A multi-target driven scientific experiment multi-scheme intelligent design method and system

Abstract

The application provides a multi-target-driven scientific experiment multi-scheme intelligent design method and system, and relates to the fields of artificial intelligence and scientific computing technology.The method comprises the following steps: obtaining a scientific hypothesis to be verified, the scientific hypothesis implying a plurality of verification sub-targets that need to be verified respectively; evaluating the coverage state of a current experiment path set on the plurality of verification sub-targets, identifying a sub-target that is not covered by the current experiment path set from the plurality of verification sub-targets as an uncovered sub-target; generating a corresponding retrieval query according to the uncovered sub-target, performing retrieval in a literature database, and extracting reference clues related to the uncovered sub-target; synthesizing a new experiment path for the uncovered sub-target based on the reference clues, and adding the new experiment path to the current experiment path set for multiple rounds of iteration to output an experiment verification scheme containing a plurality of complementary paths using different verification logics, so that complete, multi-perspective and rigorous testing of the scientific hypothesis is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and scientific computing technology, and in particular to a multi-objective driven intelligent design method and system for multiple schemes of scientific experiments. Background Technology

[0002] In the field of scientific research, verifying scientific hypotheses is a core step. Typically, a complex scientific hypothesis cannot be determined through a single experiment; instead, it requires multi-dimensional testing around several complementary verification sub-goals, such as correlation verification, causal intervention, effect quantification, and cross-condition generalization. In recent years, with the explosive growth of scientific literature data and the rapid development of artificial intelligence technology, utilizing large language models and information retrieval techniques to assist in the automatic design of scientific experimental verification schemes has become an important development trend for accelerating scientific discovery and reducing the cost of trial and error in scientific research.

[0003] Currently, existing intelligent assisted experimental design and reasoning techniques mainly rely on large language models combined with retrieval enhancement generation or agent technology. Conventional implementation schemes mainly include two modes: one is agent-based iterative retrieval, where the system alternates between literature retrieval and logical reasoning, dynamically reconstructing search terms based on the model's current local reasoning state to form a multi-step retrieval and reasoning loop to output the answer to a specific question; the other is automatic generation of experimental procedures, where, given a clear, subdivided experimental objective, the large language model directly generates the specific operational steps and a single technical route experimental procedure based on built-in pre-trained knowledge or background literature retrieved in a single retrieval.

[0004] However, the most significant technical challenge facing existing technologies in designing multiple pathways for complex scientific hypotheses is their inability to generate non-homogeneous experimental schemes with methodological differences based on gaps in the sub-goals of experimental verification. When attempting to generate multiple verification paths, existing systems rely solely on local inference states for their retrieval and generation of driving signals, lacking the ability to globally identify and track the coverage of each verification sub-goal inherent in the hypothesis. This results in a lack of exploration mechanisms for uncovered dimensions during scheme design, limiting the system to fine-tuning parameters within the same experimental paradigm. Consequently, the final set of experimental schemes exhibits high overlap in underlying methodologies, omissions in rigorous scientific verification, and an inability to achieve comprehensive and multi-perspective rigorous verification of the hypothesis. Summary of the Invention

[0005] This invention provides a multi-objective driven intelligent design method and system for multiple scientific experimental schemes, which solves the problem that existing technologies tend to make homogeneous fine-tuning within a single methodological paradigm when automatically generating experimental schemes, and cannot identify and track multiple verification dimensions implicit in scientific hypotheses, resulting in omissions or redundancies in the sub-objective dimensions of the produced experimental schemes. This invention achieves intelligent and automated design of multi-dimensional and complementary scientific experimental schemes.

[0006] This invention provides a multi-objective driven intelligent design method for multiple schemes in scientific experiments, comprising: Obtain scientific hypotheses to be verified, which implicitly contain multiple verification sub-goals that need to be verified separately; Evaluate the coverage status of the current set of experimental paths to the plurality of verification sub-targets, and identify the sub-targets that are not covered by the current set of experimental paths as uncovered sub-targets from the plurality of verification sub-targets; Based on the uncovered sub-targets, a corresponding search query is generated, and a search is performed in the literature database to extract reference clues related to the uncovered sub-targets. Based on the reference clues, a new experimental path is synthesized for the uncovered sub-target, and the new experimental path is added to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logics.

[0007] According to the method provided by the present invention, the step of evaluating the coverage status of the current experimental path set on the plurality of verification sub-targets, and identifying sub-targets not covered by the current experimental path set as uncovered sub-targets from the plurality of verification sub-targets, includes: The large language model is invoked to perform coverage audit on the current set of experimental paths to determine the coverage status of the current set of experimental paths for the multiple verification sub-targets; Based on the coverage status, the content that is still not covered by the current set of experimental paths is diagnosed from multiple evaluation levels, and the uncovered sub-targets are identified; wherein, the evaluation level includes at least one of mechanism dimension, analysis level, untested hypothesis and method family; The uncovered sub-targets are output as gap reports in natural language, which are used to characterize the missing state of the experimental scheme in the validation dimension.

[0008] According to the method provided by the present invention, the step of outputting the uncovered sub-targets as gap reports in natural language form includes: obtaining a path summary of the current experimental path set; passing in the historical gap reports generated in the previous iteration as context, and removing the verification sub-targets that have been covered by the current experimental path set from the historical gap reports in combination with the currently identified uncovered sub-targets; adding the newly identified uncovered sub-targets to generate the gap reports for the current iteration.

[0009] According to the method provided by the present invention, generating a corresponding retrieval query based on the uncovered sub-targets includes: reading the gap report and generating an exploration query for exploring new method families for the uncovered sub-targets in the gap report; For the verification sub-objectives that are already covered by the current experimental path set among the multiple verification sub-objectives, generate a deeper query to supplement the evidence strength within the existing method family; The exploratory query and the in-depth query are combined as the output of the retrieval query.

[0010] According to the method provided by the present invention, the steps of generating the exploratory query and the in-depth query further include: extracting historical query text and using it as an anti-duplication constraint; and generating the exploratory query and the in-depth query simultaneously based on the anti-duplication constraint to prevent repeated searches of already explored literature domains.

[0011] According to the method provided by the present invention, the synthesis of a new experimental path for the uncovered sub-target based on the reference cues includes: integrating the reference cues and orthogonality rules to construct a path generation instruction, wherein the orthogonality rules limit the candidate path to using a non-overlapping subset of cues with existing paths in the current experimental path set; initially generating the candidate path according to the path generation instruction; calling a large language model to determine the independence of the candidate path, and outputting a decision to retain it as a new path, merge it into the existing path, or reject it; and confirming the candidate path as the new experimental path in response to the decision to retain it as a new path.

[0012] According to the method provided by the present invention, after initially generating the candidate path according to the path generation instruction, the method further includes: extracting elements from the generated text content and constructing the candidate path into a structured data format containing intervention elements, readout elements and criterion elements; Specifically, the operations or processes applied to the experimental subjects are extracted and configured as the intervention elements in the structured data format; Extract measurable observation indicators and configure them as readout elements in the structured data format; Extract the standard content used to determine whether the scientific hypothesis is valid, and configure it as the criterion element in the structured data format so that the candidate path is transformed into an executable and verifiable standardized experimental plan.

[0013] According to the method provided by the present invention, the step of extracting reference clues related to the uncovered sub-target specifically includes: obtaining documents with normalized identifiers from the document database; extracting local content related to the uncovered sub-target from the documents; and extracting the local content and the normalized identifier of the document to which it belongs to generate the reference clues.

[0014] According to the method provided by the present invention, the step of adding the new experimental path to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logic includes: Perform rule validation on the new experimental path to determine whether the new experimental path meets the admission criteria of containing readout elements or criterion elements, being bound to the reference clue, and not overlapping with the method family of existing paths in the current experimental path set. In response to the judgment result that the admission conditions are met, the new experimental path is incorporated into the current experimental path set to obtain an updated path set; The updated path set is used as the basis to trigger the next round of iteration, and the process continues to update until all of the multiple verification sub-targets are detected to be covered. The final path set is then determined as the output of the experimental verification scheme.

