Experiment scheme generation method and system based on multi-agent cooperation and state-dependent graph

Abstract

The application discloses a kind of experimental scheme generation method and system based on multi-agent cooperation and state dependence graph, it is related to experimental scheme intelligent generation technical field, the method is executed by multiple agents cooperation, the experimental type, experimental object, detection end point and experimental constraint condition are obtained by the semantic analysis to experiment target, according to the experimental type and the experimental object from structured knowledge base retrieve to obtain structured experimental knowledge, and construct with step as node, with execution dependency as edge directed acyclic graph as experimental state dependence graph, by constraint path search on the experimental state dependence graph generates experimental step sequence satisfying execution dependency and experimental constraint condition, by literature parameter domain derivation to adjustable parameter in experimental step sequence obtains the parameter value interval of each adjustable parameter, and based on detection end point executes multilayer rule verification, finally output scientific rationality, logic rigor and executable experimental scheme.

Description

Technical Field

[0001] This application relates to the field of intelligent generation technology of experimental schemes, and in particular to a method and system for generating experimental schemes based on multi-agent cooperation and state dependency graphs. Background Technology

[0002] Currently, the automatic generation of wet laboratory protocols in drug development is mainly achieved through technologies such as large language models or knowledge graphs. Large language model-based methods treat experimental protocol design as a text generation task, with the large language model directly generating natural language descriptions of the experimental steps based on the experimental objectives. Some methods also incorporate retrieval-enhanced generation strategies or multi-model debate mechanisms to correct the generated results. Knowledge graph-based methods, on the other hand, extract experimental entities such as reagents, equipment, and operating conditions, along with their interrelationships, from scientific literature to construct structured knowledge graphs that assist large language models in generating experimental protocols.

[0003] However, the existing solutions mentioned above all have the following shortcomings: Methods based on large language models cannot explicitly model the execution dependencies between experimental steps, leading to potential sequence errors in the generated solutions. While knowledge graph-based methods achieve structured organization of knowledge, they essentially provide a semantic description of static experimental elements and do not model the execution dependencies between steps; therefore, they also cannot guarantee the correctness of the step order. Furthermore, the aforementioned solutions lack effective automated verification mechanisms and cannot automatically check the completeness of the experimental steps and the rationality of the parameters.

[0004] Therefore, there is an urgent need for an experimental scheme generation method that can explicitly model the dependencies of experimental steps and support automated verification. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method and system for generating experimental schemes based on multi-agent cooperation and state dependency graphs, including the following: Firstly, this application provides a method for generating experimental schemes based on multi-agent cooperation and state dependency graphs, which is executed collaboratively by multiple agents. The method includes: Obtain the experimental target, perform semantic parsing on the experimental target, and identify the experimental type, experimental object, detection endpoint, and experimental constraints contained in the experimental target; Based on the experimental type and the experimental object, structured experimental knowledge matching the experimental objective is retrieved from a pre-built structured knowledge base; Based on the structured experimental knowledge and the experimental type, an experimental state dependency graph is constructed; the experimental state dependency graph is a directed acyclic graph, the nodes of the experimental state dependency graph represent experimental steps and the attribute information corresponding to the experimental steps, and the edges of the experimental state dependency graph represent the execution dependency relationship between experimental steps. Based on the experimental constraints, a constraint path search is performed on the experimental state dependency graph to generate a sequence of experimental steps that satisfies the execution dependency and the experimental constraints. Based on the structured knowledge base and the experimental object, the parameter domain of the adjustable parameters in the experimental step sequence is derived to obtain the parameter value range of each adjustable parameter. Based on the detection endpoint, multi-level rule verification is performed on the experimental step sequence and the parameter value range corresponding to each experimental step to determine whether the preset verification conditions are met. When the multi-level rule verification is successful, the experimental step sequence and the parameter values ​​corresponding to each experimental step are output as the experimental scheme.

[0006] Optionally, the method further includes: When the multi-level rule verification fails, or when a user requests a change to the experimental conditions is received, the downstream influence subgraph associated with the changed step node in the experimental step sequence is located according to the experimental state dependency graph. The nodes in the experimental state dependency graph, excluding the downstream influence subgraph, are used as fixed boundaries. Under the constraints of the fixed boundary, the downstream influence subgraph is re-searched for constraint paths and deduced for parameter domains to obtain the updated sub-process. The updated sub-process is then concatenated with the fixed boundary to obtain the updated experimental scheme.

[0007] Optionally, the execution dependencies include hard dependencies, suggested dependencies, and resource mutual exclusion relationships; the hard dependencies represent the order constraints that must be followed between two experimental steps; the suggested dependencies represent the recommended order constraints between two experimental steps; and the resource mutual exclusion relationships represent the mutual exclusion constraints that arise between two experimental steps due to the sharing of experimental resources.

[0008] Optionally, the execution of multi-level rule verification includes: The execution steps rely on compliance verification to verify whether the order of each experimental step in the experimental step sequence satisfies the execution dependency relationship in the experimental state dependency graph. Perform parameter domain compliance verification to verify whether the parameter values ​​corresponding to each experimental step in the experimental step sequence are within the parameter value range. Perform condition integrity verification to verify whether the experimental step sequence contains a preset mandatory step type; Perform a target consistency verification to verify whether the initial target and the final output of the experimental step sequence are consistent.

[0009] Optionally, the step of deriving the parameter domain of the adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object to obtain the parameter value range of each adjustable parameter includes: Retrieve structured experimental knowledge that matches the experimental object from the structured knowledge base; Extract the value distribution of the corresponding adjustable parameters from the structured experimental knowledge; The value distribution is statistically aggregated to obtain the literature consensus interval for each adjustable parameter, which is used as the parameter value interval.

[0010] Optionally, the constraint path search on the experimental state dependency graph includes: Starting from the initial node of the experimental state dependency graph, candidate paths are iteratively expanded according to topological order; In each expansion step, the step nodes that have been satisfied by the predecessor nodes are included in the candidate set; Logical consistency and conditional integrity checks are performed on each candidate path in the candidate set, and candidate paths that fail the checks are eliminated. Select at least one path from the verified candidate paths as the sequence of experimental steps.

