Knowledge-based intelligent chemical discovery methods, systems, devices, and media
By combining domain knowledge graphs, large language models, and robot experimental feedback, a closed loop of scientific hypothesis generation and experimental verification is formed, solving the problem of cross-domain knowledge integration and feedback reasoning closed loop in existing systems, and realizing deep chemical discovery.
Patent Information
- Application Number
- CN202611368292.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-09-04
AI Technical Summary
Existing on-demand autonomous research systems lack cross-disciplinary knowledge integration and feedback reasoning loops in deep scientific discovery, making it difficult to implement scientific hypotheses and efficiently integrate cross-disciplinary knowledge.
By combining domain knowledge graphs, large language model reasoning, and robot experiment feedback, a closed loop of scientific hypothesis generation, experimental verification, and feedback iteration is formed, including literature conversion into knowledge graphs, cross-domain knowledge reasoning, multi-round human-machine collaborative question answering, robot experiment execution, and experiment feedback optimization.
It achieves efficient integration of cross-domain knowledge and a closed loop of feedback reasoning, which can promote the in-depth discovery of complex chemical systems by on-demand autonomous chemical research systems.
Smart Images

Figure CN122867868B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of chemical discovery technology, and in particular to a knowledge-oriented intelligent chemical discovery method, system, device, and medium. Background Technology
[0002] Current large language model-driven hierarchical multi-agent architectures, such as the ChemAgents system architecture proposed by Song et al. in 2025, and a multi-agent driven robotic AI chemist system disclosed in Chinese patent ZL202411488338.6, can automatically transform research requirements in natural language form into experimental procedures, robot execution, and computational analysis, providing a general system foundation for on-demand autonomous chemical research. However, for more complex scientific discovery tasks, simply generating experimental procedures and robot code based on task descriptions is still insufficient. Many chemical discovery problems are not simply about executing existing experiments, but require extracting domain knowledge from a large amount of scattered literature, discovering potential connections between cross-domain knowledge, further forming experimentally verifiable scientific hypotheses, and continuously refining experimental designs based on feedback from real experiments. Relying solely on a base large language model for open reasoning is easily limited by problems such as opaque knowledge sources, untraceable reasoning basis, and difficulty in implementing scientific hypotheses; relying solely on traditional literature retrieval and human experience makes it difficult to efficiently integrate cross-domain knowledge and form a feedback reasoning closed loop.
[0003] Therefore, in order to address the problem that existing on-demand autonomous scientific research systems lack cross-domain knowledge integration and feedback reasoning loops in deep scientific discovery, it is necessary to provide a method and system for automating chemical discovery tasks.
[0004] In view of this, the present invention is hereby proposed. Summary of the Invention
[0005] The purpose of this invention is to provide a knowledge-oriented intelligent chemical discovery method, system, device, and medium that can combine domain knowledge graphs, large language model reasoning, and robot experimental feedback to form an intelligent chemical discovery system with a closed loop of scientific hypothesis generation, experimental verification, and feedback iteration. This solves the problem that existing on-demand autonomous scientific research systems lack cross-domain knowledge integration and feedback reasoning closed loops in deep scientific discovery.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A knowledge-oriented intelligent chemical discovery method for achieving target chemical discovery tasks includes:
[0008] Step 1: Obtain literature from multiple fields related to the target chemical discovery task, and transform the literature from each field into a corresponding domain knowledge graph;
[0009] Step 2: Connect the knowledge graphs of each domain to the corresponding domain literature reading agents. The domain literature reading agents can perform retrieval enhancement and generation based on the corresponding domain knowledge graphs. The cross-domain knowledge reasoning agent calls the literature reading agents of each domain according to the target chemical discovery task input by the researchers. It compares, summarizes and merges the retrieval enhancement answers returned by the literature reading agents of each domain, and conducts multiple rounds of human-computer collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs and initial experimental candidate sets.
[0010] Step 3: Input the current set of experimental candidates into the on-demand autonomous chemistry research system, convert it into a robot autonomous experimental task and execute it to obtain real experimental data;
[0011] Step 4: The researchers determine whether they have obtained experimental candidates that meet the expectations or whether the total experimental budget has been exhausted. If yes, the iteration stops; otherwise, proceed to step 5.
[0012] Step 5: Incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The agent combines scientific hypotheses, experimental design, historical experimental feedback set, and cross-domain literature knowledge to generate the next round of experimental candidate set and returns to step 3.
[0013] A knowledge-guided intelligent chemical discovery system for implementing the method described in this invention includes:
[0014] The knowledge graph construction module can acquire literature from multiple fields related to the target chemical discovery task and transform it into a corresponding domain knowledge graph.
[0015] The scheme generation module can connect knowledge graphs from various fields to the corresponding literature reading agents, enabling them to perform enhanced retrieval and generation based on the corresponding graphs. The cross-domain knowledge reasoning agent calls on literature reading agents from various fields according to the target chemical discovery task input by the researchers, integrates the enhanced retrieval answers returned by the literature reading agents from various fields, and conducts multiple rounds of collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs, and an initial set of experimental candidates.
[0016] The results acquisition module can input the current set of experimental candidates into the on-demand autonomous chemistry research system, transform it into a robot autonomous experimental task and execute it to obtain real experimental data.
[0017] The judgment module can obtain the researchers' judgment on whether the experimental candidates meet the expectations or whether the total experimental budget has been exhausted; if yes, the iteration stops; if no, the feedback optimization module is started.
[0018] The feedback optimization module can incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental design, historical feedback and cross-domain literature knowledge to generate the next round of experimental candidate set and send it to the result acquisition module.
[0019] A processing apparatus, comprising:
[0020] At least one memory for storing one or more programs;
[0021] At least one processor is capable of executing one or more programs stored in the memory, such that when the processor executes one or more programs, the processor can implement the method of the present invention.
[0022] A readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the methods described in this invention.
[0023] Compared with existing technologies, the knowledge-oriented intelligent chemical discovery method, system, device, and medium provided by this invention have the following advantages:
[0024] By effectively combining literature knowledge from multiple fields, scientific hypothesis generation, candidate experiment design, robotic experiment execution, and feedback reasoning optimization, a closed-loop discovery framework of "knowledge fusion—hypothesis generation—experimental verification—feedback iteration" is constructed, which can promote the on-demand autonomous chemical research system to further extend from the execution of experimental tasks to the in-depth chemical discovery of complex chemical systems. Attached Figure Description
[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 A flowchart is provided for an embodiment of the knowledge-oriented intelligent chemical discovery method of the present invention.
[0027] Figure 2 This provides an overall flowchart of a knowledge-oriented intelligent chemical discovery method for embodiments of the present invention.
[0028] Figure 3 This diagram illustrates the candidate experiments and performance test results in six iterations of the photocatalytic hydrogen production material discovery mission in the embodiments of the present invention.
[0029] Figure 4A block diagram of a knowledge-oriented intelligent chemical discovery system is provided for embodiments of the present invention. Detailed Implementation
[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them, and do not constitute a limitation on the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.
[0031] First, the following explanations are provided for the terms that may be used in this article:
[0032] The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".
[0033] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.
[0034] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.
[0035] Unless otherwise explicitly specified or limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this document according to the specific circumstances.
[0036] The terms “center,” “longitudinal,” “lateral,” “length,” “width,” “thickness,” “up,” “down,” “front,” “back,” “left,” “right,” “vertical,” “horizontal,” “top,” “bottom,” “inner,” “outer,” “clockwise,” and “counterclockwise” indicate the current orientation or positional relationship, and are only for the convenience and simplification of description, and do not explicitly or implicitly suggest that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this document.
[0037] The technical solution provided by this invention will be described in detail below. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they shall be performed according to conventional conditions in the art or conditions recommended by the manufacturer. Reagents or instruments used in the embodiments of this invention whose manufacturers are not specified are all conventional products that can be purchased commercially.
[0038] like Figure 1 , Figure 2 As shown, this invention provides a knowledge-oriented intelligent chemical discovery method for achieving a target chemical discovery task, comprising:
[0039] Step 1: Obtain literature from multiple fields related to the target chemical discovery task, and transform the literature from each field into a corresponding domain knowledge graph;
[0040] Step 2: Connect the knowledge graphs of each domain to the corresponding domain literature reading agents. The domain literature reading agents can perform retrieval enhancement and generation based on the corresponding domain knowledge graphs. The cross-domain knowledge reasoning agent calls the literature reading agents of each domain according to the target chemical discovery task input by the researchers. It compares, summarizes and merges the retrieval enhancement answers returned by the literature reading agents of each domain, and conducts multiple rounds of human-computer collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs and initial experimental candidate sets.
[0041] Step 3: Input the current set of experimental candidates into the on-demand autonomous chemistry research system, convert it into a robot autonomous experimental task and execute it to obtain real experimental data;
[0042] Step 4: The researchers determine whether they have obtained experimental candidates that meet the expectations or whether the total experimental budget has been exhausted. If yes, the iteration stops; otherwise, proceed to step 5.
[0043] Step 5: Incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental designs, historical experimental feedback sets, and cross-domain literature knowledge to generate the next round of experimental candidate sets and returns to step 3.
[0044] Preferably, in step 1 of the above method, literature from multiple fields related to the target chemical discovery task is obtained in the following manner, including:
[0045] Obtain literature from at least two complementary knowledge areas relevant to the target chemical discovery mission.
[0046] Preferably, in step 1 of the above method, when the target chemical discovery task is the discovery of high-performance photocatalytic hydrogen production materials, the literature obtained in multiple fields related to the target chemical discovery task includes literature in the field of photocatalytic hydrogen production, literature in the field of hydrogenase, and literature in the field of enzyme-mimicking catalysis.