[0015] This invention provides a multi-objective driven intelligent design system for multiple schemes in scientific experiments, comprising: The hypothesis acquisition module is used to acquire scientific hypotheses to be verified, which implicitly contain multiple verification sub-objectives that need to be verified separately. The coverage audit module is used to evaluate the coverage status of the current experimental path set on the multiple verification sub-targets, and identify the sub-targets that are not covered by the current experimental path set as uncovered sub-targets from the multiple verification sub-targets; The retrieval and extraction module is used to generate a corresponding retrieval query based on the uncovered sub-target, perform a retrieval in the literature database, and extract reference clues related to the uncovered sub-target. The path synthesis and iteration module is used to synthesize new experimental paths for the uncovered sub-targets based on the reference clues, and add the new experimental paths to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logics.

[0016] According to the system provided by the present invention, the coverage audit module is used to: call a large language model to perform coverage audit on the current experimental path set, and determine the coverage status of the current experimental path set on the multiple verification sub-targets; Based on the coverage status, the content that is still not covered by the current set of experimental paths is diagnosed from multiple evaluation levels, and the uncovered sub-targets are identified; wherein, the evaluation level includes at least one of mechanism dimension, analysis level, untested hypothesis and method family; The uncovered sub-targets are output as gap reports in natural language, which are used to characterize the missing state of the experimental scheme in the validation dimension.

[0017] According to the system provided by the present invention, the retrieval and extraction module is configured to: read the gap report and generate an exploration query for exploring new method families for the uncovered sub-targets in the gap report; For the verification sub-objectives that are already covered by the current experimental path set among the multiple verification sub-objectives, generate a deeper query to supplement the evidence strength within the existing method family; The exploratory query and the in-depth query are combined as the output of the retrieval query.

[0018] According to the system provided by the present invention, the path synthesis and iteration module is used to: integrate the reference clues and orthogonality rules to construct path generation instructions, wherein the orthogonality rules limit the candidate paths to using a non-overlapping subset of clues with the existing paths in the current experimental path set; The candidate paths are initially generated based on the path generation instructions; The large language model is invoked to determine the independence of the candidate paths, and the decision is output as to retain them as new paths, merge them into the existing paths, or reject them. In response to the decision to retain the candidate path as a new path, the candidate path is confirmed as the new experimental path.

[0019] According to the system provided by the present invention, the retrieval and extraction module is configured to: obtain documents with normalized identifiers from the document database; extract local content related to the uncovered sub-target from the documents; and extract the local content and the normalized identifier of the document to which it belongs to generate the reference clue.

[0020] According to the system provided by the present invention, the path synthesis and iteration module is used to: perform rule verification on the new experimental path, and determine whether the new experimental path meets the admission conditions of including readout elements or criterion elements, being bound to the reference clues, and not overlapping with the method family of existing paths in the current experimental path set; In response to the judgment result that the admission conditions are met, the new experimental path is incorporated into the current experimental path set to obtain an updated path set; The updated path set is used as the basis to trigger the next round of iteration, and the process continues to update until all of the multiple verification sub-targets are detected to be covered. The final path set is then determined as the output of the experimental verification scheme.

[0021] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-objective driven intelligent design method for multiple schemes of scientific experiments as described above.

[0022] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-objective driven intelligent design method for multiple schemes of scientific experiments as described above.

[0023] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-objective driven intelligent design method for multiple schemes of scientific experiments as described above.

[0024] The present invention provides a multi-objective driven intelligent design method and system for multiple schemes in scientific experiments. By acquiring scientific hypotheses and identifying multiple implicit verification sub-objectives, it evaluates the coverage of the current path set to each sub-objective to identify uncovered sub-objectives. Then, it generates targeted search queries based on these uncovered sub-objectives and extracts corresponding reference clues. Finally, based on these reference clues, it synthesizes new experimental paths targeting uncovered sub-objectives and continuously completes them through multiple iterations. This enables proactive exploration driven by uncovered verification dimensions, ensuring that each newly generated path targets a clear, uncovered verification sub-objective. This effectively guides the system to produce a set of schemes that are complementary rather than redundant in the verification dimension, avoiding the stacking of homogeneous schemes within the same methodological paradigm. It significantly improves the recall capability for diverse experimental paradigms and ultimately achieves a complete, multi-perspective, and rigorous test of scientific hypotheses. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the multi-objective driven intelligent design method for multiple schemes in scientific experiments provided by this invention.

[0027] Figure 2 This is a schematic diagram of the overall process of the multi-objective driven intelligent design method for multiple schemes of scientific experiments provided by the present invention.

[0028] Figure 3 This is a schematic diagram of the structure of the multi-objective driven intelligent design system for scientific experiments with multiple schemes provided by the present invention.

[0029] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] In the field of automated design of scientific experimental schemes, existing technologies have mainly developed along two technical paths.

[0032] The first approach is intelligent agent-based iterative retrieval technology. Its core idea is to alternate between document retrieval and logical reasoning, dynamically reconstructing the retrieval query based on the model's current local reasoning state during the reasoning process, forming a closed loop of multi-step retrieval and reasoning, ultimately outputting an answer to a specific question. This type of method performs excellently on open-domain question answering tasks. However, when multiple experimental schemes need to be generated, because the driving signal for each retrieval comes from the information needs generated in real-time within the reasoning chain rather than the overall coverage of the method space, large language models tend to make minor adjustments at the wording level around the same method paradigm when generating multiple candidate schemes. For example, for the scheme of using optogenetics to activate a brain region, they might only replace it with a similar variant such as using chemogenetics to activate the same brain region, but they cannot break away from the existing paradigm to explore completely different verification logic.

[0033] The second approach is automated experimental procedure generation technology. Its core idea is that, given a clear experimental objective, a large language model directly generates a sequentially executed step-by-step operational procedure based on pre-trained knowledge or background literature obtained from a single search. While this method has some value in automating the writing of standardized experimental procedures, it essentially unfolds steps along a single route. It lacks the ability to globally identify the multiple verification dimensions inherent in a scientific hypothesis, and it also lacks the ability to track and complete which dimensions have been covered and which remain blank.

[0034] The common flaw in both of these technical approaches lies in the fact that the driving force for retrieval and generation both originate from local, single-level signals—the former stemming from the immediate needs of the reasoning state, and the latter from a given single experimental objective—lacking a systematic understanding of the multiple verification sub-objectives inherent in the scientific hypothesis itself. This deficiency leads to existing systems being unable to proactively identify which complementary dimensions the hypothesis needs to be tested from when faced with a complex scientific hypothesis, let alone generate differentiated experimental schemes driven by uncovered dimensions. The resulting set of schemes is either highly homogeneous within the same paradigm or lacks key verification dimensions, failing to meet the requirements of rigorous scientific testing.

[0035] To address the shortcomings of existing technologies, this invention provides a multi-objective driven intelligent design method and system for multiple scientific experimental schemes. Its core idea is to treat a scientific hypothesis as a structured object containing multiple sub-objectives that need to be verified separately. A coverage audit mechanism continuously diagnoses the coverage status of the current scheme set on each sub-objective, identifying uncovered sub-objectives as the driving force for retrieval and generation. Each iteration targets one or more clearly defined uncovered sub-objectives for literature retrieval and synthesizes corresponding experimental paths. Simultaneously, an orthogonality constraint mechanism ensures that new paths are methodologically independent of existing paths, ultimately producing multiple experimental schemes that are complementary in terms of sub-objective dimensions and differentiated in terms of methodology. The specific embodiments of this invention are described in detail below with reference to the accompanying drawings.

[0036] Before describing the technical solutions of the embodiments of the present invention, the terms and concepts involved in the embodiments of the present invention will be explained illustratively.

[0037] Scientific hypothesis: refers to a core scientific proposition that needs to be experimentally verified, and is the object of this invention. For example, a certain immune cell state driving neuroinflammation or a gene mutation causing abnormal synaptic plasticity. A scientific hypothesis typically contains multiple verification sub-goals that need to be tested from different perspectives; a single experimental approach is insufficient to achieve complete verification of the hypothesis.

