A multi-agent collaboration system and method for materials science

Abstract

The present application relates to the field of computer system technology, and in particular to a multi-agent collaboration system and method for materials science, the system comprising: a plurality of agent modules visualized by a general agent module architecture; the general agent module architecture comprises: a message queue / memory cache unit, a message listening unit, a manager large model unit and an executor large model unit connected in sequence; the manager large model unit exchanges information with the message queue / memory cache unit through a message transmission unit; the agent module comprises: an automated task planning decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experimental instruction generation module, an automated experimental verification module, and an automated experimental feedback optimization module. The present application realizes a new paradigm of materials science research by constructing a full-process system to fully automatically complete the experimental research process according to user instructions, thereby improving the efficiency of materials science research and exploration.

Description

Technical Field

[0001] The present application relates to the field of computer system technology, and in particular to a multi-agent collaboration system and method for materials science. Background Art

[0002] Research in the field of materials science usually includes multiple complex links such as material synthesis, performance testing, structural characterization and data analysis. Traditional materials science research methods rely heavily on laboratory manual operations and expert experience, and the research process is time-consuming, costly and inefficient. With the development of artificial intelligence technology, automated material research based on artificial intelligence algorithms has gradually become a trend. Its main goal is to partially or completely replace manual operations through automated experiments and artificial intelligence, thereby greatly improving research efficiency and accuracy.

[0003] The existing automation system based on artificial intelligence optimization algorithm can automatically carry out high-throughput preparation experiments and high-throughput optimization experiments, screen the optimal experimental parameters, and form an innovative development system for magnetic micro-nano materials. However, the existing automation system still relies on manual labor in experimental design, and the efficiency has not been significantly improved fundamentally; especially in terms of autonomous retrieval, precise experiments, feedback optimization and collaborative work, there are deficiencies. Summary of the invention

[0004] The embodiments of the present application provide a multi-agent collaboration system and method for materials science. By constructing a full-process system and using robots and AI systems to create a knowledge base based on simple instructions from users, the system automatically completes the experimental research process, thereby realizing a new paradigm for materials science research and improving the efficiency of materials science research and exploration.

[0005] To solve the above technical problems, in the first aspect, an embodiment of the present application provides a multi-agent collaboration system for materials science, including: a plurality of agent modules embodied by a general agent module architecture; the general agent module architecture includes: a message queue / memory cache unit, a message listening unit, a manager large model unit and an executor large model unit connected in sequence; the manager large model unit exchanges information with the message queue / memory cache unit through a message passing unit; the agent modules include: an automated task planning and decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module, and an automated experiment feedback optimization module; wherein the automated task planning and decomposition module receives and parses user instructions, decomposes complex tasks into subtasks, and distributes the subtasks to other agent modules; other agent modules include: an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module and an automated experiment feedback optimization module Optimization module; the automated knowledge retrieval module receives the task decomposition instructions distributed by the automated task planning and decomposition module, obtains the knowledge retrieval results by searching the internal knowledge base and the external data source, and transmits the knowledge retrieval results to the automated task planning and decomposition module and the automated material synthesis scheme design module; the automated material synthesis scheme design module generates and optimizes the material synthesis scheme according to the knowledge retrieval results, stores the synthesis scheme in the scheme library, and distributes the synthesis scheme to the automated experiment instruction generation module; the automated experiment instruction generation module converts the synthesis scheme into experimental instructions, stores the experimental instructions in the instruction code library, and transmits the experimental instructions to the automated experiment verification module; the automated experiment verification module executes the experimental instructions, monitors the experimental process and records the data, obtains the experimental results, stores the experimental results in the experimental log library, and transmits the experimental results to the automated experiment feedback optimization module; the automated experiment feedback optimization module provides feedback and optimization according to the experimental results and the initial task objectives, generates adjustment instructions and transmits them back to the automated task planning and decomposition module for improvement.

[0006] In some exemplary embodiments, the automated task planning decomposition module is an intelligent agent module that directly interacts with the user; the manager large model unit and the executor large model unit in the automated task planning decomposition module both include one or more task planning large language models; the manager large model unit in the automated task planning decomposition module includes a user input / output module, a message checking module, a status checking module, a message parsing and processing module, a task distribution module, and a result packaging module that are connected in sequence; the status checking module continuously monitors the user input and message queue through the message monitoring unit, and the result packaging module exchanges information with other intelligent agent modules through the message transmission unit; the executor large model unit in the automated task planning decomposition module includes a connected task decomposition planning module and a task table module, the task decomposition planning module receives the user's complete and clear requirement text output by the task distribution module, performs execution planning for the entire system, obtains the planning results, and transmits the planning results to the result packaging module through the task table module, and outputs a standardized table or json file; the message monitoring unit continuously monitors the user input and message queue, and when a clear end instruction sent by other intelligent agent modules is received, the multi-agent collaboration of the automated task planning decomposition module is completed, indicating that the user's material science research task has been completed.

[0007] In some exemplary embodiments, the automated knowledge retrieval module is an intelligent agent module that performs autonomous knowledge retrieval according to instructions; the manager large model unit and the executor large model unit in the automated knowledge retrieval module both include one or more retrieval enhanced large language models; the message monitoring unit of the automated knowledge retrieval module continuously monitors the message queue; upon receiving the subtask instruction of the automated task planning decomposition module or the optimization instruction of the automated experiment feedback optimization module, the automated knowledge retrieval module starts autonomous knowledge retrieval; upon receiving the clear end instruction sent by other intelligent agent modules, the multi-agent collaboration of the automated knowledge retrieval module is completed, indicating that the user's material science research task has been completed; the automated knowledge The retrieval-enhanced large language model of the knowledge retrieval module retrieves literature and book knowledge fragments related to user instructions in the knowledge base, independently determines the relevance of the content in the knowledge base, and eliminates low-relevance content; at the same time, the retrieval-enhanced large language model autonomously calls the online search tool, determines which search tool to use for knowledge search according to user instructions, extracts and summarizes the search results, and stores the sorted knowledge in the knowledge base; the retrieval-enhanced large language model summarizes the relevant knowledge screened out from the knowledge base and the knowledge fragments sorted out from the online search; while sending the summarized knowledge to other intelligent body modules through the message queue, the current summarized knowledge is temporarily stored in the memory cache unit.

[0008] In some exemplary embodiments, the automated material synthesis scheme design module is an intelligent agent module that calls the information of the knowledge base according to instructions and generates a material synthesis experimental scheme; the manager large model unit and the executor large model unit in the automated material synthesis scheme design module both include one or more material science large language models; the manager large model unit in the automated material synthesis scheme design module includes a message checking module, a status checking module, a message parsing processing module and a result packaging module that are connected in sequence; the message checking module continuously monitors user input and message queues through a message monitoring unit; the executor large model unit in the automated material synthesis scheme design module includes a keyword expansion module, a knowledge base retrieval module, a scheme generation module, a scheme optimization module and a scheme output module that are connected; the scheme output module is connected to the message transmission unit through the result packaging module, and the message transmission unit exchanges information with other intelligent agent modules; the keyword expansion module is used to realize the experimental scheme The keywords of the designed requirement text are expanded to improve the recall accuracy of the keywords in the knowledge base; the knowledge base retrieval module is used to extract N complete corpora related to the experimental steps from the pre-built knowledge base according to the expanded keywords; the solution generation module is used to generate a detailed experimental solution text according to the corpus information extracted by the knowledge base retrieval module, combined with the existing equipment information, according to the preset steps and format; the solution optimization module is used to optimize and check the solution according to the experimental solution text generated by the solution generation module to reduce the occurrence of errors; the solution output module is used to output the current solution to the user or subsequent intelligent agent processing, and store the solution and design requirements locally to facilitate subsequent data expansion and recording; the message monitoring unit continuously monitors the user input and message queue. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated material synthesis solution design module is completed, indicating that the user's material science research task has been completed.

