Battery material synthesis path generation method

Abstract

The invention discloses a battery material synthesis path generation method, and relates to the technical field of battery material synthesis, and the method comprises the following steps: obtaining target battery material information; searching target battery material information in a preset battery material knowledge graph to obtain most similar battery material information; if the candidate battery material information is different from the target battery material information, predicting and generating a candidate synthesis path of the target battery material through an LLM intelligent agent; and evaluating a synthesis feasibility score of the candidate synthesis path, if the synthesis feasibility score does not meet a preset threshold value, regenerating the candidate synthesis path by using a reinforcement learning method, and returning to the step of analyzing the feasibility of the candidate synthesis path until the candidate synthesis path of which the synthesis feasibility score meets the preset threshold value is generated. According to the method, the knowledge graph is retrieved, and the synthetic path is generated by using the LLM (Large Language Model) intelligent agent and the reinforcement learning method, so that the accuracy of the synthetic path is improved.

Description

Technical Field

[0001] The present application relates to the technical field of battery material synthesis, and in particular to a method for generating a battery material synthesis path. Background Art

[0002] In the research and development and synthesis of battery materials, finding efficient synthesis pathways is a key step in ensuring material performance and reducing costs. Traditional methods typically rely on the experience of the experimenter and research advances in existing literature. This approach is not only time-consuming but also involves high trial-and-error costs.

[0003] In recent years, the application of artificial intelligence technologies, particularly big data and machine learning, has provided new approaches for predicting battery material synthesis pathways. Existing synthesis pathway prediction methods primarily rely on pre-set rule libraries or deep learning-based prediction models. While these methods can provide a certain degree of accuracy in specific situations, they still have limitations. For example, the high cost of updating the rule library and the deep learning model's strong reliance on data result in a lack of flexibility and accuracy in optimizing synthesis pathways.

[0004] How to improve the accuracy of the generated battery material synthesis pathway is an urgent problem to be solved. Summary of the Invention

[0005] The main purpose of this application is to provide a method for generating a battery material synthesis path, aiming to solve the technical problem of how to improve the accuracy of the generated battery material synthesis path.

[0006] To achieve the above objectives, the present application proposes a method for generating a battery material synthesis path, the method comprising:

[0007] Obtain target battery material information;

[0008] Retrieving the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information;

[0009] If the candidate battery material information is different from the target battery material information, the candidate synthesis path of the target battery material is generated through LLM intelligent agent prediction, and the synthesis feasibility score of the candidate synthesis path is evaluated. If the synthesis feasibility score does not meet the preset threshold, the candidate synthesis path is regenerated using the reinforcement learning method, and the step of analyzing the feasibility of the candidate synthesis path is returned until a candidate synthesis path whose synthesis feasibility score meets the preset threshold is generated.

[0010] In one embodiment, the step of evaluating the synthesis feasibility score of the candidate synthesis path, and if the synthesis feasibility score does not meet a preset threshold, regenerating the candidate synthesis path using a reinforcement learning method, and returning to the step of analyzing the feasibility of the candidate synthesis path until a candidate synthesis path having the synthesis feasibility score meeting the preset threshold is generated includes:

[0011] According to the target battery material information and the candidate battery material information, predicting the reactant combination for synthesizing the target battery material through the reactant prediction model in the LLM intelligent agent;

[0012] Based on the synthesis path generation model in the LLM agent, referring to the synthesis path steps of similar battery materials in the battery material knowledge graph, the target product synthesis path corresponding to the reactant combination is predicted and generated;

[0013] Obtaining a candidate synthesis path closest to the target product synthesis path in the battery material knowledge graph, evaluating the feasibility score of the generated target product synthesis path through the synthesis path evaluation model in the LLM agent, and providing a synthesis feasibility score;

[0014] If the synthetic feasibility score is not lower than the preset threshold, the generated candidate synthetic pathway for the target product is confirmed to be a highly feasible synthetic pathway;

[0015] If the synthesis feasibility score is lower than the preset threshold, the target product synthesis path is adjusted through the context-reinforced learning model based on the reward mechanism in the LLM agent until the synthesis feasibility score is not lower than the preset threshold, and the target product synthesis path is confirmed as a candidate synthesis path.