[0011] Optionally, the method further includes: The task parsing agent performs semantic parsing on the experimental target and writes the parsed recognition result into the shared state pool. The knowledge retrieval agent retrieves the structured experimental knowledge from the structured knowledge base based on the experimental type and the experimental object in the shared state pool, and writes the retrieved structured experimental knowledge into the shared state pool. The process planning agent constructs the experiment state dependency graph based on the structured experiment knowledge and the experiment type in the shared state pool, performs the constraint path search, generates the experiment step sequence, and writes the experiment step sequence into the shared state pool. The parameter-generating agent derives the parameter domain of the adjustable parameters in the experimental step sequence based on the structured experimental knowledge in the shared state pool, and writes the parameter value range into the shared state pool. The rule verification agent performs the multi-level rule verification based on the experimental step sequence and the parameter value range in the shared state pool, and writes the verification result into the shared state pool. The local replanning agent performs the local replanning and writes the replanning result into the shared state pool.

[0012] Optionally, the task parsing agent, the knowledge retrieval agent, the process planning agent, the parameter generation agent, the rule verification agent, and the local replanning agent work collaboratively in a serial or conditionally parallel manner. The agents transmit structured information to each other by reading and writing the shared state pool. The shared state pool is a data storage space shared by all agents, used to store the recognition results, the structured experimental knowledge, the experimental state dependency graph, the experimental step sequence, the parameter value range, the verification results, and the replanning results.

[0013] Optionally, the structured knowledge base is pre-built in the following manner: Offline structured parsing is performed on experimental literature, standard experimental protocols, and historical experimental records. Entities, relations, and attributes are extracted from these materials using named entity recognition and relation extraction models to obtain structured experimental knowledge. The structured experimental knowledge is assembled according to a preset structured format and stored in the structured knowledge base; The structured experimental knowledge in the structured knowledge base is vectorized and stored in a vector database.

[0014] Secondly, this application provides an experimental scheme generation system based on multi-agent cooperation and state dependency graphs, the system comprising: The task parsing module is used to obtain the experimental target, perform semantic parsing on the experimental target, and identify the experimental type, experimental object, detection endpoint and experimental constraints contained in the experimental target; The knowledge retrieval module is used to retrieve structured experimental knowledge that matches the experimental objective from a pre-built structured knowledge base based on the experimental type and the experimental object. The process planning module is used to construct an experimental state dependency graph based on the structured experimental knowledge and the experimental type, and to perform constraint path search on the experimental state dependency graph according to the experimental constraints to generate a sequence of experimental steps that satisfy the execution dependencies and the experimental constraints. The experimental state dependency graph is a directed acyclic graph, where the nodes represent experimental steps and their corresponding attribute information, and the edges represent the execution dependencies between experimental steps. The parameter generation module is used to derive the parameter domain of the adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object, so as to obtain the parameter value range of each adjustable parameter. The rule verification module is used to perform multi-level rule verification on the experimental step sequence and the parameter value range corresponding to each experimental step according to the detection endpoint, and to determine whether the preset verification conditions are met. The output module is used to output the experimental step sequence and the parameter values ​​corresponding to each experimental step as an experimental scheme when the multi-level rule verification is successful.

[0015] Optionally, the system further includes: The local replanning module is used to locate the downstream influence subgraph associated with the changed step node in the experimental step sequence based on the experimental state dependency graph when the multi-layer rule verification fails or when a user's request to change the experimental conditions is received; the nodes in the experimental state dependency graph other than the downstream influence subgraph are used as fixed boundaries; under the constraints of the fixed boundaries, the downstream influence subgraph is re-performed with constraint path search and parameter domain derivation to obtain the updated sub-process; the updated sub-process is concatenated with the fixed boundaries to obtain the updated experimental scheme.

[0016] Optionally, the execution dependencies include hard dependencies, suggested dependencies, and resource mutual exclusion relationships; the hard dependencies represent the order constraints that must be followed between two experimental steps; the suggested dependencies represent the recommended order constraints between two experimental steps; and the resource mutual exclusion relationships represent the mutual exclusion constraints that arise between two experimental steps due to the sharing of experimental resources.

[0017] Optionally, the rule verification module performs the multi-level rule verification, including: The execution steps rely on compliance verification to verify whether the order of each experimental step in the experimental step sequence satisfies the execution dependency relationship in the experimental state dependency graph. Perform parameter domain compliance verification to verify whether the parameter values ​​corresponding to each experimental step in the experimental step sequence are within the parameter value range. Perform condition integrity verification to verify whether the experimental step sequence contains a preset mandatory step type; Perform a target consistency verification to verify whether the initial target and the final output of the experimental step sequence are consistent.

[0018] Optionally, the parameter generation module, based on the structured knowledge base and the experimental object, performs parameter domain derivation on the adjustable parameters in the experimental step sequence to obtain the parameter value range of each adjustable parameter, including: Retrieve structured experimental knowledge that matches the experimental object from the structured knowledge base; Extract the value distribution of the corresponding adjustable parameters from the structured experimental knowledge; The value distribution is statistically aggregated to obtain the literature consensus interval for each adjustable parameter, which is used as the parameter value interval.

[0019] Optionally, the process planning module performs constrained path search on the experimental state dependency graph, including: Starting from the initial node of the experimental state dependency graph, candidate paths are iteratively expanded according to topological order; In each expansion step, the step nodes that have been satisfied by the predecessor nodes are included in the candidate set; Logical consistency and conditional integrity checks are performed on each candidate path in the candidate set, and candidate paths that fail the checks are eliminated. Select at least one path from the verified candidate paths as the sequence of experimental steps.

[0020] Optionally, a task parsing module is used for the task parsing agent to perform semantic parsing on the experimental target and write the parsed recognition result into a shared state pool. The knowledge retrieval module is used by the knowledge retrieval agent to retrieve the structured experimental knowledge from the structured knowledge base according to the experimental type and the experimental object in the shared state pool, and to write the retrieved structured experimental knowledge into the shared state pool. The process planning module is used by the process planning agent to construct the experiment state dependency graph based on the structured experiment knowledge and the experiment type in the shared state pool, perform the constraint path search, generate the experiment step sequence, and write the experiment step sequence into the shared state pool. The parameter generation module is used by the parameter generation agent to derive the parameter domain of the adjustable parameters in the experimental step sequence based on the structured experimental knowledge in the shared state pool, and write the parameter value range into the shared state pool. The rule verification module is used by the rule verification agent to perform the multi-level rule verification according to the experimental step sequence and the parameter value range in the shared state pool, and write the verification result into the shared state pool. The local replanning module is used by the local replanning agent to perform the local replanning and write the replanning result into the shared state pool.