[0009] In some exemplary embodiments, the automated experiment instruction generation module is an intelligent agent module that generates executable experiment instruction codes according to the material synthesis experiment plan, combined with the existing equipment and operation function list description of the experimental platform; the manager large model unit and the executor large model unit in the automated experiment instruction generation module both include one or more code enhancement large language models; the manager large model unit in the automated experiment instruction generation module includes a message checking module, a status checking module, a message parsing processing module and a result packaging module connected in sequence; the message checking module continuously monitors user input and message queues through a message listening unit; the executor large model unit in the automated experiment instruction generation module includes a connected experiment flow table generation module, an experiment parameter filling module, a parameter checking module and a standardized json generation module; the standardized json generation module is connected to the message transmission unit through the result packaging module, and the message transmission The delivery unit exchanges information with other intelligent agent modules; the experimental flow chart generation module is used to form an experimental flow chart for operable equipment according to the experimental plan steps and the pre-agreed equipment operation functions; the experimental parameter filling module is used to achieve a complete experimental parameter mapping from the experimental plan to the operation function for each step in combination with the pre-agreed equipment reference function; the parameter checking module is used to check and verify the experimental parameters to reduce the occurrence of errors; the standardized json generation module is used to generate a standardized json file based on the checked parameters; the json file can be parsed by the automated experiment platform, and the equipment can be automatically controlled to carry out experiments in sequence; the message monitoring unit continuously monitors the user input and the message queue. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment instruction generation module is completed, indicating that the user's material science research task has been completed.

[0010] In some exemplary embodiments, the automated experiment verification module is an intelligent agent module that receives experimental instructions, automatically conducts experiments according to the contents of the experimental instructions, and records data; the automated experiment verification module includes an automated platform for executing experiments, a robot system, and one or more multimodal large language models; the multimodal large language model includes a connected loading experimental instruction module, an experimental verification module, and a result data recording and output module; the experimental verification module includes an instruction checking module, a parameter optimization module, and an experimental execution module; the loading experimental instruction module loads the executable code confirmed by the automated experimental instruction generation module through a message monitoring unit; the experimental verification module is used to receive experimental instructions, and in combination with the environment The robot and the manipulator use the environment perception information to guide the robot and the manipulator to complete the experimental verification using the various instruments of the experimental platform. At the same time, the experimental data and experimental results are collected during the experiment. The result data recording and output module is used to store the experimental process and experimental results in the experimental log library in the knowledge base after the experiment is completed, and send the experimental process and experimental results to other intelligent modules through the message queue, while temporarily storing the current experimental log in the memory cache unit. The message monitoring unit continuously monitors the user input and the message queue. When a clear end instruction is received from other intelligent modules, the multi-agent collaboration of the automated experimental instruction generation module is completed, indicating that the user's material science research task has been completed.

[0011] In some exemplary embodiments, the automated experiment feedback optimization module is an intelligent agent module that provides feedback and optimization based on the experimental results and initial task objectives, generates adjustment instructions and transmits them back to related modules for improvement; the manager large model unit and the executor large model unit of the automated experiment feedback optimization module both include one or more feedback optimization large language models; after providing feedback and optimizing the experimental results and initial task objectives, the feedback optimization large language model generates adjustment instructions, and transmits the adjustment instructions to the automated task planning decomposition module through the message queue, forming a closed-loop feedback mechanism to ensure continuous optimization and improvement of the experimental process; the message monitoring unit continuously monitors user input and message queues, and when a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment feedback optimization module is completed, indicating that the user's material science research task has been completed.

[0012] On the second aspect, the embodiment of the present application also provides a multi-agent collaboration method for materials science, which is applied to the multi-agent collaboration system for materials science of the above-mentioned embodiment, including the following steps: the user starts the multi-agent collaboration system by inputting a research task instruction, and the instruction is first received and parsed by the automated task planning decomposition module; the automated task planning decomposition module decomposes the task into subtasks and passes them to the automated knowledge retrieval module for data retrieval; the material synthesis scheme design module generates a synthesis scheme based on the retrieved data; the automated experiment instruction generation module converts the synthesis scheme into specific experimental instructions, and transmits the experimental instructions to the automated experiment verification module; the automated experiment verification module controls the experimental platform to perform actual operations and collect experimental data, obtains experimental data and experimental results, and feeds back the experimental data and experimental results to the automated experiment feedback optimization module for analysis and optimization, obtains optimization suggestions, and transmits the optimization suggestions back to the automated task planning decomposition module to form a closed-loop feedback mechanism to ensure continuous optimization and improvement of the experimental process.

[0013] In some exemplary embodiments, the automated knowledge retrieval module performs data retrieval, including the following steps: the automated knowledge retrieval module retrieves knowledge fragments from documents and books related to user instructions in the knowledge base, and autonomously calls the online search tool to determine which search tool to use for knowledge search based on the user instructions to obtain search results; extracts and summarizes the search results, and stores the sorted related knowledge in the knowledge base; summarizes the related knowledge screened out from the knowledge base and the knowledge fragments sorted out from the online search; and temporarily stores the currently summarized knowledge in the memory cache while sending the summarized knowledge to other intelligent body modules through the message queue.

[0014] In some exemplary embodiments, the material synthesis scheme design module generates a synthesis scheme based on the retrieved data, including the following steps: expanding the keywords of the experimental scheme design requirement text to improve the recall accuracy of the keywords in the knowledge base; based on the expanded keywords, extracting N complete corpora related to the experimental steps from a pre-built knowledge base; based on the corpus information, combined with the existing equipment information, and according to the preset steps and format, generating a detailed experimental scheme text; based on the experimental scheme text, optimizing and checking the scheme to reduce the occurrence of errors; outputting the checked scheme to the user or other intelligent agents for processing, and storing the scheme and design requirements locally to facilitate subsequent data expansion and recording.

[0015] The technical solution provided by the embodiments of the present application has at least the following advantages.

[0016] The embodiment of the present application provides a multi-agent collaboration system and method for materials science, the system includes multiple agent modules that are visualized by a general agent module architecture; the general agent module architecture includes: a message queue / memory cache unit, a message listening unit, a manager large model unit, and an executor large model unit that are connected in sequence; the manager large model unit interacts with the message queue / memory cache unit through a message transmission unit. The system includes six agent modules, specifically: an automated task planning and decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module, and an automated experiment feedback optimization module.

[0017] On the one hand, this application builds a full-process system, covering the construction of a professional knowledge base, the decomposition of user-instructed task planning, the design of professional material synthesis plans, the generation of experimental code, and the implementation of automated experiments. In theory, this full-process system can use robots and AI systems to build a knowledge base based on simple user instructions and fully automatically complete the experimental research process, thus truly achieving a new paradigm of scientific research.

