Method for predicting catalytic dehydrogenation performance of organic hydrogen storage liquid based on large language model

Abstract

This application discloses a method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model. This application proposes an intelligent prediction and recommendation framework that integrates a large language model and domain knowledge retrieval. The RAG module matches input conditions and catalyst information with highly relevant content in the knowledge base, and combines this with multimodal feature analysis and latent pattern reasoning using a large language model to generate predicted catalytic performance values ​​for the target LOHC system, as well as recommended catalyst composition and reaction conditions. This application is adaptable to various LOHC systems, including methylcyclohexane, demethylnaphthalene, and diphenylmethane, achieving cross-system, high-precision performance prediction and optimization.

Description

Technical Field

[0001] This application relates to the technical field of catalyst process performance prediction and recommendation and its integration with large language models, and in particular to a method for optimizing the catalytic dehydrogenation performance of organic liquid hydrogen storage by integrating large language models and domain knowledge retrieval. Background Technology

[0002] Organic liquid hydrogen storage (LOHC) technology, by reversibly storing hydrogen in liquid organic compounds, offers advantages such as high hydrogen storage density, safety and controllability, and ease of storage and transportation, making it a key pathway for large-scale hydrogen energy storage, transportation, and peak shaving. Typical LOHC systems include methylcyclohexane / toluene, demethylnaphthalene / naphthalene, and diphenylmethane / biphenyl. Catalytic dehydrogenation is the core step in releasing hydrogen, and its efficiency and selectivity are influenced by multiple factors, including catalyst type, support properties, auxiliary agents, and reaction conditions.

[0003] In practical research, LOHC dehydrogenation experiments are costly and time-consuming, and the data distribution and experimental conditions vary significantly across different systems. This results in traditional single machine learning models having insufficient accuracy and limited generalization ability in cross-system predictions. In recent years, Large Language Models (LLMs) have demonstrated the ability to jointly understand, reason about, and generate knowledge from multi-source unstructured and structured data. By introducing Retrieval Enhancement (RAG) mechanisms, high-value information from domain knowledge bases can be dynamically injected into the LLM inference process, enabling intelligent prediction and mechanism analysis in scenarios with insufficient data and missing features. However, current technologies still lack a large language model framework and prediction methods for the catalytic dehydrogenation of organic liquid hydrogen storage. Summary of the Invention

[0004] To address the challenges of sample scarcity, data heterogeneity, and insufficient cross-system generalization in predicting and recommending the catalytic dehydrogenation performance of organic liquid hydrogen storage (LOHC), this application provides a method for predicting the catalytic dehydrogenation performance of organic liquid hydrogen storage based on a large language model. This method organically combines domain knowledge retrieval with large language model reasoning, constructing an intelligent framework integrating literature retrieval, knowledge fusion, performance prediction, and condition optimization recommendation. It can accurately predict the conversion rate and selectivity of catalytic dehydrogenation in various LOHC systems and output optimized catalyst composition and reaction condition schemes based on the prediction results, thereby significantly improving R&D efficiency and reducing experimental costs.

[0005] Firstly, this application provides a method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on large language models, including: Obtain structured data, preprocess it to obtain the first data; obtain unstructured data, preprocess it to obtain the second data; A knowledge base is built based on the first and second data; The parameters input by the user are vectorized, and similarity is searched in the knowledge base to obtain the search results. The search results and the user's original input are combined to obtain the prompt engineering; the prompt engineering is input into the large language model, and the prediction results for the target reaction system are generated through reasoning; the prediction results include product conversion rate, selectivity prediction, credibility assessment and influencing factor analysis; Based on the prediction results, and using historical best records and multi-objective optimization strategies, we search for better catalyst composition and reaction conditions under feasibility constraints, and generate recommended solutions.

[0006] In one possible implementation, the structured data includes LOHC type, catalyst information, process conditions, and performance indicators; The unstructured data includes: the main text of the experimental section of academic papers, supplementary information, screenshots of tables, graphs, and spectra.