[0047] Preferably, in step 1 of the above method, after processing the acquired literature from multiple fields in the following manner, the processed literature from each field is transformed into a corresponding domain knowledge graph, including:
[0048] Document parsing tools were used to convert the literature from multiple fields into plain text format. The text was then pre-cleaned, retaining the parsable chapter titles, figure and table descriptions, and main text structure. Content irrelevant to the target chemical discovery task was then removed, and the content relevant to the target chemical discovery task was retained to obtain the pre-processed literature content.
[0049] Using a large language model, domain-specific knowledge distillation is performed on the preprocessed content of each document based on domain-specific distillation prompts. After knowledge distillation, distilled texts of each document suitable for constructing a knowledge graph are obtained.
[0050] Distilled texts of each document in the same field are compiled into a distilled literature set for that field, and then transformed into a domain knowledge graph.
[0051] Preferably, in step 1 of the above method, the distilled texts of each document in the same field are grouped into a distilled document set for that field in the following manner, and then transformed into a domain knowledge graph, including:
[0052] Let D be the collection of distillation literature in a certain field. Divide this collection of distillation literature D into multiple text segments using the following expression:
[0053] (1);
[0054] in, Let represent the i-th text segment, and n represent the total number of text segments. This represents the last text segment, where w represents the window size of the text segment; Segment() is a function that segments the distillation literature collection into blocks according to the window size.
[0055] Based on large language model prompts and graph construction tools, candidate entities and candidate relations are extracted from each text segment to form a set of triples. The expression is:
[0056] (2);
[0057] Where s represents the subject entity; r represents the relationship between entities; o represents the object entity; R_LLM represents the entity relationship recognition function based on the prompt words of the large language model. This entity relationship recognition function simultaneously identifies entities and their corresponding relationships within each text segment, so that the extracted relationship edges have clear starting entities and target entities, corresponding to the subject entity s and the object entity o in the triple, respectively.
[0058] After obtaining the set of triples for each text segment, the triples from all text segments are summarized to obtain the candidate triple set for that domain:
[0059] (3);
[0060] Where T represents the set of candidate triples extracted from all distilled texts in the domain; Distinguishing the union symbol;
[0061] Based on the set of candidate entity nodes V extracted from all triples in the candidate triple set and the set of candidate relation edges E composed of triple relations, construct a candidate knowledge graph before deduplication and merging. :
[0062] (4);
[0063] in, This represents the candidate knowledge graph before deduplication and merging; Graph() represents the operation of generating a graph structure from the candidate triple set; V represents the set of candidate entity nodes extracted from all triples in the candidate triple set; E represents the set of candidate relation edges composed of triple relations.
[0064] Generate text descriptions for the extracted entities and relations, represented as follows:
[0065] P(V, E ) = {(k, v) | k ∈ labels(V, E ), v = M_sum(D[k])} (5);
[0066] Where P represents the entity and relation description generation function; k represents the label of the entity node or relation edge; labels(V, E) represents the label set obtained from the candidate entity node set V and the candidate relation edge set E; D[k] represents the text fragment or context sentence related to label k; M_sum() represents the large language model summarization function; v represents the compact semantic description generated by the large language model;
[0067] The domain knowledge graph G is constructed by deduplicating, merging, and normalizing duplicate entities, synonyms, and similar relationships that appear repeatedly in different text fragments.
[0068] (6);
[0069] Where G represents the domain knowledge graph; and represents the set of entity nodes and the set of relation edges after deduplication and merging, respectively; P(V, E) represents the set of entity-relation semantic descriptions obtained after describing the candidate set of entity nodes and the candidate set of relation edges; Dedupe() represents the entity and relation merging function, which is used to merge duplicate or similar nodes and edges across text segments based on entity labels, relation labels and their semantic descriptions.
[0070] Preferably, in step 2 of the above method, the cross-domain knowledge reasoning agent interacts with researchers in the following manner:
[0071] The cross-domain knowledge reasoning agent receives open scientific questions or candidate research directions corresponding to the target chemical discovery task input by researchers, generates a domain decomposition query plan based on the knowledge scope involved in the open scientific questions or candidate research directions, and determines the domain literature reading agent to be invoked based on the domain decomposition query plan.
[0072] The cross-domain knowledge reasoning agent distributes the domain sub-problems in the domain decomposition query plan to the corresponding domain literature reading agents. It receives the corresponding domain retrieval enhancement answers returned by each domain literature reading agent after performing retrieval enhancement on the domain sub-problems based on the corresponding domain knowledge graph. The retrieval enhancement answers include relevant graph context, key relationship paths and reasoning basis.
[0073] The cross-domain knowledge reasoning agent compares, summarizes, and integrates the enhanced search answers for the corresponding domains returned by the document reading agents of different domains to form a comprehensive scientific answer for researchers.
[0074] After researchers confirm the research directions in the comprehensive scientific answers based on the target chemical discovery task, the cross-domain knowledge reasoning agent and researchers enter the next round of interaction, and gradually generate candidate research directions and scientific hypotheses through multiple rounds of human-computer collaborative question and answer;
[0075] A cross-domain knowledge reasoning agent transforms scientific hypotheses into experimental designs and provides an initial set of experimental candidates based on the experimental designs.
[0076] Preferably, in step 5 of the above method, real experimental data is incorporated into the historical experimental feedback set and fed back to the cross-domain knowledge reasoning agent in the following manner. The cross-domain knowledge reasoning agent then combines scientific hypotheses, experimental designs, the historical experimental feedback set, and cross-domain literature knowledge to generate the next round of experimental candidate sets, including:
[0077] The current round of real experimental data is associated with the corresponding current experimental candidate set and incorporated into the historical experimental feedback set to obtain the updated historical experimental feedback set.
[0078] The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental designs, updated historical experimental feedback sets, and cross-domain literature knowledge to compare and analyze the candidates for each round of experiments and their actual experimental data, redetermine the basis for candidate designs and subsequent recommendation directions, and generate the next round of experimental candidate sets.
[0079] See Figure 4 The present invention also provides a knowledge-oriented intelligent chemical discovery system for implementing the above-described method, comprising:
[0080] The knowledge graph construction module can acquire literature from multiple fields related to the target chemical discovery task and transform it into a corresponding domain knowledge graph.
[0081] The scheme generation module can connect knowledge graphs from various fields to the corresponding literature reading agents, enabling them to perform enhanced retrieval and generation based on the corresponding graphs. The cross-domain knowledge reasoning agent calls on literature reading agents from various fields according to the target chemical discovery task input by the researchers, integrates the enhanced retrieval answers returned by the literature reading agents from various fields, and conducts multiple rounds of collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs, and an initial set of experimental candidates.
[0082] The results acquisition module can input the current set of experimental candidates into the on-demand autonomous chemistry research system, transform it into a robot autonomous experimental task and execute it to obtain real experimental data.
[0083] The judgment module can obtain the researchers' judgment on whether the experimental candidates meet the expectations or whether the total experimental budget has been exhausted; if yes, the iteration stops; if no, the feedback optimization module is started.
[0084] The feedback optimization module can incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental design, historical feedback and cross-domain literature knowledge to generate the next round of experimental candidate set and send it to the result acquisition module.
[0085] Specifically, the knowledge graph construction module and the solution generation module are connected, and the constructed knowledge graphs of various fields can be sent to the solution generation module.
[0086] The scheme generation module and the result acquisition module are connected, and the generated set of experimental candidates can be sent to the result acquisition module.
[0087] The judgment module and the result acquisition module are connected in communication. Based on the real experimental data obtained by the result acquisition module, the judgment module can obtain the researchers' judgment on whether the experimental candidate meets the expectations, or whether the total experimental budget has been exhausted.
[0088] The feedback optimization module is connected to both the judgment module and the result acquisition module. It can be activated when the judgment module determines the result as negative, and then infers the next round of experimental candidate set and sends it to the result acquisition module.
[0089] The present invention further provides a processing apparatus, comprising:
[0090] At least one memory for storing one or more programs;
[0091] At least one processor is capable of executing one or more programs stored in the memory, such that when the processor executes one or more programs, the processor can implement the methods described above.
[0092] Another embodiment of the present invention provides a readable storage medium storing a computer program that can implement the above-described method when executed by a processor.
[0093] To more clearly demonstrate the technical solution and its effects provided by the present invention, the following detailed description of the solution provided by the embodiments of the present invention is provided with reference to specific examples.
[0094] Example 1
[0095] like Figures 1 to 3As shown, this embodiment provides a knowledge-oriented intelligent chemical discovery method for achieving target chemical discovery tasks. This method addresses open scientific questions raised by researchers and unfolds according to a closed-loop logic of "knowledge fusion—hypothesis generation—experimental verification—feedback iteration," mainly comprising four parts: First, domain-specific literature construction and knowledge graph construction, transforming scattered unstructured literature knowledge into searchable and reasonable structured knowledge resources; second, graph-enhanced retrieval-based scientific hypothesis generation and human-machine collaborative experimental candidate design, enabling the system to integrate cross-domain knowledge and form experimentally verifiable research hypotheses and candidate schemes; third, autonomous experimental scheme design and robotic experimental verification, further transforming candidate experiments into experimental procedures executable by an automated experimental platform and obtaining real experimental data; fourth, iterative judgment and experimental feedback-driven reasoning optimization, where researchers determine whether to continue iteration; if so, the system optimizes experimental strategies based on experimental data, updates candidate designs, and guides subsequent experimental iterations.