[0018] On the other hand, this application realizes the division of labor among multiple agents, and through multiple agents collaboratively handles complex materials science automation research tasks, each agent performs at its maximum efficiency in its own professional field, shares and reasonably allocates tasks, improves the overall processing power and coverage of the system, and breaks through the limitations of a single agent.

[0019] In addition, this application also provides an architecture and method for multiple agents to interact with each other. The new agent module architecture design allows it to exchange data with any other agent; compared with a single streaming system, it is more flexible in architecture. Each agent shares data in real time during the experiment, and through interactive analysis of each other's output results, dynamic feedback and optimization adjustments are performed to ensure the effectiveness and scientificity of each step and reduce manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplifications do not constitute limitations on the embodiments. Unless otherwise stated, the pictures in the drawings do not constitute proportional limitations.

[0021] Figure 1 A schematic diagram of the structure of a multi-agent collaboration system for materials science provided in one embodiment of the present application.

[0022] Figure 2 A schematic diagram of a general intelligent agent module architecture provided in accordance with an embodiment of the present application.

[0023] Figure 3An example flow chart of a general agent architecture provided for one embodiment of the present application.

[0024] Figure 4 An architectural diagram of a multi-agent collaboration system for materials science provided in accordance with one embodiment of the present application.

[0025] Figure 5 An architectural diagram of each general agent in a multi-agent collaborative system for materials science provided in one embodiment of the present application.

[0026] Figure 6 An architecture diagram of an automated task decomposition and planning agent module provided for one embodiment of the present application.

[0027] Figure 7 This is a diagram of the architecture of an intelligent agent module for designing an automated material synthesis solution provided in one embodiment of the present application.

[0028] Figure 8 An architecture diagram of an intelligent agent module for generating automated experimental instructions provided in one embodiment of the present application.

[0029] Fig. 9 A Json example diagram of an experimental step of an experimental instruction provided in an embodiment of the present application.

[0030] Fig.10 A schematic diagram of an automated experimental verification process provided in one embodiment of the present application. DETAILED DESCRIPTION

[0031] As can be seen from the background technology, the existing technology has deficiencies in autonomous retrieval, precise experiments, feedback optimization and collaborative work in materials science research.

[0032] At present, combining artificial intelligence and automation systems to carry out material science research is an important trend and direction in the current development of the field; using automation systems to carry out material science research is feasible in terms of hardware. In the past year, large models have been used to organize knowledge in the field of material science and build knowledge base question and answer libraries, which can improve the efficiency of obtaining experimental plans; based on large models, building multi-agent pipelines to improve the material science research paradigm is feasible and is also the direction of technological development.

[0033] Large Language Models (LLMs) are a type of deep learning model in the field of machine learning. Their basic network architecture is the transformer, which can generate and understand natural language by training on a large-scale corpus. In recent years, with the rapid growth of computing power and data volume, large language models have made significant progress, showing strong natural language processing capabilities, tool calling capabilities, and task planning capabilities. A number of excellent large language model products have emerged, such as ChatGPT and Tongyi Qianwen. Large language models can not only handle multiple language tasks, but also generate high-quality texts and show good performance in specific fields. In the field of materials science, large language models can efficiently process and analyze large amounts of literature data, generate research plans, experimental instructions, and optimization suggestions through natural language parsing, and greatly improve the automation and intelligence level of the research process.

[0034] There are related technologies that propose to use multi-agent pipelines for materials science research, constructing multiple large language models into pipelines to handle complex materials science research tasks, including knowledge acquisition, process generation, code generation and execution. The complete research and development process has been initially realized, which is a great improvement over traditional methods and has constructed a new research paradigm. However, the method is relatively rudimentary in structure. Due to the strong mutual dependence between multiple agents and insufficient flexibility, it is difficult to switch efficiently and work together between different tasks; a single pipeline system is not effective in handling different categories of tasks, and even requires the redesign of the pipeline and re-debugging of each agent, and its versatility is weak; a single pipeline system does not have feedback capabilities, and when large models have hallucinations, they are prone to errors (especially key steps such as generating experimental processes), and ultimately require manual intervention and adjustment, which is limited in use.

[0035] Therefore, in order to solve the above technical problems, the embodiment of the present application provides a multi-agent collaboration system and method for materials science, which includes multiple agent modules that are visualized by a general agent module architecture; the general agent module architecture includes: a message queue / memory cache unit, a message listening unit, a manager large model unit and an executor large model unit connected in sequence; the manager large model unit interacts with the message queue / memory cache unit through a message passing unit. The system includes six agent modules, specifically including: an automated task planning decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experimental instruction generation module, an automated experimental verification module, and an automated experimental feedback optimization module. The present application provides a multi-agent collaboration system and method for materials science, which aims to solve the deficiencies of the prior art in autonomous retrieval, precise experiments, feedback optimization and collaborative work in materials science research. The present application can use robots and AI systems to create a knowledge base based on simple instructions from users, and automatically complete the experimental research process, thereby truly achieving a new paradigm of scientific research.

[0036] The following will describe the various embodiments of the present application in detail with reference to the accompanying drawings. However, it will be appreciated by those skilled in the art that in the various embodiments of the present application, many technical details are provided in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solution claimed in the present application can be implemented.

[0037] refer to Figure 1 The embodiment of the present application provides a multi-agent collaboration system for materials science, including: multiple agent modules that are concretized by a general agent module architecture; the agent modules include: an automated task planning and decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module, and an automated experiment feedback optimization module; for simplicity, the above-mentioned agent modules can also be called: a task planning and decomposition module, a knowledge retrieval module, a material synthesis scheme design module, an experiment instruction generation module, an experiment verification module, and an experiment feedback optimization module.

[0038] Please continue to see Figure 1The automated task planning and decomposition module receives and parses user instructions, decomposes complex tasks into subtasks, and distributes the subtasks to other intelligent agent modules; other intelligent agent modules include: automated knowledge retrieval module, automated material synthesis scheme design module, automated experimental instruction generation module, automated experimental verification module, and automated experimental feedback optimization module. Specifically, the automated knowledge retrieval module receives the task decomposition instructions distributed by the automated task planning and decomposition module, obtains the knowledge retrieval results by searching the internal knowledge base and the external data source, and transmits the knowledge retrieval results to the automated task planning and decomposition module and the automated material synthesis scheme design module; the automated material synthesis scheme design module generates and optimizes the material synthesis scheme according to the knowledge retrieval results, stores the synthesis scheme in the scheme library, and distributes the synthesis scheme to the automated experiment instruction generation module; the automated experiment instruction generation module converts the synthesis scheme into experimental instructions, stores the experimental instructions in the instruction code library, and transmits the experimental instructions to the automated experiment verification module; the automated experiment verification module executes the experimental instructions, monitors the experimental process and records the data, obtains the experimental results, stores the experimental results in the experimental log library, and transmits the experimental results to the automated experiment feedback optimization module; the automated experiment feedback optimization module provides feedback and optimization according to the experimental results and the initial task objectives, generates adjustment instructions and transmits them back to the automated task planning and decomposition module for improvement.

