Substance scientific research experiment scheme generation method based on AI model

Abstract

The invention discloses a substance scientific research experiment scheme generation method based on an AI model, and relates to the technical field of model cooperation. The method comprises the following steps: calling a plurality of different AI model execution nodes in parallel on the basis of a directed acyclic graph architecture, so that the AI model execution nodes process scientific research questions at the same time to generate a plurality of model answers; the aggregation node stores the model answer and the corresponding session state; setting an interruption point in the calling process, displaying the model answers to the user at the interruption point, and receiving one or more of the model answers selected by the user as selected answers; the summary node uses a preset fusion cue word to guide the fusion AI model execution node to generate a comprehensive answer based on the selected answer; the problems that an existing single large model or serial multi-model scheme is uneven in answer quality and low in response efficiency in scientific research question answering, a user cannot participate in decision making, and a multi-source result lacks an efficient fusion mechanism are solved.

Description

Technical Field

[0001] This invention relates to the field of model collaboration technology, specifically to a method for generating material science experimental schemes based on AI models. Background Technology

[0002] In the generation of experimental schemes for material science research, multi-model collaboration to obtain comprehensive and accurate answers is a key step in improving research efficiency. However, in practical applications, existing technical solutions have revealed the following shortcomings: First, existing technologies rely on a limited number of models, either single or sequentially invoked. These models have inherent limitations in their knowledge scope, professional domain, and reasoning logic, failing to cover the coupled needs of complex scientific research problems across multiple dimensions such as material synthesis, characterization analysis, mechanism exploration, and process optimization. This results in incomplete and blind-spot-filled answers. Second, the current approach of sequentially invoking models leads to a linear increase in total time consumption. Furthermore, users are passively waiting throughout the process, unable to intervene or provide guidance, resulting in low response efficiency and a disconnect between user experience and overall performance. Finally, even after obtaining answers from multiple models, existing technologies lack an effective mechanism to automatically integrate multi-source information, retain core viewpoints, resolve content conflicts, and generate a logically coherent and comprehensive answer. The process still relies on tedious manual comparison and integration.

[0003] Therefore, there is an urgent need for a collaborative generation scheme that features parallel and efficient invocation, supports proactive user feedback, and can leverage the capabilities of large models to achieve intelligent fusion, in order to solve the aforementioned technical bottlenecks and improve the quality, efficiency, and controllability of material science experimental scheme generation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for generating material science experimental schemes based on AI models. This method solves the problems of inconsistent answer quality, low response efficiency, inability of users to participate in decision-making, and lack of efficient fusion mechanism for multi-source results in existing single large-scale model or serial multi-model schemes when answering scientific research questions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for generating material science experimental schemes based on AI models, comprising the following steps: The startup node receives research questions input by the user; Based on a directed acyclic graph architecture, several different AI model execution nodes are invoked in parallel, enabling the AI ​​model execution nodes to process the research problem simultaneously and generate multiple model answers. The aggregation node stores the model's answer and the corresponding session state; Set an interruption point in the call flow, and display the model answer to the user at the interruption point, and receive one or more answers selected by the user from the model answer as the selected answer; The summary node uses preset fusion prompts to guide the fusion AI model execution node to generate a comprehensive answer based on the selected answer, which serves as the final response to the research question.

[0006] Furthermore, the process of calling several different AI models in parallel based on the directed acyclic graph architecture is as follows: The startup node performs preprocessing and information initialization on the research questions. The starting node simultaneously triggers at least four AI model execution nodes to run in parallel, and each AI model execution node calls a corresponding large language model or knowledge base question-answering agent to generate an answer. The aggregation node waits for all model execution nodes to complete and then aggregates and saves all generated answers.

[0007] Furthermore, the four AI model execution nodes include a DeepSeek model execution node, a Spark model execution node, a ScienceOne model execution node, and a knowledge base proxy execution node.

[0008] Furthermore, when generating answers through the knowledge base agent execution node, the preprocessed and information-initialized research questions are queried in the knowledge base, and the content with a query score greater than a set score is used as the answer.