[0007] In one possible implementation, building a knowledge base based on the first and second data includes: Construct a unified performance database: Write the first data into structured storage and create an index by LOHC system-catalyst-condition-performance, supporting condition filtering, statistics and traceability queries; Constructing a domain knowledge graph: Entity disambiguation and synonym normalization are performed on the entities extracted from the second data to construct a relational network containing catalyst composition, preparation method, process conditions, performance results, and mechanism explanation, and store it in a graph database; causal relationship edges are established. Construct a vectorized retrieval index: Generate vector representations for the first and second data respectively, and generate a vector database; concatenate the fields of the first data into a key-value pair sequence and embed it to obtain a vector representation; for the second data, input the text fragments, table description text, and figure caption description text into the embedding model to generate vector representations, where the table description text is obtained by converting table headers, field relationships, values, and units, and the figure caption description text is obtained by parsing figure captions, experimental conditions, and result information; establish a mapping relationship between the vector representations and the original content and metadata, and write them into the vector database to support nearest neighbor retrieval, mixed retrieval, and reordering; Constructing a retrieval enhancement generation orchestration layer: Constructing a retrieval enhancement generation orchestration layer based on the LangChain framework, which performs chained orchestration on the vector retrieval module and the large language model inference module, so that when a user query is received, the fragments related to the target system are first retrieved from the vector index, and then the fragments and the user input are used together as context input to the large language model for inference.

[0008] In one possible implementation, the parameters input by the user include: LOHC type, catalyst metal, support, additive information, metal loading, reaction conditions, and user-selectable optimization objectives.

[0009] One possible implementation method for combining search results and original user input to obtain the suggested solution includes: The user's original input is used as the task input area, the retrieved evidence is organized into evidence areas according to source and type, and the output fields, units and ranges are limited in the output constraint area to form a reusable prompt template.

[0010] In one possible implementation, the method for constructing the large language model includes: The model takes a domain instruction dataset as input and outputs structured prediction results and explanations; it is trained using efficient parameter fine-tuning to build a large language model.

[0011] In one possible implementation, the process involves searching for a better catalyst composition and reaction conditions based on prediction results, historical best records, and a multi-objective optimization strategy, under feasibility constraints, and generating a recommended solution, including: Generate a set of candidate solutions: Based on the high-performance historical records and similar system migration rules in the knowledge base, generate candidate catalyst combinations and condition combinations; Multi-objective optimization and constraint filtering of candidate solutions: The candidate solutions are selected with conversion rate, selectivity and hydrogen release-related indicators as the main objective functions, and feasibility constraints are formed by combining the upper limit of temperature, pressure range, metal loading range, equipment safety boundary and availability; the Pareto optimal solution set is obtained by screening through the NSGA-II multi-objective optimization method. Output recommended solutions and evidence chains: Output the Top-N solutions from the Pareto optimal solution set, and provide the prediction performance, confidence level, key influencing factors and evidence sources for each solution to generate recommended solutions.

[0012] Secondly, this application provides a device for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model, comprising: The preprocessing module is used to acquire structured data, preprocess it to obtain the first data, and acquire unstructured data, preprocess it to obtain the second data. The knowledge base construction module is used to build a knowledge base based on the first and second data. The retrieval module is used to vectorize the parameters input by the user, perform similarity retrieval in the knowledge base, and obtain the retrieval results. The prediction module is used to merge the search results and the user's original input to obtain a prompting process; the prompting process is input into a large language model, and the model generates a prediction result for the target reaction system through reasoning; the prediction result includes product conversion rate, selectivity prediction, credibility assessment and influencing factor analysis; The generation module is used to search for better catalyst composition and reaction conditions under feasibility constraints based on prediction results, historical best records, and multi-objective optimization strategies, and to generate recommended solutions.

[0013] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model as described in the first aspect.

[0014] Fourthly, a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model as described in the first aspect.

[0015] The beneficial effects of this application are as follows: (1) The prediction method provided in this application has both performance prediction and condition recommendation functions, which can provide decision support for researchers and engineering applications.

[0016] (2) The prediction method provided in this application has strong cross-system generalization ability and supports the performance evaluation and optimization of various LOHC systems.

[0017] (3) The prediction method provided in this application still maintains high accuracy under conditions of small sample size and incomplete data.

[0018] (4) The prediction method provided in this application is highly interpretable, and the prediction results are accompanied by inferences about influencing factors and mechanisms.

[0019] (5) The prediction method provided in this application can be continuously optimized and achieves dynamic performance improvement through active learning. Attached Figure Description

[0020] Figure 1 A flowchart of prediction and optimization based on a large language model provided for embodiments of this application; Figure 2 This is a schematic diagram of the structure of the organic hydrogen storage liquid catalytic dehydrogenation performance prediction device based on a large language model provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0023] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0024] Currently, the prediction of catalytic dehydrogenation performance and condition recommendation for organic liquid hydrogen storage (LOHC) suffers from problems such as scarce samples, data heterogeneity, and insufficient generalization ability across systems. Therefore, this application provides a method for predicting the catalytic dehydrogenation performance of organic liquid hydrogen storage based on a large language model.