[0016] In one embodiment, if the synthesis feasibility score is lower than a preset threshold, adjusting the target product synthesis path through a context-reinforced learning model based on a reward mechanism in the LLM agent until the synthesis feasibility score is no lower than the preset threshold, before confirming the target product synthesis path as a candidate synthesis path includes:

[0017] Obtain historical experimental synthesis paths and their feasibility verification result datasets;

[0018] Inputting the historical experimental synthesis path into a context-reinforced learning model based on a reward mechanism, and optimizing the historical experimental synthesis path based on LLM agent prompts of the context-reinforced learning model to obtain an optimized synthesis path of the target material corresponding to the historical experimental synthesis path;

[0019] Performing a confidence analysis on the optimized synthesis path according to the feasibility verification result data set to obtain a confidence score value;

[0020] If the confidence score value does not reach the preset confidence threshold, the LLM agent prompt word of the context-reinforced learning model is optimized through the reward mechanism, and the historical experimental synthesis path is returned to the context-reinforced learning model based on the reward mechanism until the confidence score value reaches the confidence threshold, and the target product synthesis path is confirmed as a candidate synthesis path.

[0021] In one embodiment, the step of confirming the target product synthesis path as a candidate synthesis path until the synthesis feasibility score is not less than a preset threshold value includes:

[0022] Submitting the candidate synthetic pathway to a pre-set high-throughput experimental platform for synthetic experimental verification;

[0023] Recording the experimental process and experimental results returned by the high-throughput experimental platform to verify the candidate synthetic pathway;

[0024] The successfully verified target battery material synthesis path is converted into a knowledge subgraph, and the battery material knowledge graph is updated.

[0025] In one embodiment, the step of converting the successfully verified target battery material synthesis path into a knowledge subgraph and updating the battery material knowledge graph includes:

[0026] Perform integrity check on the newly added knowledge subgraph, and use the LLM agent to supplement the missing nodes and relationships in the knowledge subgraph;

[0027] Performing a correctness check on the newly added knowledge subgraph, and automatically repairing erroneous data in the knowledge subgraph through the LLM agent;

[0028] The newly added knowledge subgraph is integrated into the battery material knowledge graph to eliminate redundant knowledge subgraphs, nodes and relationships in the battery material knowledge graph.

[0029] In one embodiment, the step of searching the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information includes:

[0030] Analyze target battery material information;

[0031] If the target battery material information is a chemical expression, convert the target battery material information into a chemical structure fingerprint vector, calculate the Euclidean distance between the chemical structure fingerprint vector and the chemical structure fingerprint vector of the relevant battery material in the preset battery material knowledge graph, and return the most similar material synthesis path subgraph;

[0032] If the target battery material information is an attribute feature vector, the Euclidean distance between the attribute feature vector and the battery material attribute feature vector in the battery material knowledge graph is calculated, and the most similar material synthesis path subgraph is returned.

[0033] One or more technical solutions proposed in this application have at least the following technical effects:

[0034] Compared with the related art, which mainly relies on a preset rule base or a prediction model based on deep learning, the present application obtains target battery material information; retrieves the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information; if the candidate battery material information is different from the target battery material information, the candidate synthesis path of the target battery material is generated by LLM agent prediction, and the synthesis feasibility score of the candidate synthesis path is evaluated. If the synthesis feasibility score does not meet the preset threshold, the candidate synthesis path is regenerated using the reinforcement learning method, and the step of analyzing the feasibility of the candidate synthesis path is returned until the candidate synthesis path whose synthesis feasibility score meets the preset threshold is generated. It can be understood that the present application uses information based on the battery material knowledge graph to calculate similar battery materials, and predicts reactant combinations and synthesis paths through LLM agents. When the target battery material information is received, the battery material knowledge graph and LLM agent predict a candidate synthesis path with high feasibility, thereby improving the accuracy of the generated battery material synthesis path. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0036] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0037] Figure 1 A schematic diagram of a process flow diagram provided for Example 1 of the method for generating a synthetic pathway for battery materials of this application;

[0038] Figure 2 A schematic diagram of the process flow provided for Example 2 of the method for generating a synthetic pathway for battery materials of this application;

[0039] Figure 3 A schematic diagram of the process flow provided in Example 3 of the method for generating a synthetic pathway for battery materials of this application;

[0040] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0041] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0042] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0043] The main solutions of the embodiments of this application are:

[0044] Obtain target battery material information;

[0045] Retrieving the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information;

[0046] If the candidate battery material information is different from the target battery material information, a candidate synthesis path for the target battery material is generated through LLM agent prediction;

[0047] Evaluate the synthesis feasibility score of the candidate synthesis path; if the synthesis feasibility score does not meet the preset threshold, regenerate the candidate synthesis path using the reinforcement learning method, and return to the step of analyzing the feasibility of the candidate synthesis path until a candidate synthesis path is generated whose synthesis feasibility score meets the preset threshold.