[0038] Validation sub-objectives refer to different types of testing dimensions in the scientific validation process, i.e., from which angle the scientific hypothesis is experimentally tested. These include, but are not limited to, correlation validation, causal intervention, effect quantification, control exclusion, cross-condition generalization, mechanism and pathway characterization, and reversibility assessment. Different scientific hypotheses imply different sets of sub-objectives. This invention identifies these sub-objectives and uses them as drivers for scheme design.

[0039] Coverage status: refers to the degree to which the current set of experimental paths satisfies each validation sub-objective. For a given validation sub-objective, if at least one path in the current set can provide experimental evidence for the validation dimension represented by the sub-objective, then the sub-objective is marked as covered; conversely, if no path responds to the sub-objective, or its response remains at a general level without the three elements of actionable intervention, readout, and criterion, then the sub-objective is marked as uncovered or with weak evidence.

[0040] Retrieval query: refers to the query statement used to retrieve relevant literature in a literature database. In this invention, the retrieval query is generated based on uncovered sub-targets and includes two types: exploratory queries, generated for uncovered sub-targets to discover new method families; and in-depth queries, generated for covered but weakly evidenced sub-targets to supplement the strength of evidence within existing method families. Both types of queries are subject to anti-duplication constraints to avoid repeatedly searching already explored areas.

[0041] Reference cues refer to local content extracted from retrieved literature that is relevant to the uncovered sub-targets. Each cue includes normalized identifiers of the source literature, literature excerpts, method descriptions, measurement metrics, and contextual information. Reference cues are the direct source of material for synthesizing new experimental pathways.

[0042] A new experimental path refers to an executable experimental verification scheme synthesized for a specific uncovered sub-target. Each new experimental path is structured into a three-element format: intervention—readout—criteria. The intervention records the operations or treatments applied to the experimental subject, the readout records the measurable observation indicators or endpoints, and the criteria record the standards for determining whether a hypothesis is supported or refuted. New paths must undergo orthogonality testing and rule verification before being added to the current path set.

[0043] Mechanism dimension: One of the assessment dimensions of coverage audit, referring to the type of biological or physical mechanism of action involved in the experimental protocol. Examples include different levels of action mechanisms such as signaling pathways, molecular interactions, cell behavior, and circuit connections. Coverage audit, through diagnosis at the mechanism dimension, determines whether the current set of pathways covers the main mechanistic levels involved in the hypothesis.

[0044] Analysis level: One of the assessment levels in coverage audit, referring to the organizational level of the research subjects involved in the experimental protocol. Examples include molecular level, cellular level, tissue level, organ level, individual level, and population level. Coverage audit, through diagnosis of the analysis level, determines whether the current set of paths has tested the hypothesis at different organizational levels.

[0045] Untested hypotheses: One of the assessment dimensions of coverage audit, referring to intermediate propositions or auxiliary hypotheses implicit in scientific assumptions that have not yet been experimentally tested within the current set of experimental pathways. Coverage audit, through the diagnosis of untested hypotheses, identifies potential hypotheses in the current set of pathways that are assumed to be true but have not yet been experimentally verified.

[0046] Method family: One of the assessment dimensions in audit coverage, referring to a set of experimental methods with the same core verification logic. For example, optogenetics, chemogenetics, and pharmacological interventions belong to the causal intervention method family, while functional magnetic resonance imaging, two-photon calcium imaging, and fiber optic photometric recording belong to the functional readout method family. Different method families have fundamentally different verification logics, while different schemes within the same method family are substantially the same in their underlying methodology. The purpose of orthogonality constraints is to ensure that new paths belong to a method family different from existing paths.

[0047] Gap Briefing: This refers to the natural language diagnostic text output by the large language model as a methodological meta-analyst, approximately 120-200 words. This briefing describes content not yet covered by the current path set, including unanswered mechanism dimensions, unaddressed analysis levels, untested hypotheses, and yet-to-emerge method families—in other words, uncovered validation sub-goals. The gap briefing only outputs diagnoses of the gaps and cannot provide search queries or keywords; it is generated independently for each round to ensure real-time diagnostics.

[0048] Deepening queries: These are retrieval queries generated for validation sub-goals that have been covered but for which evidence is weak. Unlike exploratory queries, which target new method families, deepening queries supplement the retrieval within the same method family by quantifying information such as indicators, validation benchmarks, sample size, control settings, and failure modes, thereby strengthening the evidentiary strength of existing pathways.

[0049] Orthogonality rules are a set of rules used to constrain the generation of new paths and ensure methodological independence between paths. These rules stipulate that candidate paths must use a non-overlapping subset of clues with existing paths, and that verification logic between paths should be orthogonal to the existing path family. If a new clue cannot support orthogonal paths, an empty list is returned, and existing paths cannot be restated or slightly renamed. Orthogonality rules are a core constraint to prevent homogenization of solutions.

[0050] Judgment Decision: This refers to one of three decision outcomes output by the orthogonal judge large language model for candidate paths, including retaining as a new path, i.e., retaining the candidate path as a methodologically orthogonal to the existing path; merging into an existing path, i.e., merging the candidate path into the existing path if it belongs to the same family as an existing path and can supplement its evidence; and rejecting, i.e. rejecting the candidate path if it belongs to the same family as an existing path and has no new value.

[0051] Criterion elements: These refer to the criterion portion of the three elements of an experimental pathway, recording the specific standards for determining whether a scientific hypothesis is supported or refuted by the experiment. For example, if the level of inflammatory factors increases more than twice that of the control group after optogenetic activation and the statistical test is significant, then the hypothesis is supported; otherwise, if there is no significant difference before and after the intervention, the hypothesis is not valid. Criterion elements make the experimental pathway testable, meaning that a third party can independently determine whether the experimental results support or refute the hypothesis based on these criteria.

[0052] The method provided in this invention can be executed by a computer device, including but not limited to electronic devices with computing and storage capabilities such as servers, personal computers, workstations, and cloud computing platforms. The computer device is equipped with at least one processor and a memory, the memory storing an executable computer program. The processor executes the steps described in this invention by calling the program. The computer device is also equipped with a communication interface for data interaction with external literature databases. Preferably, the computer device is deployed with a large language model inference environment for performing natural language understanding and generation tasks such as coverage auditing, retrieval planning, path synthesis, and orthogonal referencing. The large language model can be deployed locally on the computer device or can obtain inference results by calling a remote large language model service interface through the communication interface. The computer device is also equipped with a literature database access module for searching and obtaining full text or abstracts of literature from one or more literature data sources such as local literature databases, PubMed, arXiv, and SemanticScholar. The steps described in this invention can be completed collaboratively by different functional modules in the aforementioned computer device, or can be executed sequentially by a single processor.

[0053] Figure 1 This is a flowchart illustrating a multi-objective-driven intelligent design method for multiple schemes in scientific experiments provided in an embodiment of the present invention. The method includes: Step 101: Obtain the scientific hypothesis to be verified, which implicitly contains multiple verification sub-goals that need to be verified separately.

[0054] The scientific hypothesis is a core scientific proposition input by the user that needs to be verified experimentally, such as an immune cell state driving neuroinflammation or a gene mutation causing abnormal synaptic plasticity. After acquiring the hypothesis, the device performs semantic parsing to identify multiple verification sub-goals implicit in the hypothesis at the scientific verification level. Verification sub-goals are different types of testing dimensions in the scientific verification process, including but not limited to correlation verification, causal intervention, effect quantification, control exclusion, cross-condition generalization, mechanism pathway characterization, and reversibility assessment. Different scientific hypotheses imply different sets of sub-goals; for example, a claim focusing on causality may emphasize causal intervention and control exclusion, while a claim focusing on descriptive relationships may emphasize correlation verification and cross-condition generalization.

[0055] The device uses natural language understanding technology to identify the set of sub-targets that the current hypothesis should cover, and uses this set as a reference benchmark for subsequent coverage audits. Initially, the current experimental path set can be empty or it can include the initial scheme provided by the user; the device will iteratively complete the set based on this.

[0056] Step 102: Evaluate the coverage status of the current experimental path set on the multiple verification sub-targets, and identify the sub-targets that are not covered by the current experimental path set as uncovered sub-targets.