[0096] Specifically, the method includes the following steps:
[0097] Step 1: Obtain literature from multiple fields related to the target chemical discovery task, and transform the literature from each field into a corresponding domain knowledge graph;
[0098] Step 2: Connect the knowledge graphs of each domain to the corresponding domain literature reading agents. The domain literature reading agents can perform retrieval enhancement and generation based on the corresponding domain knowledge graphs. The cross-domain knowledge reasoning agent calls the literature reading agents of each domain according to the target chemical discovery task input by the researchers. It compares, summarizes and merges the retrieval enhancement answers returned by the literature reading agents of each domain, and conducts multiple rounds of human-computer collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs and initial experimental candidate sets.
[0099] Step 3: Input the current set of experimental candidates into the on-demand autonomous chemistry research system, convert it into a robot autonomous experimental task and execute it to obtain real experimental data;
[0100] Step 4: The researchers determine whether they have obtained experimental candidates that meet the expectations or whether the total experimental budget has been exhausted. If yes, the iteration stops; otherwise, proceed to step 5.
[0101] Step 5: Incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental designs, historical experimental feedback sets, and cross-domain literature knowledge to generate the next round of experimental candidate sets and returns to step 3.
[0102] First, literature from multiple fields is acquired around the target scientific question, and the literature texts are parsed, distilled, and structured. Original literature typically contains a large amount of content unrelated to the target scientific question, such as author information, references, general introductions, and some experimental details. To improve the quality of subsequent knowledge extraction, this embodiment employs a large language model to perform domain-specific distillation on the literature, retaining information related to material structure, reaction mechanisms, active sites, electron or proton transfer, and structure-property relationships. Subsequently, entities and relationships are extracted based on the distilled text, and knowledge graphs for each relevant field are constructed, transforming scattered literature knowledge into searchable and reasonable structured knowledge resources.
[0103] Secondly, domain knowledge graphs are integrated into the graph-enhanced retrieval framework, and intelligent agents for reading literature across multiple domains are constructed. Each domain-specific intelligent agent is connected to the corresponding domain knowledge graph, enabling it to retrieve relevant entities, relationships, and graph paths based on the query, and generate enhanced responses based on the search results. Compared to ordinary text retrieval, graph-enhanced retrieval can explicitly return knowledge paths, making the reasoning process of the large language model more traceable. Building upon this, a cross-domain knowledge reasoning agent acts as the upper-level coordinating agent, calling upon different domain-specific intelligent agents to aggregate cross-domain literature knowledge and conduct collaborative question-and-answer sessions with researchers. After researchers propose scientific questions or candidate directions, the cross-domain knowledge reasoning agent breaks down the questions into multiple domain-specific sub-questions, calling upon the corresponding domain-specific intelligent agents for each, and integrating the returned results into candidate research directions and scientific hypotheses. After confirmation by researchers, the cross-domain knowledge reasoning agent further transforms the scientific hypotheses into experimental designs and provides an initial set of experimental candidates based on the experimental designs.
[0104] Then, the candidate experiments are input into the on-demand autonomous chemistry research system ChemAgents. ChemAgents then calls upon the experimental design agent and the robot operation agent to generate experimental procedures and robot control programs, and completes experimental verification within the automated laboratory. Specifically, the experimental design agent generates an executable experimental procedure in the automated laboratory based on the candidate experimental scheme. The robot operation agent further translates the experimental procedure into robot control code and drives the automated experimental platform to complete candidate sample preparation, characterization, and performance testing. This allows the scientific hypotheses and experimental designs formed in the previous stage to be transformed into real physical experiments, and obtains real experimental data that can be used for hypothesis verification.
[0105] Finally, researchers determine whether to continue iteration based on real experimental data. If so, the real experimental data is returned as feedback to the cross-domain knowledge reasoning agent to optimize the experimental strategy and guide subsequent candidate recommendations and closed-loop experimental iterations. Specifically, experimental data is incorporated into the historical experimental feedback set and fed back to the cross-domain knowledge reasoning agent. The agent combines scientific hypotheses, experimental design, historical experimental feedback sets, and cross-domain literature knowledge to generate the next round of experimental candidate sets, which are then returned to the robotic experimental verification step for verification by the autonomous chemistry research system ChemAgents.
[0106] Through the above process, a closed-loop discovery process is formed, from cross-domain knowledge construction, scientific hypothesis generation, robot experiment verification to feedback reasoning optimization.
[0107] In this embodiment, the discovery of high-performance photocatalytic hydrogen production materials is used as an example to illustrate the knowledge-oriented intelligent chemical discovery method described above.
[0108] This task selected three related fields as knowledge sources: photocatalytic hydrogen production, hydrogenases, and enzyme-mimetic catalysis. Photocatalytic hydrogen production provides knowledge on artificial catalytic materials, performance influencing factors, and structural optimization strategies; hydrogenases offer inspiration from the active centers, proton transfer, and electron transfer mechanisms in naturally occurring, highly efficient catalytic systems; and enzyme-mimetic catalysis provides bridging knowledge between natural catalytic mechanisms and the design of artificial catalytic materials. Through graphical modeling and enhanced reasoning of the knowledge in these fields, scientific hypotheses regarding the regulation of the catalytic microenvironment were formed and further transformed into candidate material design and robotic experimental tasks for subsequent closed-loop experimental verification.
[0109] Step 1: Acquire and extract literature knowledge from multiple fields, and construct a domain knowledge graph:
[0110] (11) Literature acquisition and screening in multiple fields:
[0111] The first step in knowledge-driven intelligent chemical discovery is to construct a literature database across multiple fields related to the target scientific question. For complex chemical discovery tasks, literature from a single field often fails to provide complete design guidance. Taking the discovery of high-performance photocatalytic hydrogen production materials as an example, literature on artificial photocatalytic materials contains a wealth of reaction systems, performance indicators, and material design experience, but it may not systematically reveal the microenvironment regulation mechanisms in highly efficient catalytic systems in the natural biological world. Literature related to hydrogenases contains the active center structure, proton transfer, and electron transfer mechanisms in natural catalytic systems, but this knowledge needs to be further transformed into design principles for artificial materials. Literature on enzyme-mimicking catalysis connects natural enzyme inspiration and artificial catalyst construction, providing an intermediate bridge for cross-disciplinary knowledge transfer. Therefore, this step constructs three complementary literature databases around the target scientific question, corresponding to the fields of photocatalytic hydrogen production, hydrogenases, and enzyme-mimicking catalysis, respectively. These databases collectively cover artificial catalytic systems, natural catalytic mechanisms, and biomimetic design strategies, helping to simultaneously capture existing domain knowledge and potential research opportunities.
[0112] Specifically, this step first uses "photocatalytic hydrogen production," "hydrogenase," and "enzyme-mimetic catalysis" as search keywords to retrieve relevant literature from the Web of Science database, up to November 30, 2024. The initial search yielded 35,673 relevant articles across three fields. To improve relevance and control the scale of subsequent text processing, the search results were first deduplicated based on DOI. Due to the large number of original search results and the differences in research objects, document types, and information density across different fields, this invention further employs a field-specific screening strategy, filtering by document type, research content, and mechanism relevance based on title and abstract to ensure the accuracy and processability of subsequent knowledge extraction.
[0113] For the fields of photocatalytic hydrogen production and hydrogenases, the screening strategy prioritizes original research papers published in high-impact peer-reviewed journals, requiring that the literature contain information at the mechanistic or process level, such as excited-state dynamics, electron or charge transfer, interfacial reaction steps, experimental performance indicators, and structure-performance relationships. Non-English literature, editorials, conference abstracts, patents, and purely device-level studies lacking mechanistic content are excluded from subsequent processing. For the field of enzyme-mimicking catalysis, given the relatively small size of the initial search results and the fact that review articles can more effectively summarize structure-function relationships, representative materials, and design frameworks, the strategy prioritizes representative review articles that reflect the field's terminology and design principles.
[0114] After the above deduplication and screening, 1261 articles in the field of photocatalytic hydrogen production, 348 articles in the field of hydrogenases, and 30 review articles in the field of enzyme-mimicking catalysis were retained. Subsequently, based on the list of DOIs obtained from the screening, the corresponding PDF articles were automatically downloaded in batches using the literature management software Zotero, providing input for subsequent literature distillation of large language models and knowledge graph construction.
[0115] The above literature construction process ensures both the coverage of cross-domain knowledge and the control of the computational load and noise level of subsequent large-scale text processing.
[0116] (12) Text preprocessing and document distillation:
[0117] Original literature texts typically contain a large amount of information irrelevant to the target scientific question, such as author and institution information, references, grant acknowledgments, general background descriptions, and experimental details unrelated to the mechanism. Directly using the original text for entity recognition and relation extraction would increase noise in knowledge extraction, leading to numerous irrelevant nodes, duplicate entities, or low-value relations in the knowledge graph. Therefore, this step introduces text preprocessing and large language model-based literature distillation before knowledge graph construction.
[0118] First, the PDF documents were converted to UTF-8 plain text format using the open-source document parsing tool Docling, and the text was initially cleaned to retain as many parsable chapter titles, figure captions, and main text structure as possible. Due to the complex formatting structure of scientific literature, the parsed text may contain headers and footers, reference numbers, figure captions, formula symbols, and irrelevant paragraphs. Therefore, it is necessary to further remove information irrelevant to the target scientific question, retaining only the text content related to mechanisms, materials, reaction processes, and performance discussions. Subsequently, by calling a large language model (such as GPT-4o-mini) API, domain-specific knowledge distillation was performed on the plain text document content. Distillation prompts not only required the removal of author information, institutional information, references, and non-mechanistic descriptions, but also required the standardization of key chemical entities and symbol expressions to reduce ambiguity in subsequent entity merging and relation extraction.