[0048] In this embodiment, the present application uses a battery material synthesis path generation device as the execution body. For the convenience of description, it is specifically described below as "device".

[0049] Since existing technologies mainly rely on preset rule bases or prediction models based on deep learning;

[0050] This application provides a solution based on the battery material knowledge graph and LLM agent, which obtains the target battery material information, parses its chemical expression or property characteristics, calculates the information of similar battery materials, and then predicts the reactant combination and synthesis path, and performs multi-dimensional scoring through the synthesis path evaluation model. If the path score is lower than the preset threshold, it is adjusted by the context reinforcement learning model. The predicted synthesis path is submitted to the high-throughput experimental platform for verification, and the successful path updates the knowledge graph. The path that fails to pass the verification is optimized by the reinforcement learning model and marked as a negative case. The knowledge graph is also updated, thereby continuously optimizing the generation logic of the path generation model to ensure the accuracy and practicality of the generated synthesis path. This process realizes the automation of the entire process from the acquisition of target material information to the generation of synthesis path, experimental verification, path optimization and knowledge graph update, thereby improving the efficiency and success rate of battery material research and development.

[0051] Based on this, the present invention provides a method for generating a battery material synthesis path, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the method for generating a battery material synthesis path of the present application.

[0052] In this embodiment, the battery material synthesis path generation method includes steps S10 to S40:

[0053] Step S10, obtaining target battery material information;

[0054] It should be noted that the target battery material information is the specific information of the battery material desired to be synthesized, which can be a chemical expression or a property feature vector.

[0055] Step S20, searching the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information;

[0056] It should be noted that the preset battery material knowledge graph is a structure that stores a large amount of battery material information, including the material's chemical expression, structural fingerprint, property characteristics and other information, and is used to retrieve materials similar to the target battery material.

[0057] It is understandable that, assuming that the target battery material is lithium iron phosphate (LiFePO4), its chemical expression is LiFePO4. The device first converts the chemical expression of LiFePO4 into a structural fingerprint vector. Then, in the preset battery material knowledge graph, the Euclidean distance between the structural fingerprint vector and other battery materials stored in the graph is calculated. The device returns a knowledge subgraph of the k battery materials and their synthesis paths that are most similar to the LiFePO4 structure. These similar materials may include different lithium iron phosphate variants or other similar phosphate materials. In this way, the device can quickly find the material and its synthesis path that are most similar to the target material, providing a basis for subsequent synthesis path generation.

[0058] This process demonstrates the device's ability to use knowledge graphs for efficient retrieval and precise matching, providing strong support for the design of synthesis pathways for battery materials.

[0059] In a feasible embodiment, the step of searching the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information includes:

[0060] Analyze target battery material information;

[0061] If the target battery material information is a chemical expression, convert the target battery material information into a chemical structure fingerprint vector, calculate the Euclidean distance between the chemical structure fingerprint vector and the chemical structure fingerprint vector of the relevant battery material in the preset battery material knowledge graph, and return the most similar material synthesis path subgraph;

[0062] If the target battery material information is an attribute feature vector, the Euclidean distance between the attribute feature vector and the battery material attribute feature vector in the battery material knowledge graph is calculated, and the most similar material synthesis path subgraph is returned.

[0063] It should be noted that the chemical structure fingerprint vector converts the chemical expression of the target battery material into a numerical one-dimensional vector representation for calculating structural similarity.

[0064] The property feature vector is a feature vector constructed by the performance parameters of the material (such as capacity, cycle stability, etc.) and is used to calculate performance similarity.

[0065] Euclidean distance is used to measure the similarity between two vectors. The smaller the distance, the higher the similarity.

[0066] The material synthesis path subgraph is the synthesis path information of a certain battery material stored in the knowledge graph, including reactants, operation steps, etc.