[0021] Optionally, the task parsing module, the knowledge retrieval module, the process planning module, the parameter generation module, the rule verification module, and the local replanning module work collaboratively in a serial or conditional parallel manner. The modules exchange structured information by reading and writing to the shared state pool. The shared state pool is a data storage space shared by all modules, used to store the identification results, the structured experimental knowledge, the experimental state dependency graph, the experimental step sequence, the parameter value range, the verification results, and the replanning results.

[0022] Optionally, the structured knowledge base is pre-built in the following manner: Offline structured parsing is performed on experimental literature, standard experimental protocols, and historical experimental records. Entities, relations, and attributes are extracted from these materials using named entity recognition and relation extraction models to obtain structured experimental knowledge. The structured experimental knowledge is assembled according to a preset structured format and stored in the structured knowledge base; The structured experimental knowledge in the structured knowledge base is vectorized and stored in a vector database.

[0023] Thirdly, this application provides an apparatus comprising a memory and a processor, the memory for storing instructions or code, and the processor for executing the instructions or code to cause the apparatus to perform the experimental scheme generation method based on multi-agent cooperation and state dependency graph described in any of the implementations of the first aspect.

[0024] Fourthly, this application provides a computer-readable storage medium storing code, wherein when the code is executed, a device running the code implements the experimental scheme generation method based on multi-agent cooperation and state dependency graph described in any of the implementations of the first aspect.

[0025] This application provides a method for generating experimental schemes based on multi-agent cooperation and state dependency graphs. When executing the method, firstly, the experimental target is acquired and semantically parsed to identify the experimental type, experimental object, detection endpoint, and experimental constraints contained in the experimental target. Then, based on the experimental type and the experimental object, structured experimental knowledge matching the experimental target is retrieved from a pre-built structured knowledge base. Next, based on the structured experimental knowledge and the experimental type, a directed acyclic graph (DAG) is constructed as the experimental state dependency graph, with steps as core nodes, attribute information corresponding to the experimental steps as surrounding nodes, and execution dependencies as edges. Constraint path search is performed on the experimental state dependency graph according to the experimental constraints to generate an experimental step sequence that satisfies the execution dependencies and experimental constraints. Then, based on the structured knowledge base and the experimental object, parameter domain derivation is performed on the adjustable parameters in the experimental step sequence to obtain the parameter value range of each adjustable parameter. Finally, based on the detection endpoint, multi-level rule verification is performed on the experimental step sequence and the parameter value range corresponding to each experimental step. After successful verification, the experimental step sequence and the parameter values ​​corresponding to each experimental step are output as the experimental scheme.

[0026] In this way, by having multiple agents collaboratively execute the above steps, each agent works together to complete the entire process from target parsing to solution output, making the experimental solution generation process modular and controllable. Simultaneously, by constructing an explicit directed acyclic graph to model the execution dependencies between experimental steps, and performing constraint path search on the directed acyclic graph to generate a sequence of experimental steps that satisfy dependency constraints and conditions, the generation of experimental solutions transforms from text synthesis into a structured workflow planning problem, fundamentally avoiding the problem of incorrect experimental step order. By deriving parameter value ranges from the adjustable parameters in the experimental step sequence and performing multi-layer rule verification based on the detection endpoint, the generated experimental solutions meet execution requirements in terms of parameter rationality and logical integrity, avoiding the problems of missing key steps and parameter out-of-control issues. Thus, through a structured generation method combining multi-agent collaboration and state dependency graphs, experimental solutions with scientific rationality, logical rigor, and executability can be generated. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart illustrating an experimental scheme generation method based on multi-agent cooperation and state dependency graphs provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of an experimental scheme generation system based on multi-agent cooperation and state dependency graph provided in an embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0030] It should be noted that the experimental scheme generation method provided in this application is executed collaboratively by multiple intelligent agents. Each intelligent agent refers to a software entity with independent functional modules. Data is transferred between agents through a shared state pool, and no scheme-level debate or voting occurs between them. The method described in this application can be applied to wet experimental process design in the field of drug development, such as biochemical experiments with strict operational sequence requirements, such as cytotoxicity experiments, Western blotting experiments, and PCR experiments.

[0031] Figure 1 This is a flowchart illustrating an experimental scheme generation method based on multi-agent cooperation and state dependency graphs, provided in an embodiment of this application. Combined with... Figure 1 As shown, the experimental scheme generation method based on multi-agent cooperation and state dependency graph provided in this application embodiment may include: S101. Obtain the experimental target, perform semantic parsing on the experimental target, and identify the experimental type, experimental object, detection endpoint, and experimental constraints contained in the experimental target.

[0032] The user inputs the experimental objective, such as "to detect the half-maximal inhibitory concentration (IC50) of a certain compound on a certain cell type." Semantic parsing refers to the process of extracting key information from the user input using natural language processing technology. The experiment type refers to the category of the experiment, such as cytotoxicity assay, Western blotting assay, or PCR assay. The experimental object refers to the biological or chemical sample targeted in the experiment, such as a specific cell line or protein. The endpoint refers to the final metric to be measured, such as cell viability, protein expression level, or gene amplification curve. Experimental constraints refer to the explicitly stated limiting requirements by the user, such as temperature range, incubation time, or reagent specifications. Through semantic parsing, the system transforms ambiguous user requirements into structured parameters, providing clear input for subsequent steps. This step enables the system to accurately understand the user's experimental needs, providing clear input information for subsequent steps and avoiding errors in the experimental design due to misunderstandings of the requirements.

[0033] S102. Based on the experiment type and the experiment object, retrieve the structured experimental knowledge that matches the experiment objective from the pre-built structured knowledge base.

[0034] The structured knowledge base is pre-built using the following methods: Offline structured parsing is performed on experimental literature, standard experimental protocols, and historical experimental records. Named entity recognition (NAME) and relation extraction (REM) models are used to extract entities, relationships, and attributes from these sources, resulting in structured experimental knowledge. Specifically, the NAME model identifies key experimental elements in the literature, such as "steps," "actions," "materials," "parameters," and "conditions," while the REM model identifies logical relationships between these elements, such as "dependencies," "outputs," and "consumptions." To avoid error accumulation in end-to-end joint extraction, a pipelined extraction strategy is employed, sequentially performing entity recognition, relation classification, and assembly according to a preset structured format. Each stage has a confidence threshold; extraction results below the threshold are automatically marked as "pending confirmation," manually verified, and then added back to the database, forming a structured knowledge base with quality labels.