[0119] To address the specific knowledge characteristics of different fields, this step designed domain-specific distillation prompts. For the field of photocatalytic hydrogen production, the prompts are required to retain information related to the reaction pathway, photogenerated charge separation, catalytic sites, material composition, structural regulation, and performance relationships, while removing content irrelevant to the hydrogen production mechanism. For the field of hydrogenases, the prompts are required to retain information related to the enzyme active site, metal sulfide clusters, catalytic cycle, proton transfer, electron transfer, and structure-function relationships. For the field of enzyme-mimicking catalysis, the prompts are required to retain information related to enzyme-mimicking structure design, active site simulation, functional groups, structural motifs, and catalytic mechanisms. Simultaneously, the prompts also require standardized expression of key entities, such as unifying the terminology for hydrogen, platinum nanoparticles, metal sulfide clusters, and specific framework materials, to reduce duplicate nodes and entity ambiguity in subsequent knowledge graphs.
[0120] After distillation using a large language model, each document is transformed into mechanism-related text more suitable for knowledge graph construction. This step compresses redundant information in the original documents while retaining key knowledge relevant to the generation of scientific hypotheses, including material composition, active sites, reaction pathways, electron or proton transfer processes, and structure-property relationships. This allows for the provision of clearly structured, mechanism-related, and low-noise text fragments for subsequent knowledge graph construction while maintaining cross-domain knowledge coverage, thereby improving the accuracy of entity relation extraction and the quality of knowledge graph construction.
[0121] (13) Domain knowledge graph construction:
[0122] After document distillation, the distilled text is further transformed into a domain knowledge graph. Let D be the set of distilled documents in a specific domain. First, this set is divided into multiple text fragments to facilitate fine-grained entity recognition, relation extraction, and subsequent retrieval. The expression is:
[0123] (1);
[0124] in, Let represent the i-th text segment, and n represent the total number of text segments. The last text segment is represented by `w`, which represents the size of the text segmentation window. `Segment()` is a function that segments the distilled literature collection into text segments according to the window size. Each text segment serves as the basic processing unit for entity recognition and relation extraction, reducing the contextual redundancy and information dilution problems caused by directly inputting long documents into a large language model.
[0125] Subsequently, based on large language model prompts and the open-source graph construction tool LightRAG, candidate entities and candidate relations are extracted from each text segment to form a set of triples. The expression is:
[0126] (2);
[0127] Here, s represents the subject entity, r represents the relationship between entities, and o represents the object entity. For example, in task-related text, materials, active sites, functional groups, electron transfer processes, proton transport pathways, and performance indicators can all be considered entities; "promote," "inhibit," "connect," "participate," and "enhance" can be considered relationships between entities. R_LLM represents an entity relationship recognition function based on large language model prompts. This function simultaneously identifies entities and their corresponding relationships within each text segment, ensuring that the extracted relationship edges have clear starting and target entities, corresponding to the subject entity s and the object entity o in the triple, respectively.
[0128] After obtaining the set of triples for each text segment, the system summarizes the triples from all text segments to obtain the candidate triple set for that domain:
[0129] (3);
[0130] Where T represents the set of candidate triples extracted from all distilled texts in the domain.
[0131] Furthermore, the system uses the candidate entity node set V, extracted from all triples in the candidate triple set, and the candidate relation edge set E, composed of triple relations, as follows:
[0132] (4);
[0133] in, Graph represents the candidate knowledge graph before deduplication and merging; V represents the process of generating the graph structure from the set of triples; E represents the set of candidate entity nodes extracted from all triples in the candidate triple set; and E represents the set of candidate relation edges composed of triple relations.
[0134] This transforms knowledge, originally scattered throughout natural language text, into a computable "entity-relationship-entity" structure. Nodes represent entities in fields such as materials, mechanisms, structural primitives, and performance indicators; edges represent semantic relationships and mechanistic connections between entities.
[0135] To improve the efficiency of graph retrieval and large language model inference, and to enhance the interpretability of graph nodes and relations, text descriptions are further generated for the extracted entities and relations. These descriptions not only possess symbolic labels but also semantic information that can be understood by large language models, and can be represented as follows:
[0136] P(V, E) = {(k, v) | k ∈ labels(V, E), v = M_sum(D[k])} (5);
[0137] Where P represents the entity and relation description generation function; k represents the label of the entity node or relation edge; labels(V, E) represents the label set obtained from the candidate entity node set V and the candidate relation edge set E; D[k] represents the text fragment or context sentence related to label k; M_sum represents the large language model summarization function; and v represents the compact semantic description generated by the large language model. This step compresses the relevant descriptions scattered in different document fragments into a semantic summary oriented towards retrieval and reasoning, so that each node and relation edge not only retains graph structure information but also has an interpretable text representation.
[0138] Finally, entities that appear repeatedly, synonyms, and similar relationships in different text fragments are merged, deduplicated, and normalized to ultimately construct the domain knowledge graph G:
[0139] (6);
[0140] Where G represents the domain knowledge graph; and represents the set of entity nodes and the set of relation edges after deduplication and merging, respectively; P(V, E) represents the set of entity-relation semantic descriptions obtained after describing the candidate entity nodes and relation edges; Dedupe represents the entity and relation merging function, which is used to merge duplicate or similar nodes and edges across text segments based on entity labels, relation labels and their semantic descriptions.
[0141] Through the above processing, a more compact domain knowledge graph is obtained, reducing duplicate nodes and redundant relationships caused by cross-document extraction, and improving the efficiency of subsequent graph traversal and retrieval. The final domain knowledge graph G not only preserves the entity, relation, and triple structure, but also retains the corresponding semantic context through entity and relation descriptions, providing a knowledge foundation for subsequent graph-enhanced retrieval and cross-domain knowledge reasoning.
[0142] Based on the above process, this step constructed knowledge graphs for three fields: photocatalytic hydrogen production, hydrogenases, and enzyme-mimicking catalysis. Ultimately, the photocatalytic hydrogen production knowledge graph contained 2612 triplets, the hydrogenase knowledge graph contained 914 triplets, and the enzyme-mimicking catalysis knowledge graph contained 597 triplets. The photocatalytic hydrogen production knowledge graph mainly includes information on artificial photocatalytic materials, reaction processes, photogenerated charge behavior, and performance indicators; the hydrogenase knowledge graph mainly includes information on the active center of natural enzymes, metal clusters, hydrogen bond networks, and proton / electron transfer; and the enzyme-mimicking catalysis knowledge graph mainly includes information on enzyme-mimicking structure design, functional groups, catalytic site simulation, and artificial system construction strategies. Table 1 provides statistical information on literature and knowledge graphs in multiple fields.
[0143] Table 1 shows the statistical results of literature and knowledge graphs in multiple fields.
[0144]
[0145] Through the aforementioned knowledge graph construction process, unstructured knowledge scattered across multiple domains is organized into entities, relationships, and graph paths. In subsequent scientific question answering and hypothesis generation, relevant entities and their relationship paths can be retrieved around the query question, enabling large language model reasoning to be based on structured knowledge evidence. This improves the traceability and reliability of cross-domain knowledge fusion and scientific hypothesis generation. The resulting three domain knowledge graphs provide a cross-domain knowledge foundation for the knowledge-oriented intelligent chemical discovery system.
[0146] Step 2, Enhanced Graph Retrieval, Scientific Hypothesis Generation, and Experimental Design:
[0147] (21) Design of a domain literature reading agent based on graph-enhanced retrieval:
[0148] In existing technologies, document reading agents are mainly used to retrieve and mine task-related structured information from document databases, providing document knowledge support for experimental design and robotic experiment execution. However, when facing deep scientific discovery tasks, ordinary document retrieval and information mining methods still have certain limitations. On the one hand, traditional retrieval results are usually returned in the form of document fragments or abstracts, making it difficult to explicitly express the relationships and mechanisms between entities; on the other hand, relying solely on document retrieval tools is insufficient to fully mine deep knowledge connections across documents and fields, limiting the knowledge basis and reasoning depth of the system in generating scientific hypotheses. Therefore, this invention further expands the document reading agent based on the construction of a domain knowledge graph. It integrates the domain knowledge graph obtained through document distillation, entity relation extraction, and structured modeling into the corresponding domain document reading agent, enabling the domain document reading agent to perform enhanced retrieval based on the corresponding domain knowledge graph. This transforms it from an ordinary document retrieval and information mining tool into a domain document reading agent with domain knowledge organization, relation retrieval, and professional question-answering capabilities.
[0149] Specifically, this embodiment uses the open-source graph retrieval enhancement tool LightRAG to construct three domain-specific document reading agents, corresponding to the fields of photocatalytic hydrogen production, hydrogenase, and enzyme-mimicking catalysis, respectively. In the system implementation, all three domain-specific document reading agents use GPT-4o-mini as the base large language model and are connected to the corresponding domain knowledge graph and graph enhancement retrieval module. Each domain-specific document reading agent can retrieve relevant entities, relationships, and local graph paths in the corresponding knowledge graph based on the input question, and input the search results as context into the large language model. Subsequently, the large language model generates an enhanced response based on the graph retrieval context, thereby strengthening the correlation between the response and the graph evidence. Compared to directly asking questions to the base large language model, this graph enhancement retrieval method can constrain the large language model's response within the constructed domain knowledge scope, making the response process more traceable, domain-consistent, and professionally targeted.
[0150] During the retrieval process, LightRAG supports a two-tiered graph retrieval system: entity-level and topic / summary-level. Entity-level retrieval is used to locate nodes directly related to the query, such as materials, structural units, active sites, functional groups, and performance indicators. Topic / summary-level retrieval, on the other hand, aggregates higher-level mechanism descriptions and local knowledge subgraphs related to the query. Through this two-tiered retrieval approach, the domain literature reading agent can obtain both fine-grained evidence of entity relationships and a more complete domain background and mechanistic context, thereby improving the completeness and interpretability of the domain response.