[0057] Specifically, the device acquires all currently generated experimental paths, performs semantic analysis on each path, and determines the type of verification sub-target that the path responds to. Based on this, the device compares all verification sub-targets identified in step 101 with the sub-targets covered by the current path, and counts whether each sub-target is covered by at least one path. For covered sub-targets, the device can further evaluate whether they possess sufficient experimental details, such as whether they include explicit intervention methods, measurable readout metrics, and criteria for determining the validity of the hypothesis. For sub-targets that do not have any path response or whose response lacks executable experimental details, the device marks them as uncovered sub-targets. This coverage evaluation process ensures that the device can perceive the blank areas of the current scheme set in the verification dimension in real time, providing accurate driving force for subsequent targeted retrieval and generation.

[0058] Step 103: Generate a corresponding search query based on the uncovered sub-target, perform a search in the literature database, and extract reference clues related to the uncovered sub-target.

[0059] The generation of search queries is based on uncovered sub-targets. Specifically, the device designs search statements for uncovered verification dimensions, ensuring that the search results provide directly relevant literature support for the synthesis of experimental protocols within that dimension. Literature databases include, but are not limited to, local literature repositories, PubMed, arXiv, and Semantic Scholar. After relevance screening, the device extracts relevant local content from the selected literature as reference clues. Each clue includes at least the identification information of the source literature and a description of the technical content related to that sub-target.

[0060] This retrieval process ensures that the literature information obtained in each round directly serves to fill the coverage gaps in the current scheme set, avoiding invalid retrievals that overlap with the verification dimensions of existing schemes.

[0061] Step 104: Based on the reference clues, synthesize new experimental paths for the uncovered sub-targets, and add the new experimental paths to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logics.

[0062] Specifically, the device uses the technical content from the reference clues as material and synthesizes one or more candidate paths using a large language model. Each candidate path describes an experimental verification method for the uncovered sub-target. During the synthesis process, the device ensures that the generated new path is logically distinct from the existing paths, making the different paths methodologically independent, thus forming a complementary rather than redundant set of solutions. After the new path is generated, the device adds it to the current experimental path set, forming an updated path set. Subsequently, based on the updated set, the device returns to step 102 to re-execute the coverage audit and enter the next iteration. This iterative process continues until the device detects that all verification sub-targets have been fully covered. At this point, the device stops iterating and outputs the final path set as the experimental verification solution.

[0063] Through the aforementioned multi-round iterative mechanism, the equipment gradually improves the coverage of the solution set on various verification sub-objectives, and finally produces multiple experimental paths that can comprehensively test scientific hypotheses from multiple complementary dimensions.

[0064] Through steps 101-104 above, the device is driven by multiple verification sub-goals implied by the scientific hypothesis. It identifies uncovered sub-goals through coverage auditing, and generates targeted search queries, extracts reference clues, and synthesizes new experimental paths for these uncovered sub-goals. Through multiple rounds of iteration, it continuously fills in the blind spots of the scheme set in each verification dimension, so that the final output of multiple experimental paths complement each other in the verification dimension and naturally differentiates in methodology. This effectively avoids the problem of stacking homogeneous schemes within the same paradigm, and realizes a complete and rigorous test of the scientific hypothesis from multiple dimensions, significantly reducing the trial and error costs for researchers in the experimental scheme design stage.

[0065] Specifically, in step 102, the device invokes a large language model to perform a coverage audit on the current set of experimental paths. This large language model, acting as a methodological meta-analyst, receives the current set of all paths, the original research question, and the scientific hypothesis as input. It performs semantic analysis on each existing path to determine the type of validation sub-goal it addresses, and based on this, determines the coverage status of the current set of experimental paths for all validation sub-goals. Based on this coverage status, the large language model diagnoses the content still not covered by the current set of experimental paths from multiple evaluation levels, thereby identifying the uncovered sub-goals. These evaluation levels include at least one of the following: mechanism dimension, analysis level, untested hypothesis, and method family. Among them, the mechanism dimension refers to the type of biological or physical mechanism of action involved in the experimental protocol, such as different levels of action mechanisms like signaling pathways, molecular interactions, cell behavior, and circuit connections. The device uses the mechanism dimension to determine whether the current set of pathways covers the main mechanism levels involved in the hypothesis. The analysis level refers to the organizational level of the research object involved in the experimental protocol, such as the molecular level, cellular level, tissue level, organ level, individual level, and population level. The device uses the analysis level to determine whether the current set of pathways has tested the hypothesis at different organizational levels. Untested hypotheses refer to the current experimental pathways. The set of untested intermediate propositions or auxiliary hypotheses implicit under scientific assumptions are identified by the device through the diagnosis of untested hypotheses. The method family refers to a set of experimental methods with the same core verification logic. For example, optogenetics, chemogenetics, and pharmacological interventions belong to the causal intervention method family, while functional magnetic resonance imaging, two-photon calcium imaging, and fiber optic photometric recording belong to the functional readout method family. The device uses the diagnosis of method families to determine whether the current path set covers different types of method categories at the verification logic level.

[0066] After diagnosing the uncovered content, the device outputs a gap report in natural language. This gap report is a diagnostic text of approximately 120 to 200 words, used to characterize the missing information in the validation dimension of the experimental design. Specifically, it includes uncovered content such as unresponsive mechanism dimensions, unaddressed analysis levels, untested hypotheses, and unexplored method families. When generating the gap report, the device explicitly requires the large language model not to generate search queries or keywords, but only to output the diagnosis of the gap. Furthermore, it prohibits pasting gap reports generated in the previous round, ensuring that each report is generated independently for the current path set to accurately reflect the current coverage gap status. This gap report will serve as the direct basis for generating search queries in subsequent steps, driving the device to perform targeted searches for the uncovered sub-targets.

[0067] Through the aforementioned coverage audit process, the device can perceive the coverage status of the current solution set on each verification sub-target in real time and accurately, and output the uncovered sub-targets in the form of a structured gap report. This transforms the identification of which sub-targets are not covered into a precise driving force for subsequent retrieval and generation, effectively avoiding repeated exploration of covered directions and improving the efficiency of multiple iterations.

[0068] Further, in step 102, when outputting the uncovered sub-targets as gap reports in natural language form, the device generates the gap report for the current round in the following manner: The device first obtains a path summary of the current set of experimental paths. This summary encapsulates the core content of all currently generated experimental paths, including the validation sub-objectives targeted by each path, the method family used, and a summary of the three elements of intervention, readout, and criteria. The device then feeds the historical gap report generated in the previous iteration into the large language model as context, enabling the model to understand the previously identified uncovered sub-objectives and their evolution. Based on this, the device updates the historical gap report by incorporating the uncovered sub-objectives identified in the current iteration: removing validation sub-objectives already covered by the current set of experimental paths and adding newly identified uncovered sub-objectives, thus generating the current iteration's gap report. During generation, the device requires the large language model not to paste the original text of the previous iteration's report, ensuring that each iteration's report is generated independently based on the actual situation of the current set of paths.

[0069] Through the above method, the device achieves incremental updates to the gap report. Each round of gap reports inherits historical diagnostic information, removes previously filled content, and adds newly discovered blind spots, ensuring that the gap report always reflects the true coverage gap status at the current moment. This mechanism avoids redundant calculations and information omissions caused by the device re-diagnosing from scratch in each round. It also prevents the identification of new blind spots from being delayed by covered targets occupying diagnostic space in historical gap reports, thus guaranteeing the diagnostic accuracy and timeliness of the gap report and providing a reliable basis for the accurate generation of subsequent retrieval queries.

[0070] Through the aforementioned coverage audit process, the device can perceive the coverage status of the current solution set on each verification sub-target in real time and accurately, and output the uncovered sub-targets in the form of a structured gap report. This transforms the identification of which sub-targets are not covered into a precise driving force for subsequent retrieval and generation, effectively avoiding repeated exploration of covered directions and improving the efficiency of multiple iterations.

