Multi-agent driven intelligent experiment design system, method, equipment and medium

Abstract

The present invention discloses a multi-agent-driven intelligent experiment design system, method, equipment, and medium, belonging to the field of chemical experiment design systems. The system includes a scheme design unit and a step generation unit. Through debate and collaboration between multiple agents driven by different large models, combined with the reasoning and optimization capabilities of multiple large models, it can automatically generate and optimize chemical experiment design schemes and experimental steps. During this process, the system continuously iterates and optimizes to ensure the efficiency, scientific nature, and operability of the experimental scheme. The system can effectively solve design problems in complex chemical experiments, promote the development of intelligent chemical laboratories, and improve the efficiency and accuracy of experimental design.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical experiment design systems, and in particular to an intelligent experiment design system and method driven by multiple agents based on multiple models. Background Art

[0002] As the complexity and scale of chemical experiments continue to expand, traditional laboratories face numerous challenges in operational efficiency, resource allocation, and manual intervention, especially in high-demand fields such as substance and materials discovery and drug development. Research in these areas often involves complex, multi-step processes and the collection and processing of large amounts of data. Consequently, chemical laboratories are increasingly in need of automated and intelligent technologies, hoping to accelerate scientific research and reduce costs by reducing human error and improving experimental efficiency.

[0003] In recent years, breakthroughs have been made in artificial intelligence technology, particularly in multi-agent systems driven by a single large model. These systems have opened up new possibilities for intelligent and automated autonomous chemical laboratories. Leveraging their powerful natural language processing and adaptive learning capabilities, large language models can efficiently analyze experimental designs, optimize experimental procedures, and dynamically adjust them based on actual needs. Multi-agent systems, through collaborative division of labor and resource optimization, significantly improve the efficiency of experimental management. These systems can not only process multiple experimental tasks in parallel but also dynamically optimize resource allocation through intelligent scheduling, thereby shortening experimental cycles and improving overall efficiency.

[0004] However, multi-agent systems driven by a single large model still have limitations in dealing with the diversity and scalability of experimental design. Since these systems usually rely on a single type of large model to handle independent process links, any complex cross-requirements or subtle changes may exceed the processing capabilities of the large model, thus affecting the quality and effectiveness of the experimental design. In addition, a single large model lacks an effective collaboration and communication mechanism, making it difficult to comprehensively solve problems through collective discussions from multiple perspectives. For the highly complex and uncertain field of chemical experiments, the limitations of a single large model often lead to a one-sided design of experimental plans, which cannot comprehensively consider the interrelationships and potential impacts of various factors, and may ultimately affect the accuracy and success rate of the experiment. Therefore, how to further improve the collaboration capabilities between multiple agents and enhance the flexibility and depth of the system on the basis of the existing multi-agent system driven by a single large model has become a key issue in improving intelligent chemical experiments.

[0005] In view of this, the present invention is proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-agent driven intelligent experiment design system, method, equipment and medium, which can effectively respond to the challenges in complex chemical experiments through mutual debate and collaboration between multiple agents driven by different large models, combined with the optimization capabilities of multiple large models, and automatically generate efficient, scientific and accurate experimental design plans and experimental steps, thereby promoting the development of intelligent chemical laboratories, realizing more efficient and more accurate scientific research work, and well solving the above-mentioned technical problems existing in the prior art.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A multi-agent driven intelligent experiment design system, comprising:

[0009] In the design unit, multiple debate agents driven by different large models engage in multiple rounds of debate on experimental design questions posed by human researchers, resulting in the experimental design solutions proposed by each debate agent. Scoring agents then evaluate the consistency of each answer in each round. Once the scores reach a predetermined value, voting agents then evaluate the multiple answers based on three dimensions: correctness, completeness, and rationality. The debate agent with the highest average score is selected as the optimal experimental design solution.

[0010] The step generation unit communicates with the scheme design unit and can take the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme as input. Multiple debater agents driven by different large models combine each other's answers with the provided list of chemical equipment workstations to conduct multiple rounds of debate on the optimal experimental design scheme, and obtain the answers of the global optimization experimental steps that match the optimal experimental design scheme given by each debater agent. After the scoring agent 2 scores the consistency of the answers in each round and the scores reach a predetermined value, the voting agent 2 scores multiple answers from the two dimensions of correctness and rationality, and selects the answer of the debater agent with a higher average score as the global optimization experimental step. Supplementary explanations and local optimization are performed on each experimental step of the global optimization experimental step.

[0011] A multi-agent driven intelligent experiment design method, using the system of the present invention, comprises:

[0012] Design Steps: The system's design unit uses multiple debate agents driven by different large models to conduct multiple rounds of debate on the experimental design questions posed by human researchers, obtaining the responses of each debate agent to the experimental design plan. Scoring agents then evaluate the consistency of each answer in each round. Once the scores reach a predetermined value, voting agents then evaluate each answer based on correctness, completeness, and rationality. The debate agent with the highest average score is selected as the optimal experimental design plan.

[0013] Experimental step generation step: The system's step generation unit takes the list of standard equipment workstations in a chemical laboratory and the optimal experimental design plan as input. Multiple debater agents driven by different large models conduct multiple rounds of debate on the optimal experimental design plan based on each other's answers and the provided list of chemical equipment workstations, and obtain the answers of each debater agent to the globally optimized experimental steps that match the optimal experimental design plan. After the scoring agent 2 scores the consistency of the answers in each round and the score reaches the predetermined value, the voting agent 2 scores each answer from the two dimensions of correctness and rationality, and selects the answer of the debater agent with the higher average score as the globally optimized experimental step. Supplementary explanations and local optimization are performed on each globally optimized experimental step.