[0151] Let the input question be q, and the corresponding domain knowledge graph be q. The graph-enhanced retrieval process of a domain literature reading agent can be represented as:
[0152] C = R (q, G) (7;
[0153] a = M (q, C) (8;
[0154] Here, R represents the knowledge graph retrieval function, which internally includes entity-level retrieval and topic / summary-level retrieval; C represents the contextual information retrieved from the knowledge graph, including relevant entities, relationships, local graph paths, and descriptions of topical mechanisms related to the query question; M represents the large language model question-answering function; and a represents the retrieval-enhanced answer generated based on the knowledge graph context. The above process demonstrates that the output of the domain literature reading agent is not directly and freely generated by the large language model, but rather obtained through a combination of "question input—knowledge graph retrieval—context enhancement—large language model answer."
[0155] The three domain-specific literature-reading agents have relatively clear functional boundaries. The agent for photocatalytic hydrogen production primarily answers questions related to artificial photocatalytic materials, performance influencing factors, photogenerated charge separation, and active sites. The agent for hydrogenases primarily answers questions related to the active centers of natural hydrogenases, proton transport channels, and electron transfer processes. The agent for enzyme-mimicking catalysis primarily answers questions related to biomimetic structural design, the introduction of functional groups, and strategies for constructing artificial catalytic systems. This domain-specific division avoids the knowledge confounding issues caused by a single agent simultaneously handling multiple complex domains and provides specialized knowledge sources for subsequent cross-domain knowledge fusion and scientific hypothesis generation.
[0156] (22) Cross-domain knowledge reasoning agents and scientific hypothesis generation methods:
[0157] Simply constructing multiple domain-specific literature reading agents is insufficient for generating complex scientific hypotheses. Scientific discovery tasks typically require the comparison, transfer, and fusion of knowledge from different domains. For example, the proton transfer mechanism in natural enzymes needs to be mapped into the design of artificial photocatalytic materials, and the structural design strategies in enzyme-mimicking catalysis need to be further used to explain how to inspire the transformation of natural systems into synthetic and testable candidate materials. Therefore, this embodiment, based on three domain-specific literature reading agents, further constructs a cross-domain knowledge reasoning agent. This agent acts as a higher-level coordinating agent, interacting with researchers and managing, calling upon, and synthesizing the responses from different domain-specific literature reading agents. The cross-domain knowledge reasoning agent, together with the multiple domain-specific literature reading agents, constitutes a multi-agent collaborative system for cross-domain knowledge fusion and scientific hypothesis generation. Its structure and workflow are as follows: Figure 2 As shown in the figure. In the system implementation, the cross-domain knowledge reasoning agent uses GPT-4o as the base large language model.
[0158] The workflow of the cross-domain knowledge reasoning agent includes scientific question understanding, requirement decomposition, domain agent invocation, cross-domain answer fusion, and scientific hypothesis generation. First, researchers propose open scientific questions or candidate research directions. The cross-domain knowledge reasoning agent generates a domain decomposition query plan based on the knowledge scope involved in the question and determines which domain literature reading agents need to be invoked. Subsequently, the cross-domain knowledge reasoning agent distributes sub-questions to the corresponding domain literature reading agents. The photocatalytic hydrogen production, hydrogenase, and enzyme-mimicking catalysis domain literature reading agents generate corresponding domain-enhanced answers based on graph-enhanced retrieval, returning relevant graph context, key relationship paths, and reasoning basis. Finally, the cross-domain knowledge reasoning agent compares, summarizes, and fuses the enhanced retrieval answers returned by different domain literature reading agents to form a comprehensive scientific answer for researchers. After researchers confirm the research directions in the comprehensive scientific answer based on their target chemical discovery task, the cross-domain knowledge reasoning agent and researchers enter the next round of interaction, gradually generating candidate research directions and scientific hypotheses through multiple rounds of human-computer collaborative question-and-answer sessions.
[0159] The above cross-domain collaborative reasoning process can be further represented as:
[0160] (9);
[0161] (10);
[0162] (11);
[0163] (12);
[0164] in, , and These represent knowledge graphs in the fields of photocatalytic hydrogen production, hydrogenases, and enzyme-mimicking catalysis, respectively. , and This represents a graph retrieval function for the corresponding field; , and q represents the large language model question-and-answer function in the intelligent agent for reading literature in the corresponding field; q represents the original scientific question raised by the researcher. , and This represents the domain sub-problems obtained by decomposing a cross-domain knowledge reasoning agent. , and This represents the contextual information obtained from the knowledge graph retrieval in the corresponding domain. , and M_cross represents the enhanced response for retrieval across three domains; M_cross represents the large language model inference function of the cross-domain knowledge reasoning agent; U represents the feedback, follow-up questions, and confirmation information from researchers in multiple rounds of question answering; and H represents the scientific hypothesis obtained from the fusion of cross-domain knowledge.
[0165] In this process, researchers do not directly provide the final experimental plan, but rather act as problem proposers, direction confirmers, and judges of low-confidence results in the human-machine collaborative process. The cross-domain knowledge reasoning agent is responsible for understanding the scientific question, planning the multi-agent invocation process, synthesizing domain knowledge to answer, and gradually forming experimentally verifiable scientific hypotheses through interaction with researchers. Through this human-machine collaborative approach, the system can leverage large language model agents to improve the efficiency of cross-domain knowledge integration and hypothesis generation, while retaining the selection of research directions, confirmation of key hypotheses, and final research decisions in the hands of researchers, thereby enhancing the reliability and controllability of the scientific discovery process.
[0166] (23) Generation of the human-machine collaboration hypothesis in the task of discovering photocatalytic hydrogen production materials:
[0167] Through multi-round human-machine collaborative question-and-answer sessions with a cross-domain knowledge-reasoning agent, the system gradually formed scientific hypotheses for the discovery of high-performance photocatalytic hydrogen production materials from knowledge graphs in three domains: photocatalytic hydrogen production, hydrogenases, and enzyme-mimicking catalysis. First, researchers posed questions to the cross-domain knowledge-reasoning agent, requesting research directions related to the structural design of photocatalytic hydrogen production materials, with particular attention to the inspiration that enzyme structural composition provides for the catalytic process. After integrating knowledge from different domains, the cross-domain knowledge-reasoning agent provided candidate research directions from the perspectives of material composition and metal centers, morphology and surface structure, and structural modification and stability. In particular, the cross-domain knowledge-reasoning agent pointed out that the protein framework surrounding the metal center in hydrogenases plays a crucial role in maintaining structural stability and promoting proton and electron transfer; this direction was identified by researchers as a key area for future exploration.
[0168] Subsequently, the researchers further requested the cross-domain knowledge reasoning agent to provide specific structural design schemes around "simulating the protein framework of enzymes to promote proton and electron transfer in the photocatalytic process." Based on the spectral search results, the cross-domain knowledge reasoning agent proposed schemes such as band structure regulation and heterostructure construction, proton transfer pathway construction, introduction of redox-active cofactors, and nanostructure and porous structure design. The system further pointed out that rapid proton transfer can be promoted by constructing hydrogen bond networks and utilizing amino acid residues as proton relays to simulate the local environment in natural systems that is conducive to proton migration. After confirmation by the researchers, "constructing proton transfer pathways and introducing amino acid residues as proton relays" was identified as the core idea for subsequent research. Table 2 presents the key reasoning information of the human-machine collaborative hypothesis generation process in the photocatalytic hydrogen production material discovery task.
[0169] Table 2 shows the process of generating the human-machine collaboration hypothesis in the task of discovering photocatalytic hydrogen production materials.
[0170] Based on the above human-machine collaborative question-and-answer process, the scientific hypothesis formed in the task of discovering photocatalytic hydrogen production materials can be summarized as follows: introducing functional groups similar to amino acid residues into photocatalysts and constructing a proton transfer microenvironment close to the catalytic active center is expected to promote water activation, proton transfer and charge utilization, thereby improving the performance of photocatalytic hydrogen production.
[0171] (24) Transformation from scientific hypothesis to experimental design and initial set of experimental candidates:
[0172] Scientific hypotheses can only proceed to the subsequent robot experimental verification stage after they are further transformed into experimental designs and based on these designs, forming an initial set of executable experimental candidates. The purpose of this process is to gradually transform scientific hypotheses in natural language form into selectable, verifiable, and executable experimental designs and specific experimental candidates, rather than remaining research ideas at the textual level.
[0173] During the experimental candidate design process, researchers raise specific design questions related to the candidate construction process. A cross-domain knowledge reasoning agent invokes relevant domain knowledge graphs and domain literature reading agents based on the questions to generate candidate suggestions and reasons for recommendation. Researchers then make judgments and confirmations at key nodes to ensure that the candidate designs not only conform to the aforementioned scientific assumptions but also meet the feasibility requirements for subsequent experimental verification. This forms a candidate design chain with each round of "question formulation—knowledge retrieval—agent recommendation—human confirmation" as the basic unit, which gradually converges through multiple rounds of interaction.
[0174] In the mission to discover photocatalytic hydrogen production materials, the cross-domain knowledge reasoning agent, with the participation of researchers, further transformed the scientific hypothesis formed in the previous section—"introducing functional groups similar to amino acid residues into photocatalysts to construct a proton transfer microenvironment close to the catalytic active center"—into experimental candidate design problems such as the catalytic active center, photocatalytic material type, specific material platform, functionalization strategy, and candidate small molecule types. Table 3 presents the human-machine collaborative candidate design chain in the mission to discover photocatalytic hydrogen production materials.
[0175] Table 3 shows the experimental candidate design chain for the mission of discovering photocatalytic hydrogen production materials.