[0009] Furthermore, the DeepSeek model execution node, Spark model execution node, and ScienceOne model execution node all generate answers based on corresponding preset model prompts.

[0010] Furthermore, the knowledge base agent execution node generates answers based on preset knowledge base prompts.

[0011] Furthermore, the preset knowledge base prompts cause the knowledge base agent to use a numbered citation format in the answer and provide a list of cited references at the end of the answer.

[0012] Furthermore, the preset fusion prompts are used to guide the execution node of the fusion AI model to extract information and viewpoints from the selected answer; identify and merge duplicate content; and organize information points into a structured comprehensive answer according to logical relationships.

[0013] Furthermore, after generating a comprehensive answer, historical dialogue records related to the research question are loaded, and corresponding dialogue memories are built for each AI model execution node that will be invoked.

[0014] Furthermore, the process of constructing the corresponding dialogue memory is as follows: For historical dialogue records, if there is a fused and generated historical comprehensive answer in the historical dialogue record, then the historical comprehensive answer will be assigned as the assistant's reply content to all AI model execution nodes; If no historical comprehensive answer exists, then each AI model execution node extracts its own original answer generated in the historical dialogue record as the assistant's reply content; The user's questions in each historical dialogue record are combined with the assistant's responses assigned to the corresponding AI model execution node to form a corresponding dialogue memory.

[0015] The present invention has the following beneficial effects: This invention enables parallel invocation of multiple models using a directed acyclic graph, reducing the total time consumption from the cumulative time of a serial mode to approximately the time of a single invocation, thus significantly improving response efficiency. A user interruption feedback mechanism is introduced after parallel invocation, allowing researchers to intuitively compare the answers from different models and independently select the content to be fused, transforming users from passive recipients into decision-makers and enhancing the relevance and controllability of the solution. Abandoning complex traditional fusion algorithms, carefully designed prompts directly drive a large model to intelligently integrate the user-selected answers, automatically completing information extraction, deduplication, logical organization, and contradiction handling. This generates a more comprehensive, accurate, and structured final answer than a single model, simplifying system implementation while ensuring high-quality output.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0017] Figure 1 This is a flowchart of the material science experiment scheme generation method based on AI model of the present invention.

[0018] Figure 2 This is a schematic diagram of the directed acyclic graph architecture in the material science experiment scheme generation method based on the AI ​​model of the present invention. Detailed Implementation

[0019] Please see Figures 1-2 This invention provides a technical solution: a method for generating material science experimental schemes based on AI models, comprising the following steps: S1, the startup node receives research questions input by the user.

[0020] S2. Based on a directed acyclic graph architecture, several different AI model execution nodes are called in parallel, so that the AI ​​model execution nodes can process the research problem at the same time to generate multiple model answers.

[0021] It should be noted that the process of calling several different AI model execution nodes in parallel is as follows: The startup node performs preprocessing and information initialization on the research questions. The starting node simultaneously triggers at least four AI model execution nodes to run in parallel, and each AI model execution node calls a corresponding large language model or knowledge base question-answering agent to generate an answer. The aggregation node waits for all model execution nodes to complete and then aggregates and saves all generated answers.

[0022] Furthermore, the four AI model execution nodes include the DeepSeek model execution node, the Spark model execution node, the ScienceOne model execution node, and the knowledge base agent execution node.

[0023] The DeepSeek model execution node calls the DeepSeek large model, the Spark model execution node calls the iFlytek Spark model, the ScienceOne model execution node calls the Science One large model, and the knowledge base agent execution node calls the knowledge base question-answering agent. All four models run simultaneously, maximizing the utilization of computing resources, and timeout control is supported; the timeout of one model does not affect other models.

[0024] When generating answers through the knowledge base agent execution node, the research question after preprocessing and initializing information is queried in the knowledge base, and the content with a query score greater than the set score is used as the answer.

