CN-121998102-A - Material research system and method based on multi-agent game

CN121998102ACN 121998102 ACN121998102 ACN 121998102ACN-121998102-A

Abstract

The application relates to the technical field of computer systems, in particular to a material research system and method based on multi-agent game, wherein the system comprises a man-machine interaction module, a scheme generation module, a review module, an arbitration module, a scheme execution module and a knowledge base module, wherein the man-machine interaction module generates task instructions according to user requirements; the system comprises a scheme generation module, a review module, an arbitration module, a scheme execution module and a knowledge base module, wherein the scheme generation module is used for generating various candidate schemes, the review module is used for conducting multi-view cross review and risk interception on the candidate schemes, the arbitration module is used for processing opinion conflicts among multiple agents, the scheme execution module is used for executing experiments and collecting data, and the knowledge base module is used for providing static knowledge support and closed loop backflow of dynamic experiment data for the system. The application converts multi-objective conflict, knowledge uncertainty and theoretical and practical gaps in material research and development into a structured and computable dynamic game process, thereby systematically improving the comprehensiveness of scheme generation, the rigor of review and the success rate of experimental conversion.

Inventors

YU XUEFENG
XU WENHE
SHI TONGYU
LI YUTANG

Assignees

深圳先进技术研究院
中国科学院深圳先进技术研究院

Dates

Publication Date: 20260508
Application Date: 20260408

Claims (10)

1. The material research system based on the multi-agent game is characterized by comprising a man-machine interaction module, a scheme generation module, a review module, an arbitration module, a scheme execution module and a knowledge base module: the man-machine interaction module is used for generating task instructions according to user requirements; the scheme generation module is used for generating a plurality of candidate schemes respectively corresponding to different technical routes according to the task instruction; The evaluation module is used for performing multi-view cross evaluation and risk interception on the candidate schemes to obtain evaluation results of multiple agents; the arbitration module is used for processing opinion conflicts among the multiple agents according to the review results of the multiple agents and outputting a final decision; The scheme execution module is used for executing experiments and collecting data according to the final decision; the knowledge base module is used for providing static knowledge support and closed loop reflux of dynamic experimental data for the system.
2. The multi-agent game-based material research system of claim 1, wherein the human-machine interaction module comprises a receiving unit, a converting unit, a display unit and a human intervention interface; the receiving unit is used for receiving the research and development requirements input by a user, wherein the research and development requirements comprise performance indexes, cost budget and process limitation; The conversion unit is used for converting unstructured natural language and parameters in the research and development requirements into standardized task instructions which can be identified in the system; The display unit is used for visually displaying the decision process of the system, and presenting the review opinion interaction and the scheme iteration path among the multiple agents in the review module to a user so that the research and development process has transparency; the manual intervention interface is used for allowing a user to conduct final strategic decisions or parameter adjustment through manual intervention under the condition that an agreed decision cannot be achieved inside the system.
3. The multi-agent game-based material research system of claim 1 wherein the scenario generation module comprises an initial scenario generation unit and a screening unit; the initial scheme generating unit is used for designing candidate material schemes comprising raw material formulas, reaction paths and process parameters through reasoning and combination based on the existing scientific principles and knowledge reserves; The screening unit is used for primarily screening the rationality of the chemical structure of the candidate material scheme, so that the generated candidate scheme is ensured to have basic effectiveness in the theoretical aspect.
4. The multi-agent game-based material research system of claim 1 wherein the review module comprises a plurality of domain-specific perspective evaluation units each responsible for independent verification of candidate solutions in terms of scientific principles, engineering feasibility, and safety compliance; each evaluation unit not only outputs a quantified score, but also proposes a modification suggestion or an objection to the found defect.
5. The multi-agent game-based material research system of claim 1 wherein the arbitration module comprises a discrimination unit; When different evaluation units give out the conflicting evaluation results to the same scheme, the judgment unit automatically judges according to a preset decision logic or weight strategy, and a scheme with the optimal comprehensive benefit is selected; For complex cases involving critical risk or divergence exceeding a system threshold, the arbitration module triggers a suspension mechanism, routes decision weights to the human-machine interaction module to request manual confirmation.
6. The multi-agent game-based material research system of claim 1 wherein the protocol execution module is capable of dynamically selecting an adapted execution mode based on the complexity of the experimental protocol, operational fineness requirements, and current laboratory hardware resource configuration.
7. The material research system based on multi-agent game according to claim 1, wherein the knowledge base module is responsible for knowledge storage, management and service of the whole system, wherein static scientific knowledge and experimental data dynamically generated in the running process of the system are integrated in the knowledge base module; The knowledge base module provides retrieval service for the generation and review module through an indexing technology, so that the knowledge base module can refer to relevant historical experience in decision making, and can continuously update data storage of the knowledge base module by continuously collecting new experimental feedback data, thereby supporting self-correction and capability improvement of a system decision model.
8. A material research method based on multi-agent gaming, applied to the material research system based on multi-agent gaming as defined in any one of claims 1 to 7, comprising the steps of: S1, a user inputs a target requirement of material research and development through a man-machine interaction module, and a system analyzes a user instruction to generate a task instruction; step S2, based on the task instruction, a scheme generating module starts a design flow to generate a plurality of candidate schemes respectively corresponding to different technical routes; S3, each agent in the review module reviews the candidate schemes, and potential loopholes and risks of the schemes are found out from different dimensions, so that high-strength antagonism verification is initiated on the schemes; Step S4, when the scheme generating module submits the iterated scheme, the arbitration module firstly carries out initial judgment according to the grading model, if the main indexes of the scheme are not converged to the preset threshold value, the game is not balanced yet, the system keeps iterating in step S3, when the arbitration module judges that the scheme is towards perfection and the game process presents a convergence trend, a final-examination confirming mechanism is started to verify whether the system really reaches a Nash equilibrium state, the arbitration module submits the scheme to all the review agents again for review, and only when all the opponents do not propose critical objections and give a feasible conclusion consistently, the system confirms the game to achieve consistency consensus, and the arbitration module locks the scheme immediately and flows to a subsequent link; Step S5, the scheme execution module receives the final decision locked by the game, selects automatic execution or generates a standard operation procedure to guide manual execution according to experimental conditions, and acquires and formats data and a final result in the experimental process and then returns the data and the final result to the knowledge base module; Step S6, the final experimental result is fed back to the user through a human-computer interaction interface, the user carries out final value judgment, if the user confirms that the research and development target is achieved according to the measured data, the task is formally ended, if the user considers that the result does not reach the expectation, new feedback comments can be input, the system injects the new feedback into the game environment as new constraint conditions, the scheme generating module is restarted, the step S2 is skipped, the current balance state is broken, and a new round of countermeasure optimization cycle is started until the optimal solution meeting the user requirement is obtained.
9. The material research method based on multi-agent gaming of claim 8, wherein in step S2, the solution generating module is used as a generating party in the game, and starts the design flow according to the requirement of the task instruction; The scheme generating module firstly carries out wide-area retrieval on a knowledge base, retrieves the existing historical experimental cases, academic documents, professional books, patent data and other unstructured knowledge contents, then utilizes a large model to carry out comprehensive reasoning on the information, constructs one or more initial candidate schemes in a limited search space, and submits the initial candidate schemes to the arbitration module as targets of first-round games.
10. The multi-agent game-based material research method of claim 8, wherein in step S3, the arbitration module introduces the candidate solution into a review environment, each agent in the review module acts as an opponent, attempting to find potential vulnerabilities and risks of the solution from different dimensions, thereby initiating a high-intensity opponent verification of the solution; The arbitration module plays a role of a game manager in the process, is responsible for collecting scattered antagonism opinions and carrying out aggregation and duplication removal, balances benefits of all parties according to strategy weights when conflict exists in review results of different dimensions, forms unified correction feedback, carries out targeted completion and optimization on a scheme according to the feedback, and puts a new scheme into a review environment again, so that dynamic game circulation that the generator continuously perfects defense and the opponents continuously find vulnerabilities is formed.