[0067] For example, referring to Figure 2, assuming that the target battery material is lithium iron phosphate (LiFePO4), and its chemical expression is LiFePO4. The device first parses the target battery material information and determines that it is a chemical expression. Then, the chemical expression of LiFePO4 is converted into a structural fingerprint vector. In the preset battery material knowledge graph, the device calculates the Euclidean distance between the structural fingerprint vector and other battery materials stored in the graph. The device returns the k battery materials that are most similar to the LiFePO4 structure and their synthesis path subgraphs. These similar materials may include different lithium iron phosphate variants or other similar phosphate materials.

[0068] If the target battery material information is an attribute feature vector, such as a capacity of 160 mAh / g, a cycle stability of 90%, etc., the device calculates the Euclidean distance between the attribute feature vector and the battery material stored in the knowledge graph, and returns the most similar material synthesis path subgraph.

[0069] This process demonstrates the device's ability to quickly retrieve similar materials using structural fingerprints or property feature vectors, providing an efficient foundation for subsequent synthesis path generation. In this way, the device can leverage the rich data in the knowledge graph to quickly find the materials most similar to the target material and their synthesis paths, thereby accelerating the design of synthesis paths for new materials.

[0070] For example, after the user inputs the fitted battery material representation information (such as chemical expressions, property descriptions, etc.), the similar material retrieval agent searches the knowledge graph for the most similar materials and their synthesis pathways. The agent supports the following two retrieval methods:

[0071] Structural fingerprint-based structural similarity retrieval: First, the chemical expression of the target battery material to be generated is converted into a numerical one-dimensional vector representation of its structural fingerprint. Then, the Euclidean distance between this vector and the battery material structural fingerprint stored in the knowledge graph and the battery material structural fingerprint to be generated is calculated, thereby obtaining the synthesis path knowledge subgraph of the k battery materials with the most similar structures (i.e., the smallest distance).

[0072] Material performance similarity retrieval based on multiple attributes: construct a feature vector representing the attribute value of the battery material to be generated based on its attribute value, and then calculate the Euclidean distance between this vector and the battery material attributes stored in the knowledge graph and the battery material attributes to be generated, so as to obtain the synthesis path knowledge subgraph of k battery materials with the most similar material properties (i.e., the smallest distance).

[0073] If a battery material with identical structural fingerprints and properties is found, the synthesis path is directly output and the process jumps to performing synthetic experimental verification. Otherwise, the knowledge subgraph of the k most similar materials found is used as the initial knowledge subgraph for generating the synthesis path of the target battery material. The system core module then uses the synthesis path generation agent to generate the synthesis path of the target battery material.

[0074] Step S30: If the candidate battery material information is different from the target battery material information, a candidate synthesis path of the target battery material is generated through LLM agent prediction;

[0075] It should be noted that the candidate synthesis path is a battery material synthesis path generated by the LLM agent and needs to be verified through experiments.

[0076] Step S40, evaluating the synthesis feasibility score of the candidate synthesis path. If the synthesis feasibility score does not meet the preset threshold, regenerating the candidate synthesis path using the reinforcement learning method, and returning to the step of analyzing the feasibility of the candidate synthesis path until a candidate synthesis path is generated whose synthesis feasibility score meets the preset threshold.

[0077] In a feasible embodiment, the step of evaluating the synthesis feasibility score of the candidate synthesis path, and if the synthesis feasibility score does not meet a preset threshold, regenerating the candidate synthesis path using a reinforcement learning method, and returning to the step of analyzing the feasibility of the candidate synthesis path until a candidate synthesis path having the synthesis feasibility score meeting the preset threshold is generated includes:

[0078] According to the target battery material information and the candidate battery material information, predicting the reactant combination for synthesizing the target battery material through the reactant prediction model in the LLM intelligent agent;

[0079] Based on the synthesis path generation model in the LLM agent, referring to the synthesis path steps of similar battery materials in the battery material knowledge graph, the target product synthesis path corresponding to the reactant combination is predicted and generated;

[0080] Obtaining a candidate synthesis path closest to the target product synthesis path in the battery material knowledge graph, evaluating the feasibility score of the generated target product synthesis path through the synthesis path evaluation model in the LLM agent, and providing a synthesis feasibility score;

[0081] If the synthetic feasibility score is not lower than the preset threshold, the generated candidate synthetic pathway for the target product is confirmed to be a highly feasible synthetic pathway;

[0082] If the synthesis feasibility score is lower than the preset threshold, the target product synthesis path is adjusted through the context-reinforced learning model based on the reward mechanism in the LLM agent until the synthesis feasibility score is not lower than the preset threshold, and the target product synthesis path is confirmed as a candidate synthesis path.