[0014] A processing device comprising:

[0015] at least one memory for storing one or more programs;

[0016] At least one processor is capable of executing one or more programs stored in the memory. When the one or more programs are executed by the processor, the processor is enabled to implement the method described in the present invention.

[0017] A readable storage medium stores a computer program, which can implement the method described in the present invention when the computer program is executed by a processor.

[0018] Compared with the existing technology, the intelligent scientist system, method, device and medium for driving multiple agents with multiple models provided by the present invention have the following beneficial effects:

[0019] By adopting multiple intelligent agents driven by different large models and using an iterative optimization process of debate optimization and scoring selection, the optimal experimental design plan can be obtained. By optimizing the experimental steps, the accuracy and feasibility of the experimental steps can be gradually improved, and ultimately a complete, feasible and efficient experimental step plan can be generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 A block diagram of the multi-agent driven intelligent experiment design system provided in an embodiment of the present invention.

[0022] Figure 2 This is a processing flow chart of the scheme design unit of the multi-agent driven intelligent experiment design system provided in an embodiment of the present invention.

[0023] Figure 3 A processing flow chart of the step generation unit of the multi-agent driven intelligent experiment design system provided in an embodiment of the present invention.

[0024] Figure 4 This is a processing flow chart of the supplementary description generation module and the local optimization and scoring module of the solution generation unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the specific content of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments, and do not constitute a limitation of the present invention. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0026] First, the following terms may be used in this article:

[0027] The term “and / or” means that either or both of them can be realized at the same time. For example, X and / or Y includes both “X” or “Y” and “X and Y”.

[0028] The terms "include," "comprises," "contains," "has," or other similar expressions should be interpreted as non-exclusive. For example, "including certain technical features (such as raw materials, components, ingredients, carriers, dosage forms, materials, dimensions, parts, components, mechanisms, devices, steps, procedures, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products, or manufactured articles)" should be interpreted as including not only the technical features explicitly listed, but also other technical features known in the art that are not explicitly listed.

[0029] The term "consisting of" excludes any technical features not explicitly listed. If used in a claim, this term renders the claim closed, excluding any technical features other than those explicitly listed, except for conventional impurities associated with them. If this term appears only in a clause of a claim, it limits only the elements explicitly listed in that clause; elements listed in other clauses are not excluded from the claim as a whole.

[0030] Unless otherwise specified or limited, the terms "mounted," "connected," "connect," and "fixed" should be interpreted broadly. For example, they can refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediary; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in this document based on specific circumstances.

[0031] When concentration, temperature, pressure, size or other parameters are expressed in the form of a numerical range, the numerical range should be understood to specifically disclose all ranges formed by the pairing of any upper limit, lower limit, or preferred value within the numerical range, regardless of whether the range is explicitly stated. For example, if a numerical range of "2 to 8" is stated, the numerical range should be interpreted as including ranges of "2 to 7," "2 to 6," "5 to 7," "3 to 4 and 6 to 7," "3 to 5 and 7," "2 and 5 to 7," etc. Unless otherwise specified, the numerical ranges stated herein include both their endpoints and all integers and fractions within the numerical range.

[0032] The terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings and are only for the convenience and simplification of description, and do not explicitly or implicitly indicate that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore should not be understood as a limitation to this document.

[0033] The scheme provided by the present invention is described in detail below. The contents not described in detail in the examples of the present invention belong to the prior art known to professionals in this field. If specific conditions are not specified in the examples of the present invention, they are carried out according to conventional conditions in the field or conditions recommended by the manufacturer. If the manufacturer of the reagents or instruments used in the examples of the present invention is not specified, they are all conventional products that can be purchased commercially.

[0034] like Figure 1 As shown, an embodiment of the present invention provides a multi-agent driven intelligent experiment design system, comprising:

[0035] In the design unit, multiple debate agents driven by different large models engage in multiple rounds of debate on experimental design questions posed by human researchers, resulting in the experimental design solutions proposed by each debate agent. Scoring agents then evaluate the consistency of each answer in each round. Once the scores reach a predetermined value, voting agents then evaluate the multiple answers based on three dimensions: correctness, completeness, and rationality. The debate agent with the highest average score is selected as the optimal experimental design solution.

[0036] The step generation unit communicates with the scheme design unit and can take the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme as input. Multiple debater agents driven by different large models combine each other's answers with the provided list of chemical equipment workstations to conduct multiple rounds of debate on the optimal experimental design scheme, and obtain the answers of the global optimization experimental steps that match the optimal experimental design scheme given by each debater agent. After the scoring agent 2 scores the consistency of the answers in each round and the scores reach a predetermined value, the voting agent 2 scores multiple answers from the two dimensions of correctness and rationality, and selects the answer of the debater agent with a higher average score as the global optimization experimental step. Supplementary explanations and local optimization are performed on each experimental step of the global optimization experimental step.

[0037] Preferably, in the above system, the scheme design unit includes:

[0038] Initial response module, debate and global optimization module and consensus scoring and voting module; among them,

[0039] The preliminary response module consists of Debater Agent 1 and Debater Agent 2, each driven by a different large model and equipped with only the context memory of the module. These two Debater Agents can provide preliminary responses to the experimental design questions input by the human researcher and derive preliminary experimental design solutions from each Debater Agent.

[0040] The debate and global optimization module is composed of debater agent No. 3 and debater agent No. 4, which are driven by different large models and only have the context memory of the module. Both are in communication connection with the preliminary response module. The two debater agents can conduct multiple rounds of debate and optimization on the preliminary experimental design scheme given by the preliminary response module. In each round of debate, each debater agent analyzes, discusses and gives optimization suggestions to improve the preliminary experimental design scheme from a global perspective. During each round of optimization, each debater agent refers to existing feedback information and combines it with newly emerging relevant data to make adjustments in order to obtain a more optimal experimental design scheme.