[0071] Further, in step 103, the device generates corresponding search queries based on the uncovered sub-targets identified in step 102. Specifically, the device first reads the gap report, which describes the validation sub-targets not yet covered by the current path set in natural language. For the uncovered sub-targets in the gap report, the device generates an exploratory query to explore new method families. This exploratory query explicitly targets validation dimensions not covered by any existing path, aiming to discover novel experimental paradigms and method categories corresponding to the sub-target. Simultaneously, for validation sub-targets already covered by the current experimental path set, the device generates a refinement query to supplement the evidence strength within existing method families. This refinement query further retrieves information such as quantifiable readout metrics, validation benchmarks, sample size, control settings, and failure modes within the same method family to strengthen the evidence strength of existing paths. The device outputs both the exploratory and refinement queries as search queries in the literature database.

[0072] During the generation of exploratory and in-depth queries, the device extracts historical query text and uses it as an anti-duplication constraint. Specifically, the device obtains the search query text used in previous iterations and injects this historical query text as a constraint into the current query generation process. Based on this anti-duplication constraint, the device simultaneously generates exploratory and in-depth queries, prohibiting repeated searches of already explored literature domains, thereby avoiding the waste of computing resources and information redundancy caused by repeated searches in the same direction in multiple iterations.

[0073] Through the aforementioned query generation methods, the device implements a dual-track parallel retrieval strategy of exploration and refinement. On the one hand, exploratory queries continuously expand the family of new methods to fill coverage gaps; on the other hand, refinement queries solidify the evidence foundation of existing paths, ensuring that the retrieval process balances breadth and depth. Simultaneously, the introduction of anti-duplication constraints ensures that the device continuously explores new literature neighborhoods during iteration, preventing search results from repeating across different rounds, significantly improving retrieval efficiency and information increment in multiple iterations.

[0074] After generating a search query, the device searches the literature database and extracts reference clues related to the uncovered sub-targets from the returned literature. Specifically, the device retrieves literature with normalized identifiers from the literature database, extracts local content related to the uncovered sub-targets from these documents, and combines this local content with the normalized identifiers of the respective documents to generate reference clues. Each reference clue includes at least the normalized identifier of the source document, a document excerpt, a method description, and measurement indicators. The normalized identifiers include, but are not limited to, PMID, DOI, arXiv ID, or local database identifiers. The literature-anchoring characteristics of the reference clues provide a clear and traceable basis for the subsequent experimental path of synthesis, laying the foundation for the verifiability and reproducibility of the experimental design.

[0075] Through the above retrieval and extraction process, the device transforms the uncovered sub-targets described in the gap report into actual retrieval behavior, obtains direct materials from massive scientific literature to support the synthesis of experimental schemes under the sub-target, and ensures that incremental information can be obtained in each round of retrieval through anti-duplication constraints, providing accurate and continuously updated knowledge sources for subsequent path synthesis.

[0076] Further, in step 103, the device generates a corresponding search query based on the identified uncovered sub-targets and performs a search in the literature database to extract reference clues related to the uncovered sub-targets. Specifically, the device retrieves the retrieved literature from the literature database. These literatures are all accompanied by normalized identifiers, including but not limited to PMID, DOI, arXiv ID, or local database identifiers. Each paper carries the aforementioned normalized identifiers and their metadata. The device extracts local content related to the uncovered sub-targets from the retrieved literatures. This local content includes fragmented information such as specific method descriptions, experimental designs, measurement indicators, and result data corresponding to the sub-target. Based on this, the device extracts this local content and the normalized identifier of the corresponding literature to generate reference clues, so that each reference clue is bound to its source literature. During the extraction process, the device establishes an alias mapping, corresponding the abbreviations used in the literature within the steps to their global identifiers, ensuring the integrity of the clue tracing link.

[0077] Through the extraction method described above, each reference clue generated by the device is anchored to a source document with a standardized identifier, ensuring that the subsequent synthesized experimental path has a complete literature tracing link. When reviewing the final experimental design, researchers can directly trace back to the original literature through this standardized identifier to verify the experimental evidence on which the design is based. This significantly enhances the credibility and verifiability of the design, while providing researchers with a convenient tracing method, facilitating in-depth research and adjustments based on the design.

[0078] Further, in step 104, the device synthesizes a new experimental path for the uncovered sub-target based on the reference cues extracted in step 103. Specifically, the device integrates reference cues and orthogonality rules to construct a path generation instruction, which is input into the large language model to drive it to generate candidate paths. The orthogonality rule restricts the candidate paths to using non-overlapping subsets of cues with existing paths in the current experimental path set; that is, the documentary evidence upon which the new path is based must not overlap with the documentary evidence upon which existing paths are based in core content, thereby ensuring that the new path is distinguished from existing paths at the knowledge source level. The device initially generates candidate paths according to this path generation instruction, and the candidate paths describe an experimental verification method for the uncovered sub-target.

[0079] After initially generating candidate paths, the device extracts elements from the generated text content, constructing the candidate paths into a structured data format containing intervention elements, readout elements, and criterion elements. Specifically, the device extracts the operations or treatments applied to the experimental subjects from the generated text, configuring them as intervention elements in the structured data format, such as optogenetic activation of a brain region or drug injection of a receptor antagonist; it extracts measurable observation indicators, configuring them as readout elements in the structured data format, such as changes in calcium signal fluorescence intensity or inflammatory factor expression levels; and it extracts the criteria used to determine whether a scientific hypothesis is valid, configuring them as criterion elements in the structured data format, such as if the change in the readout indicator relative to the control group exceeds a preset threshold and is statistically significant, then the hypothesis is supported; otherwise, the hypothesis is not valid. Through the above structured extraction, the candidate paths are transformed from natural language descriptions into standardized experimental protocols that are executable and verifiable. Each path includes a clear intervention method, a quantifiable measurement endpoint, and reproducible judgment criteria.

[0080] Subsequently, the device invokes a large language model to determine the independence of candidate paths. This large language model acts as an orthogonal judge, comparing candidate paths with existing paths in batches to determine, from a methodological perspective, whether the two are substantially independent, and outputting one of three decision options: First, retain as a new path, meaning the candidate path and existing paths are substantially orthogonal methodologically and are retained; second, merge into an existing path, meaning the candidate path belongs to the same method family as an existing path and can supplement its evidence, and its clues are merged into that existing path; third, reject, meaning the candidate path and existing paths substantially belong to the same family and offer no new value, and are rejected. Upon receiving the decision, the device, responding to the decision to retain as a new path, confirms the candidate path as a new experimental path.

[0081] Through the aforementioned orthogonality constraints and structured extraction mechanisms, the device ensures that each newly generated path is methodologically independent of existing paths, breaking away from the homogeneous generation model of fine-tuning wording within the same paradigm in existing technologies. This enables true diversity in the set of solutions at the method family level. Simultaneously, the structured format of the intervention-readout-criteria three-element approach ensures that each path is executable and verifiable. Third-party researchers can directly reproduce experiments and independently judge the results based on this structured scheme, greatly enhancing the scientific rigor and practical value of the solution.

[0082] The device adds the confirmed new experimental path to the current experimental path set, forming an updated path set. Then, based on the updated set, the device returns to step 102 to re-execute the coverage audit, entering the next iteration. Through the coordinated operation of the above steps, the device achieves a complete automated process from scientific hypothesis input to multi-dimensional complementary experimental scheme output, effectively solving the shortcomings of scheme homogenization and missing verification dimensions in existing technologies.

[0083] Further, in step 104, before adding a new experimental path to the current experimental path set, the device first performs rule verification on the new experimental path. Specifically, the device determines whether the new experimental path meets preset admission criteria, which include: whether the new experimental path contains readout elements or criterion elements, i.e., whether the path has measurable observation indicators or judgment criteria for determining whether a hypothesis is valid; whether the new experimental path is bound to a reference clue, i.e., whether the path is anchored to source literature with normalized identification; and whether the method family of the new experimental path overlaps with that of existing paths in the current experimental path set, where a method family refers to a set of experimental methods with the same core verification logic. The device determines whether the candidate path and the existing path belong to the same method family by comparing the method type, verification logic, and core operations used by the candidate path and the existing path. When a new experimental path meets all of the above admission criteria, the device confirms it as a valid new path; if it does not meet the criteria, the device refuses to merge the path into the path set.