[0035] The structured experimental knowledge is assembled according to a preset structured format and stored in the structured knowledge base. The structured experimental knowledge is represented as: Protocol=(N,E,T_N,T_E,A); where N is the set of experimental entity nodes, and the node type T_N includes steps, actions, materials, parameters, and conditions; E is the set of relationship edges between entities, and the edge type T_E includes execution dependencies, outputs, and consumptions; A is an attribute mapping used to assign experiment-specific attribute values ​​to each node and edge, such as step number, action type, material specifications, parameter values, conditional predicates, etc.

[0036] The structured experimental knowledge in the structured knowledge base is vectorized and encoded, and then stored in a vector database. Specifically, the extracted knowledge units are vectorized and encoded using an embedding model, and then stored in a Milvus vector database.

[0037] During retrieval, the system converts the experimental objective into a query vector, performs an approximate nearest neighbor index retrieval in the vector database, recalls a set of candidate structured experimental knowledge, and then uses a Cross-Encoder to reorder the relevance. Based on the reordering result, the system returns the structured experimental knowledge. Through this step, the system can obtain structured experimental knowledge directly related to the current experimental objective from the knowledge base, providing a reliable factual basis for the subsequent construction of the dependency graph.

[0038] S103. Based on the structured experimental knowledge and the experimental type, construct an experimental state dependency graph.

[0039] The experimental state dependency graph is a directed acyclic graph. The nodes of the experimental state dependency graph represent experimental steps and the attribute information corresponding to the experimental steps, and the edges of the experimental state dependency graph represent the execution dependency relationships between experimental steps.

[0040] A directed acyclic graph (DAG) is a graph structure where there are no directed cycles. It is suitable for representing task sequences with partial order relationships and can guarantee that there are no cyclic dependencies between experimental steps. Nodes are extracted from the structured experimental knowledge returned by S102: entity nodes of type "step" are used as core step nodes, and their associated material and parameter nodes via consumption edges, as well as their associated output resource nodes via output edges, are aggregated into the internal attributes of that step node; entity nodes of type "condition" are mapped to step execution condition attributes. Each node inherits the experiment-specific attribute values ​​from the attribute graph model and adds a completion status flag as a runtime dynamic attribute. The edge set directly inherits the extracted execution dependency edges from the attribute graph model and maps them to three types of execution dependencies according to the constraint strength: for explicitly marked sequential relationships in the structured experimental knowledge, they are mapped to hard dependency edges; for non-mandatory but recommended sequential relationships, they are mapped to suggested dependency edges; for competition or mutual exclusion relationships arising from shared output or consumption resources, they are mapped to resource mutual exclusion edges.

[0041] After the skeleton is generated, the system automatically performs two checks: first, loop detection, which immediately interrupts and alerts if a logical contradiction is found, triggering a review and correction of the source data; second, connectivity checks, ensuring that all target output nodes are reachable from the starting node, avoiding isolated steps. Once the checks pass, the dependency graph is frozen to the current version. The system implements version management for the dependency graph, generating a new version snapshot after each update of the structured knowledge base. The current graph version is locked during experimental process planning to prevent execution errors caused by changes in the graph structure. By constructing an experimental state dependency graph, the system represents the experimental process in an explicit graph structure, making the dependencies between steps clear and providing a structured planning space for subsequent constraint path searches.

[0042] S104. Based on the experimental constraints, perform constraint path search on the experimental state dependency graph to generate an experimental step sequence that satisfies the execution dependency relationship and the experimental constraints.

[0043] The system starts from the initial node and iteratively expands candidate paths based on the topology of the current dependency graph. Each step only includes steps whose predecessor nodes have already satisfied the requirements. Topological order refers to a linear ordering of nodes in a directed acyclic graph, such that for every directed edge u→v, node u is always preceding node v. Iterative expansion according to topological order ensures that each step is considered only after all its predecessor steps have been completed. For each candidate path, the system performs multi-dimensional hard constraint checks. These hard constraint checks are constraints that cannot be compensated for by subsequent adjustments and must be strictly satisfied during the path search phase. They include logical consistency checks and conditional integrity checks. Logical consistency checks whether the preconditions of each step in the path are satisfied by the output states of the preceding steps; conditional integrity checks whether the path covers all key output types required by the experimental objective. If a candidate path fails any hard constraint check, it is directly eliminated; only when a path passes all hard constraint checks is it retained as a feasible candidate path for further processing. The system retains multiple feasible paths and prunes invalid branches through a width-constrained search, ultimately outputting several sequences of experimental steps that satisfy all constraints for the user to choose from. If a critical step is found missing during the search, the system automatically matches condition-action pairs from a pre-defined rule base to complete the process. The recursive completion depth is limited; exceeding this limit stops the process and triggers an alarm. All automatically completed nodes are explicitly marked in the output and can only proceed to the next execution stage after user confirmation, preventing the system from unknowingly altering the workflow.

[0044] The execution dependencies include hard dependencies, suggested dependencies, and resource exclusivity. Hard dependencies represent a mandatory sequential constraint between two experimental steps; suggested dependencies represent a recommended sequential constraint between two experimental steps; and resource exclusivity represents a mutual exclusion constraint between two experimental steps due to shared experimental resources. For example, in cytotoxicity experiments, "cell seeding" must be performed before "cell culture," which is a hard dependency; "cell culture" and "drug administration" are usually hard dependencies because drug administration must be performed after cell adhesion culture is complete.

[0045] The constraint path search on the experimental state dependency graph includes: Starting from the initial node of the experimental state dependency graph, candidate paths are iteratively expanded according to topological order. The initial node refers to a step node without a predecessor node; for example, in a cytotoxicity experiment, "cell seeding" is usually the initial node.

[0046] In each expansion step, step nodes whose predecessor nodes have all been satisfied are included in the candidate set. That is, a step can only be added to the candidate path if all its predecessor steps have been selected and completed.