[0041] The consensus scoring and voting module is composed of a scoring agent and a voting agent, and is communicated with the debate and global optimization module. The scoring agent can score each round of debate and the optimized answers of the two debater agents in the debate and global optimization module from three dimensions: correctness, completeness and rationality, with a scoring range of 0 to 10 points. If the score in a certain round reaches 10 points, it is confirmed that the two debater agents have reached a consensus on the current experimental design plan, and the debate process is stopped. The voting agent takes the average score of the two debater agents No. 3 and No. 4 as the average, and finally selects the experimental design plan of the debater agent with the higher average score as the optimal experimental design plan.

[0042] Preferably, in the above system, the step generation unit includes:

[0043] Experimental steps generation module, supplementary description generation module and local optimization and scoring module; among them,

[0044] The experimental step generation module is composed of debater agent No. 5, debater agent No. 6, debater agent No. 7, debater agent No. 8, scoring agent No. 2, and voting agent No. 2, all driven by different large models. It is connected to the scheme design unit in communication. Debater agent No. 5 and debater agent No. 6 can use the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme given by the scheme design unit as input, and make a preliminary response to the optimal experimental design scheme in combination with the other party's answer and the provided list of chemical equipment workstations. Debater agent No. 7 and debater agent No. 8 conduct multiple rounds of debate to obtain the global optimized experimental steps matching the optimal experimental design scheme given by each debater agent. Scoring agent No. 2 scores the consistency of the answers of debater agent No. 7 and debater agent No. 8 in each round and after the scores reach a predetermined value, voting agent No. 2 scores the two answers from the two dimensions of correctness and rationality, and finally selects the answer of the debater agent with a higher average score as the global optimized experimental step;

[0045] The supplementary description generation module uses a supplementary description agent, which is in communication with the experimental step generation module and can extract supplementary description information for each experimental step from the global optimization experimental steps given by the experimental step generation module and the experimental design questions given by human researchers;

[0046] The local optimization and scoring module is composed of a host agent, a radical agent and a conservative agent driven by different large models. It is communicated with the supplementary description generation module and the experimental step generation module. It can combine the supplementary description information of each experimental sub-step provided by the supplementary description generation module. The host agent, the radical agent and the conservative agent use a multi-agent optimization method based on Markov chain to perform local optimization on each experimental step after global optimization one by one until all experimental steps are optimized.

[0047] Preferably, in the above system, in the local optimization and scoring module, the moderator agent, the radical agent, and the conservative agent use a Markov chain-based multi-agent optimization method to perform local optimization on each experimental step after global optimization one by one in the following manner:

[0048] During the local optimization process, there are two states: the initial state and the normal state. The initial state is the global optimization experimental step provided by the experimental step generation module. The normal state includes two different discussion modes that are carried out alternately, initiated by radical agents or conservative agents respectively. The former expands and optimizes the current experimental step in the order of radical agent-conservative agent-host agent; the latter enhances the stability and feasibility of the experimental step in the order of conservative agent-radical agent-host agent.

[0049] During each round of optimization, the host agent collects the experimental steps of the initial state and the two-party answers obtained according to one of the discussion modes to form three optimization plans, and scores and votes based on the rationality of the optimization plans. The plan with the highest score is selected as the new initial state to enter the next round of optimization, forming a state transfer based on the Markov chain. The optimization process continues until the host agent's score for the current experimental step reaches a predetermined standard of less than or equal to 1 point, that is, the experimental step is considered to have completed local optimization. Subsequently, the optimization process enters the next experimental step until the entire experimental step is optimized.

[0050] Preferably, in the above system, the large model driving debater agent No. 1, debater agent No. 2, debater agent No. 3, debater agent No. 4, debater agent No. 5, debater agent No. 6, debater agent No. 7, debater agent No. 8, radical agent and conservative agent is any one of llama3.0-70b and llama3.1-70b-instruct, and the large models used by different debater agents in the same module are different;

[0051] The large model that drives the voting and scoring agent and the host agent is GPT-4o.

[0052] See also Figure 2、 Figure 3 、 Figure 4 The embodiment of the present invention provides a multi-agent driven intelligent experiment design method, which adopts the above-mentioned system and includes:

[0053] Design Steps: The system's design unit uses multiple debate agents driven by different large models to conduct multiple rounds of debate on the experimental design questions posed by human researchers, obtaining the responses of each debate agent to the experimental design plan. Scoring agents then evaluate the consistency of each answer in each round. Once the scores reach a predetermined value, voting agents then evaluate each answer based on correctness, completeness, and rationality. The debate agent with the highest average score is selected as the optimal experimental design plan.

[0054] Experimental step generation step: The system's step generation unit takes the list of standard equipment workstations in a chemical laboratory and the optimal experimental design plan as input. Multiple debater agents driven by different large models conduct multiple rounds of debate on the optimal experimental design plan based on each other's answers and the provided list of chemical equipment workstations, and obtain the answers of each debater agent to the globally optimized experimental steps that match the optimal experimental design plan. After the scoring agent 2 scores the consistency of the answers in each round and the score reaches the predetermined value, the voting agent 2 scores each answer from the two dimensions of correctness and rationality, and selects the answer of the debater agent with the higher average score as the globally optimized experimental step. Supplementary explanations and local optimization are performed on each globally optimized experimental step.