[0039] The multi-agent collaboration system for materials science proposed in this application is mainly to build a multi-agent collaboration system, providing a highly automated and intelligent experimental system and method in the field of materials science. Through the collaborative work of multiple agents, a new paradigm of materials science research is realized, thereby improving the efficiency of materials science research and exploration. The multi-agent collaboration system of this application includes six agent modules with specific functions and cooperation capabilities, and each agent module operates independently and can make autonomous decisions and collaborate with each other; each agent is based on multiple large language models (LLMs) and related tools and subsystems.

[0040] like Figure 2 As shown, the general agent module architecture includes: a message queue / memory cache unit, a message listening unit, a manager large model unit and an executor large model unit connected in sequence; the manager large model unit exchanges information with the message queue / memory cache unit through a message passing unit. The multi-agent collaboration system for materials science provided in this application is visualized by the general agent module architecture to realize the functions of overall task planning, material science professional knowledge base construction, experimental plan generation, experimental code generation and execution, and experimental data feedback optimization.

[0041] Figure 3 An example flow of a general agent architecture is shown. Figure 3The messages involved in message monitoring and message transmission refer to the messages transmitted between agent modules. Since the transmitted messages are generally long, message queues / memory caches can be used to prevent loss or overflow; in order to ensure that the specified message of agent A is accurately obtained by agent B, the two parties can agree on a specified topic name and a specified data format; the messages transmitted in the message queue can contain data results, or text output by a large model, or other forms such as a path to a shared file.

[0042] At the physical level, communication between intelligent agents can be achieved through the data bus within the same computer system or through network communication between different computer systems. Data transmission can be achieved using TCP / IP.

[0043] The manager model mainly plays the role of a manager in the intelligent body, responsible for external message monitoring and message transmission, and is also responsible for communicating with the executor model group in the intelligent body. Its core is a large language model that is given a specific manager role, which realizes the ability of message parsing, task distribution, and result packaging. The periphery includes program modules such as message reception, message transmission, and status monitoring. Its working method is: when a message instruction is received through message monitoring, it first checks whether the message instruction can be processed by the intelligent body, otherwise it returns; the second step is to perform message parsing and processing, and check the current system status at the same time to decide whether to distribute the task to the executor model group; finally, it receives the results of the executor model group, packages them as required, and transmits messages.

[0044] Since the executor big model can include multiple big language models, the executor big model can also be called the executor big model group. The executor big model mainly plays the role of the executor in the intelligent body. It is composed of several series of big language models, tools, knowledge bases, etc., and its role is to specifically realize a specific function. Figure 3 This is just an example, and the actual situation may be different. For example, the executor model can answer knowledge in the field of material science, generate experimental plans, generate experimental codes, etc.

[0045] Based on the general intelligent agent module architecture, this application provides a solution and its organizational structure including six intelligent agent modules with specific functions and cooperation capabilities, such as Figure 4 As shown in the figure. These intelligent agent modules are: task planning decomposition module, data collection and processing module (i.e. knowledge retrieval module), material synthesis scheme design module, experimental instruction generation module, experimental verification module and experimental feedback optimization module. Specifically, the specific functions of each intelligent agent module are as follows.

[0046] (1) Task planning and decomposition module: Receives and parses user instructions, decomposes complex tasks into subtasks, and distributes them to various sub-modules for processing.

[0047] (2) Data collection and processing module: Receives task decomposition instructions, searches the internal knowledge base and external data sources (such as the Internet, academic paper database, etc.), and transmits the knowledge data back to the task planning decomposition module and other execution modules.

[0048] (3) Material synthesis scheme design module: Generate and optimize the material synthesis scheme based on the knowledge retrieval results, store it in the scheme library, and distribute it to the experimental instruction generation module.

[0049] (4) Experimental instruction generation module: converts the synthesis scheme into executable experimental instructions, stores them in the instruction code library, and transmits them to the experimental verification module.

[0050] (5) Experimental verification module: executes experimental instructions, monitors the experimental process and records data, stores it in the experimental log library, and sends the experimental results to the feedback optimization module.

[0051] (6) Experimental feedback optimization module: Based on the experimental results and initial task objectives, feedback and optimization are performed, and adjustment instructions are generated and sent back to the relevant intelligent agent modules for improvement.

[0052] Figure 5 The specific structural diagram of the multi-agent collaboration system for materials science provided by this application is shown. Figure 5 As shown in the figure, the multi-agent collaborative system uses six agents that are concretized by the general agent module architecture to perform overall task planning, material science professional knowledge base construction, experimental plan generation, experimental code generation and execution, and experimental data feedback optimization functions.

[0053] The execution method flow of the system is as follows: after the user inputs the research task, the task planning module decomposes the task into multiple subtasks and interacts with other intelligent agent modules through message queues and memory caches. The knowledge retrieval module retrieves relevant information from the internal knowledge base and external data sources to provide data support for the material synthesis scheme design module. The material synthesis scheme design module generates and optimizes the synthesis scheme, passes it to the code enhancement module, and converts the scheme into experimental instructions. The experimental verification module executes the experiment and collects experimental data, and the experimental feedback optimization module analyzes and optimizes according to the experimental results. Each module realizes data sharing and real-time communication through message queues and memory caches to form a closed-loop feedback mechanism. The entire system assists in completing material science research tasks in an efficient, automated and intelligent manner, significantly improving the efficiency and accuracy of research.

[0054] In some embodiments, the automated task planning decomposition module is an agent module that interacts directly with the user. Figure 6The schematic diagram of the structure of the automated task planning decomposition module is shown. This module understands the user's instructions and formulates all the subtasks that the entire system needs to perform under the requirements of completing the user's instructions, the order of subtask execution, and which agent module will perform the subtasks. After the subtasks are decomposed, the subtasks are distributed to the relevant agent modules through the message queue, and the current model output response is temporarily stored in the memory cache. Continuously monitor user input and message queues. When a clear end instruction is received from other agent modules, this round of multi-agent collaboration is completed, indicating that the user's material science research task has been completed.

[0055] The manager large model unit and the executor large model unit in the automated task planning decomposition module both include one or more task planning large language models; Figure 6 As shown, the manager large model unit in the automated task planning decomposition module includes a user input / output module, a message checking module, a status checking module, a message parsing and processing module, a task distribution module and a result packaging module connected in sequence; the status checking module continuously monitors the user input and the message queue through the message monitoring unit, and the result packaging module exchanges information with other intelligent agent modules through the message transmission unit; among which, the other intelligent agent modules include: an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module and an automated experiment feedback optimization module; the executor large model unit in the automated task planning decomposition module includes a connected task decomposition planning module and a task table module, the task decomposition planning module receives the user's complete and clear requirement text output by the task distribution module, performs execution planning for the entire system, obtains the planning results, and transmits the planning results to the result packaging module through the task table module, and outputs a standardized table or json file; the message monitoring unit continuously monitors the user input and the message queue, and when a clear end instruction sent by other intelligent agent modules is received, the multi-agent collaboration of the automated task planning decomposition module is completed, indicating that the user's material science research task has been completed.

[0056] For details, please continue to see Figure 6, the manager model in the automated task planning decomposition module can be a system containing a large language model and several subroutines. Among them, the workflow of message checking is: check whether the user input instruction is related to the overall system function. If it is related, go to the next step; if it is not related, reply the user with the default answer and do not process it later. The workflow of status checking is: before distributing instructions to the executor, check the status of the entire system and the executor. If there is a task in execution, prompt to wait, otherwise go to the next step. The workflow of message parsing processing is: interactively process the user's input instructions using a large language model to clarify the user's needs until the user confirms and goes to the next step. The workflow of task distribution is: pass the parsed message, that is, the user's complete and clear demand text, to the executor model to perform execution planning for the entire system. In the message monitoring part, when the clear end instruction sent by other modules is received, this round of multi-agent collaboration is completed, indicating that the user's material science research task has been completed.