[0176] Specifically, in selecting the catalytic active center, the cross-domain knowledge reasoning agent recommends platinum nanoparticles (Pt NPs) as the active center for hydrogen production. In selecting the material support, after systematically comparing various material types such as metal oxides, metal-organic frameworks (MOFs), covalent organic frameworks (COFs), semiconductor quantum dots, organic dyes, metal complexes, graphitic carbon nitride, MXene, high-entropy oxides, and natural photosensitive proteins, MOFs are recommended as a modifiable porous photocatalytic material support. In selecting a specific material platform, the cross-domain knowledge reasoning agent further compares Zr-based MOFs, Cu-based MOFs, and the MIL series MOFs, and recommends UiO-67, which has good stability and a large pore size, as the structural platform. UiO-67 is a typical Zr-based MOF, a porous crystalline material formed by the self-assembly of Zr metal ions and organic ligands through coordination bonds. It possesses good structural stability and post-modification space, making it suitable as a support for introducing functionalized small molecules and constructing local microenvironments. Regarding functionalization strategies, after comparing pre-synthetic modification, post-synthetic modification (PSM), and defect engineering-assisted modification, the cross-domain knowledge reasoning agent recommends the defect engineering-assisted post-synthetic modification strategy to introduce small molecule functional groups similar to amino acids into the UiO-67 framework environment.
[0177] After completing the candidate design chain shown in Table 3, the candidate small molecule types need to be further transformed into specific experimental candidates that can be used for synthesis and testing. Based on the screening criteria provided by the cross-domain knowledge reasoning agent, candidate molecules must simultaneously meet the requirements of being relevant to the scientific hypothesis, being able to be introduced into the UiO-67-based material system, and being suitable for subsequent automated experimental verification. Ultimately, researchers, combining the laboratory's existing chemical reagent library, identified a range of experimentally viable candidates comprising 31 specific small molecules. These candidate molecules mainly cover small molecule skeletons containing monocarboxylic acid anchoring groups, nitrogen-containing sites, and those with similar local structural features to amino acids. They can be introduced into Pt@UiO-67-based photocatalysts through post-synthetic modification to regulate the local proton transfer microenvironment near platinum nanoparticles and to verify the impact of this microenvironment on hydrogen production performance from different structural types. Based on the aforementioned scientific hypothesis and experimental design, and combined with cross-domain literature knowledge, the cross-domain knowledge reasoning agent provides an initial set of experimental candidates as task input for the subsequent on-demand autonomous chemical research system.
[0178] The aforementioned candidate experimental design method demonstrates that the cross-domain knowledge reasoning agent not only undertakes the task of generating scientific hypotheses but also further plays a bridging role between scientific hypotheses and experimental candidates. Based on the reasoning results of the domain knowledge graph and the large language model, the cross-domain knowledge reasoning agent proposes candidate design suggestions, while researchers confirm key nodes in the candidate system construction process. This avoids the uncontrollable risks that may arise from fully automated candidate generation and reduces the workload of researchers in sifting through a large amount of literature and candidate systems one by one. Through this process, scientific hypotheses are further transformed into a set of executable experimental candidates, making them task inputs that the on-demand autonomous chemistry research system ChemAgents can accept, providing a candidate foundation for subsequent automated preparation, performance testing, and closed-loop feedback.
[0179] Step 3, Independent Experiment Verification:
[0180] Robot Autonomous Experiment:
[0181] After obtaining the initial set of experimental candidates in step 2, the current set is input into the on-demand autonomous chemistry research system ChemAgents, and further transformed into robotic autonomous experimental tasks to obtain real experimental data. This process is based on the existing ChemAgents hierarchical multi-agent architecture and automated laboratory resources. The automated laboratory includes two robots and 20 different types of experimental workstations, capable of supporting automated synthesis, characterization, and performance testing tasks. Therefore, in this embodiment, the set of experimental candidates formed in the previous stage is used as the task input for ChemAgents, and the experimental design agent and the robot operation agent complete the candidate sample preparation, material processing, performance testing, and experimental data feedback.
[0182] Specifically, the task management agent in ChemAgents first receives candidate experimental tasks and calls the experiment design agent based on the task objectives. The experiment design agent combines candidate experimental information and an experimental protocol library to generate an executable experimental protocol for the automated laboratory. Subsequently, the robot operation agent further translates the experimental protocol into robot control code and drives the automated experimental platform to complete candidate sample preparation and performance testing. After the experiment is completed, the system obtains the real experimental data of the candidate samples and provides this data to researchers for iteration termination judgment. If researchers decide to continue iteration, the current round of real experimental data is associated with the corresponding current experimental candidate set and included in the historical experimental feedback set, which is then fed back to the cross-domain knowledge reasoning agent. Thus, the system can transform the experimental design and experimental candidate set formed in the previous stage into real physical experiments and obtain real experimental data for subsequent feedback reasoning and candidate updates.
[0183] In the mission to discover photocatalytic hydrogen production materials, the experimental design agent invokes an experimental protocol library to match and combine a set of experimental protocols related to candidate experiments. These protocols mainly include the synthesis of platinum nanoparticles, the assembly of Pt@UiO-67, the post-synthetic modification of Pt@UiO-67-X, and standardized photocatalytic hydrogen production testing, where X represents a specific candidate small molecule. These protocols cover key aspects such as candidate material preparation, functionalization modification, photocatalytic reaction, and hydrogen production performance testing. Based on these protocols, the experimental design agent assembles a complete experimental workflow. Subsequently, the robot-operated agent compiles this workflow into executable code, driving robots and related experimental workstations in the automated laboratory to complete experimental operations such as solid sample introduction, liquid sample introduction, heating, stirring, centrifugation, washing, drying, photocatalytic reaction, and gas chromatography detection. Ultimately, this achieves the preparation of the Pt@UiO-67-X candidate sample and the testing of its hydrogen production performance.
[0184] Step 4, Iteration termination check:
[0185] After obtaining the current round of real experimental data in step 3, researchers determine whether satisfactory experimental candidates have been obtained or whether the total experimental budget has been exhausted based on the real experimental data. If satisfactory experimental candidates have been obtained or the total experimental budget has been exhausted, the iteration stops; if neither is satisfied, step 5 is executed, and the current round of real experimental data is used for feedback inference optimization.
[0186] Step 5, Feedback Reasoning Optimization:
[0187] This embodiment designs an inference optimization method based on feedback from real robot experimental data. In the first experiment, the initial experimental candidate set obtained in step 2... After step 3, the data is input into the ChemAgents on-demand autonomous chemistry research system, which then performs robotic autonomous experiments to obtain corresponding real experimental data. Researchers based their findings on real experimental data. Perform the iteration termination judgment according to step 4; when it is determined that iteration needs to continue, the initial experimental candidate set will be... Corresponding real experimental data The data was correlated and included in the historical experimental feedback set as the first round of experimental feedback.
[0188] During the 3rd round of feedback reasoning, the system has obtained the historical experimental feedback set. :
[0189] (13);
[0190] in, This represents the set of experimental candidates accumulated up to round 3, along with their corresponding real experimental data. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental designs, and updated historical experimental feedback sets. By combining cross-disciplinary literature knowledge, we compare and analyze the candidate experiments from previous rounds and their actual experimental data, redetermine the basis for candidate designs and subsequent recommendation directions, and generate the next round of candidate experiment sets. .
[0191] Subsequently, the candidate set for the next round of experiments will be... Returning to step 3, the data is input again into the ChemAgents on-demand autonomous chemistry research system, transformed into a robotic autonomous experimental task, and executed to obtain the next round of real experimental data. Researchers then use this new real experimental data to make judgments according to step 4. If iteration continues, the real experimental data from this round is associated with the corresponding set of experimental candidates and included in the historical experimental feedback set, and step 5 is executed again.
[0192] Through the above process, a closed-loop iterative process is formed: "initial experimental candidate generation—autonomous robot experimentation—researcher judgment—experimental feedback update—reasoning optimization—next round of experimental candidate generation." As the historical experimental feedback set accumulates, the cross-domain knowledge reasoning agent can continuously adjust the candidate design basis and subsequent recommendation direction, focusing subsequent experiments on more promising structural categories. Iteration stops when researchers determine, based on real experimental data from a certain round, that they have obtained experimental candidates that meet expectations, or when the total experimental budget is exhausted.
[0193] Results and analysis of the mission to discover photocatalytic hydrogen production materials:
[0194] In the discovery of photocatalytic hydrogen production materials, the method of this invention involves multiple rounds of autonomous robotic experiments and feedback reasoning optimization. Based on the experimental candidate set generated by a cross-domain knowledge reasoning agent, the system selects corresponding candidate samples for automated preparation and performance testing round by round. Researchers make iterative termination decisions based on the real experimental data from each round. After determining whether to continue iteration, the current round's real experimental data is included in the historical experimental feedback set and fed back to the cross-domain knowledge reasoning agent to generate the next round's experimental candidate set. The experiment uses standardized photocatalytic hydrogen production test conditions, with full-spectrum irradiation, water as the hydrogen source, and Pt@UiO-67 without the introduction of candidate small molecules as the benchmark material. Under the same automated test conditions, the hydrogen production rate of the benchmark material Pt@UiO-67 is 0.06. This provides a quantitative reference for improving the performance of subsequent candidate materials.