[0083] It should be noted that the LLM agent refers to an intelligent system that uses the LLM (Large Language Model) agent as its core component, combined with modules such as planning, memory, and tool calling, and can autonomously perceive the environment, make decisions, and perform complex tasks.

[0084] The reactant prediction model is a model that predicts the reactant combination required to synthesize the target battery material based on the target battery material information and candidate battery material information.

[0085] The synthesis path generation model is a model that generates the synthesis path of the target battery material based on the reactant combination and candidate battery material information.

[0086] The synthetic pathway evaluation model is a model that performs multi-dimensional feasibility scoring on the generated synthetic pathways.

[0087] The context-enhanced learning model is a model that can optimize the contextual prompt words that generate synthesis paths through deep learning.

[0088] For example, when no complete and consistent battery material synthesis path information is found, the LLM agent module for generating the target battery material synthesis path is started, and a valid target battery material synthesis path is generated according to the following steps:

[0089] Based on the chemical expression or property description of the retrosynthetic battery material input by the user and the synthesis path knowledge subgraphs of the k most similar battery materials in the searched knowledge graph, the reactant combination of the target battery material is predicted.

[0090] Based on the different reactant combinations selected and the synthesis experimental steps of k similar battery materials in the searched knowledge graph, the synthesis experimental operation steps of the target battery material are generated, and the adoptable synthesis experimental schemes are elaborated in detail.

[0091] A knowledge subgraph matching algorithm is used to retrieve a successful synthetic pathway from the knowledge graph that best matches the automatically generated pathway, and the similarity between the two pathways is calculated. Furthermore, a feasibility analysis is performed on all generated synthetic pathways based on factors such as reactant availability, experimental time expenditure, and experimental difficulty. Based on a user-defined feasibility score threshold, pathways with feasibility exceeding that threshold are output, along with their feasibility scores and supporting explanations.

[0092] If the synthesis path similarity is greater than the set threshold, the generated synthesis path is considered feasible and will be output to the high-throughput experimental platform for experimental verification. Otherwise, it is considered invalid and a context-based reinforcement learning method based on a reward mechanism is used to regenerate a new synthesis path using the output synthesis path.

[0093] In a feasible embodiment, if the synthesis feasibility score is lower than a preset threshold, the target product synthesis path is adjusted by the context-reinforced learning model based on the reward mechanism in the LLM agent until the synthesis feasibility score is not lower than the preset threshold. Before the step of confirming the target product synthesis path as a candidate synthesis path includes:

[0094] Obtain historical experimental synthesis paths and their feasibility verification result datasets;

[0095] Inputting the historical experimental synthesis path into a context-reinforced learning model based on a reward mechanism, and optimizing the historical experimental synthesis path based on the large language model prompt words of the context-reinforced learning model to obtain an optimized synthesis path of the target material corresponding to the historical experimental synthesis path;

[0096] Performing a confidence analysis on the optimized synthesis path according to the feasibility verification result data set to obtain a confidence score value;

[0097] If the confidence score value does not reach the preset confidence threshold, the large language model prompt word of the context-reinforced learning model is optimized through the reward mechanism, and the historical experimental synthesis path is returned to the context-reinforced learning model based on the reward mechanism until the confidence score value reaches the confidence threshold, and the target product synthesis path is confirmed as a candidate synthesis path.

[0098] It should be noted that the reward mechanism is a mechanism used to evaluate and optimize the synthesis path generation model. The reward mechanism can guide the generation of more feasible synthesis paths.

[0099] For example, suppose the device generates multiple synthetic pathways to be verified and submits them to the high-throughput experimental platform for synthetic experiments. The experimental results show that one of the pathways fails verification due to the reaction temperature being set too high, resulting in substandard product purity.

[0100] After the device identifies the failed validation path, it activates a reward mechanism and a context-reinforced learning model. Based on historical successes and failures, the reward mechanism assesses the path's shortcomings and provides optimization suggestions, such as lowering the reaction temperature. The context-reinforced learning model uses these suggestions to adjust the generation logic of the synthesis path generation model and optimize the contextual prompts to regenerate an optimized synthesis path.