[0084] In response to the judgment result that the admission conditions are met, the device incorporates the new experimental path into the current experimental path set, obtaining an updated path set. Using this updated path set as a basis, the device triggers the next iteration, i.e., returns to step 102 to re-execute the coverage audit, and begins a new round of coverage status assessment, query generation, and path synthesis. This iterative process continues, with the device performing a coverage audit on the current path set in each iteration to check whether all verification sub-objectives have been sufficiently covered. When the device detects that multiple verification sub-objectives have been covered, it stops iterating and determines the final path set as the experimental verification scheme output.

[0085] Through the aforementioned rule verification and iteration mechanisms, the device ensures that each new path incorporated into the path set possesses executability, traceability, and methodological independence. This effectively prevents invalid solutions lacking experimental details, literature support, or homogeneity with existing paths from entering the set, guaranteeing that each path in the final experimental verification scheme has independent verification value and complete experimental elements. Simultaneously, the continuous iteration mechanism ensures the completeness of the scheme set in the verification dimension. The device only stops iterating after all sub-objectives are covered, avoiding omissions in verification dimensions due to premature termination. These verification and iteration mechanisms collectively ensure that the final output experimental scheme set not only meets the standards of executability, verifiability, and traceability in terms of quality, but also achieves comprehensive verification of scientific hypotheses from multiple dimensions in terms of coverage completeness.

[0086] The following is combined Figure 2 The method provided in the embodiments of the present invention will be described in detail. Figure 2This is a schematic diagram of the overall process of the multi-objective driven intelligent design method for scientific experiments provided by the present invention. The method includes three major stages: verification sub-objective identification, closed-loop iterative optimization, and multi-scheme output and evaluation.

[0087] In the sub-target identification stage, the device first acquires the scientific hypothesis to be verified, along with its corresponding research questions and constraints. This scientific hypothesis is a core scientific proposition input by the user that needs to be verified experimentally, such as an immune cell state driving neuroinflammation or a gene mutation causing abnormal synaptic plasticity. Constraints include limitations on experimental subjects, resources, time, or ethical review requirements, used for scheme selection and adaptation in the subsequent pathway synthesis stage. After acquiring the hypothesis, the device performs semantic parsing to identify multiple verification sub-targets implicit in the scientific verification process. Verification sub-targets are different types of testing dimensions in the scientific verification process, including but not limited to correlation verification, causal intervention, effect quantification, control exclusion, cross-condition generalization, mechanism pathway characterization, and reversibility assessment. For example, for the hypothesis that an immune cell state drives neuroinflammation, its implicit verification sub-targets include at least the correlation between the immune cell state and neuroinflammation (correlation verification), the causal driving effect of the state on inflammation (causal intervention), the quantification magnitude of the inflammatory response (effect measurement), the presence of confounding factors (control exclusion), and the robustness of the mechanism across different brain regions or strains (cross-condition generalization). The device leverages the semantic understanding capabilities of pre-built scientific validation ontology or large language models to automatically identify the set of sub-targets that should be covered based on the specific content of the hypothesis. Initially, the current experimental path set can be empty or contain initial schemes provided by the user.

[0088] After completing the identification of the verified sub-targets, the device enters the closed-loop iterative optimization stage. This stage is carried out in multiple iterations, with each iteration executing three core steps in sequence: sub-target coverage audit, gap-driven iterative retrieval, and path synthesis.

[0089] In the sub-target coverage audit phase, the device assesses the coverage status of the current set of experimental paths for each validation sub-target, identifying uncovered sub-targets. Specifically, the device invokes a large language model to perform a coverage audit on the current set of experimental paths. This large language model, acting as a methodological meta-analyst, receives the entire current set of paths, the original research question, and the scientific hypothesis as input. It performs semantic analysis on each existing path to determine the type of validation sub-target it responds to, and based on this, determines the coverage status of the current set of experimental paths for all validation sub-targets. Based on this coverage status, the large language model diagnoses the content still not covered by the current set of experimental paths from multiple assessment levels, thereby identifying uncovered sub-targets. These assessment levels include at least one of the following: mechanism dimension, analysis level, untested hypothesis, and method family. Mechanism dimension refers to the type of biological or physical mechanism involved in the experimental protocol, such as signaling pathways, molecular interactions, cell behavior, and circuit connections; analysis level refers to the organizational level of the research object involved in the experimental protocol, such as molecular level, cellular level, tissue level, and individual level; untested hypothesis refers to intermediate propositions or auxiliary hypotheses implicit under scientific hypotheses that have not yet been experimentally tested in the current set of experimental pathways; method family refers to a set of experimental methods with the same core verification logic. For example, optogenetics, chemogenetics, and pharmacological interventions belong to the causal intervention method family, while functional magnetic resonance imaging and two-photon calcium imaging belong to the functional readout method family.

[0090] After diagnosing the uncovered content, the device outputs the uncovered sub-targets as a gap report in natural language. When generating the gap report, the device explicitly requires the large language model not to provide search queries or keywords, but only to output the diagnosis of the gap, and not to paste the gap report generated in the previous round, ensuring that each round's report is generated independently for the current path set. Specifically, the device first obtains the path summary of the current experimental path set, inputs the historical gap report generated in the previous iteration as context into the large language model, and updates the content of the historical gap report based on the uncovered sub-targets identified in the current round: removing the validation sub-targets already covered by the current experimental path set from the historical gap report, and adding the newly identified uncovered sub-targets in this round, thus generating the gap report for the current round. This gap report is a natural language diagnostic text, approximately 120 to 200 words, used to characterize the missing state of the experimental scheme in the validation dimension, specifically including content that is still uncovered, such as unresponsive mechanism dimensions, uninvolved analysis levels, untested hypotheses, and method families that have not yet appeared.

[0091] In the gap-driven iterative retrieval phase, the device generates corresponding retrieval queries based on the uncovered sub-targets identified in the coverage audit phase, and performs searches in the literature database to extract reference clues related to the uncovered sub-targets. Specifically, the device first reads the gap summary and generates an exploratory query for exploring new method families for the uncovered sub-targets in the gap summary. This exploratory query is limited to single-line plain text, no more than 320 characters, and uses only searchable scientific terms. Simultaneously, for the verification sub-targets already covered by the current path set, the device generates a deepening query to supplement the strength of evidence within existing method families. The device outputs both the exploratory and deepening queries as retrieval queries and performs a search in the literature database. During the generation of exploratory and deepening queries, the device extracts historical query text and uses it as an anti-duplication constraint to prevent repeated searches of already explored literature domains. The device extracts reference clues related to the uncovered sub-targets from the retrieved literature. Specifically, the device obtains literature with normalized identifiers from the literature database, extracts local content related to the uncovered sub-targets from these literatures, and extracts the local content along with the normalized identifiers of the corresponding literature to generate reference clues, ensuring that each reference clue is bound to its source literature. The device establishes an alias mapping during the extraction process, matching the abbreviations used in the document within each step with their global identifiers, ensuring the integrity of the tracing chain.

[0092] In the path synthesis stage, the device synthesizes new experimental paths for uncovered sub-targets based on reference cues extracted in the retrieval stage. Specifically, the device integrates reference cues and orthogonality rules to construct path generation instructions. The orthogonality rules stipulate that candidate paths and existing paths in the current experimental path set should use non-overlapping subsets of cues, and that verification logic orthogonal to existing path families should be used between paths. If a new cue cannot support orthogonal paths, an empty list is returned, and existing paths must not be restated or slightly renamed. The device initially generates candidate paths based on these path generation instructions. After initially generating candidate paths, the device extracts elements from the generated text content, constructing the candidate paths into a structured data format containing intervention elements, readout elements, and criterion elements.

[0093] The device adds the confirmed new experimental path to the current experimental path set. Before adding it, it first performs rule validation on the new experimental path to determine whether it meets the admission criteria of containing readout elements or criterion elements, being bound to reference clues, and not overlapping with the method family of existing paths. When the new experimental path meets all of the above admission criteria, the device formally merges it into the current experimental path set, obtaining an updated path set. Based on this updated set, it triggers the next iteration, returning to the sub-target coverage audit stage for re-execution. This iterative process continues until the device detects that all verification sub-targets have been covered, at which point it stops iterating and enters the multi-scheme output and evaluation stage, determining the final path set as the experimental verification scheme output.