[0025] The DeepSeek, Spark, and ScienceOne model execution nodes all generate answers based on corresponding preset model hints. The knowledge base agent execution node generates answers based on preset knowledge base hints. These preset knowledge base hints cause the knowledge base agent to use a numbered citation format in the answers and provide a list of cited references at the end of the answers.

[0026] It should be noted that DeepSeek large models, iFlytek Spark models, and scientific large models all use the following preset model prompts: Provides accurate, professional, and timely scientific knowledge responses, ensuring scientific rigor and practical applicability; quickly interprets user query intent, with particular expertise in identifying research-related questions involving materials synthesis, experimental methods, process flows, instrument operation, and data analysis; integrates diverse knowledge sources to provide comprehensive and thorough solutions; uses standardized terminology and formal language to maintain academic rigor and accuracy; and organizes complex information into coherent, professional, and highly readable responses, including necessary formulas (using LaTeX).

[0027] The goal is to accurately understand users' questions and provide expert-level answers regarding research details, including material preparation, characterization techniques, reaction mechanisms, performance testing, and process parameters; generate responses from the knowledge base corresponding to the model to ensure the accuracy and scientific validity of the information; maintain a professional, academically rigorous, and practical response; and provide a friendly self-introduction for simple identity inquiries.

[0028] Constraints include: responses must be based entirely on the corresponding knowledge base and not rely on specific literature databases; they must be objective and neutral, providing objective and scientific information and avoiding subjective speculation; and responses should not begin with phrases such as "based on literature," "based on literature," or similar expressions.

[0029] The output format is as follows: Language: Uses professional, clear written language that matches the user's query language—Chinese queries are answered in Chinese, and English queries are answered in English; Structure: Presents answers in complete sentences and paragraphs to maintain logical flow; Formulas: Displays equations using standard LaTeX format; Provides only information that directly addresses the problem to maintain conciseness and clarity.

[0030] Example Explanation: Clearly identify typical research questions, such as how to prepare XX, the synthetic method of XX material, the selection of catalysts for XX reaction, or the operating parameters of XX instrument, and answer them appropriately. Only questions unrelated to research, academia, science, chemistry, physics, or biology—such as "Will it rain tomorrow?", "Do you think artificial intelligence is dangerous?", or "Recommend a restaurant?"—will trigger a rejection response.

[0031] The knowledge base question-answering agent's preset knowledge base prompts include relevant professional knowledge and, combined with potentially available knowledge base content, provide complete and accurate answers to user questions; the goal is to follow the user's input through each piece of information...<user_query> The instructions and requirements proposed by the label.

[0032] Whenever a user submits a question, relevant knowledge snippets retrieved from the local knowledge base are provided. These snippets will include...<rag_knowledge> Tag and answer based on the following principles: Use your own professional knowledge as the primary basis for your answer; when providing knowledge base content, prioritize the information from the knowledge base; if there is a conflict between the knowledge base content and your own knowledge, the knowledge base content shall prevail; ensure the completeness and professionalism of your answer; directly answer the user's question without mentioning whether knowledge base content was used; if the knowledge base content is irrelevant to the question, answer directly based entirely on your own professional knowledge and ensure the answer is detailed; do not suggest supplementary literature or other valid materials; the answer must be complete and detailed, with a length of no less than 1000 characters.

[0033] Organize your responses using Markdown format. Ensure the content is comprehensive, logically clear, professionally designed, and easy to understand. Use standard LaTeX format for mathematical formulas where necessary. You can enhance clarity by adding charts, formula metadata, or other visual elements.

[0034] In terms of communication, the candidate is required to quickly analyze query intent and be adept at identifying scientific research-related questions involving materials synthesis, experimental methods, process flow, instrument operation, and data analysis; be proficient in processing structured literature data (text, charts, formula metadata) and extract the most relevant information from databases; possess multi-turn dialogue capabilities and be able to handle complex queries that require integrating the content of multiple documents to answer; and ensure that the response language is consistent with the language used to ask the question—respond in Chinese if the user asks the question in Chinese, and in English if the user asks the question in English.