[0047] For each candidate path in the candidate set, perform logical consistency checks and conditional integrity checks, and eliminate candidate paths that fail the checks. The logical consistency check checks whether the preconditions of each step in the path are satisfied by the output state of the preceding step; the conditional integrity check checks whether the path covers all the key output types required by the experimental objective.

[0048] At least one path is selected from the validated candidate paths as the experimental step sequence. The system can sort the validated candidate paths based on the prior support frequency in the literature search results, or allow the user to choose. Through this step, the system generates the experimental step sequence while satisfying all dependency constraints and user constraints, fundamentally eliminating the possibility of incorrect sequence.

[0049] S105. Based on the structured knowledge base and the experimental object, perform parameter domain derivation on the adjustable parameters in the experimental step sequence to obtain the parameter value range of each adjustable parameter.

[0050] Adjustable parameters refer to parameters in experimental procedures that require user confirmation of specific values, such as incubation time, reagent concentration, centrifugation speed, and temperature. Parameter domain derivation refers to synthesizing the commonly used value ranges for each parameter from existing structured experimental knowledge.

[0051] The step involves deriving the parameter domain for adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object, to obtain the parameter value range for each adjustable parameter, including: The system retrieves structured experimental knowledge matching the experimental object from the structured knowledge base. Specifically, it retrieves structured experimental knowledge matching the current experimental objective, action type, and material system from the structured knowledge base.

[0052] Extract the value distribution of the corresponding adjustable parameters from the structured experimental knowledge.

[0053] Statistical aggregation of the value distribution yields a consensus interval for each adjustable parameter, which serves as the parameter value interval. This interval reflects the commonly used parameter range in publicly available experimental protocols for similar experiments, independent of any historical experimental results, costs, or time consumption records. Users can view the generated consensus interval and independently set specific parameter values ​​within it. Once set, the parameter values ​​are locked and executed along with the experimental step sequence. Through this step, the system transforms parameter recommendations into a documented and clearly defined interval presentation. Users set parameters independently within the range, and the system only performs locking and verification, mitigating the risk of parameter loss of control.

[0054] S106. Based on the detection endpoint, perform multi-level rule verification on the experimental step sequence and the parameter value range corresponding to each experimental step to determine whether the preset verification conditions are met.

[0055] The pre-defined verification conditions refer to the set of judgment rules that the experimental scheme must meet. These verification conditions include step dependency compliance, parameter domain compliance, condition integrity, and goal consistency. Specifically, the step dependency compliance condition requires that the order of each experimental step in the scheme must satisfy the temporal relationship of all hard dependency edges in the experimental state dependency graph; the parameter domain compliance condition requires that the parameter values ​​of each experimental step in the scheme are within the consensus range of the literature; the condition integrity condition requires that the scheme contains all the mandatory step types corresponding to the experimental type; and the goal consistency condition requires that the initial goal and the final output of the scheme are consistent. Multi-level rule verification means that the above four conditions are verified sequentially, and the scheme is considered to have passed verification only when all verification conditions are met.

[0056] The execution of multi-level rule verification includes: The execution steps rely on compliance verification to verify whether the order of the experimental steps in the experimental step sequence satisfies the execution dependency relationship in the experimental state dependency graph. This layer of verification checks whether the sequence of steps strictly satisfies the temporal relationship of all hard dependency edges in the dependency graph, and whether there are any missing or reversed predecessor steps.

[0057] Perform parameter domain compliance verification to verify whether the parameter values ​​corresponding to each experimental step in the experimental step sequence are within the parameter value range. This layer of verification checks whether the parameter values ​​of each step in the scheme exceed the abnormal values ​​of the parameter value range.

[0058] The system performs a condition integrity verification, checking whether the experimental step sequence contains a preset mandatory step type. It matches mandatory step nodes with key output types one by one according to the experimental type template; any missing steps are considered incomplete.

[0059] Perform a consistency verification of the execution objectives, verifying whether the initial objective and the final output of the experimental step sequence are consistent. The system checks whether the initial objective, final output, and selected step chain of the scheme are self-consistent, and whether there is any deviation between the objective and the execution path.

[0060] If any of the four layers of verification fails, the experimental design is deemed not to meet the preset verification conditions. The verification results are output in the form of a structured conflict report, listing the conflict level, involved steps, and the rules applied. Through this step, the system performs multi-layered progressive verification after the design is generated, ensuring that the design meets the execution requirements at the logical, parameter, structural, and objective levels.

[0061] S107. When the multi-layer rule verification is successful, output the experimental step sequence and the parameter values ​​corresponding to each experimental step as the experimental scheme.

[0062] Once all four layers of verification in S106 pass, the experimental step sequence and the corresponding parameter values ​​for each step are output in a structured data format as the final structured experimental workflow scheme. This experimental scheme fully records the entire operation process from the initial steps to the final output, as well as the quantitative parameters of each step, allowing users to execute experiments at the experimental platform according to the experimental step sequence and the corresponding parameter values ​​for each step. Through this step, the system completes the complete transformation from the user-input experimental objective to a structured and executable experimental scheme, and the output experimental scheme combines scientific rationality, logical rigor, and executability.

[0063] In practical applications, user-inputted experimental conditions may need adjustment, or system-generated solutions may need correction due to failed verification. If a full regeneration is performed each time, all confirmed preliminary steps will be lost, compromising the continuity and consistency of the experimental solution. To address this, this application also provides a partial replanning mechanism. When the above situations occur, only the affected parts are replanned to retain the confirmed steps.

[0064] Specifically, in one implementation of this application embodiment, the method further includes: When the multi-level rule verification fails, or when a user requests a change to the experimental conditions is received, the downstream influence subgraph associated with the changed step node in the experimental step sequence is located based on the experimental state dependency graph. A changed node refers to a step node that the user requests to modify or whose attributes have changed. The downstream influence subgraph is the set of all subsequent step nodes reachable from the changed node along the dependency path. The system determines the scope of the downstream influence subgraph by traversing all reachable paths from the changed node in the experimental state dependency graph.

[0065] The nodes in the experimental state dependency graph, excluding the downstream influence subgraph, are designated as fixed boundaries. Fixed boundaries refer to the parts that remain unchanged during local replanning; their step sequences and parameter values ​​are frozen and do not participate in the replanning process.