[0101] At the same time, the device marks the failed verification path as a negative case and adds it to the battery material knowledge graph. The negative case contains detailed failure reasons and optimization suggestions, helping the device avoid repeating errors when generating similar paths in the future.

[0102] This process demonstrates how the device optimizes the generation logic through reward mechanisms and context-reinforced learning models, and uses negative cases to update the knowledge graph, thereby continuously improving the feasibility and reliability of the synthesis path.

[0103] It is understood that the reward mechanism is trained and continuously learned using a dataset consisting of data from historical validation experiments and their verification results. The trained and tuned reward mechanism is then used to optimize the LLM prompts when regenerating synthesis paths whose feasibility falls below the target threshold, guiding the LLM to regenerate a more feasible target battery synthesis path. In this way, by introducing historically generated synthesis paths as contextual cases and implementing contextual reinforcement learning based on the reward mechanism, the path generation capability of the synthesis path generation LLM agent is improved.

[0104] In a feasible embodiment, the step of confirming the target product synthesis path as a candidate synthesis path until the synthesis feasibility score is not lower than a preset threshold value includes:

[0105] Submitting the candidate synthetic pathway to a pre-set high-throughput experimental platform for synthetic experimental verification;

[0106] Recording the experimental process and experimental results returned by the high-throughput experimental platform to verify the candidate synthetic pathway;

[0107] The successfully verified target battery material synthesis path is converted into a knowledge subgraph, and the battery material knowledge graph is updated.

[0108] It should be noted that a high-throughput experimental platform is an automated platform that can quickly perform multiple experiments to verify the feasibility and effectiveness of the synthetic pathway.

[0109] A knowledge subgraph is structured data representing a single experimental protocol and its results, and is used to update and enrich the knowledge graph.

[0110] For example, assume that the device generates multiple synthetic pathways to be verified and submits these pathways to the high-throughput experimental platform for synthetic experiments. The high-throughput experimental platform automatically executes these synthetic experiments, recording the detailed experimental process (such as reaction conditions, operation steps, etc.) and experimental results (such as product purity, performance indicators, etc.).

[0111] After the experiment is complete, the device verifies the results to identify which synthesis pathways can successfully synthesize the target material. For successfully verified synthesis pathways, the device converts them into knowledge subgraphs and updates the battery materials knowledge graph. The knowledge subgraph contains detailed information about the synthesis pathway, such as reactants, reaction conditions, operation steps, and experimental results.

[0112] This process demonstrates the device's ability to rapidly verify synthetic pathways through a high-throughput experimental platform and to continuously enrich and optimize the information in the knowledge graph by converting successfully verified pathways into knowledge subgraphs. This not only improves the efficiency of new material synthesis but also enhances the device's ability to generate more reliable synthetic pathways in the future.

[0113] In a feasible embodiment, the step of converting the successfully verified target battery material synthesis path into a knowledge subgraph and updating the battery material knowledge graph includes:

[0114] Perform integrity check on the newly added knowledge subgraph, and use the LLM agent to supplement the missing nodes and relationships in the knowledge subgraph;

[0115] Performing a correctness check on the newly added knowledge subgraph, and automatically repairing erroneous data in the knowledge subgraph through the LLM agent;

[0116] The newly added knowledge subgraph is integrated into the battery material knowledge graph to eliminate redundant knowledge subgraphs, nodes and relationships in the battery material knowledge graph.

[0117] It should be noted that the integrity check is to check whether the data in the knowledge subgraph is complete and whether there are any missing or errors.

[0118] For example, assume that in an experiment, the device automatically generates a knowledge subgraph containing detailed information of the synthesis path, such as reactants, reaction conditions (temperature, time), and experimental results (product purity, performance indicators, etc.).

[0119] Before adding this knowledge subgraph to the battery materials knowledge graph, the device first performed a completeness check. This check revealed that the experimental steps were missing information about the order in which the reactants were mixed, which could cause confusion for other researchers trying to replicate the experiment. The device then used an LLM agent to automatically complete the information about the reactant mixing order based on the existing knowledge graph and records of similar experiments, correcting the missing data.