[0094] Through the aforementioned closed-loop iterative optimization mechanism, the device ensures that each new path incorporated into the path set is executable, traceable, and methodologically independent, effectively preventing invalid solutions lacking experimental details, literature support, or homogeneous with existing paths from entering the set. The continuous iteration mechanism ensures the completeness of the solution set in the validation dimension; the device only stops iterating after all sub-objectives have been covered, avoiding omissions in validation dimensions due to premature termination. The final output set of experimental solutions is complementary in the validation dimension and naturally differentiated in terms of methodology, effectively avoiding the problem of stacking homogeneous solutions within the same paradigm. This enables a complete and rigorous examination of scientific hypotheses from multiple dimensions, significantly reducing the trial-and-error costs for researchers in the experimental design phase.

[0095] To more clearly demonstrate the technical effects of the multi-objective driven closed-loop iterative optimization mechanism of this invention in practical applications, a detailed explanation is provided below using a specific scientific experimental design application scenario.

[0096] In a specific application embodiment, the scientific hypothesis received by the system of this invention relates to the study of Fibulin-2 (Fbln2) and the blood-brain barrier (BBB). The scientific hypothesis to be verified acquired by the device is: "The brain vascular tree realizes a hierarchical, multi-scale control architecture, in which the structural heterogeneity of the perivascular matrix acts as a distributed regulatory layer—encoding spatial information and regulating the response of local endothelial cells to systemic and neural signals." The corresponding research question is: "Does the regional heterogeneity of Fibulin-2 distribution in the extracellular matrix of brain blood vessels shape the spatial heterogeneity of blood-brain barrier integrity?" The device analyzes this complex hypothesis and identifies several implicit verification sub-goals, including at least: verification of the correlation between the spatial distribution of Fbln2 and BBB permeability, verification of the causal intervention of perturbing Fbln2 to change BBB heterogeneity, verification of the mechanism pathway of Fbln2 regulating endothelial response, verification of cross-scale generalization from nanoscale to macroscale, and verification of the reconfigurability of Fbln2 gradient reconstructed in vitro.

[0097] After the first round of iterations, the device generated in vivo validation pathways based on gene ablation and pathways based on multi-scale correlation imaging. In the second round of iterations, the device's coverage audit module output a gap report, indicating a mechanistic coverage deficiency in the current pathway set: "Severely focused on genetic and pharmacological interventions, but neglecting nanoscale biophysical mechanisms and direct mechanotransduction pathways." Based on this gap, the device simultaneously generated two types of queries: One type of query is the exploratory query, used to explore new method families. The search text is: "AFM nanomechanical mapping of vascular basement membrane rheology, mechanotransduction between elastin network and cytoskeleton, in situ proteomics analysis of Fbln2 and fibronectin in brain microvascular niches." Based on the reference clues obtained from this query, the instrument successfully synthesized a new experimental pathway belonging to a completely new method family: interactionome characterization based on in situ nearest neighbor-dependent cell surface proteomics (filling the sub-goals of mechanism pathway and cross-scale generalization validation).

[0098] Another type is the in-depth query, used to supplement the strength of evidence within the existing method family. The search text is: "Measurable readouts, validation benchmarks, and failure modes for 3D-related light sheet imaging of dextran tracer leakage, as well as experimental controls and sample sizes for conditional gene ablation of the extracellular matrix in the blood-brain barrier model." Based on this query, the device supplements the existing light sheet imaging pathway in the first round with quantitative benchmarks such as optical calibration, transforming it into an executable scheme with rigorous indicators.

[0099] In the third iteration, the device's coverage audit module updated the gap report again, pointing out that "there is still a lack of in vivo biomechanical imaging for mapping macroscopic stiffness gradients in living tissues, and a lack of computational multi-scale modeling." The device generated a new round of exploratory queries, precisely locating "magnetic resonance elastography (MRE) mapping macroscopic brain stiffness gradients" and "multi-scale computational modeling of blood-brain barrier permeability," thereby recalling literature clues in the fields of macroscopic imaging and computational biology, and synthesizing corresponding new experimental pathways; at the same time, the shear stress control and hydrogel gradient verification parameters in the existing microfluidic chip pathways were further refined through in-depth queries.

[0100] As can be seen from the above embodiments, this invention ensures the breadth and completeness of sub-target coverage through exploratory queries, while ensuring the depth of evidence for covered sub-targets (supplementing measurable and verifiable experimental benchmarks and failure modes) through in-depth queries. These two aspects constitute a collaborative mechanism within sub-targets, enabling the system to systematically generate a set of high-quality experimental verification schemes with mutually orthogonal methodologies and mutually reinforcing evidence chains around multiple verification sub-targets under a single hypothesis. This overcomes the technical deficiency of existing large language models that generate homogeneous models within a single paradigm.

[0101] The following describes the multi-objective driven intelligent design system for multiple scientific experiments provided by the embodiments of the present invention. The multi-objective driven intelligent design system for multiple scientific experiments described below can be referred to in correspondence with the multi-objective driven intelligent design method for multiple scientific experiments described above.

[0102] This invention provides a multi-objective driven intelligent design system for multiple schemes in scientific experiments. See [link to relevant documentation]. Figure 3 ,include: Assume that the acquisition module 310 is used to acquire the scientific hypothesis to be verified, wherein the scientific hypothesis implicitly contains multiple verification sub-objectives that need to be verified separately; The coverage audit module 320 is used to evaluate the coverage status of the current experimental path set on the multiple verification sub-targets, and identify the sub-targets that are not covered by the current experimental path set from the multiple verification sub-targets as uncovered sub-targets; The retrieval and extraction module 330 is used to generate a corresponding retrieval query based on the uncovered sub-target, perform a retrieval in the literature database, and extract reference clues related to the uncovered sub-target. The path synthesis and iteration module 340 is used to synthesize a new experimental path for the uncovered sub-target based on the reference clues, and add the new experimental path to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logics.

[0103] Furthermore, the coverage audit module 320 is used to: call the large language model to perform coverage audit on the current experimental path set, and determine the coverage status of the current experimental path set on the multiple verification sub-targets; Based on the coverage status, the content that is still not covered by the current set of experimental paths is diagnosed from multiple evaluation levels, and the uncovered sub-targets are identified; wherein, the evaluation level includes at least one of mechanism dimension, analysis level, untested hypothesis and method family; The uncovered sub-targets are output as gap reports in natural language, which are used to characterize the missing state of the experimental scheme in the validation dimension.

[0104] Furthermore, the retrieval and extraction module 330 is used to: read the gap report and generate an exploration query for exploring new method families for the uncovered sub-targets in the gap report; For the verification sub-objectives that are already covered by the current experimental path set among the multiple verification sub-objectives, generate a deeper query to supplement the evidence strength within the existing method family; The exploratory query and the in-depth query are combined as the output of the retrieval query.

[0105] Furthermore, the path synthesis and iteration module 340 is used to: integrate the reference clues and orthogonality rules to construct path generation instructions, wherein the orthogonality rules limit the candidate paths and the existing paths in the current experimental path set to use non-overlapping subsets of clues; The candidate paths are initially generated based on the path generation instructions; The large language model is invoked to determine the independence of the candidate paths, and the decision is output as to retain them as new paths, merge them into the existing paths, or reject them. In response to the decision to retain the candidate path as a new path, the candidate path is confirmed as the new experimental path.

[0106] Furthermore, the retrieval and extraction module 330 is used to: obtain documents with standardized identifiers from the document database; extract local content related to the uncovered sub-target from the documents; and extract the local content and the standardized identifier of the document to which it belongs to generate the reference clue.

[0107] Furthermore, the path synthesis and iteration module 340 is used to: perform rule verification on the new experimental path, and determine whether the new experimental path meets the admission conditions of containing readout elements or criterion elements, being bound to the reference clues, and not overlapping with the method family of existing paths in the current experimental path set. In response to the judgment result that the admission conditions are met, the new experimental path is incorporated into the current experimental path set to obtain an updated path set; The updated path set is used as the basis to trigger the next round of iteration, and the process continues to update until all of the multiple verification sub-targets are detected to be covered. The final path set is then determined as the output of the experimental verification scheme.

[0108] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions from the memory 830 to execute a multi-objective-driven intelligent design method for scientific experiments with multiple schemes.