[0055] Preferably, in the above method, the scheme design step includes:

[0056] Through the preliminary response module of the scheme design unit, Debater Agent 1 and Debater Agent 2, which are driven by different large models that constitute the module and only have the context memory of the module, make preliminary responses to the experimental design questions input by the human researcher and obtain preliminary experimental design schemes given by each debater agent;

[0057] Through the debate and global optimization module of the scheme design unit, Debater Agents 3 and 4, which are driven by different large models that make up the module and only have context memory of the module, conduct multiple rounds of debate and optimization on the preliminary experimental design scheme given by the preliminary response module. In each round of debate, each debater agent analyzes, discusses and gives optimization suggestions to improve the preliminary experimental design scheme from a global perspective. During each round of optimization, each debater agent refers to existing feedback information and combines it with newly emerging relevant data to make adjustments in order to obtain a more optimal experimental design scheme.

[0058] Through the consensus scoring of the scheme design unit and the scoring agent 1 of the voting module, each round of debate and the optimized answers of the two debater agents in the debate and global optimization module are scored from three dimensions: correctness, completeness and rationality, with a scoring range of 0 to 10 points. If the score in a certain round reaches 10 points, it is confirmed that the two debater agents have reached a consensus on the current experimental design scheme, and the debate process is stopped. The consensus scoring and voting agent 1 of the voting module will use the average score of the two debater agents No. 3 and No. 4 to finally select the experimental design scheme of the debater agent with the higher average score as the optimal experimental design scheme.

[0059] Preferably, in the above method, the experimental step generating step includes:

[0060] Debater agent No. 5 and debater agent No. 6, which are driven by different large models of the experimental step generation module of the step generation unit, take the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme given by the scheme design unit as input, and make a preliminary response to the optimal experimental design scheme in combination with the other party's answer and the provided list of chemical equipment workstations. Debater agent No. 7 and debater agent No. 8 conduct multiple rounds of debate to obtain the global optimized experimental steps matching the optimal experimental design scheme given by each debater agent. Scoring agent No. 2 scores the consistency of the answers of debater agent No. 7 and debater agent No. 8 in each round and after the scores reach a predetermined value, voting agent No. 2 scores the two answers from the two dimensions of correctness and rationality, and finally selects the answer of the debater agent with the higher average score as the global optimized experimental step;

[0061] Extracting supplementary explanation information for each experimental step from the global optimization experimental steps given by the experimental step generation module and the experimental design problem given by the human researcher through the supplementary explanation intelligent agent of the supplementary explanation generation module of the step generation unit;

[0062] Through the local optimization of the step generation unit combined with the supplementary explanation information of each experimental sub-step provided by the supplementary explanation generation module, the host agent, radical agent and conservative agent use the multi-agent optimization method based on Markov chain to perform local optimization on each experimental step after global optimization one by one until all experimental steps are optimized. The host agent, radical agent and conservative agent driven by different large models from the scoring module,

[0063] Preferably, in the above method, the host agent, radical agent, and conservative agent use a Markov chain-based multi-agent optimization method to perform local optimization on each experimental step after global optimization one by one in the following manner:

[0064] During the local optimization process, there are two states: the initial state and the normal state. The initial state is the global optimization experimental step provided by the experimental step generation module. The normal state includes two different discussion modes that are carried out alternately, initiated by radical agents or conservative agents respectively. The former expands and optimizes the current experimental step in the order of radical agent-conservative agent-host agent; the latter enhances the stability and feasibility of the experimental step in the order of conservative agent-radical agent-host agent.

[0065] During each round of optimization, the host agent collects the experimental steps of the initial state and the two-party answers obtained according to one of the discussion modes to form three optimization plans, and scores and votes based on the rationality of the optimization plans. The plan with the highest score is selected as the new initial state to enter the next round of optimization, forming a state transfer based on the Markov chain. The optimization process continues until the host agent's score for the current experimental step reaches a predetermined standard of less than or equal to 1 point, that is, the experimental step is considered to have completed local optimization. Subsequently, the optimization process enters the next experimental step until the entire experimental step is optimized.

[0066] An embodiment of the present invention further provides a processing device, comprising:

[0067] at least one memory for storing one or more programs;

[0068] At least one processor can execute one or more programs stored in the memory, and when the one or more programs are executed by the processor, the processor can implement the above method.

[0069] The embodiments of the present invention further provide a readable storage medium storing a computer program, which can implement the above method when executed by a processor.

[0070] In summary, it can be seen that the system and method provided by the embodiments of the present invention, by adopting multiple intelligent agents driven by different large models, through the iterative optimization process of debate optimization and scoring selection, can obtain the optimal experimental design plan, optimize the experimental steps, gradually improve the accuracy and feasibility of the experimental steps, and ultimately generate a complete, feasible and efficient experimental step plan.

[0071] In order to more clearly demonstrate the technical solution and technical effects provided by the present invention, the solution provided by the embodiment of the present invention is described in detail with reference to specific embodiments below.

[0072] Example 1

[0073] This embodiment provides a multi-agent driven intelligent experiment design system based on a multi-model, including: a scheme design unit and a step generation unit; specifically:

[0074] (1) Scheme design unit:

[0075] See also Figure 2 The main function of this unit is to automatically generate and optimize the optimal experimental design based on the chemical experimental problems proposed by human researchers. This unit includes the following core modules:

[0076] Initial Response Module: This module consists of multiple agents based on different types of large models. For example, Debater Agent 1 and Debater Agent 2, driven by different large models and equipped only with contextual memory for this module, are responsible for providing preliminary responses to research questions or experimental design requirements input by human researchers. Each agent, based on its expertise and domain knowledge, proposes a preliminary design solution. At this stage, the system can also integrate online search tools to obtain the latest information in the relevant field in real time to enrich and refine the initial response. The results of the initial response provide a foundation for subsequent debate and solution optimization.