[0066] Under the constraints of the fixed boundary, the downstream influence subgraph is re-performed with constraint path search and parameter domain derivation to obtain an updated sub-process. Specifically, the system re-invokes the process planning agent to perform constraint path search and the parameter generation agent to perform parameter domain derivation to generate a new sub-process.

[0067] The updated sub-processes are then concatenated with the fixed boundaries to obtain the updated experimental scheme. After the updated sub-processes pass rule verification, they are seamlessly integrated into the frozen portion of the original scheme to form the updated complete scheme. Throughout the process, unaffected confirmed steps are fully preserved, and users only need to focus on and confirm the changes to the local areas. Through this local replanning mechanism, the system can accurately modify local areas based on verification results or user needs, solving the problem of lost preceding steps caused by the need for a full regeneration when modifying any experimental condition in existing technologies, thus reducing the cognitive burden and modification costs for users.

[0068] The above embodiments provide a detailed description of the method steps for generating experimental schemes provided in this application. The method described in this application is executed collaboratively by multiple intelligent agents; that is, each of the aforementioned steps is executed by an intelligent agent undertaking a different sub-task. The agents transfer data through a shared state pool to cooperate in completing the entire experimental scheme generation process. The following provides a detailed description of the multi-agent collaborative scheduling method in this application's embodiments.

[0069] Specifically, in one implementation of this application embodiment, the method further includes: The task parsing agent performs semantic parsing on the experimental target and writes the parsed recognition results into a shared state pool. The shared state pool is a data storage space shared by all agents, using a key-value pair structure to store the input and output data of each agent, supporting dynamic read and write operations across rounds. After completing the parsing, the task parsing agent stores the identified experimental type, experimental object, detection endpoint, and experimental constraints in a structured form into the shared state pool.

[0070] The knowledge retrieval agent retrieves the structured experimental knowledge from the structured knowledge base based on the experimental type and the experimental object in the shared state pool, and writes the retrieved structured experimental knowledge into the shared state pool.

[0071] The process planning agent constructs the experiment state dependency graph based on the structured experiment knowledge and the experiment type in the shared state pool, performs the constraint path search, generates the experiment step sequence, and writes the experiment step sequence into the shared state pool.

[0072] The parameter-generating agent derives the parameter domain of the adjustable parameters in the experimental step sequence based on the structured experimental knowledge in the shared state pool, and writes the parameter value range into the shared state pool.

[0073] The rule verification agent performs the multi-level rule verification based on the experimental step sequence and parameter value range in the shared state pool, and writes the verification result into the shared state pool.

[0074] The local replanning agent performs the local replanning and writes the replanning result into the shared state pool.

[0075] In this manner, the agents do not communicate directly; instead, they transmit and share data by reading and writing to a shared state pool. The task parsing agent, knowledge retrieval agent, process planning agent, parameter generation agent, rule verification agent, and local replanning agent work collaboratively in a serial or conditional parallel manner. The agents transmit structured information by reading and writing to the shared state pool. In a serial manner, one agent completes its task and writes the result to the shared state pool before the next agent reads data from the pool for processing. In a conditional parallel manner, when the input data of multiple agents is ready, these agents can execute simultaneously to improve processing efficiency. The agents do not engage in scheme-level debates, mutual criticism, voting, or merging. The shared state pool is a shared data storage space for all agents, used to store the recognition results, the structured experimental knowledge, the experimental state dependency graph, the experimental step sequence, the parameter value range, the verification results, and the replanning results. Through the multi-agent collaborative scheduling method described above, the system decomposes the complex experimental scheme generation task into multiple independent sub-tasks, with each agent undertaking a single responsibility, thus enabling the entire process to have good maintainability and scalability.

[0076] The embodiments described above in this application transform the design of wet experimental schemes for drug development from a traditional text generation problem into a structured workflow problem involving planning, verification, and local dynamic repair. Multiple intelligent agents collaborate to complete the entire process from task parsing, knowledge retrieval, process planning, parameter generation to rule verification and local repair by dividing roles and sharing states. This gives the generation of experimental schemes structured characteristics that allow for planning, verification, and repair. By constructing an explicit experimental state dependency graph to model the sequential constraints between experimental steps, and performing constraint path search on the experimental state dependency graph to generate an experimental step sequence that satisfies the execution dependencies and experimental constraints, the generation of experimental schemes is transformed from free text synthesis to path search and deterministic verification under strict dependency constraints, fundamentally avoiding the problem of incorrect experimental step order. By performing multi-level rule verification on the experimental step sequence and the parameter value ranges corresponding to each experimental step based on the detection endpoint, and locating the downstream influence subgraph and replanning only the steps within the subgraph when verification fails, automatic verification and precise repair are achieved during the scheme generation process, avoiding the omission of key steps and parameter out-of-control issues. It also solves the problem in existing technologies where modifying any experimental condition requires a full regeneration, leading to the loss of preceding steps. The technical solution provided in this application can generate experimental schemes with scientific rationality, logical rigor, and executability, achieving near-expert-level design performance. Furthermore, the system is a pure software system, operating only in the design stage before experimental execution, without involving hardware device control or real-time data acquisition, making it flexible in deployment and highly universal.

[0077] The above are some specific implementations of an experimental scheme generation method based on multi-agent cooperation and state dependency graphs provided in this application. Based on this, this application also provides a corresponding system. The system provided in this application will be described below from the perspective of functional modularization.

[0078] Figure 2 This is a schematic diagram of an experimental scheme generation system based on multi-agent cooperation and state dependency graphs, provided in an embodiment of this application. (Combined with...) Figure 2 As shown in the figure, the experimental scheme generation system 200 based on multi-agent cooperation and state dependency graph provided in this application includes: The task parsing module 210 is used to obtain the experimental target, perform semantic parsing on the experimental target, and identify the experimental type, experimental object, detection endpoint and experimental constraints contained in the experimental target; The knowledge retrieval module 220 is used to retrieve structured experimental knowledge that matches the experimental objective from a pre-built structured knowledge base based on the experimental type and the experimental object. The process planning module 230 is used to construct an experimental state dependency graph based on the structured experimental knowledge and the experimental type, and to perform constraint path search on the experimental state dependency graph according to the experimental constraints to generate a sequence of experimental steps that satisfy the execution dependencies and the experimental constraints; the experimental state dependency graph is a directed acyclic graph, the nodes of the experimental state dependency graph represent experimental steps and the attribute information corresponding to the experimental steps, and the edges of the experimental state dependency graph represent the execution dependencies between experimental steps. The parameter generation module 240 is used to perform parameter domain derivation on the adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object, so as to obtain the parameter value range of each adjustable parameter. The rule verification module 250 is used to perform multi-level rule verification on the experimental step sequence and the parameter value range corresponding to each experimental step according to the detection endpoint, and to determine whether the preset verification conditions are met. The output module 260 is used to output the experimental step sequence and the parameter values ​​corresponding to each experimental step as an experimental scheme when the multi-level rule verification is passed.