[0120] Next, the device uses knowledge graph fusion technology to compare and fuse the newly generated knowledge subgraph with existing knowledge subgraphs for similar materials in the knowledge graph. The device discovers that a knowledge subgraph for the new target material already exists in the knowledge graph, containing similar reaction conditions and experimental steps. Using fusion technology, the device integrates the duplicate information in the two knowledge subgraphs, removes redundancy, and updates the knowledge graph with the new experimental results and optimized steps.

[0121] This process demonstrates how the device ensures that the data in the knowledge graph is accurate, complete and non-redundant through integrity checking and knowledge graph fusion technology, thereby continuously improving the quality and practicality of the knowledge graph.

[0122] It is understood that in order to effectively utilize high-throughput experimental results data to enhance the domain knowledge graph and the LLM agent's ability to automatically generate battery material synthesis paths, the detailed process and results of high-throughput synthesis experiments are recorded and stored, and converted into structured data and stored in a relational database as historical archive data of historical battery material synthesis processes. At the same time, a knowledge subgraph representing the experimental plan and results is invented based on each experimental process, so that each synthesis experiment operation process and its results are added as prior knowledge to the existing knowledge graph.

[0123] Since the fitting materials, reactants, synthesis steps, etc. of each experiment are missing, repeated or similar, a technology that supports knowledge subgraph inspection, completion and fusion is proposed. First, the integrity and correctness of the knowledge subgraph composed of the data recorded in each synthesis experiment process is checked to find missing or incorrect data. The LLM intelligent agent driven by the existing knowledge graph is used to automatically complete or correct the missing or incorrect experimental record data to obtain a knowledge subgraph that has passed the inspection. Secondly, the knowledge graph fusion technology is used to add each newly generated knowledge subgraph to the existing knowledge graph, continuously enriching the existing battery material synthesis experiment records stored in the knowledge graph, and enhancing the performance of the knowledge graph in improving the LLM intelligent agent.

[0124] This embodiment provides a method for generating a battery material synthesis path. This method uses a battery material knowledge graph to calculate information about similar battery materials and uses an LLM agent to predict reactant combinations and synthesis paths. Upon receiving target battery material information, the battery material knowledge graph and LLM agent predict reactant combinations and the target power synthesis path, enabling accurate generation of the battery material synthesis path.

[0125] For example, in order to help understand the implementation process of the battery material synthesis path generation method obtained by combining this embodiment with the above embodiment 1, please refer to Figure 3 , Figure 3A brief flow chart of a method for generating a battery material synthesis pathway is provided, specifically:

[0126] The LLM agent system architecture proposed in this application for battery material synthesis path generation includes the following components:

[0127] 1. LLM agent for similar material retrieval: Based on the battery material information fitted into the user input, such as chemical expressions and property descriptions, it uses methods such as structure fingerprints and material properties to search for the most similar knowledge subgraphs of materials and their synthesis paths in the knowledge graph.

[0128] 2. Generate the LLM intelligent agent module for the target battery material synthesis path. This module is the core component of this system and includes four LLM intelligent agents: reactant selection, operation step generation, synthesis scheme evaluation, and context-reinforced learning based on reward mechanism.

[0129] 3. High-throughput experiments to verify the synthesis paths generated by the LLM agent. These verification synthesis experiments are automatically completed based on the high-throughput experimental platform.

[0130] 4. The Experimental Process Recording and Knowledge Graph Update Module includes an experimental record database, a knowledge subgraph construction agent, and a knowledge graph update agent. The system automatically writes descriptions of the experimental process and results into the experimental record database. It then uses the knowledge subgraph construction and knowledge graph update agents to construct a knowledge subgraph for each newly added synthetic path verification experiment and adds this subgraph to the existing knowledge graph, expanding the domain knowledge graph.

[0131] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the method for generating a synthesis path for battery materials of the present application. Simple transformations in more forms based on this technical concept are all within the scope of protection of the present application.

[0132] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for generating a battery material synthesis path, characterized in that: The method includes: Obtain target battery material information; Retrieving the target battery material information in a preset battery material knowledge graph to obtain the most similar battery material information; If the candidate battery material information is different from the target battery material information, a candidate synthesis path for the target battery material is generated through LLM agent prediction; Evaluate the synthesis feasibility score of the candidate synthesis path; if the synthesis feasibility score does not meet the preset threshold, regenerate the candidate synthesis path using the reinforcement learning method, and return to the step of analyzing the feasibility of the candidate synthesis path until a candidate synthesis path is generated whose synthesis feasibility score meets the preset threshold.