[0035] The constraints are as follows: Source of content: Please make comprehensive use of the provided literature knowledge base content and your own professional knowledge to provide a complete answer; Citation format: Use the numbered citation format (e.g., [1], [2]) to clearly distinguish between content cited from the database and knowledge generated by the model; References: When the content from the database is integrated into the reply, a list of references including the titles of the literature must be provided at the end of the reply; Objectivity: Only present the original information from the source and avoid adding subjective interpretations, comments or inferences; Prohibited opening phrases: Replies must not begin with statements such as "based on literature" or "based on literature".

[0036] Carefully break down the question, clarify the core query intent and keywords, and pay special attention to whether the question involves scientific research scenarios such as material preparation, chemical reaction, experimental steps, instrument use or data modeling; combine the question and the retrieved content (if any) to construct a complete answer framework; organically integrate the retrieved information with the corresponding knowledge to form a complete and accurate answer; arrange the information into a logically clear and comprehensive coherent reply; insert a numbered source citation mark (e.g., [1], [2]) after each key information point in the reply, and a list of references must be provided at the end; the citation numbers should be arranged in order from 1 to n; each reference must occupy a separate line.

[0037] S3. Once all AI model execution nodes have completed outputting answers or timed out, the aggregation node saves the model answers and the corresponding session states.

[0038] Before the model answer and the corresponding session state are sent to the breakpoint, all states are saved to persistent storage (Redis / database).

[0039] S4. Set an interruption point in the call flow, and display the model answer to the user at the interruption point, and receive one or more of the model answers selected by the user as the selected answer.

[0040] The breakpoint, also known as the selection node, displays the generated model answers to the user, allowing them to choose one or more. The user then marks the selected model answers as the chosen ones. Users have ample time to review each model answer, enabling flexible selection and customized output. Furthermore, there is no wasted computation during the breakpoint, and multiple iterations are supported.

[0041] S5. The summary node uses preset fusion prompts to guide the fusion AI model execution node to generate a comprehensive answer based on the selected answer, which serves as the final answer to the research question.

[0042] The preset fusion prompts are used to guide the execution nodes of the fusion AI model to extract information and viewpoints from the selected answer; identify and merge duplicate content; and organize information points into a structured comprehensive answer according to logical relationships.

[0043] First, selected answers are collected and constructed into a fused input, which is then output to the fused AI model. In this embodiment, the fused AI model is the DeepSeek model. The prompt words for the DeepSeek model when used as a fused AI model are: You are a professional intelligent summarization expert, responsible for integrating and refining answers from multiple large models to generate a comprehensive, accurate, and well-structured integrated response. Requirements include extracting key information and core viewpoints from the answers provided by multiple large models; identifying and merging duplicate content, retaining only the most valuable information; organizing scattered information points into a structured and easy-to-understand answer; and filtering for highly accurate content while eliminating contradictory or low-quality information.

[0044] The work objective is to synthesize answers from multiple models to produce a more comprehensive and accurate response than a single model; maintain professionalism and academic rigor to ensure the scientific validity of the information; deliver concise, logically clear, and user-friendly outputs; the workflow includes carefully reading the answers of each major model to understand its core viewpoints and supporting arguments; identifying points of consensus, differences, and complementary information among the answers from different models; organizing valuable information points into coherent answers based on logical relationships; and refining language to ensure that the answers are professional, accurate, and easy to understand.

[0045] Output requirements include using the same language as the user's question (answer Chinese questions in Chinese and English questions in English); using paragraph or bullet point format to highlight key content; covering all key points in the model's answer without omitting important information; using professional terminology and adhering to academic standards, and presenting formulas in LaTeX format when necessary; and basing the output on facts and scientific evidence, avoiding subjective assumptions.