[0057] The core of the executor model in the automated task planning decomposition module can be a single large language model, which gives the ability to carry out task planning through built-in prompt words. Its input is the user's complete and clear demand text, and the output is a standardized table or json file, which corresponds to which intelligent agent modules in the system need to be used in order to meet the user's needs, as well as the corresponding input instructions and output requirements when using these intelligent agent modules.

[0058] In some embodiments, the automated knowledge retrieval module is an agent module that performs autonomous knowledge retrieval according to instructions. The manager big model unit and the executor big model unit in the automated knowledge retrieval module both include one or more retrieval enhanced big language models; the message listening unit of the automated knowledge retrieval module continuously listens to the message queue; when receiving the subtask instruction of the automated task planning decomposition module or the optimization instruction of the automated experiment feedback optimization module, the automated knowledge retrieval module starts autonomous knowledge retrieval; when receiving the clear end instruction sent by other intelligent agent modules, the multi-agent collaboration of the automated knowledge retrieval module is completed, indicating that the user's material science research task has been completed; the retrieval enhanced big language model of the automated knowledge retrieval module retrieves the literature and book knowledge fragments related to the user's instructions in the knowledge base, autonomously judges the relevance of the content in the knowledge base, and eliminates the content with low relevance; at the same time, the retrieval enhanced big language model autonomously calls the online search tool, determines which search tool to use for knowledge search according to the user's instructions, extracts and summarizes the search results, and stores the sorted knowledge in the knowledge base; the retrieval enhanced big language model summarizes the relevant knowledge screened out from the knowledge base and the knowledge fragments sorted out from the online search; while sending the summarized knowledge to other intelligent agent modules through the message queue, the currently summarized knowledge is temporarily stored in the memory cache unit.

[0059] Specifically, upon receiving a subtask instruction from the automated task planning decomposition module or an optimization instruction from the automated feedback optimization module, the automated knowledge retrieval module begins autonomous knowledge retrieval through two main channels.

[0060] (1) The retrieval-enhanced large language model of the knowledge retrieval module first searches the knowledge base for knowledge fragments such as documents and books that are related to the user's instructions. The retrieval-enhanced large language model will autonomously determine the relevance of the content in the knowledge base and eliminate low-relevance content.

[0061] (2) Since the content in the knowledge base may be outdated, in order to make up for the lack of timeliness of the knowledge base, the retrieval-enhanced large language model will autonomously call online search tools and determine which search tool to use for knowledge search based on user instructions. In order to address the problem of disordered search engine results, the retrieval-enhanced large language model will extract and summarize the search results and store the sorted relevant knowledge in the knowledge base.

[0062] In some embodiments, the automated material synthesis scheme design module is an intelligent agent module that calls the information in the knowledge base according to instructions and generates a material synthesis experimental scheme. Figure 7 The schematic diagram of the structure of the automatic material synthesis scheme design module is shown. The manager big model unit and the executor big model unit in the automatic material synthesis scheme design module both include one or more material science big language models.

[0063] Specifically, the manager model of the automated material synthesis scheme design module is mainly responsible for communicating with other external demander modules. After receiving the instruction to design the experimental scheme and the demand text, it will convey it to the executor model to design the experimental scheme; finally, the design results will be returned to the demander as required. The demander here can be a user or other related intelligent agent modules, such as the task decomposition and planning module, or the experimental code generation module.

[0064] The executor model of the automated material synthesis scheme design module is mainly responsible for converting the experimental scheme design requirement text into a detailed experimental scheme, experimental steps, and experimental parameters.

[0065] like Figure 7 As shown, the manager large model unit in the automated material synthesis scheme design module includes a message checking module, a status checking module, a message parsing and processing module and a result packaging module connected in sequence; the message checking module continuously monitors user input and message queues through a message monitoring unit; the executor large model unit in the automated material synthesis scheme design module includes a keyword expansion module, a knowledge base retrieval module, a scheme generation module, a scheme optimization module and a scheme output module connected to each other; the scheme output module is connected to the message transmission unit through the result packaging module, and the message transmission unit exchanges information with other intelligent body modules; the keyword expansion module is used to expand the keywords of the requirement text of the experimental scheme design, thereby improving the recall accuracy of the keywords in the knowledge base; the knowledge base retrieval module is used to retrieve the keywords according to the expanded keywords , extracting N complete corpora related to the experimental steps from the pre-built knowledge base; the solution generation module is used to generate a detailed experimental solution text according to the corpus information extracted by the knowledge base retrieval module, combined with the existing equipment information, according to the preset steps and format; the solution optimization module is used to optimize and check the experimental solution text generated by the solution generation module to reduce the occurrence of errors; the solution output module is used to output the current solution to the user or subsequent intelligent agent processing, and store the solution and design requirements locally to facilitate subsequent data expansion and recording; the message monitoring unit continuously monitors the user input and message queue. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated material synthesis solution design module is completed, indicating that the user's material science research task has been completed.

[0066] The following will Figure 7 Each step / submodule is described in detail.

[0067] Keyword expansion: It can be a large language model designed based on prompt words1, which mainly expands the keywords of the experimental design requirement text to prompt its recall accuracy in the knowledge base. For example, black scale is expanded to BP, Black Phosphorus, etc.

[0068] Knowledge base retrieval and knowledge base: Based on keywords, using RAG and other methods, N complete corpora related to experimental steps are extracted from the pre-built knowledge base;

[0069] Solution generation: It can be a large language model designed based on prompt words 2. It is mainly based on the corpus information in the previous step, combined with the existing equipment information, and generates a detailed experimental solution text according to the preset steps and format;

[0070] Solution optimization: This can be a large language model designed based on prompt words 3. It is mainly based on the detailed experimental solution text in the previous step to achieve solution optimization and inspection to reduce the occurrence of errors;

[0071] Solution output: On the one hand, the current solution is output to the user or subsequent intelligent agent processing, and on the other hand, the solution + design requirements are stored locally for subsequent data expansion and recording. When the solution is output to the user, it can also be labeled based on the user's feedback.

[0072] In some embodiments, the automated experiment instruction generation module is an intelligent agent module that generates executable experiment instruction codes according to the material synthesis experiment plan and in combination with the equipment and operation function list descriptions of the experimental platform; Figure 8 The schematic diagram of the structure of the automatic experiment instruction generation module is shown. The manager big model unit and the executor big model unit in the automatic experiment instruction generation module both include one or more code enhanced big language models.

[0073] The manager model of the automated experiment instruction generation intelligent agent module is mainly responsible for communicating with other external demand-side modules. When receiving a request for experimental instruction generation, the obtained material synthesis experiment plan will be conveyed to the executor model to generate experimental instructions and output them in standard json format. After completion, the results will be transmitted to the demander; the json file of the experimental instructions is stored in the shared memory in the form of a file for the experiment of the experimental platform to call; it can also be displayed to the user for secondary confirmation.