[0109] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the multi-objective driven intelligent design method for multiple schemes of scientific experiments provided by the above methods.

[0111] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-objective driven intelligent design method for multiple schemes of scientific experiments provided by the above methods.

[0112] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown 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 this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0113] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective driven intelligent design method for multiple schemes in scientific experiments, characterized in that, include: Obtain scientific hypotheses to be verified, which implicitly contain multiple verification sub-goals that need to be verified separately; Evaluate the coverage status of the current set of experimental paths to the plurality of verification sub-targets, and identify the sub-targets that are not covered by the current set of experimental paths as uncovered sub-targets from the plurality of verification sub-targets; Based on the uncovered sub-targets, a corresponding search query is generated, and a search is performed in the literature database to extract reference clues related to the uncovered sub-targets. Based on the reference clues, a new experimental path is synthesized for the uncovered sub-target, and the new experimental path is added to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logics.

2. The method according to claim 1, characterized in that, The step of evaluating the coverage status of the current experimental path set for the multiple verification sub-targets, and identifying sub-targets not covered by the current experimental path set as uncovered sub-targets, includes: The large language model is invoked to perform coverage audit on the current set of experimental paths to determine the coverage status of the current set of experimental paths for the multiple verification sub-targets; Based on the coverage status, the content that is still not covered by the current set of experimental paths is diagnosed from multiple evaluation levels, and the uncovered sub-targets are identified; wherein, the evaluation level includes at least one of mechanism dimension, analysis level, untested hypothesis and method family; The uncovered sub-targets are output as gap reports in natural language, which are used to characterize the missing state of the experimental scheme in the validation dimension.

3. The method according to claim 2, characterized in that, The step of outputting the uncovered sub-targets as a gap report in natural language includes: Obtain a path summary of the current set of experimental paths; The historical gap report generated in the previous iteration is passed in as context, and combined with the currently identified uncovered sub-targets, the verification sub-targets that have been covered by the current experimental path set are removed from the historical gap report; Add the newly identified uncovered sub-targets to generate the gap report for the current round.

4. The method according to claim 2, characterized in that, Generate corresponding search queries based on the uncovered sub-targets, including: Read the gap report and generate an exploration query for exploring new method families for the uncovered sub-targets in the gap report; For the verification sub-objectives that are already covered by the current experimental path set among the multiple verification sub-objectives, generate a deeper query to supplement the evidence strength within the existing method family; The exploratory query and the in-depth query are combined as the output of the retrieval query.

5. The method according to claim 4, characterized in that, The steps of generating the exploratory query and the detailed query also include: Extract historical query text and use it as an anti-duplicate constraint; Based on the anti-duplication constraint, the exploratory query and the in-depth query are generated simultaneously to prevent repeated searches of already explored literature domains.

6. The method according to claim 1, characterized in that, The synthesis of a new experimental path for the uncovered sub-target based on the reference clues includes: The path generation instructions are constructed by integrating the reference clues and orthogonality rules, wherein the orthogonality rules limit the candidate paths to using a non-overlapping subset of clues with the existing paths in the current experimental path set; The candidate paths are initially generated based on the path generation instructions; The large language model is invoked to determine the independence of the candidate paths, and the decision is output as to retain them as new paths, merge them into the existing paths, or reject them. In response to the decision to retain the candidate path as a new path, the candidate path is confirmed as the new experimental path.

7. The method according to claim 6, characterized in that, After initially generating the candidate paths according to the path generation instructions, the process further includes: The generated text content is subjected to element extraction, and the candidate path is constructed into a structured data format containing intervention elements, readout elements and criterion elements; Specifically, the operations or processes applied to the experimental subjects are extracted and configured as the intervention elements in the structured data format; Extract measurable observation indicators and configure them as readout elements in the structured data format; Extract the standard content used to determine whether the scientific hypothesis is valid, and configure it as the criterion element in the structured data format, so that the candidate path is transformed into an executable and verifiable standardized experimental plan.

8. The method according to claim 1, characterized in that, The extraction of reference clues related to the uncovered sub-target specifically includes: Retrieve documents with standardized identifiers from the aforementioned document database; Extract local content related to the uncovered sub-target from the document; The reference clues are generated by extracting the local content and the standardized identifiers of the relevant documents.

9. The method according to claim 1, characterized in that, The process of adding the new experimental path to the current experimental path set for multiple iterations to produce an experimental verification scheme containing multiple complementary paths using different verification logic includes: Perform rule validation on the new experimental path to determine whether the new experimental path meets the admission criteria of containing readout elements or criterion elements, being bound to the reference clue, and not overlapping with the method family of existing paths in the current experimental path set. In response to the judgment result that the admission conditions are met, the new experimental path is incorporated into the current experimental path set to obtain an updated path set; The updated path set is used as the basis to trigger the next round of iteration, and the process continues to update until all of the multiple verification sub-targets are detected to be covered. The final path set is then determined as the output of the experimental verification scheme.

10. A multi-objective driven intelligent design system for multiple schemes in scientific experiments, characterized in that, include: The hypothesis acquisition module is used to acquire scientific hypotheses to be verified, which implicitly contain multiple verification sub-objectives that need to be verified separately. The coverage audit module is used to evaluate the coverage status of the current experimental path set on the multiple verification sub-targets, and identify the sub-targets that are not covered by the current experimental path set as uncovered sub-targets from the multiple verification sub-targets; The retrieval and extraction module is used to generate a corresponding retrieval query based on the uncovered sub-target, perform a retrieval in the literature database, and extract reference clues related to the uncovered sub-target. The path synthesis and iteration module is used to synthesize new experimental paths for the uncovered sub-targets based on the reference clues, and add the new experimental paths to the current experimental path set for multiple rounds of iteration to produce an experimental verification scheme containing multiple complementary paths using different verification logics.

11. The system according to claim 10, characterized in that, The coverage audit module is used to: call the large language model to perform coverage audit on the current experimental path set, and determine the coverage status of the current experimental path set on the multiple verification sub-targets; Based on the coverage status, the content that is still not covered by the current set of experimental paths is diagnosed from multiple evaluation levels, and the uncovered sub-targets are identified; wherein, the evaluation level includes at least one of mechanism dimension, analysis level, untested hypothesis and method family; The uncovered sub-targets are output as gap reports in natural language, which are used to characterize the missing state of the experimental scheme in the validation dimension.

12. The system according to claim 11, characterized in that, The retrieval and extraction module is used to: read the gap report and generate an exploration query for exploring new method families for the uncovered sub-targets in the gap report; For the verification sub-objectives that are already covered by the current experimental path set among the multiple verification sub-objectives, generate a deeper query to supplement the evidence strength within the existing method family; The exploratory query and the in-depth query are combined as the output of the retrieval query.

13. The system according to claim 10, characterized in that, The path synthesis and iteration module is used to: integrate the reference clues and orthogonality rules to construct path generation instructions, wherein the orthogonality rules limit the candidate paths to using non-overlapping subsets of clues with the existing paths in the current experimental path set; The candidate paths are initially generated based on the path generation instructions; The large language model is invoked to determine the independence of the candidate paths, and the decision is output as to retain them as new paths, merge them into the existing paths, or reject them. In response to the decision to retain the candidate path as a new path, the candidate path is confirmed as the new experimental path.

14. The system according to claim 10, characterized in that, The retrieval and extraction module is used to: retrieve documents with standardized identifiers from the document database; Extract local content related to the uncovered sub-target from the document; The reference clues are generated by extracting the local content and the standardized identifiers of the relevant documents.

15. The system according to claim 10, characterized in that, The path synthesis and iteration module is used to: perform rule verification on the new experimental path, and determine whether the new experimental path meets the admission conditions of containing readout elements or criterion elements, being bound to the reference clues, and not overlapping with the method family of existing paths in the current experimental path set. In response to the judgment result that the admission conditions are met, the new experimental path is incorporated into the current experimental path set to obtain an updated path set; The updated path set is used as the basis to trigger the next round of iteration, and the process continues to update until all of the multiple verification sub-targets are detected to be covered. The final path set is then determined as the output of the experimental verification scheme.

Images