[0046] Constraints include integrating only the content from the provided large model answers; not adding external information; presenting different viewpoints truthfully if there are obvious contradictions between the answers from multiple models; maintaining the scientific validity and accuracy of the original answers; not distorting their original meaning; and ensuring that the answers are concise and avoid redundancy and repetition.

[0047] After generating a comprehensive answer, historical dialogue records related to the research question are loaded, and corresponding dialogue memories are built for each AI model execution node to be invoked. The specific process of building the corresponding dialogue memories is as follows: For historical dialogue records, if there is a fused and generated historical comprehensive answer in the historical dialogue record, then the historical comprehensive answer will be assigned as the assistant's reply content to all AI model execution nodes; If no historical comprehensive answer exists, then each AI model execution node extracts its own original answer generated in the historical dialogue record as the assistant's reply content; The user's questions in each historical dialogue record are combined with the assistant's responses assigned to the corresponding AI model execution node to form a corresponding dialogue memory.

[0048] The execution sequence of this embodiment is as follows: First, the participants in the sequence diagram (11 in total): 1. User - End user; 2. API - HTTP entry gateway; 3. GraphService - Graph service coordination layer; 4. Database - Database storage; 5. PolyMindAgent - PolyMind multi-model agent; 6. DeepseekNode - DeepSeek model caller; 7. SparkNode - Spark model caller; 8. ScienceoneNode - ScienceOne model caller; 9. KnowledgeNode - Knowledge base query node; 10. ChoiceNode - User selection breakpoint node; 11. SummaryNode - Answer fusion node.

[0049] There are 7 execution phases in total, as follows: [Phase 1] Request Reception and Initialization: Participant interaction: User → API Gateway → Graph Service → Database; Key operations: The user sends a POST / graph / polymind / chat request, which includes query, chatId, and userId.

[0050] APIGateway receives the request and passes it to GraphService.

[0051] The GraphService initializes the execution context: it generates a unique callId (UUID); creates a conversation record in the Database; loads historical session messages into chatMemory; and constructs the input and RunnableConfig.

[0052] State changes: state={ chatId:"chat_xxx", callId:"call_uuid", query:"user's problem", chatMemory: [Message History List] }

[0053] Phase 2: Agent Startup Participant interaction: GraphService → PolyMindAgent Key operations: GraphService calls PolyMindAgent.execute(); PolyMindAgent starts the DAG state graph; the execution flow begins, branching to the four model nodes and the breakpoint.

[0054] Graph topology branches: PolyMindAgent→├→DeepseekNode ├→SparkNode ├→ScienceoneNode └→KnowledgeNode (parallel branch).

[0055] [Phase 3] 4 models are executed in parallel: Participant interaction: 4 model nodes execute in true parallel. Key features: DeepseekNode: Calls the DashScopeDeepSeek API and stores the returned result in `state.polymind_deepseek`; SparkNode: Calls the iFlytek Spark API and stores the returned result in `state.polymind_spark`; ScienceoneNode: Calls the ScienceOne model API and stores the returned result in `state.polymind_scienceone`; KnowledgeNode: Queries the internal knowledge base and stores the returned result in `state.polymind_knowledge`.

[0056] Parallelism: Four nodes execute simultaneously without blocking each other; each node has independent timeout control (usually 30 seconds); all results are simultaneously aggregated into the state.

[0057] State changes: state={ polymind_deepseek:"DeepSeek's response text", polymind_spark:"Spark's response text", polymind_scienceone:"ScienceOne's answer text", polymind_knowledge: "The text of the knowledge answer" }

[0058] Phase 4: Aggregation and State Preservation Participant interaction: PolyMindAgent → Database Key operations: Complete all 4 models and aggregate the results; save the complete state to Database or Redis; prepare for subsequent interruption recovery; the saved content includes: all state variables + the answers of the 4 models.