[0077] Debate and Global Optimization Module: This module consists of multiple agents based on different large-scale models. For example, Debater Agents 3 and 4, driven by different large-scale models and equipped only with contextual memory from this module, engage in multiple rounds of debate and optimization around preliminary design proposals, leveraging their respective expertise and data. During each round, the agents analyze, discuss, and propose optimization suggestions, focusing on improving the design from a global perspective. During the optimization process, the agents not only consider existing feedback but also adapt to emerging relevant data to optimize the experimental design.

[0078] Consensus Scoring and Voting Module: This module consists of Scoring Agent 1 and Voting Agent 1. After each round of debate and optimization, Scoring Agent 1 evaluates the optimized design solution, scoring it on a scale of 0 to 10. If the score reaches 10 in a given round, it indicates that all debate agents have reached consensus on the design solution, and the debate process stops and the voting phase begins. At this point, Voting Agent 1 selects the optimal design solution based on the scores and optimization suggestions of the participating agents, which becomes the final experimental design for the system.

[0079] In the above-mentioned unit work, when a human researcher raises a question, the preliminary response module uses two debater agents driven by different large models, namely Debater Agent 1 and Debater Agent 2, to make preliminary responses in round 0 to generate a preliminary experimental design plan; after each round of debate, the scoring agent scores the consistency of the two debaters' debate views, and the initial score defaults to 0 points, which serves as the debate scoring benchmark; subsequently, the input prompt words of Debater Agent 1 and Debater Agent 2 will be changed to the prompt words of the debate and global optimization module, and the memory of the preliminary response will be retained, and the first round of debate will officially begin as Debater Agent 3 and Debater Agent 4.

[0080] In each round of formal debate, each debater receives the other debater's previous point of view as a reference for their current round, using dialectical thinking to globally optimize a new experimental design. Subsequently, a scoring agent determines the consistency of the two debaters' views and assigns a score. If the score does not reach 10 points, the above process is repeated for a new round of debate. If the score reaches 10 points, indicating that the two debaters' views are consistent, the voting agent selects the more correct, complete, and reasonable view as the final optimal experimental design after global optimization.

[0081] (2) Step generation unit:

[0082] See also Figure 3 and Figure 4 The main function of this unit is to automatically generate and optimize the optimal experimental steps based on the optimal experimental design plan obtained by the experimental design plan generation and optimization unit, making each step more accurate and operational. This unit includes the following modules:

[0083] Experimental Step Generation Module: This module consists of multiple agents based on different large-scale models, such as Debater Agents 5, 6, 7, and 8, as well as Scoring Agent 2 and Voting Agent 2, all driven by different large-scale models. It is responsible for generating the optimal experimental steps based on the final experimental design. This process is similar to the preliminary response and global optimization process of the previous framework. Based on its domain knowledge, research experience, and the provided list of chemical equipment workstations, the large-scale model proposes a preliminary design of the experimental steps. The generated experimental steps cover every operational process, experimental parameters, and required equipment, ensuring the feasibility and scientificity of the steps.

[0084] The Supplementary Notes Generation Module utilizes a Supplementary Notes Agent to generate detailed supplementary notes for each experimental step. These notes not only include detailed instructions for each experimental step, but also cover potential experimental risks, precautions, and common errors, helping researchers improve experimental success rates and avoid potential operational errors. The Supplementary Notes Module analyzes historical experimental data and literature to ensure that the supplementary notes for each step are accurate, reliable, and actionable.

[0085] Local Optimization and Scoring Module: This module utilizes multiple agents based on different macro models, such as a moderator agent, radical agents, and conservative agents driven by different macro models, to locally optimize each experimental step after global optimization. This stage utilizes a multi-agent optimization approach based on a Markov chain. The optimization process involves multiple agents, with radical agents and conservative agents each proposing different optimization solutions. Radical agents typically propose more innovative or bold optimization solutions, while conservative agents advocate for cautious and conservative improvements. The moderator agent collects and evaluates the optimization results of both agents, critically examining both rational and irrational aspects. Ultimately, it forms its own opinion and conducts multiple rounds of optimization and adjustment based on a set scoring criteria (typically ranging from 0 to 5). Based on the Markov chain definition, the globally optimized experimental steps are defined as the initial state. In addition, two normal states are defined, which can be viewed as two different discussion modes. The first normal state is the discussion initiated by the radical agent, following the order radical-conservative-moderator. This mode aims to diverge and expand upon existing content before further analyzing the previous experimental steps. The second regular state is initiated by the conservative agent, with the sequence being conservative, radical, and then moderator. This mode tends to reinforce the acceptance of the previous experimental steps before introducing divergent content. The optimization process in this embodiment continuously oscillates between these two debate modes to reach the optimal judgment. This process continues iteratively until the score for each experimental step reaches a predetermined standard (typically ≤1), indicating that the step has been fully optimized, ensuring the operability and reliability of the experimental step.

[0086] The function of the step generation unit of the present embodiment is to generate and globally optimize the experimental steps. Among them, the input of the experimental step generation module is the experimental design scheme after global optimization and the equipment workstation list of the chemical laboratory. In the preliminary response stage, the No. 5 debater intelligent agent and the No. 6 debater intelligent agent will respectively receive the experimental design scheme and the provided chemical laboratory standard equipment workstation list as input, generate the experimental steps of the preliminary response, and then the No. 7 debater intelligent agent and the No. 8 debater intelligent agent will launch a formal debate. During the debate, the intelligent agent debaters will combine the other party's views and the provided chemical equipment workstation list to carry out the debate on the question and optimize their own new experimental step answers. Finally, the scoring intelligent agent 2 will judge the consistency of the two debaters' views. After the score is equal to 10 points, the voting intelligent agent 2 will select the more reasonable and more consistent experimental steps as the globally optimized experimental steps.