[0074] The executor model of the automated experiment instruction generation agent module is mainly responsible for converting the experimental plan steps in the text into a json file that can be executed by the automated experiment platform. It can be composed of one or more large language models in series, and realize the corresponding text processing function through the specified prompt words.

[0075] like Figure 8 As shown, the manager large model unit in the automated experiment instruction generation module includes a message checking module, a status checking module, a message parsing and processing module and a result packaging module connected in sequence; the message checking module continuously monitors the user input and the message queue through the message monitoring unit; the executor large model unit in the automated experiment instruction generation module includes a connected experiment flow table generation module, an experiment parameter filling module, a parameter checking module and a standardized json generation module; the standardized json generation module is connected to the message passing unit through the result packaging module, and the message passing unit exchanges information with other intelligent body modules; the experiment flow table generation module is used to form an experiment flow table for operable equipment according to the steps of the experiment plan and in combination with the pre-agreed equipment operation functions; that is, what steps the experiment is divided into and what equipment needs to be operated in each step. The experimental parameter filling module is used to achieve a complete experimental parameter mapping from the experimental plan to the operation function for each step, combined with the pre-agreed equipment reference function; the parameter checking module is used to check and verify the experimental parameters to reduce the occurrence of errors; the standardized json generation module is used to generate a standardized json file based on the checked parameters; the json file can be parsed by the automated experiment platform, and the equipment can be automatically controlled to conduct experiments in sequence; the message monitoring unit continuously monitors user input and message queues, and when a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment instruction generation module is completed, indicating that the user's material science research task has been completed. The Json example of an experimental step of an experimental instruction is as follows: Fig. 9 shown.

[0076] It should be noted that a Json file is a file of simple data structures and objects stored in the JavaScript Object Notation format, which is a standard data exchange format and is mainly used to transmit data between Web applications and servers.

[0077] Fig.10 The structure diagram of the automated experiment verification module is shown. In some embodiments, the automated experiment verification module is an intelligent agent module that receives experimental instructions, automatically conducts experiments according to the contents of the experimental instructions, and records data; the automated experiment verification module includes an automated platform for executing experiments, a robotic system, and one or more multimodal large language models; Fig.10As shown, the multimodal large language model includes a connected loading experiment instruction module, an experiment verification module and a result data recording and output module; wherein the experiment verification module includes an instruction checking module, a parameter optimization module and an experiment execution module; the loading experiment instruction module loads the executable code confirmed by the automatic experiment instruction generation module through the message monitoring unit; the experiment verification module is used to receive the experiment instruction, and in combination with the environmental perception information, guide the robot and the mechanical arm to use a variety of instruments of the experimental platform to complete the experiment verification; at the same time, collect the experimental data and experimental results during the experiment; the result data recording and output module is used to store the experiment process and the experimental results in the experiment log library in the knowledge base after the experiment is completed, and send the experiment process and the experimental results to other intelligent agent modules through the message queue, while temporarily storing the current experiment log in the memory cache unit; the message monitoring unit continuously monitors the user input and the message queue, and when receiving the clear end instruction sent by other intelligent agent modules, the multi-agent collaboration of the automatic experiment instruction generation module is completed, indicating that the user's material science research task has been completed.

[0078] In some embodiments, the automated experiment feedback optimization module is an intelligent agent module that provides feedback and optimization based on the experimental results and initial task objectives, generates adjustment instructions and transmits them back to related modules for improvement; the manager large model unit and the executor large model unit of the automated experiment feedback optimization module both include one or more feedback optimization large language models; after providing feedback and optimizing the experimental results and initial task objectives, the feedback optimization large language model generates adjustment instructions, and transmits the adjustment instructions to the automated task planning decomposition module through the message queue, forming a closed-loop feedback mechanism to ensure continuous optimization and improvement of the experimental process; the message monitoring unit continuously monitors user input and message queues, and when a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment feedback optimization module is completed, indicating that the user's material science research task has been completed.

[0079] The embodiment of the present application also provides a multi-agent collaboration method for materials science, which is applied to the multi-agent collaboration system for materials science described in the above embodiment, and includes the following steps: the user starts the multi-agent collaboration system by inputting a research task instruction, and the instruction is first received and parsed by the automated task planning decomposition module; the automated task planning decomposition module decomposes the task into subtasks and passes them to the automated knowledge retrieval module for data retrieval; the material synthesis scheme design module generates a synthesis scheme based on the retrieved data; the automated experiment instruction generation module converts the synthesis scheme into specific experimental instructions, and transmits the experimental instructions to the automated experiment verification module; the automated experiment verification module controls the experimental platform to perform actual operations and collect experimental data, obtains experimental data and experimental results, and feeds back the experimental data and experimental results to the automated experiment feedback optimization module for analysis and optimization, obtains optimization suggestions, and transmits the optimization suggestions back to the automated task planning decomposition module to form a closed-loop feedback mechanism to ensure continuous optimization and improvement of the experimental process.

[0080] In some embodiments, the automated knowledge retrieval module performs data retrieval, including the following steps: the automated knowledge retrieval module retrieves knowledge fragments from documents and books related to user instructions in the knowledge base, and autonomously calls the online search tool to determine which search tool to use for knowledge search based on the user instructions to obtain search results; extracts and summarizes the search results, and stores the sorted related knowledge in the knowledge base; summarizes the related knowledge screened out from the knowledge base and the knowledge fragments sorted out from the online search; and temporarily stores the currently summarized knowledge in the memory cache while sending the summarized knowledge to other intelligent body modules through the message queue.

[0081] In some embodiments, a material synthesis scheme design module generates a synthesis scheme based on the retrieved data, including the following steps: expanding the keywords of the experimental scheme design requirement text to improve the recall accuracy of the keywords in the knowledge base; based on the expanded keywords, extracting N complete corpora related to the experimental steps from a pre-built knowledge base; based on the corpus information, combined with the existing equipment information, generating a detailed experimental scheme text according to preset steps and formats; based on the experimental scheme text, optimizing and checking the scheme to reduce the occurrence of errors; outputting the checked scheme to the user or other intelligent agents for processing, and storing the scheme and design requirements locally for subsequent data expansion and recording.

[0082] The multi-agent collaboration framework provided in this application is used for material science experiments. By using agents with different roles, each module works together to achieve full-process intelligence from task planning, knowledge retrieval, experimental design to experimental execution and feedback optimization. In addition, this application provides a general agent module architecture, including a manager large model and an executor large model. The manager large model plays the role of a manager in the agent, responsible for external message monitoring and message delivery, and is also responsible for communicating with the executor large model group inside the agent. The executor large model group plays the role of an executor in the agent. It consists of several large language models, tools, knowledge bases, etc. connected in series, and its role is to specifically implement a specific function.

[0083] This application realizes fully automated management of experimental tasks and goal-oriented intelligent control through multi-agent task coordination, data sharing and intelligent dialogue. This application uses large language models in material science research, and enhances the actual effectiveness of large language models in scientific research by using large prediction models to implement specific technologies in knowledge retrieval, program design, experimental execution and feedback optimization.