[0059] Phase 5: Breakpoints and User Selection Participant interaction: ChoiceNode⟷User(

INTERRUPT

[0060] [Stage 6] Answer Integration: Participant interaction: SummaryNode → LLM Key operations: SummaryNode extracts the user-selected answers from the state; constructs a fusion Prompt (POLYMIND_SUMMARY_PROMPT), which includes: multiple answers selected by the user; fusion method and output format requirements; calls the LLM fusion model (usually DeepSeek); generates a comprehensive answer and stores it in `state.answer`.

[0061] The responsibilities of a fusion prompt include: extracting key points; removing duplicate content; integrating complementary information; explaining different viewpoints; organizing the answer logically; and citing the sources of information.

[0062] [Phase 7] Result Output and Data Persistence: Participant interaction: SummaryNode → GraphService → Database → User Key operations: SummaryNode returns the final comprehensive answer; raphService receives the complete answer; pushes it to the User via SSE (Server-SentEvents); simultaneously saves it asynchronously to the Database: updates the call_history table, populates the answer field; records the execution time and status; saves the complete dialogue history; and displays the final answer to the User in real time.

[0063] A dual-table structure is used to manage conversation history and session metadata: The `chat_history` table (session table) manages session-level metadata. One `chatId` corresponds to one session, which includes multiple conversation rounds, as detailed in the table below: ; Index design: Primary key index: id; Business indexes: user_id, status.

[0064] The `call_history` table (dialogue table) stores detailed information about each user question and model response. One `callId` corresponds to one round of dialogue, as shown in the table below: ; Instructions for using the PolyMindAgent field: 1. The `content` field (output content for multiple models): Purpose: To store the raw output of four models; Format: JSON object; structure: { "deepseek": "The response content of the DeepSeek model..." "spark":"The response from the Spark model...", "scienceone": "The response content of the ScienceOne model..." "knowledge": "KnowledgeAgent's response..." }; Note: After PolyMindAgent completes the parallel calls to 4 models (stage 4), the data is immediately stored in the content field.

[0065] 2. The `answer` field (the merged answer) Purpose: To store the final answer selected by the user after fusion, and to build the chatMemory for the next round.

[0066] Format: Plain text string.

[0067] Two scenarios: Case 1: User selects fusion → answer = fused comprehensive answer; Case 2: User does not select fusion → answer = "" (empty string).

[0068] Note: After the user completes their selection (stage 6), the merged answer is stored.

[0069] 3. chatMemory loading rules: If `answer` is not empty, use `answer` as the AssistantMessage; if `answer` is empty, extract the model answer selected by the user from the content and concatenate them into the AssistantMessage.

[0070] Ensure that historical memories are available for each round of dialogue.

[0071] Field description: query: Stores the user's original question; content: PolyMind Dedicated - Stores the raw output (JSON format) of 4 models; Answer: PolyMind Dedicated - Final Answer After Storage Convergence (for chatMemory).

[0072] Index design: Primary key index: `id`; Foreign key index: `chat_id` (used for quickly querying all conversations under a certain session); Composite index: `(chat_id, create_time)` (used for loading historical records in chronological order).

[0073] Data table relationships: One session (chat_history) contains multiple conversations (call_history). Example data relationships: chat_history: ; call_history: ; A complete example of this embodiment is as follows: First round of dialogue (no historical memory); User action: The user asks a question for the first time.

[0074] Request: POST / graph / polymind / chat { "chatId":"chat_001", "query":"What is Fischer-Tropsch synthesis?" "userId":"user_1" } 1. Create a conversation log: INSERTINTOcall_history(id,chat_id,query,status,error,create_time) VALUES('call_001_01','chat_001','What is Fischer-Tropsch synthesis?',0,0,NOW()); 2. Load historical memories: SELECT * FROM call_history WHEREchat_id='chat_001'ANDstatus=0 ORDERBYcreate_timeASC; Result: [](empty list, no history for the first conversation).

[0075] 3. Construct Agent Input: input={ "chatId":"chat_001", "callId":"call_001_01", "query":"What is Fischer-Tropsch synthesis?" "chatMemory":[] / / Empty list }; 4. Execute PolyMindAgent: 4 models respond in parallel; user selects to merge; generate the final answer: "Fischer-Tropsch synthesis is an important chemical process...".