[0087] like Figure 4 As shown, the local optimization and scoring module in the step generation unit of this embodiment receives the experimental steps optimized by the experimental step generation module and the chemical questions raised by human researchers, in order to gradually optimize the experimental steps locally.

[0088] After receiving the experimental steps, the supplementary explanation agent based on the large model extracts the supplementary explanation information of each step from the experimental steps and chemical problems, and forms the input of the agent debate with the corresponding steps for use by the debaters.

[0089] During the formal debate, two standard processes were implemented. First, the radical agent initiated the optimization process, with the radical agent and then the conservative agent optimizing the sub-steps separately. The conservative agent also incorporated the optimization results of the radical agent. After a round of debate, each agent produced optimized sub-steps, which, together with the original sub-steps, constituted three outputs and were then submitted to the moderator agent. The moderator agent scored the sub-steps based on factors such as effectiveness and potential defects, and selected the sub-step with the lowest score as the new input sub-step for the next round. If the score exceeded 1, it meant that the local optimization was not yet complete. At this point, a new sub-step would be used as input, and the conservative agent would initiate the next round of optimization, repeating the above steps. If the score was less than or equal to 1, it meant that the local optimization of the sub-step was complete, and the local optimization of the next sub-step would begin, and the process would continue until all sub-steps were locally optimized.

[0090] Example 2

[0091] This embodiment provides a multi-agent intelligent experiment design system based on a multi-model, wherein the scheme design unit is an experimental design scheme generation and optimization framework. Figure 2As shown in the example, a human researcher first poses a chemistry experiment design question, such as "How to synthesize a high-entropy alloy catalyst?" Debater Agents 1 and 2 in the design unit then make preliminary responses in Round 0 based on their respective knowledge and domain experience, generating their own experimental design plans. At this point, Scoring Agent 1 assigns a baseline score (default 0) to each debater's preliminary design plan, which serves as a starting point for subsequent optimization.

[0092] Afterward, Debater Agents 3 and 4 each received the initial responses from Debater Agents 1 and 2, officially launching the first round of debate. In each round, each debater received the other's design proposal from the previous round. Each debater analyzed the other's proposals and, combining their own knowledge, globally optimized the original design. After each round of optimization, each debater generated a proposed experimental design. Scoring Agent 1 then scored the consistency of the two debater's output. If the score was less than 10, the next round of optimization would be iterated. When Scoring Agent 1 scored 10, indicating a high degree of agreement between the two debater's proposals, Voting Agent 1 would select the more reasonable and comprehensive design as the final experimental design.

[0093] Through this multi-round debate and optimization process, this embodiment can effectively integrate opinions from multiple parties and generate a more efficient and feasible experimental design plan through continuous revision and feedback.

[0094] The experimental step generation module in the step generation unit of this embodiment. Figure 3 As shown in the figure, this module first receives the final experimental design output by the design unit and, combined with the list of chemical equipment workstations, conducts a global debate on the generation of experimental steps. The debate process is consistent with the experimental design generation and optimization framework. Two debater agents, Debater 5 and Debater 6, provide preliminary responses to the design to generate preliminary experimental steps. Subsequently, Debater 7 and Debater 8 each receive the memory of the preliminary responding agent, officially launching the first round of debate. In each round, Scoring Agent 2 evaluates the consistency of the two debater agents' opinions. When the score is equal to 10, Voting Agent 2 selects the most reasonable and feasible experimental step as the final experimental step output.

[0095] In this way, this embodiment can ensure the integrity and feasibility of the experimental steps and a high degree of consistency with the experimental design plan.

[0096] The supplementary description generation module and the local optimization and scoring module in the step generation unit of this embodiment. Figure 4As shown in the figure, the Supplementary Notes Generation Module first receives the final experimental steps generated by the Experimental Steps Generation Module. Combined with the chemistry questions posed by the human researcher, the Supplementary Notes Agent generates supplementary notes for each sub-step, covering operational procedures, experimental risks, precautions, and possible sources of error. These supplementary notes, along with the original experimental steps, are used as input for subsequent local optimization by the Local Optimization and Scoring Module.

[0097] Starting from the initial step, the radical agent initiates a debate, engaging in in-depth discussions with the conservative agent on substeps with the goal of achieving precise local optimization of the substeps. During this process, the two debaters, the radical agent and the conservative agent, take turns speaking, with the conservative agent needing to absorb the former's views and use them as a reference for its own optimization. After each round of debate, the substeps optimized by the two debaters are combined with the input substeps to form a list of candidate steps for subsequent voting and scoring. The moderator agent then scores the three optimized substeps based on their plausibility and feasibility, selecting the most plausible one. If a substep scores more than 1, further optimization is required. The conservative agent initiates a new round of debate, using the selected substep as the input for the next round of iteration. If the score is less than or equal to 1, optimization of the substep is complete, and the radical agent initiates a debate on the next substep. This continues until all substeps are optimized, ultimately forming a complete experimental procedure. Through this alternating debate method, we ensure that each round of debate can optimize the sub-steps from different perspectives in order to achieve the best experimental steps.

[0098] Through this iterative optimization process, the accuracy and feasibility of the experimental steps can be gradually improved, and ultimately a complete, feasible and efficient experimental step plan can be generated.

[0099] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a program. The program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0100] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or any form of implication that the information constitutes prior art already known to those skilled in the art.