[0084] Compared with the prior art, the multi-agent collaboration system and method for materials science provided by the present application has the following advantages: on the one hand, compared with the existing multi-agent pipeline technology for materials science based on large language models, the present application shows significant advantages in many aspects, which are combined with the specific technical solutions of the present application, thereby achieving a high degree of automation and intelligence in the field of materials science research, and improving research efficiency and accuracy. On the other hand, the present application significantly improves the flexibility and dynamic adjustment capabilities of the system through a multi-agent collaboration framework. Compared with the multi-agent pipeline mode, the agent collaboration system can dynamically allocate and adjust tasks according to task requirements, avoiding the problems of interdependence and lack of flexibility between agents in the pipeline mode. Through the coordination of the automated task planning decomposition module, the work between the agents is more coordinated and consistent, effectively reducing the delays and errors that may occur during task conversion and information transmission.

[0085] In addition, this application has significant advantages in the implementation of technical solutions, including improving system flexibility, enhancing knowledge retrieval timeliness, optimizing material synthesis schemes, standardizing experimental instruction generation, precise experimental monitoring, and full-process feedback optimization. These advantages combined have enabled this application to achieve significant improvements in the automation and intelligence of materials science research, and compared with existing multi-agent pipeline technology, it has demonstrated better performance and application prospects.

[0086] Based on the above technical solution, the embodiment of the present application provides a multi-agent collaboration system and method for materials science, which includes multiple agent modules that are visualized by a general agent module architecture; the general agent module architecture includes: a message queue / memory cache unit, a message listening unit, a manager large model unit, and an executor large model unit that are connected in sequence; the manager large model unit interacts with the message queue / memory cache unit through a message transmission unit. The system includes six agent modules, specifically: an automated task planning and decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module, and an automated experiment feedback optimization module.

[0087] On the one hand, this application builds a full-process system, covering the construction of a professional knowledge base, the decomposition of user-instructed task planning, the design of professional material synthesis plans, the generation of experimental code, and the implementation of automated experiments. In theory, this full-process system can use robots and AI systems to build a knowledge base based on simple user instructions and fully automatically complete the experimental research process, thus truly achieving a new paradigm of scientific research.

[0088] On the other hand, this application realizes the division of labor among multiple agents, and through multiple agents collaboratively handles complex materials science automation research tasks, each agent performs at its maximum efficiency in its own professional field, shares and reasonably allocates tasks, improves the overall processing power and coverage of the system, and breaks through the limitations of a single agent.

[0089] In addition, this application also provides an architecture and method for multiple agents to interact with each other. The new agent module architecture design allows it to exchange data with any other agent; compared with a single streaming system, it is more flexible in architecture. Each agent shares data in real time during the experiment, and through interactive analysis of each other's output results, dynamic feedback and optimization adjustments are performed to ensure the effectiveness and scientificity of each step and reduce manual intervention.

[0090] Those skilled in the art can understand that the above-mentioned embodiments are specific examples for implementing the present application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present application. Any person skilled in the art can make their own changes and modifications without departing from the spirit and scope of the present application, so the scope of protection of the present application shall be based on the scope defined in the claims.

Claims

1. A multi-agent collaborative system for materials science, characterized in that: include: Multiple agent modules embodied by a general agent module architecture; The general agent module architecture includes: a message queue / memory cache unit, a message listening unit, a manager large model unit and an executor large model unit connected in sequence; the manager large model unit exchanges information with the message queue / memory cache unit through a message transmission unit; The intelligent agent module includes: an automated task planning and decomposition module, an automated knowledge retrieval module, an automated material synthesis scheme design module, an automated experiment instruction generation module, an automated experiment verification module and an automated experiment feedback optimization module; wherein, The automated task planning and decomposition module receives and parses user instructions, decomposes complex tasks into subtasks, and distributes the subtasks to other intelligent agent modules; other intelligent agent modules include: the automated knowledge retrieval module, the automated material synthesis scheme design module, the automated experiment instruction generation module, the automated experiment verification module, and the automated experiment feedback optimization module; The automated knowledge retrieval module receives the task decomposition instruction distributed by the automated task planning decomposition module, obtains knowledge retrieval results by searching the internal knowledge base and the external data source, and transmits the knowledge retrieval results to the automated task planning decomposition module and the automated material synthesis scheme design module; The automated material synthesis scheme design module generates and optimizes the material synthesis scheme according to the knowledge retrieval results, stores the scheme in a scheme library, and distributes the scheme to the automated experiment instruction generation module; The automated experiment instruction generation module converts the scheme into an experiment instruction, stores the experiment instruction in an instruction code library, and transmits the experiment instruction to the automated experiment verification module; The automated experiment verification module executes the experimental instructions, monitors the experimental process and records data, obtains the experimental results, stores the experimental results in the experimental log library, and transmits the experimental results to the automated experiment feedback optimization module; The automated experiment feedback optimization module performs feedback and optimization based on the experimental results and the initial task objectives, generates adjustment instructions and transmits them back to the automated task planning decomposition module for improvement.

2. The multi-agent collaborative system for materials science according to claim 1 is characterized in that: The automated task planning decomposition module is an intelligent agent module that directly interacts with the user; The manager large model unit and the executor large model unit in the automated task planning decomposition module each include one or more task planning large language models; The manager model unit in the automated task planning decomposition module includes a user input / output module, a message checking module, a status checking module, a message parsing and processing module, a task distribution module, and a result packaging module connected in sequence; the status checking module continuously monitors user input and message queues through a message monitoring unit, and the result packaging module interacts with other intelligent agent modules through a message transmission unit; The executor large model unit in the automated task planning and decomposition module includes a connected task decomposition planning module and a task table module. The task decomposition planning module receives the user's complete and clear requirement text output by the task distribution module, performs execution planning of the entire system, obtains the planning result, and transmits the planning result to the result packaging module through the task table module, and outputs a standardized table or json file; The message monitoring unit continuously monitors user input and message queues. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated task planning decomposition module is completed, indicating that the user's material science research task has been completed.

3. The multi-agent collaboration system for materials science according to claim 1 is characterized in that: The automated knowledge retrieval module is an intelligent agent module that performs autonomous knowledge retrieval according to instructions; The manager large model unit and the executor large model unit in the automated knowledge retrieval module each include one or more retrieval enhanced large language models; The message monitoring unit of the automated knowledge retrieval module continuously monitors the message queue; upon receiving the subtask instruction of the automated task planning decomposition module or the optimization instruction of the automated experiment feedback optimization module, the automated knowledge retrieval module starts autonomous knowledge retrieval; upon receiving the clear end instruction sent by other intelligent agent modules, the multi-agent collaboration of the automated knowledge retrieval module is completed, indicating that the user's material science research task has been completed; The retrieval-enhanced large language model of the automated knowledge retrieval module retrieves the literature and book knowledge fragments related to the user's instructions in the knowledge base, autonomously determines the relevance of the content in the knowledge base, and removes the content with low relevance; at the same time, the retrieval-enhanced large language model autonomously calls the online search tool, determines which search tool to use for knowledge search according to the user's instructions, extracts and summarizes the search results, and stores the sorted knowledge in the knowledge base; The retrieval-enhanced large language model will summarize the relevant knowledge screened out from the knowledge base and the knowledge fragments sorted out from the online search; while sending the summarized knowledge to other intelligent agent modules through the message queue, the current summarized knowledge will be temporarily stored in the memory cache unit.