[0079] In one implementation of this application, the system further includes: The local replanning module is used to locate the downstream influence subgraph associated with the changed step node in the experimental step sequence based on the experimental state dependency graph when the multi-layer rule verification fails or when a user's request to change the experimental conditions is received; the nodes in the experimental state dependency graph other than the downstream influence subgraph are used as fixed boundaries; under the constraints of the fixed boundaries, the downstream influence subgraph is re-performed with constraint path search and parameter domain derivation to obtain the updated sub-process; the updated sub-process is concatenated with the fixed boundaries to obtain the updated experimental scheme.

[0080] In one implementation of this application, the execution dependency relationship includes hard dependency relationship, suggested dependency relationship and resource mutual exclusion relationship; the hard dependency relationship represents the order constraint that must be followed between two experimental steps; the suggested dependency relationship represents the recommended order constraint between two experimental steps; the resource mutual exclusion relationship represents the mutual exclusion constraint between two experimental steps due to sharing experimental resources.

[0081] In one implementation of this application, the rule verification module performs the multi-layer rule verification, including: The execution steps rely on compliance verification to verify whether the order of each experimental step in the experimental step sequence satisfies the execution dependency relationship in the experimental state dependency graph. Perform parameter domain compliance verification to verify whether the parameter values ​​corresponding to each experimental step in the experimental step sequence are within the parameter value range. Perform condition integrity verification to verify whether the experimental step sequence contains a preset mandatory step type; Perform a target consistency verification to verify whether the initial target and the final output of the experimental step sequence are consistent.

[0082] In one implementation of this application, the parameter generation module performs parameter domain derivation on the adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object, to obtain the parameter value range of each adjustable parameter, including: Retrieve structured experimental knowledge that matches the experimental object from the structured knowledge base; Extract the value distribution of the corresponding adjustable parameters from the structured experimental knowledge; The value distribution is statistically aggregated to obtain the literature consensus interval for each adjustable parameter, which is used as the parameter value interval.

[0083] In one implementation of this application, the process planning module performs constraint path search on the experimental state dependency graph, including: Starting from the initial node of the experimental state dependency graph, candidate paths are iteratively expanded according to topological order; In each expansion step, the step nodes that have been satisfied by the predecessor nodes are included in the candidate set; Logical consistency and conditional integrity checks are performed on each candidate path in the candidate set, and candidate paths that fail the checks are eliminated. Select at least one path from the verified candidate paths as the sequence of experimental steps.

[0084] In one implementation of this application, the task parsing module is used to perform semantic parsing on the experimental target by the task parsing agent and write the parsed recognition result into a shared state pool. The knowledge retrieval module is used by the knowledge retrieval agent to retrieve the structured experimental knowledge from the structured knowledge base according to the experimental type and the experimental object in the shared state pool, and to write the retrieved structured experimental knowledge into the shared state pool. The process planning module is used by the process planning agent to construct the experiment state dependency graph based on the structured experiment knowledge and the experiment type in the shared state pool, perform the constraint path search, generate the experiment step sequence, and write the experiment step sequence into the shared state pool. The parameter generation module is used by the parameter generation agent to derive the parameter domain of the adjustable parameters in the experimental step sequence based on the structured experimental knowledge in the shared state pool, and write the parameter value range into the shared state pool. The rule verification module is used by the rule verification agent to perform the multi-level rule verification according to the experimental step sequence and the parameter value range in the shared state pool, and write the verification result into the shared state pool. The local replanning module is used by the local replanning agent to perform the local replanning and write the replanning result into the shared state pool.

[0085] In one implementation of this application, the task parsing module, the knowledge retrieval module, the process planning module, the parameter generation module, the rule verification module, and the local replanning module work collaboratively in a serial or conditional parallel manner. The modules transmit structured information to each other by reading and writing the shared state pool. The shared state pool is a data storage space shared by all modules, used to store the identification results, the structured experimental knowledge, the experimental state dependency graph, the experimental step sequence, the parameter value range, the verification results, and the replanning results.

[0086] In one implementation of this application, the structured knowledge base is pre-constructed in the following manner: Offline structured parsing is performed on experimental literature, standard experimental protocols, and historical experimental records. Entities, relations, and attributes are extracted from these materials using named entity recognition and relation extraction models to obtain structured experimental knowledge. The structured experimental knowledge is assembled according to a preset structured format and stored in the structured knowledge base; The structured experimental knowledge in the structured knowledge base is vectorized and stored in a vector database.

[0087] This application also provides corresponding devices and computer storage media for implementing the solutions provided in this application.

[0088] The device includes a memory and a processor. The memory stores instructions or code, and the processor executes the instructions or code to cause the device to perform the method described in any embodiment of this application.

[0089] The computer storage medium stores code, and when the code is run, the device running the code implements the method described in any embodiment of this application.

[0090] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as a read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0091] It is understood that in the specific embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved need to obtain user permission or consent when the above embodiments of this application are applied to specific products or technologies, and the collection, use and processing of related data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0093] It should also be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0094] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating experimental schemes based on multi-agent cooperation and state dependency graphs, characterized in that, The method is executed collaboratively by multiple intelligent agents, and the method includes: Obtain the experimental target, perform semantic parsing on the experimental target, and identify the experimental type, experimental object, detection endpoint, and experimental constraints contained in the experimental target; Based on the experimental type and the experimental object, structured experimental knowledge matching the experimental objective is retrieved from a pre-built structured knowledge base; Based on the structured experimental knowledge and the experimental type, an experimental state dependency graph is constructed; the experimental state dependency graph is a directed acyclic graph, the nodes of the experimental state dependency graph represent experimental steps and the attribute information corresponding to the experimental steps, and the edges of the experimental state dependency graph represent the execution dependency relationship between experimental steps. Based on the experimental constraints, a constraint path search is performed on the experimental state dependency graph to generate a sequence of experimental steps that satisfies the execution dependency and the experimental constraints. Based on the structured knowledge base and the experimental object, the parameter domain of the adjustable parameters in the experimental step sequence is derived to obtain the parameter value range of each adjustable parameter. Based on the detection endpoint, multi-level rule verification is performed on the experimental step sequence and the parameter value range corresponding to each experimental step to determine whether the preset verification conditions are met. When the multi-level rule verification is successful, the experimental step sequence and the parameter values ​​corresponding to each experimental step are output as the experimental scheme.