Claims

1. A multi-agent driven intelligent experiment design system, characterized by: include: In the design unit, multiple debate agents driven by different large models engage in multiple rounds of debate on experimental design questions posed by human researchers, resulting in the experimental design solutions proposed by each debate agent. Scoring agents then evaluate the consistency of each answer in each round. Once the scores reach a predetermined value, voting agents then evaluate the multiple answers based on three dimensions: correctness, completeness, and rationality. The debate agent with the highest average score is selected as the optimal experimental design solution. The step generation unit communicates with the scheme design unit and can take the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme as input. Multiple debater agents driven by different large models combine each other's answers with the provided list of chemical equipment workstations to conduct multiple rounds of debate on the optimal experimental design scheme, and obtain the answers of the global optimization experimental steps that match the optimal experimental design scheme given by each debater agent. After the scoring agent 2 scores the consistency of the answers in each round and the scores reach a predetermined value, the voting agent 2 scores multiple answers from the two dimensions of correctness and rationality, and selects the answer of the debater agent with a higher average score as the global optimization experimental step. Supplementary explanations and local optimization are performed on each experimental step of the global optimization experimental step.

2. The multi-agent driven intelligent experiment design system according to claim 1, characterized in that: The scheme design unit includes: Initial response module, debate and global optimization module and consensus scoring and voting module; among them, The preliminary response module consists of Debater Agent 1 and Debater Agent 2, each driven by a different large model and equipped with only the context memory of the module. These two Debater Agents can provide preliminary responses to the experimental design questions input by the human researcher and derive preliminary experimental design solutions from each Debater Agent. The debate and global optimization module is composed of debater agent No. 3 and debater agent No. 4, which are driven by different large models and only have the context memory of the module. Both are in communication connection with the preliminary response module. The two debater agents can conduct multiple rounds of debate and optimization on the preliminary experimental design scheme given by the preliminary response module. In each round of debate, each debater agent analyzes, discusses and gives optimization suggestions to improve the preliminary experimental design scheme from a global perspective. During each round of optimization, each debater agent refers to existing feedback information and combines it with newly emerging relevant data to make adjustments in order to obtain a more optimal experimental design scheme. The consensus scoring and voting module is composed of a scoring agent and a voting agent, and is communicated with the debate and global optimization module. The scoring agent can score each round of debate and the optimized answers of the two debater agents in the debate and global optimization module from three dimensions: correctness, completeness and rationality, with a scoring range of 0 to 10 points. If the score in a certain round reaches 10 points, it is confirmed that the two debater agents have reached a consensus on the current experimental design plan, and the debate process is stopped. The voting agent takes the average score of the two debater agents No. 3 and No. 4 as the average, and finally selects the experimental design plan of the debater agent with the higher average score as the optimal experimental design plan.

3. The multi-agent driven intelligent experiment design system according to claim 1, characterized in that: The step generation unit includes: Experimental steps generation module, supplementary description generation module and local optimization and scoring module; among them, The experimental step generation module is composed of debater agent No. 5, debater agent No. 6, debater agent No. 7, debater agent No. 8, scoring agent No. 2, and voting agent No. 2, all driven by different large models. It is connected to the scheme design unit in communication. Debater agent No. 5 and debater agent No. 6 can use the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme given by the scheme design unit as input, and make a preliminary response to the optimal experimental design scheme in combination with the other party's answer and the provided list of chemical equipment workstations. Debater agent No. 7 and debater agent No. 8 conduct multiple rounds of debate to obtain the global optimization experimental steps matching the optimal experimental design scheme given by debater agent No. 7 and debater agent No.

8. Scoring agent No. 2 scores the consistency of the answers of debater agent No. 7 and debater agent No. 8 in each round and after the scores reach a predetermined value, voting agent No. 2 scores the two answers from the two dimensions of correctness and rationality, and finally selects the answer of the debater agent with a higher average score as the global optimization experimental step; The supplementary description generation module uses a supplementary description agent, which is in communication with the experimental step generation module and can extract supplementary description information for each experimental step from the global optimization experimental steps given by the experimental step generation module and the experimental design questions given by human researchers; The local optimization and scoring module is composed of a host agent, a radical agent and a conservative agent driven by different large models. It is communicated with the supplementary description generation module and the experimental step generation module. It can combine the supplementary description information of each experimental sub-step provided by the supplementary description generation module. The host agent, the radical agent and the conservative agent use a multi-agent optimization method based on Markov chain to perform local optimization on each experimental step after global optimization one by one until all experimental steps are optimized.

4. The multi-agent driven intelligent experiment design system according to claim 3, characterized in that: In the local optimization and scoring module, the host agent, radical agent, and conservative agent use a multi-agent optimization method based on Markov chains to perform local optimization on each experimental step after global optimization. The following method is used: During the local optimization process, there are two states: the initial state and the normal state. The initial state is the global optimization experimental step provided by the experimental step generation module. The normal state includes two different discussion modes that are carried out alternately, initiated by radical agents or conservative agents respectively. The former expands and optimizes the current experimental step in the order of radical agent-conservative agent-host agent; the latter enhances the stability and feasibility of the experimental step in the order of conservative agent-radical agent-host agent. During each round of optimization, the host agent collects the experimental steps of the initial state and the two-party answers obtained according to one of the discussion modes to form three optimization plans, and scores and votes based on the rationality of the optimization plans. The plan with the highest score is selected as the new initial state to enter the next round of optimization, forming a state transfer based on the Markov chain. The optimization process continues until the host agent's score for the current experimental step reaches a predetermined standard of less than or equal to 1 point, that is, the experimental step is considered to have completed local optimization. Subsequently, the optimization process enters the next experimental step until the entire experimental step is optimized.