[0195] Figure 3 The candidate experiments and performance test results are presented in six rounds of closed-loop iteration. Figure 3Sample number 30 corresponds to the ultimately discovered optimal sample, Pt@UiO-67-8-quinolinecarboxylate. From Figure 3 The hydrogen production rate, indicated by the vertical axis, shows that the introduction of different candidate small molecules significantly affected the performance of the Pt@UiO-67-X material, demonstrating that the candidate molecular structure has an experimentally distinguishable influence on local microenvironment regulation and hydrogen production performance. In the first round of iterations, the cross-domain knowledge reasoning agent preferentially recommended small molecules rich in nitrogen sites, capable of hydrogen bonding, and potentially promoting water activation and proton transfer, including 2-pyridinecarboxylic acid, pyridazine-4-carboxylic acid, imidazole-2-carboxylic acid, 4-pyrazolecarboxylic acid, and L-histidine. Robotically automated preparation and testing results showed that Pt@UiO-67-L-histidine and Pt@UiO-67-4-pyrazolecarboxylic acid performed better, with Pt@UiO-67-L-histidine (… Figure 3 The hydrogen production rate of the sample (numbered 19) reached 0.24. It is approximately four times that of the reference material Pt@UiO-67.
[0196] Based on the updated historical experimental feedback set, the cross-domain knowledge reasoning agent, combining the aforementioned scientific hypotheses, experimental design, and cross-domain literature knowledge, further compared the structural characteristics and hydrogen production performance differences of different candidate samples, and adjusted the subsequent recommendations to focus on candidate small molecules with more nitrogen-containing sites and structural configurations more suitable for constructing local microenvironments. In the second iteration, the system recommended 1H-benzimidazole-5-carboxylic acid and L-arginine as candidate molecules. The hydrogen production performance of the corresponding materials was 2-3 times that of the benchmark materials, further supporting the important role of nitrogen-containing sites and local proton transfer environment in improving hydrogen production performance. Subsequently, in the third to fifth iterations, the system continued to explore structural types such as amino acid derivatives, nitrogen-containing cyclic molecules, and amino-substituted aromatic molecules. Experimental results showed that the better-performing candidate molecules usually have nitrogen-containing sites, hydrogen bonding ability, or certain structural adaptability, indicating that the local interactions formed by functional groups near the catalytic active site may be an important factor affecting hydrogen production performance.
[0197] In the sixth iteration, the cross-domain knowledge reasoning agent further focused its candidate recommendation direction on quinoline-based small molecules with nitrogen-containing aromatic structures. These molecules possess nitrogen-containing interaction sites, aromatic conjugated structures, and relatively fixed spatial configurations, which may help regulate the local microenvironment near platinum nanoparticles and promote improved hydrogen production performance. Experimental results showed that Pt@UiO-67-8-quinolinecarboxylic acid exhibited the best hydrogen production activity, with a hydrogen production rate reaching 2.04. The hydrogen production performance of the Pt@UiO-67-based photocatalyst is approximately 34 times that of the baseline material Pt@UiO-67, and significantly higher than that of candidate samples obtained in previous iterations. This result indicates that the local microenvironment formed after the introduction of 8-quinoline carboxylic acid can significantly improve the hydrogen production performance of the Pt@UiO-67-based photocatalyst. Combining the changes in candidate recommendation directions across six iterations, it can be seen that the cross-domain knowledge reasoning agent can combine scientific hypotheses, experimental design, historical experimental feedback sets, and cross-domain literature knowledge to gradually adjust the candidate design basis and subsequent recommendation directions, shifting the search focus from general nitrogen-containing small molecules to more promising quinoline candidate structures, thereby effectively guiding experimental candidate design and identifying high-performance candidate materials.
[0198] The results of the robot's autonomous experiments were used not only to screen candidate samples with superior performance but also to verify the rationality of the scientific hypotheses generated by the system. The core scientific hypothesis formed in this embodiment is that introducing functional groups similar to amino acid residues near the catalytic active site helps to construct a local microenvironment that promotes water activation, proton transfer, and charge utilization. Experimental results show that the superior candidate samples often possess nitrogen-containing structures, hydrogen bonding capabilities, or structural features conducive to local proton migration. In particular, the 8-quinoline carboxylic acid-modified sample exhibited a significantly higher hydrogen production rate than the baseline material and other candidate samples. This result is consistent with the microenvironment regulation hypothesis generated by the system, indicating that this hypothesis can provide effective guidance for the selection of candidate functionalized molecules.
[0199] From the perspective of experimental search efficiency, the system completed 15 candidate experiments during 6 rounds of feedback inference optimization, corresponding to the experimental design, automated preparation, and performance testing of 15 Pt@UiO-67-X catalytic materials. Compared to the experimentally viable candidate range of 31 small molecules, the system did not employ an exhaustive verification approach. Instead, guided by feedback from real experiments, it gradually narrowed the search scope and focused subsequent experiments on more promising candidate structure types. These results demonstrate that the feedback inference optimization mechanism can improve candidate screening efficiency, reduce the number of invalid experiments, and accelerate the discovery of high-performance candidate materials within a limited experimental budget.
[0200] In summary, the knowledge-driven intelligent chemical discovery method and system can transform scientific hypotheses generated through cross-domain knowledge reasoning into a set of executable experimental candidates, and then perform autonomous robotic experimental verification through ChemAgents. Furthermore, by incorporating historical experimental feedback sets, the cross-domain knowledge reasoning agent can compare and analyze the experimental candidates and their actual experimental data from each round, and update the candidate design basis and subsequent recommended directions round by round. This process couples graph-enhanced knowledge reasoning, candidate experimental design, robotic experimental execution, and feedback reasoning optimization, realizing a closed-loop process from scientific hypothesis generation and candidate material recommendation to the discovery of high-performance photocatalytic hydrogen production materials, providing methodological support for the verification of scientific hypotheses and the discovery of high-performance candidates in complex chemical systems.
[0201] This embodiment of the knowledge-oriented intelligent chemical discovery method and system, based on the existing ChemAgents hierarchical multi-agent architecture, introduces domain knowledge graphs, graph-enhanced retrieval, large language model agent reasoning, and robot experimental feedback to construct a closed-loop discovery framework of "knowledge fusion—hypothesis generation—experimental verification—feedback iteration." Around this framework, this embodiment first constructs literature from multiple domains for the target scientific question. Through steps such as literature parsing, domain-specific distillation, text segmentation, triple extraction, entity and relation description generation, and deduplication and merging, the domain knowledge scattered in unstructured literature is transformed into a searchable and reasonable knowledge graph. Subsequently, based on LightRAG, three domain literature reading agents—photocatalytic hydrogen production, hydrogenase, and enzyme-mimicking catalysis—are constructed, enabling them to obtain domain knowledge evidence through two-layer graph retrieval at the entity layer and topic / abstract layer, and generate retrieval-enhanced answers supported by graph context. Building upon this, a cross-domain knowledge reasoning agent is further constructed. Through domain question decomposition, domain agent invocation, cross-domain answer fusion, and human-machine collaborative confirmation, cross-domain knowledge reasoning and scientific hypothesis generation for open scientific questions are achieved. Subsequently, the scientific hypothesis was transformed into an experimental design, and an initial set of experimental candidates was formed based on the experimental design. This set was used as the task input for the ChemAgents on-demand autonomous chemistry research system. The experimental design agent and the robot operation agent completed the generation of experimental protocols, robot code conversion, automated preparation, performance testing, and result feedback. Finally, a feedback reasoning optimization method based on real experimental data from the robot was designed. After the researchers decided to continue iteration, the current round of real experimental data was associated with the corresponding set of experimental candidates and included in the historical experimental feedback set. This enabled the cross-domain knowledge reasoning agent to combine scientific hypotheses, experimental designs, historical experimental feedback sets, and cross-domain literature knowledge to redetermine the basis for candidate designs and subsequent recommended directions, generate the next round of experimental candidate sets, and guide subsequent experimental iterations. Through validation in the autonomous discovery task of photocatalytic hydrogen production materials, scientific hypotheses about the regulation of catalytic microenvironment were formed from literature knowledge in multiple fields. Through 6 rounds of closed-loop experimental iterations and 15 candidate experiments, the optimal-performing Pt@UiO-67-8-quinolinecarboxylic acid was discovered. The results show that the knowledge-oriented intelligent chemical discovery method and system of the present invention can effectively combine literature knowledge from multiple fields, scientific hypothesis generation, candidate experiment design, robotic experiment execution and feedback reasoning optimization, and promote the on-demand autonomous chemical research system to further extend from experimental task execution to in-depth scientific discovery of complex chemical systems.
[0202] Example 2
[0203] See Figure 4 This embodiment provides a knowledge-oriented intelligent chemical discovery system for implementing the method of Embodiment 1 above, comprising:
[0204] The system comprises a map construction module, a scheme generation module, a result acquisition module, a judgment module, and a feedback optimization module; among which...
[0205] The knowledge graph construction module communicates with the scheme generation module and is used to acquire literature from multiple fields related to the target chemical discovery task and transform them into corresponding domain knowledge graphs.
[0206] The scheme generation module communicates with the graph construction module to connect the knowledge graphs of various fields constructed by the graph construction module to the corresponding domain literature reading agents, enabling the literature reading agents of each domain to perform retrieval enhancement generation based on the corresponding graphs; the cross-domain knowledge reasoning agent calls the corresponding domain literature reading agents according to the researchers' input, integrates the retrieval enhancement answers returned by the literature reading agents of each domain, and conducts multiple rounds of collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs and initial experimental candidate sets;
[0207] The results acquisition module communicates with the scheme generation module and is used to input the current set of experimental candidates into the on-demand autonomous chemical research system, transform it into a robot autonomous experimental task and execute it to obtain real experimental data.
[0208] The judgment module communicates with the result acquisition module and the feedback optimization module respectively. It is used to obtain the researchers' judgment on whether the experimental candidates meet the expectations or whether the total experimental budget has been exhausted using real experimental data. If yes, the iteration stops; if no, the feedback optimization module is started.
[0209] The feedback optimization module communicates with the result acquisition module and is used to incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental design, historical feedback and cross-domain literature knowledge to generate the next round of experimental candidate set and send it to the result acquisition module.
[0210] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0211] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.
