Let R be the hyperbolic tangent function of the average token length L based on the content of the thought tag. think = tanh(L / λ), where λ is the preset scaling factor; R format To enforce a strict time constraint between thinking and action, meaning that at each intermediate step, the thought process must strictly precede the content invoked by the tool; R syntax The verification was determined through a two-layer check: the first layer is the existence check of the tool registry; the second layer is the parameter integrity and semantic resolvability check based on the Pydantic schema, and the semantic resolvability check includes the verification of the resolvability of the generated CIF crystal structure file. The core knowledge model is used to perform in-depth analysis, interpretation, and reasoning on the preliminary results, including physical plausibility verification, and to make iterative decisions based on the results of the in-depth analysis and physical plausibility verification. The physical plausibility verification includes: verification of the parsingability of the crystal structure file (CIF), verification of the plausibility of the crystal geometric parameters, and verification of the convex hull energy E of thermodynamic stability. hull Threshold verification; The materials science toolset comprises software tools and database interfaces encapsulated in accordance with a standardized model context protocol; the materials science toolset includes multiple tools that are deployed as services that can be invoked via APIs. The underlying computing resource layer provides physical computing support for the research system.

2. The lightweight materials science intelligent research system with dual-model collaboration according to claim 1, characterized in that, The underlying computing resource layer includes a central processing unit, a graphics processing unit, memory, and storage.

3. A dual-model collaborative intelligent research method for lightweight materials science, the method being implemented based on the dual-model collaborative intelligent research system for lightweight materials science as described in any one of claims 1 to 2, characterized in that, Includes the following steps: Step S1: Receive the research task and use the execution coordination model to parse the research task to obtain the task intent; Step S2: The execution coordination model autonomously plans and executes preliminary operations based on the task intent to obtain preliminary results; Step S3: The execution coordination model submits the obtained preliminary results and the research task to the knowledge core model; the knowledge core model performs in-depth analysis, interpretation and reasoning on the preliminary results and the research task; the in-depth analysis includes physical rationality verification of the preliminary results, and obtains in-depth analysis results containing the physical rationality verification results; Step S4: Based on the deep analysis results and the physical rationality verification results, the knowledge core model makes iterative decisions; If the knowledge core model determines that the current information is sufficient to generate the final answer, the knowledge core model will directly integrate all the information to form a well-organized final report containing scientific insights, and output it to the user, thus ending the process; If the knowledge core model determines that further computation or information supplementation is needed to solve the problem, the knowledge core model will generate one or a series of new, more specific sub-task instructions. Step S5: The newly generated subtask instruction is returned to the execution coordination model, and the system flow returns to step S2. The execution coordination model parses the newly generated subtask instruction, calls the corresponding tool to execute it, and submits the newly obtained results to the knowledge core model for analysis. After closed-loop feedback and iterative execution, the process continues until the knowledge core model determines that all necessary information has been collected and a final conclusion can be formed.

4. The intelligent research method for lightweight materials science based on dual-model synergy according to claim 3, characterized in that, The execution coordination model autonomously plans and executes preliminary operations based on the task intent, obtaining preliminary results, including: If the task intent is open exploration, the execution coordination model calls information retrieval tools to collect background knowledge; If the task intent is specific calculation, the execution coordination model selects and invokes the corresponding calculation tools to obtain structured execution results.

5. The intelligent research method for lightweight materials science based on dual-model synergy according to claim 3, characterized in that, The method for constructing the core knowledge model includes: By automating data processing workflows, multi-source heterogeneous data is integrated to construct datasets; Based on a lightweight base model that combines parametric reasoning capabilities and cost-effectiveness, and the dataset mentioned above, the knowledge core model is fine-tuned with full parameter supervision.

6. The intelligent research method for lightweight materials science based on dual-model synergy according to claim 5, characterized in that, The lightweight base model includes one or more of the Qwen2 series, Llama 2 series, Qwen3 series, Llama 3 series, and ChatGLM4.

7. The intelligent research method for lightweight materials science based on dual-model synergy according to claim 5, characterized in that, By automating data processing workflows, multi-source heterogeneous data is integrated to construct datasets, including: Crystal structure and computational performance data are obtained from public crystallography databases and linked to academic literature using digital object identifiers; Extracting unstructured text from documents using natural language processing techniques, revealing synthesis methods and application scenarios; By employing a "generate-distillate" strategy, a high-quality instruction fine-tuning question-answer dataset containing full-link information on structure, performance, synthesis, and application is constructed.

8. The intelligent research method for lightweight materials science based on dual-model synergy according to claim 3, characterized in that, The method for constructing the execution coordination model includes: Build a standardized model context protocol platform, encapsulate all external tools into a unified API interface service, and clearly define the function, input parameters and output format of each tool; Choose a computationally efficient, lightweight base model and train it using a reinforcement learning algorithm; Design a composite reward function that includes the correctness of tool call format, syntactic executableness, problem-solving efficiency, and the accuracy of the final answer. Through multiple rounds of autonomous tool call training on a large number of tasks, the model learns how to plan complex tool sequences, handle tool errors, and efficiently complete the specified task.

9. The intelligent research method for lightweight materials science based on dual-model synergy according to claim 8, characterized in that, The lightweight base model includes one or more of the Qwen2 series, Llama 2 series, Qwen3 series, Llama 3 series, and ChatGLM4.

Images