4. The multi-agent collaboration system for materials science according to claim 1, characterized in that: The automated material synthesis scheme design module is an intelligent agent module that calls the information in the knowledge base according to the instructions and generates a material synthesis experimental scheme; The manager big model unit and the executor big model unit in the automated material synthesis scheme design module both include one or more material science big language models; The manager model unit in the automated material synthesis scheme design module includes a message checking module, a status checking module, a message parsing processing module and a result packaging module connected in sequence; the message checking module continuously monitors user input and message queues through a message monitoring unit; The executor large model unit in the automated material synthesis scheme design module includes a keyword expansion module, a knowledge base retrieval module, a scheme generation module, a scheme optimization module and a scheme output module connected to each other; the scheme output module is connected to the message transmission unit through the result packaging module, and the message transmission unit exchanges information with other intelligent agent modules; The keyword expansion module is used to expand the keywords of the requirement text of the experimental scheme design, thereby improving the recall accuracy of the keywords in the knowledge base; The knowledge base retrieval module is used to extract N complete corpora related to the experimental steps from the pre-built knowledge base according to the expanded keywords; The scheme generation module is used to generate a detailed experimental scheme text according to the corpus information extracted by the knowledge base retrieval module, combined with the existing equipment information, according to the preset steps and format; The scheme optimization module is used to optimize and check the scheme according to the experimental scheme text generated by the scheme generation module to reduce the occurrence of errors; The solution output module is used to output the current solution to the user or subsequent agent for processing, and store the solution and design requirements locally to facilitate subsequent data expansion and recording; The message monitoring unit continuously monitors user input and message queues. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated material synthesis scheme design module is completed, indicating that the user's material science research task has been completed.

5. The multi-agent collaboration system for materials science according to claim 1 is characterized in that: The automated experiment instruction generation module is an intelligent agent module that generates executable experiment instruction codes according to the material synthesis experiment plan and the equipment and operation function list description of the experimental platform; The manager large model unit and the executor large model unit in the automated experiment instruction generation module each include one or more code enhanced large language models; The manager model unit in the automated experiment instruction generation module includes a message checking module, a status checking module, a message parsing processing module and a result packaging module connected in sequence; the message checking module continuously monitors user input and message queues through a message monitoring unit; The executor large model unit in the automated experiment instruction generation module includes a connected experiment flow table generation module, an experiment parameter filling module, a parameter checking module and a standardized json generation module; the standardized json generation module is connected to the message transmission unit through the result packaging module, and the message transmission unit exchanges information with other intelligent agent modules; The experimental flow chart generation module is used to form an experimental flow chart for operating equipment according to the experimental plan steps and the pre-agreed equipment operation functions; The experimental parameter filling module is used to realize a complete experimental parameter mapping from the experimental scheme to the operation function for each step in combination with the pre-agreed equipment reference function; The parameter checking module is used to check and verify the experimental parameters to reduce the occurrence of errors; The standardized json generation module is used to generate a standardized json file according to the checked parameters; the json file can be parsed by the automated experiment platform, and the equipment can be automatically controlled to carry out experiments in sequence; The message monitoring unit continuously monitors user input and message queues. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment instruction generation module is completed, indicating that the user's material science research task has been completed.

6. The multi-agent collaboration system for materials science according to claim 1, characterized in that: The automated experiment verification module is an intelligent agent module that receives the experimental instructions, automatically conducts the experiment according to the contents of the experimental instructions, and records the data; The automated experiment verification module includes an automated platform for executing experiments, a robotic system, and one or more multimodal large language models; The multimodal large language model includes a connected loading experiment instruction module, an experiment verification module and a result data recording and output module; the experiment verification module includes an instruction checking module, a parameter optimization module and an experiment execution module; The loading experiment instruction module loads the executable code confirmed by the automated experiment instruction generation module through the message monitoring unit; The experimental verification module is used to receive experimental instructions and, in combination with environmental perception information, guide the robot and the mechanical arm to use a variety of instruments on the experimental platform to complete experimental verification; at the same time, collect experimental data and experimental results during the experiment; The result data recording and output module is used to store the experimental process and experimental results in the experimental log library in the knowledge base after the experiment is completed, and send the experimental process and experimental results to other intelligent agent modules through the message queue, while temporarily storing the current experimental log in the memory cache unit; The message monitoring unit continuously monitors user input and message queues. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment instruction generation module is completed, indicating that the user's material science research task has been completed.

7. The multi-agent collaboration system for materials science according to claim 1 is characterized in that: The automated experiment feedback optimization module is an intelligent agent module that performs feedback and optimization based on the experimental results and initial task objectives, generates adjustment instructions and transmits them back to related modules for improvement; The manager large model unit and the executor large model unit of the automated experiment feedback optimization module both include one or more feedback optimization large language models; After the feedback optimization large language model provides feedback and optimizes the experimental results and the initial task objectives, it generates adjustment instructions, and transmits the adjustment instructions to the automated task planning decomposition module through the message queue, forming a closed-loop feedback mechanism to ensure continuous optimization and improvement of the experimental process; The message monitoring unit continuously monitors user input and message queues. When a clear end instruction is received from other intelligent agent modules, the multi-agent collaboration of the automated experiment feedback optimization module is completed, indicating that the user's material science research task has been completed.

8. A multi-agent collaboration method for materials science, applied to a multi-agent collaboration system for materials science as claimed in any one of claims 1 to 7, characterized in that: The following steps are involved: The user starts the multi-agent collaborative system by inputting the research task instruction, which is first received and parsed by the automated task planning decomposition module; The automated task planning decomposition module decomposes the task into subtasks and passes them to the automated knowledge retrieval module for data retrieval; The material synthesis scheme design module generates a synthesis scheme based on the retrieved data; The automated experiment instruction generation module converts the synthesis scheme into specific experiment instructions and transmits the experiment instructions to the automated experiment verification module; The automated experiment verification module controls the experimental platform to perform actual operations and collect experimental data, obtains experimental data and experimental results, and feeds the experimental data and experimental results back to the automated experiment feedback optimization module for analysis and optimization, obtains optimization suggestions, and transmits the optimization suggestions back to the automated task planning and decomposition module, forming a closed-loop feedback mechanism to ensure continuous optimization and improvement of the experimental process.

9. The multi-agent collaboration method for materials science according to claim 8, characterized in that: The automated knowledge retrieval module performs data retrieval, including the following steps: The automated knowledge retrieval module searches the knowledge base for knowledge fragments of documents and books related to the user's instructions, and autonomously calls the online search tool to determine which search tool to use for knowledge search according to the user's instructions to obtain search results; Extract and summarize the search results, and store the sorted relevant knowledge in the knowledge base; Combine the relevant knowledge selected from the knowledge base and the knowledge fragments sorted out from online searches to summarize; While sending the summarized knowledge to other intelligent agent modules through the message queue, the current summarized knowledge is temporarily stored in the memory cache.

10. The multi-agent collaboration method for materials science according to claim 8, characterized in that: The material synthesis scheme design module generates a synthesis scheme according to the retrieved data, including the following steps: Expand the keywords of the experimental design requirement text to improve the recall accuracy of keywords in the knowledge base; Based on the expanded keywords, N complete corpora related to the experimental steps are extracted from the pre-built knowledge base; Based on the corpus information, combined with the existing equipment information, and according to the preset steps and formats, a detailed experimental plan text is generated; Based on the experimental protocol text, optimize and check the protocol to reduce the occurrence of errors; The checked plan is output to the user or other intelligent agents for processing, and the plan and design requirements are stored locally to facilitate subsequent data expansion and recording.

Images