[0076] 5. Save your answer: UPDATEcall_history SETanswer='Fischer-Tropsch synthesis is an important chemical process...', WHEREid='call_001_01'; Database status: call_history: id chat_id query answer What is Fischer-Tropsch synthesis? Fischer-Tropsch synthesis is...

[0077] Second round of dialogue (no historical memory): User action: The user continues to ask questions.

[0078] Request: POST / graph / polymind / chat { "chatId":"chat_001", "query":"What are the main catalysts?" "userId":"user_1" } 1. Create a new conversation record INSERTINTOcall_history(id,chat_id,query,status,error,create_time) VALUES('call_001_02','chat_001','What are the main catalysts?',0,0,NOW()); 2. Load model-specific historical memories SELECT * FROM call_history WHEREchat_id='chat_001'ANDstatus=0 ORDERBYcreate_timeASC; result: [ {id:'call_001_01',query:'What is Fischer-Tropsch synthesis?',answer:'',content:'{"deepseek":"...","spark":"...",...}'}, {id:'call_001_02',query:'What are the main catalysts?',answer:null} ] 3. Convert to a model-specific message list modelMemories={ "deepseek":[ UserMessage("What is Fischer-Tropsch synthesis?") AssistantMessage("deepseek's response to Fischer-Tropsch synthesis...") ], "spark":[ UserMessage("What is Fischer-Tropsch synthesis?") AssistantMessage("Spark's response to Fischer-Tropsch synthesis...") ], "scienceone":[ UserMessage("What is Fischer-Tropsch synthesis?") AssistantMessage("scienceone's response to Fischer-Tropsch synthesis...") ], "knowledge":[ UserMessage("What is Fischer-Tropsch synthesis?") AssistantMessage("knowledge's response to Fischer-Tropsch synthesis...") ] } 4. Construct Agent Input input={ "chatId":"chat_001", "callId":"call_001_02", "query":"What are the main catalysts?" "chatMemory":[Previous round's history] / / Includes context } 5. Execute PolyMindAgent: The model uses chatMemory to understand the context; it knows that the user is asking "catalysts for Fischer-Tropsch synthesis"; and generates the answer: "Fischer-Tropsch synthesis mainly uses iron-based catalysts...".

[0079] 6. Save your answers: UPDATEcall_history; SETanswer='Fischer-Tropsch synthesis mainly uses iron-based catalysts...'; WHEREid='call_001_02'; Specific application examples Scenario: Collaborative answering of research questions.

[0080] User question: "What are the main mechanisms and industrial applications of Fischer-Tropsch synthesis?" Execution process: 1. Startup (polymind_start): Problem preprocessing and information initialization.

[0081] 2. Parallel Invocation: DeepSeek: Basic chemical principles of Fischer-Tropsch synthesis; ScienceOne: Latest research progress and mechanisms; Spark: Industrial application cases and processes; Knowledge: Relevant literature and patent information.

[0082] 3. User selection: The system presents: DeepSeek: "Fischer-Tropsch synthesis converts CO and H2 into..." through iron-based catalysts; ScienceOne: "Recent studies have shown that the activity of high-temperature iron catalysts..."; Spark: "Currently, most Fischer-Tropsch synthesis plants worldwide use the Shell process..."; Knowledge: "More than 2,500 related patents have been retrieved...".

[0083] User selection: [DeepSeek, ScienceOne, Spark].

[0084] 4. Summary and Conclusion: Input Fusion Prompt: { "deepseek":"Fischer-Tropsch synthesis is the process of converting CO and H2 into..." using iron-based catalysts." "scienceone": "Recent studies have shown that the activity of high-temperature iron catalysts...", "spark": "Currently, most Fischer-Tropsch synthesis plants worldwide use Shell's process..." } Large model fusion output: Fischer-Tropsch synthesis is an important chemical process, and its main mechanisms include: [Basic Principles] The synthesis gas (CO+H2) is converted into hydrocarbon compounds using iron-based catalysts... (A mechanism explanation combining DeepSeek and ScienceOne).