2. The method according to claim 1, characterized in that, The method further includes: When the multi-level rule verification fails, or when a user requests a change to the experimental conditions is received, the downstream influence subgraph associated with the changed step node in the experimental step sequence is located according to the experimental state dependency graph. The nodes in the experimental state dependency graph, excluding the downstream influence subgraph, are used as fixed boundaries. Under the constraints of the fixed boundary, the downstream influence subgraph is re-searched for constraint paths and deduced for parameter domains to obtain the updated sub-process. The updated sub-process is then concatenated with the fixed boundary to obtain the updated experimental scheme.

3. The method according to claim 1, characterized in that, The execution dependencies include hard dependencies, suggested dependencies, and resource mutual exclusion relationships; the hard dependencies represent the order constraints that must be followed between two experimental steps; the suggested dependencies represent the recommended order constraints between two experimental steps; and the resource mutual exclusion relationships represent the mutual exclusion constraints between two experimental steps due to sharing experimental resources.

4. The method according to claim 1, characterized in that, The execution of multi-level rule verification includes: The execution steps rely on compliance verification to verify whether the order of each experimental step in the experimental step sequence satisfies the execution dependency relationship in the experimental state dependency graph. Perform parameter domain compliance verification to verify whether the parameter values ​​corresponding to each experimental step in the experimental step sequence are within the parameter value range. Perform condition integrity verification to verify whether the experimental step sequence contains a preset mandatory step type; Perform a target consistency verification to verify whether the initial target and the final output of the experimental step sequence are consistent.

5. The method according to claim 1, characterized in that, The step involves deriving the parameter domain for adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object, to obtain the parameter value range for each adjustable parameter, including: Retrieve structured experimental knowledge that matches the experimental object from the structured knowledge base; Extract the value distribution of the corresponding adjustable parameters from the structured experimental knowledge; The value distribution is statistically aggregated to obtain the literature consensus interval for each adjustable parameter, which is used as the parameter value interval.

6. The method according to claim 1, characterized in that, The constraint path search on the experimental state dependency graph includes: Starting from the initial node of the experimental state dependency graph, candidate paths are iteratively expanded according to topological order; In each expansion step, the step nodes that have been satisfied by the predecessor nodes are included in the candidate set; Logical consistency and conditional integrity checks are performed on each candidate path in the candidate set, and candidate paths that fail the checks are eliminated. Select at least one path from the verified candidate paths as the sequence of experimental steps.

7. The method according to claim 2, characterized in that, The method further includes: The task parsing agent performs semantic parsing on the experimental target and writes the parsed recognition result into the shared state pool. The knowledge retrieval agent retrieves the structured experimental knowledge from the structured knowledge base based on the experimental type and the experimental object in the shared state pool, and writes the retrieved structured experimental knowledge into the shared state pool. The process planning agent constructs the experiment state dependency graph based on the structured experiment knowledge and the experiment type in the shared state pool, performs the constraint path search, generates the experiment step sequence, and writes the experiment step sequence into the shared state pool. The parameter-generating agent derives the parameter domain of the adjustable parameters in the experimental step sequence based on the structured experimental knowledge in the shared state pool, and writes the parameter value range into the shared state pool. The rule verification agent performs the multi-level rule verification based on the experimental step sequence and the parameter value range in the shared state pool, and writes the verification result into the shared state pool. The local replanning agent performs the local replanning and writes the replanning result into the shared state pool.

8. The method according to claim 7, characterized in that, The task parsing agent, the knowledge retrieval agent, the process planning agent, the parameter generation agent, the rule verification agent, and the local replanning agent work collaboratively in a serial or conditionally parallel manner. The agents transmit structured information to each other by reading and writing to the shared state pool. The shared state pool is a shared data storage space for all agents, used to store the recognition results, the structured experimental knowledge, the experimental state dependency graph, the experimental step sequence, the parameter value range, the verification results, and the replanning results.

9. The method according to claim 1, characterized in that, The structured knowledge base is pre-built using the following methods: Offline structured parsing is performed on experimental literature, standard experimental protocols, and historical experimental records. Entities, relations, and attributes are extracted from these materials using named entity recognition and relation extraction models to obtain structured experimental knowledge. The structured experimental knowledge is assembled according to a preset structured format and stored in the structured knowledge base; The structured experimental knowledge in the structured knowledge base is vectorized and stored in a vector database.

10. An experimental scheme generation system based on multi-agent cooperation and state dependency graphs, characterized in that, include: The task parsing module is used to obtain the experimental target, perform semantic parsing on the experimental target, and identify the experimental type, experimental object, detection endpoint and experimental constraints contained in the experimental target; The knowledge retrieval module is used to retrieve structured experimental knowledge that matches the experimental objective from a pre-built structured knowledge base based on the experimental type and the experimental object. The process planning module is used to construct an experimental state dependency graph based on the structured experimental knowledge and the experimental type, and to perform constraint path search on the experimental state dependency graph according to the experimental constraints to generate a sequence of experimental steps that satisfy the execution dependencies and the experimental constraints. The experimental state dependency graph is a directed acyclic graph, where the nodes represent experimental steps and their corresponding attribute information, and the edges represent the execution dependencies between experimental steps. The parameter generation module is used to derive the parameter domain of the adjustable parameters in the experimental step sequence based on the structured knowledge base and the experimental object, so as to obtain the parameter value range of each adjustable parameter. The rule verification module is used to perform multi-level rule verification on the experimental step sequence and the parameter value range corresponding to each experimental step according to the detection endpoint, and to determine whether the preset verification conditions are met. The output module is used to output the experimental step sequence and the parameter values ​​corresponding to each experimental step as an experimental scheme when the multi-level rule verification is successful.

Images