5. A multi-agent driven intelligent experiment design method, characterized in that: The system according to any one of claims 1 to 4, comprising: Design Steps: The system's design unit uses multiple debate agents driven by different large models to conduct multiple rounds of debate on the experimental design questions posed by human researchers, obtaining the responses of each debate agent to the experimental design plan. Scoring agents then evaluate the consistency of each answer in each round. Once the scores reach a predetermined value, voting agents then evaluate each answer based on correctness, completeness, and rationality. The debate agent with the highest average score is selected as the optimal experimental design plan. Experimental step generation step: The system's step generation unit takes the list of standard equipment workstations in a chemical laboratory and the optimal experimental design plan as input. Multiple debater agents driven by different large models conduct multiple rounds of debate on the optimal experimental design plan based on each other's answers and the provided list of chemical equipment workstations, and obtain the answers of each debater agent to the globally optimized experimental steps that match the optimal experimental design plan. After the scoring agent 2 scores the consistency of the answers in each round and the score reaches the predetermined value, the voting agent 2 scores each answer from the two dimensions of correctness and rationality, and selects the answer of the debater agent with the higher average score as the globally optimized experimental step. Supplementary explanations and local optimization are performed on each globally optimized experimental step.

6. The multi-agent driven intelligent experiment design method according to claim 5, characterized in that: The scheme design steps include: Through the preliminary response module of the scheme design unit, Debater Agent 1 and Debater Agent 2, which are driven by different large models that constitute the module and only have the context memory of the module, make preliminary responses to the experimental design questions input by the human researcher and obtain preliminary experimental design schemes given by each debater agent; Through the debate and global optimization module of the scheme design unit, Debater Agents 3 and 4, which are driven by different large models that make up the module and only have context memory of the module, conduct multiple rounds of debate and optimization on the preliminary experimental design scheme given by the preliminary response module. In each round of debate, each debater agent analyzes, discusses and gives optimization suggestions to improve the preliminary experimental design scheme from a global perspective. During each round of optimization, each debater agent refers to existing feedback information and combines it with newly emerging relevant data to make adjustments in order to obtain a more optimal experimental design scheme. Through the consensus scoring of the scheme design unit and the scoring agent 1 of the voting module, each round of debate and the optimized answers of the two debater agents in the debate and global optimization module are scored from three dimensions: correctness, completeness and rationality, with a scoring range of 0 to 10 points. If the score in a certain round reaches 10 points, it is confirmed that the two debater agents have reached a consensus on the current experimental design scheme, and the debate process is stopped. The consensus scoring and voting agent 1 of the voting module will use the average score of the two debater agents No. 3 and No. 4 to finally select the experimental design scheme of the debater agent with the higher average score as the optimal experimental design scheme.

7. The multi-agent driven intelligent experiment design method according to claim 5, characterized in that: The experimental step generation step includes: Debater agent No. 5 and debater agent No. 6, which are driven by different large models of the experimental step generation module of the step generation unit, take the list of standard equipment workstations in the chemical laboratory and the optimal experimental design scheme given by the scheme design unit as input, and make a preliminary response to the optimal experimental design scheme in combination with the other party's answer and the provided list of chemical equipment workstations. Debater agent No. 7 and debater agent No. 8 conduct multiple rounds of debates to obtain the global optimized experimental steps matching the optimal experimental design scheme given by debater agent No. 7 and debater agent No.

8. Scoring agent No. 2 scores the consistency of the answers of debater agent No. 7 and debater agent No. 8 in each round and after the scores reach a predetermined value, voting agent No. 2 scores the two answers from the two dimensions of correctness and rationality, and finally selects the answer of the debater agent with a higher average score as the global optimized experimental step; Extracting supplementary explanation information for each experimental step from the global optimization experimental steps given by the experimental step generation module and the experimental design problem given by the human researcher through the supplementary explanation intelligent agent of the supplementary explanation generation module of the step generation unit; Through the local optimization of the step generation unit combined with the supplementary description information of each experimental sub-step given by the supplementary description generation module, the host agent, radical agent and conservative agent use a multi-agent optimization method based on Markov chain to locally optimize each experimental step after global optimization one by one until all experimental steps are optimized.

8. The multi-agent driven intelligent experiment design method according to claim 7, characterized in that: In the method, the host agent, radical agent, and conservative agent use a Markov chain-based multi-agent optimization method to perform local optimization on each experimental step after global optimization as follows: During the local optimization process, there are two states: the initial state and the normal state. The initial state is the global optimization experimental step provided by the experimental step generation module. The normal state includes two different discussion modes that are carried out alternately, initiated by radical agents or conservative agents respectively. The former expands and optimizes the current experimental step in the order of radical agent-conservative agent-host agent; the latter enhances the stability and feasibility of the experimental step in the order of conservative agent-radical agent-host agent. During each round of optimization, the host agent collects the experimental steps of the initial state and the two-party answers obtained according to one of the discussion modes to form three optimization plans, and scores and votes based on the rationality of the optimization plans. The plan with the highest score is selected as the new initial state to enter the next round of optimization, forming a state transfer based on the Markov chain. The optimization process continues until the host agent's score for the current experimental step reaches a predetermined standard of less than or equal to 1 point, that is, the experimental step is considered to have completed local optimization. Subsequently, the optimization process enters the next experimental step until the entire experimental step is optimized.

9. A processing device, characterized in that include: at least one memory for storing one or more programs; At least one processor is capable of executing one or more programs stored in the memory, and when the one or more programs are executed by the processor, the processor is able to implement the method according to any one of claims 5 to 8.

10. A readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the method according to any one of claims 5 to 8 can be implemented.

Images