Claims
1. A knowledge-oriented intelligent chemical discovery method, characterized in that, To achieve the target chemical discovery task, including: Step 1: Obtain literature from multiple fields related to the target chemical discovery task, and convert the literature from each field into corresponding field knowledge graphs to obtain the knowledge graphs for each field. Step 2: Connect the knowledge graphs of each domain obtained in Step 1 to the corresponding domain literature reading agents. The domain literature reading agents can perform retrieval enhancement and generation based on the corresponding domain knowledge graphs. The cross-domain knowledge reasoning agent calls the literature reading agents of each domain according to the target chemical discovery task input by the researchers. It compares, summarizes and merges the retrieval enhancement answers returned by the literature reading agents of each domain, and conducts multiple rounds of human-computer collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs and initial experimental candidate sets. Step 3: Input the current set of experimental candidates into the on-demand autonomous chemistry research system, convert it into a robot autonomous experimental task and execute it to obtain real experimental data; Step 4: The researchers determine whether they have obtained experimental candidates that meet the expectations or whether the total experimental budget has been exhausted. If yes, the iteration stops; otherwise, proceed to step 5. Step 5: Incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental designs, historical experimental feedback sets, and cross-domain literature knowledge to generate the next round of experimental candidate sets and returns to step 3.
2. The knowledge-oriented intelligent chemical discovery method according to claim 1, characterized in that, In step 1, literature from multiple fields related to the target chemical discovery task is obtained in the following manner, including: Obtain literature from at least two complementary knowledge areas relevant to the target chemical discovery mission.
3. The knowledge-oriented intelligent chemical discovery method according to claim 2, characterized in that, In step 1, when the target chemical discovery task is the discovery of high-performance photocatalytic hydrogen production materials, the literature obtained in multiple fields related to the target chemical discovery task includes literature in the field of photocatalytic hydrogen production, literature in the field of hydrogenase, and literature in the field of enzyme-mimicking catalysis.
4. The knowledge-guided intelligent chemical discovery method according to any one of claims 1-3, characterized in that, In step 1, the acquired literature from multiple fields is processed in the following manner, and then the processed literature from each field is transformed into a corresponding domain knowledge graph, including: Document parsing tools were used to convert the literature from multiple fields into plain text format. The text was then pre-cleaned, retaining the parsable chapter titles, figure and table descriptions, and main text structure. Content irrelevant to the target chemical discovery task was then removed, and the content relevant to the target chemical discovery task was retained to obtain the pre-processed literature content. Using a large language model, domain-specific knowledge distillation is performed on the preprocessed content of each document based on domain-specific distillation prompts. After knowledge distillation, distilled texts of each document suitable for constructing a knowledge graph are obtained. Distilled texts of each document in the same field are compiled into a distilled literature set for that field, and then transformed into a domain knowledge graph.
5. The knowledge-oriented intelligent chemical discovery method according to claim 4, characterized in that, In step 1, the distilled texts of each document in the same field are grouped into a distilled document set for that field in the following manner, and then transformed into a domain knowledge graph, including: Let D be the collection of distillation literature in a certain field. Divide this collection of distillation literature D into multiple text segments using the following expression: (1); in, Let represent the i-th text segment, and n represent the total number of text segments. This represents the last text segment, where w represents the window size of the text segment; Segment() is a function that segments the distillation literature collection into blocks according to the window size. Based on large language model prompts and graph construction tools, candidate entities and candidate relations are extracted from each text segment to form a set of triples. The expression is: (2); Where s represents the subject entity; r represents the relationship between entities; o represents the object entity; R_LLM represents the entity relationship recognition function based on the prompt words of the large language model. This entity relationship recognition function simultaneously identifies entities and their corresponding relationships within each text segment, so that the extracted relationship edges have clear starting entities and target entities, corresponding to the subject entity s and the object entity o in the triple, respectively. After obtaining the set of triples for each text segment, the triples from all text segments are summarized to obtain the candidate triple set for that domain: ( 3); Where T represents the set of candidate triples extracted from all distilled texts in the domain; Distinguishing the union symbol; Based on the set of candidate entity nodes V extracted from all triples in the candidate triple set and the set of candidate relation edges E composed of triple relations, construct a candidate knowledge graph before deduplication and merging. : (4); in, This represents the candidate knowledge graph before deduplication and merging; Graph() represents the operation of generating a graph structure from the candidate triple set; V represents the set of candidate entity nodes extracted from all triples in the candidate triple set; E represents the set of candidate relation edges composed of triple relations. Generate text descriptions for the extracted entities and relations, represented as follows: P(V, E ) = {(k, v) | k ∈ labels(V, E ), v = M_sum(D[k])} (5); Where P represents the entity and relation description generation function; k represents the label of the entity node or relation edge; labels(V, E) represents the label set obtained from the candidate entity node set V and the candidate relation edge set E; D[k] represents the text fragment or context sentence related to label k; M_sum() represents the large language model summarization function; v represents the compact semantic description generated by the large language model; The domain knowledge graph G is constructed by deduplicating, merging, and normalizing duplicate entities, synonyms, and similar relationships that appear repeatedly in different text fragments. (6); Where G represents the domain knowledge graph; and represents the set of entity nodes and the set of relation edges after deduplication and merging, respectively; P(V, E) represents the set of entity-relation semantic descriptions obtained after describing the candidate set of entity nodes and the candidate set of relation edges; Dedupe() represents the entity and relation merging function, which is used to merge duplicate or similar nodes and edges across text segments based on entity labels, relation labels and their semantic descriptions.
6. The knowledge-guided intelligent chemical discovery method according to any one of claims 1-3, characterized in that, In step 2, the cross-domain knowledge reasoning agent interacts with researchers in the following manner: The cross-domain knowledge reasoning agent receives open scientific questions or candidate research directions corresponding to the target chemical discovery task input by researchers, generates a domain decomposition query plan based on the knowledge scope involved in the open scientific questions or candidate research directions, and determines the domain literature reading agent to be invoked based on the domain decomposition query plan. The cross-domain knowledge reasoning agent distributes the domain sub-problems in the domain decomposition query plan to the corresponding domain literature reading agents. It receives the corresponding domain retrieval enhancement answers returned by each domain literature reading agent after performing retrieval enhancement on the domain sub-problems based on the corresponding domain knowledge graph. The retrieval enhancement answers include relevant graph context, key relationship paths and reasoning basis. The cross-domain knowledge reasoning agent compares, summarizes, and integrates the enhanced search answers for the corresponding domains returned by the document reading agents of different domains to form a comprehensive scientific answer for researchers; After researchers confirm the research directions in the comprehensive scientific answers based on the target chemical discovery task, the cross-domain knowledge reasoning agent and researchers enter the next round of interaction, and gradually generate candidate research directions and scientific hypotheses through multiple rounds of human-computer collaborative question and answer; A cross-domain knowledge reasoning agent transforms scientific hypotheses into experimental designs and provides an initial set of experimental candidates based on the experimental designs.
7. The knowledge-guided intelligent chemical discovery method according to any one of claims 1-3, characterized in that, In step 5, real experimental data is incorporated into the historical experimental feedback set and fed back to the cross-domain knowledge reasoning agent in the following manner. The cross-domain knowledge reasoning agent then combines scientific hypotheses, experimental designs, the historical experimental feedback set, and cross-domain literature knowledge to generate the next round of experimental candidate sets, including: The current round of real experimental data is associated with the corresponding current experimental candidate set and incorporated into the historical experimental feedback set to obtain the updated historical experimental feedback set. The cross-domain knowledge reasoning agent combines scientific hypotheses, experimental designs, updated historical experimental feedback sets, and cross-domain literature knowledge to compare and analyze the candidates for each round of experiments and their actual experimental data, redetermine the basis for candidate designs and subsequent recommendation directions, and generate the next round of experimental candidate sets.
8. A knowledge-guided intelligent chemical discovery system for implementing the method of any one of claims 1-7, characterized in that, include: The knowledge graph construction module can acquire literature from multiple fields related to the target chemical discovery task and transform it into a corresponding domain knowledge graph. The scheme generation module can connect knowledge graphs from various fields to the corresponding literature reading agents, enabling them to perform enhanced retrieval and generation based on the corresponding graphs. The cross-domain knowledge reasoning agent calls on literature reading agents from various fields according to the target chemical discovery task input by the researchers, integrates the enhanced retrieval answers returned by the literature reading agents from various fields, and conducts multiple rounds of collaborative question and answer with the researchers to generate scientific hypotheses, experimental designs, and an initial set of experimental candidates. The results acquisition module can input the current set of experimental candidates into the on-demand autonomous chemistry research system, transform it into a robot autonomous experimental task and execute it to obtain real experimental data. The judgment module can obtain the researchers' judgment on whether the experimental candidates meet the expectations or whether the total experimental budget has been exhausted; if yes, the iteration stops; if no, the feedback optimization module is started. The feedback optimization module can incorporate real experimental data into the historical experimental feedback set and feed it back to the cross-domain knowledge reasoning agent. This cross-domain knowledge reasoning agent combines scientific hypotheses, experimental design, historical feedback, and cross-domain literature knowledge to generate the next round of experimental candidate set and send it to the result acquisition module.
9. A processing device, characterized in that, include: At least one memory for storing one or more programs; At least one processor is capable of executing one or more programs stored in the memory, such that when the one or more programs are executed by the processor, the processor can perform the method according to any one of claims 1-7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it can implement the method described in any one of claims 1-7.
Citation Information
Patent Citations
Robot artificial intelligence chemical home system driven by multiple agents
CN119005347A
Text inference method based on limited semantic dependency analysis
CN102360346A
Knowledge recall and innovation hypothesis generation method based on solid waste cementing material large language model
CN120492586A