Description

Material research system and method based on multi-agent game Technical Field The application relates to the technical field of computer systems, in particular to a material research system and method based on multi-agent game. Background The research paradigm of materials science is undergoing a process of transition from "trial and error" to "data driven" and "knowledge driven". Traditional material research and development relate to complicated links such as literature investigation, scheme design, synthesis experiments, characterization tests and the like, and the material research and development method is long in period, high in cost and extremely dependent on personal experience and intuition of scientific researchers. In recent years, the explosion of Large Language Models (LLM) has brought new opportunities for material science. LLM is based on massive text training, and has powerful logic reasoning, code generation and tool calling capabilities. LLM is combined with an automatic laboratory, and an AI scientist is constructed to automatically complete the whole process of material research and development, so that the LLM becomes the current forefront research hotspot. The method can not only liberate human beings from repeated labor, but also explore a wide chemical space which is difficult for human beings to intuitively reach. In the prior art, when artificial intelligence is applied to material research and development, the following limitations generally exist, so that the feasibility, reliability and engineering landing success rate of a design scheme are not high: (1) And the problem that the multi-objective optimization is difficult to dynamically cooperate. The material design scheme is required to simultaneously balance a plurality of targets which are frequently conflicting with each other, such as performance, cost, safety, process feasibility and the like. The existing method mostly adopts fixed weights to carry out linear weighting or sequential optimization, and cannot realize dynamic and self-adaptive weighing and decision making according to specific contents of schemes, so that an optimization result is often stiff, and the necessary flexible compromise in a real engineering scene is difficult to reflect. (2) The scheme evaluation mechanism lacks substantial countermeasures, which results in the problem of insufficient defect detection depth. The current multi-agent-based system adopts a sequential execution or cooperation verification mode, so that the target consistency of an evaluation module and a generation module is too high, and an effective balancing and challenge mechanism is lacked. This makes it difficult for the assessment to stay on surface compliance inspection, to make deep questions and stress tests on the underlying assumptions, intrinsic logic and potential risks of the solution, and for errors and "hallucinations" in the generated content to be difficult to discover and correct in time. (3) And the problems of disjoint theoretical prediction and physical experiment verification and weak closed loop feedback effectiveness. The existing virtual-real combination system generally only uses experimental data for indirect optimization of model parameters, and cannot be constructed as a strong constraint and decisive verification link for a virtual design scheme. This results in a system that tends to preferentially fit historical or simulation data, but is insensitive to feedback that deviates from the expected failure results that occurs in real experiments, and cannot effectively drive the design strategy to make fundamental adjustments in the direction that engineering can achieve. In summary, the application provides an innovative antagonistic multi-agent architecture based on arbitration centers, which aims to create a next-generation intelligent material research and development platform with self-error correction capability, multi-objective collaborative decision capability and virtual-real closed-loop evolution capability. Disclosure of Invention The embodiment of the application provides a material research system and a method based on multi-agent game, which convert multi-objective conflict, knowledge uncertainty, theoretical and practical differences in material research and development into a structured and computable dynamic game process, thereby systematically improving the comprehensiveness of scheme generation, the rigor of review and the success rate of experimental conversion. In order to solve the technical problems, in a first aspect, the embodiment of the application provides a multi-agent game-based material research system, which comprises a man-machine interaction module, a scheme generation module, a review module, an arbitration module, a scheme execution module and a knowledge base module, wherein the man-machine interaction module is used for generating task instructions according to user requirements, the scheme generation module is used f