2. The method according to claim 1, wherein The step of evaluating the synthesis feasibility score of the candidate synthesis path, and if the synthesis feasibility score does not meet a preset threshold, regenerating the candidate synthesis path using a reinforcement learning method, and returning to the step of analyzing the feasibility of the candidate synthesis path until a candidate synthesis path having the synthesis feasibility score meeting the preset threshold is generated comprises: According to the target battery material information and the candidate battery material information, predicting the reactant combination for synthesizing the target battery material through the reactant prediction model in the LLM intelligent agent; Based on the synthesis path generation model in the LLM agent, referring to the synthesis path steps of similar battery materials in the battery material knowledge graph, the target product synthesis path corresponding to the reactant combination is predicted and generated; Obtaining a candidate synthesis path closest to the target product synthesis path in the battery material knowledge graph, evaluating the feasibility score of the generated target product synthesis path through the synthesis path evaluation model in the LLM agent, and providing a synthesis feasibility score; If the synthetic feasibility score is not lower than the preset threshold, the generated candidate synthetic pathway for the target product is confirmed to be a highly feasible synthetic pathway; If the synthesis feasibility score is lower than the preset threshold, the target product synthesis path is adjusted through the context-reinforced learning model based on the reward mechanism in the LLM agent until the synthesis feasibility score is not lower than the preset threshold, and the target product synthesis path is confirmed as a candidate synthesis path.

3. The method according to claim 2, wherein If the synthesis feasibility score is lower than a preset threshold, the target product synthesis path is adjusted by a context-reinforced learning model based on a reward mechanism in the LLM agent until the synthesis feasibility score is not lower than the preset threshold. Before the step of confirming the target product synthesis path as a candidate synthesis path includes: Obtain historical experimental synthesis paths and their feasibility verification result datasets; Inputting the historical experimental synthesis path into a context-reinforced learning model based on a reward mechanism, and optimizing the historical experimental synthesis path based on LLM agent prompts of the context-reinforced learning model to obtain an optimized synthesis path of the target material corresponding to the historical experimental synthesis path; Performing a confidence analysis on the optimized synthesis path according to the feasibility verification result data set to obtain a confidence score value; If the confidence score value does not reach the preset confidence threshold, the LLM agent prompt word of the context-reinforced learning model is optimized through the reward mechanism, and the historical experimental synthesis path is returned to the context-reinforced learning model based on the reward mechanism until the confidence score value reaches the confidence threshold, and the target product synthesis path is confirmed as a candidate synthesis path.

4. The method according to claim 2, wherein The step of confirming the target product synthesis path as a candidate synthesis path until the synthesis feasibility score is not lower than a preset threshold value includes: Submitting the candidate synthetic pathway to a pre-set high-throughput experimental platform for synthetic experimental verification; Recording the experimental process and experimental results returned by the high-throughput experimental platform to verify the candidate synthetic pathway; The successfully verified target battery material synthesis path is converted into a knowledge subgraph, and the battery material knowledge graph is updated.

5. The method according to claim 4, wherein The step of converting the successfully verified target battery material synthesis path into a knowledge subgraph and updating the battery material knowledge graph includes: Perform integrity check on the newly added knowledge subgraph, and use the LLM agent to supplement the missing nodes and relationships in the knowledge subgraph; Performing a correctness check on the newly added knowledge subgraph, and automatically repairing erroneous data in the knowledge subgraph through the LLM agent; The newly added knowledge subgraph is integrated into the battery material knowledge graph to eliminate redundant knowledge subgraphs, nodes and relationships in the battery material knowledge graph.

6. The method according to claim 1, wherein The step of searching the target battery material information in the preset battery material knowledge graph to obtain the most similar battery material information includes: Analyze target battery material information; If the target battery material information is a chemical expression, convert the target battery material information into a chemical structure fingerprint vector, calculate the Euclidean distance between the chemical structure fingerprint vector and the chemical structure fingerprint vector of the relevant battery material in the preset battery material knowledge graph, and return the most similar material synthesis path subgraph; If the target battery material information is an attribute feature vector, the Euclidean distance between the attribute feature vector and the battery material attribute feature vector in the battery material knowledge graph is calculated, and the most similar material synthesis path subgraph is returned.

Images