[0085] [Latest Developments] Recent studies have shown that high-temperature iron catalysts exhibit superior performance under nitride support... (from ScienceOne).

[0086] [Industrial Applications] Currently, the mainstream processes used globally include: 1. Shell process: fixed bed reactor, temperature 300-350°C...; 2. Sasol process: gas phase fluidized bed...; (from actual application of Spark)

[0087] 5. Output: The front end displays the integrated answer in real time; users can ask further questions to supplement their needs.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating material science experimental schemes based on AI models, characterized in that, Includes the following steps: The startup node receives research questions input by the user; Based on a directed acyclic graph architecture, several different AI model execution nodes are invoked in parallel, enabling the AI ​​model execution nodes to process the research problem simultaneously and generate multiple model answers. The aggregation node stores the model's answer and the corresponding session state; Set an interruption point in the call flow, and display the model answer to the user at the interruption point, and receive one or more answers selected by the user from the model answer as the selected answer; The summary node uses preset fusion prompts to guide the fusion AI model execution node to generate a comprehensive answer based on the selected answer, which serves as the final response to the research question.

2. The method for generating a material science experimental scheme based on an AI model according to claim 1, characterized in that, The process of calling several different AI models in parallel based on the directed acyclic graph architecture is as follows: The startup node performs preprocessing and information initialization on the research questions. The starting node simultaneously triggers at least four AI model execution nodes to run in parallel, and each AI model execution node calls a corresponding large language model or knowledge base question-answering agent to generate an answer. The aggregation node waits for all model execution nodes to complete and then aggregates and saves all generated answers.

3. The method for generating a material science experimental scheme based on an AI model according to claim 2, characterized in that, The four AI model execution nodes include the DeepSeek model execution node, the Spark model execution node, the ScienceOne model execution node, and the knowledge base proxy execution node.

4. The method for generating a material science experimental scheme based on an AI model according to claim 3, characterized in that, When generating answers through the knowledge base agent execution node, the research question after preprocessing and initializing information is queried in the knowledge base, and the content with a query score greater than the set score is used as the answer.

5. The method for generating a material science experimental scheme based on an AI model according to claim 3, characterized in that, The DeepSeek model execution node, Spark model execution node, and ScienceOne model execution node all generate answers based on corresponding preset model hint words.

6. The method for generating a material science experimental scheme based on an AI model according to claim 3, characterized in that, The knowledge base agent execution node generates answers based on preset knowledge base prompts.

7. The method for generating a material science experimental scheme based on an AI model according to claim 6, characterized in that, The preset knowledge base prompts cause the knowledge base agent to use a numbered citation format in the answer and provide a list of cited references at the end of the answer.

8. The method for generating a material science experimental scheme based on an AI model according to claim 1, characterized in that, The preset fusion prompts are used to guide the execution node of the fusion AI model to extract information and viewpoints from the selected answer; identify and merge duplicate content; and organize information points into a structured comprehensive answer according to logical relationships.

9. The method for generating a material science experimental scheme based on an AI model according to claim 1, characterized in that, After generating a comprehensive answer, the system loads historical dialogue records related to the research question and builds corresponding dialogue memories for each AI model execution node that will be invoked.

10. The method for generating a material science experimental scheme based on an AI model according to claim 9, characterized in that, The process of constructing the corresponding dialogue memory is as follows: For historical dialogue records, if there is a fused and generated historical comprehensive answer in the historical dialogue record, then the historical comprehensive answer will be assigned as the assistant's reply content to all AI model execution nodes; If no historical comprehensive answer exists, then each AI model execution node extracts its own original answer generated in the historical dialogue record as the assistant's reply content; The user's questions in each historical dialogue record are combined with the assistant's responses assigned to the corresponding AI model execution node to form a corresponding dialogue memory.

Images