[0025] See Figure 1 The method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model, provided in the embodiments of this application, includes: S101. Obtain structured data and preprocess it to obtain the first data; obtain unstructured data and preprocess it to obtain the second data.

[0026] In one possible implementation, the structured data includes: LOHC type, catalyst information, process conditions, and performance indicators.

[0027] Furthermore, the LOHC types include: methylcyclohexane / toluene system, demethylnaphthalene / methylnaphthalene system, diphenylmethane and its hydride system, decahydronaphthalene / naphthalene system, perhydrodibenzyltoluene / dibenzyltoluene system, and other aromatic or alicyclic organic hydrogen storage systems with reversible hydrogenation-dehydrogenation characteristics.

[0028] Furthermore, the catalyst information includes: the type and ratio of active metals, the type of support, the type and content of additives or promoters, the metal loading, the preparation method, and a characterization summary.

[0029] Furthermore, the process conditions include: reactor type, raw material information, and operating parameters.

[0030] Furthermore, the performance indicators include: substrate conversion rate, target product selectivity, and byproduct distribution.

[0031] In one possible implementation, the unstructured data includes: the experimental section of an academic paper, supplementary information, table screenshots, graphs (such as conversion-temperature, selectivity-time), and spectra (GC / GC-MS, NMR, XPS, etc.).

[0032] In one possible implementation, S101 includes: S101a. Standardize the units and dimensions of the structured data, handle missing and anomaly data, and encode features. Then construct structured records according to a unified field standard to obtain a relational database table. Each record contains at least four types of fields: "system-catalyst-condition-performance" and is accompanied by a source identifier and timestamp.

[0033] Furthermore, the missing and anomaly handling includes: marking missing fields; and removing or correcting data that exceeds physical boundaries or has obvious input errors.

[0034] Furthermore, the feature encoding includes dictionary encoding of discrete variables and normalization of continuous variables.

[0035] Furthermore, the unified field standard includes: For the LOHC system, a pairwise naming convention of hydride / dehydrogenation product should be uniformly adopted for the fields. The catalyst information fields should uniformly adopt the description format of metal composition – support – metal loading – additives; The units and value ranges of process condition fields such as temperature, pressure, space velocity, and hydrogen partial pressure are uniformly defined. Standardized definitions and units such as conversion rate and target product selectivity are uniformly adopted for performance indicator fields; The naming rules, data types, and unit systems of the above fields are pre-defined to ensure that data from different sources can be aligned and compared in the same relational database.

[0036] S101b: After parsing the unstructured data, divide it into multiple segments according to paragraph-figure caption-table rows and columns. Extract entity-relationship-attribute triples from the segments. Each segment is accompanied by metadata tags such as system, catalyst, condition, and indicator, as well as the source location. The source location includes at least the unique identification information of the original document (such as document number, DOI or patent number), page number or chapter number, and the specific location identifier of the segment in the document (such as paragraph number, figure / table number and its row and column index). Store the data in a document-segment-metadata structure.

[0037] Specifically, in S101b, a large language model is used to extract entity-relation-attribute triples from fragments of unstructured data. Specifically, for paragraph fragments, a named entity recognition and relation extraction method based on a large language model is used to identify information such as LOHC system, catalyst composition, process conditions, and performance indicators from the text. For table fragments, a table recognition and field alignment method is used to map the table to a field structure consistent with the first data. For figure caption fragments (image / spectral image segments), an image parsing and text recognition method is used to extract coordinate axes, peak positions, legends, and key values, and align them with the text captions. It should be noted that in the entity-relationship-attribute triple, the entity refers to objects such as the LOHC system, catalyst, process conditions, or performance indicators; the relationship refers to the semantic connection between these entities (e.g., "achieving a certain performance under certain conditions"); and the attribute refers to the specific numerical value and unit directly related to the entity or relationship (e.g., temperature, pressure, conversion rate, selectivity, etc.). For example, from "at 300..." o At C, the conversion rate of 1 wt% Pt / H-MOR for the dehydrogenation of methylcyclohexane is 95%. A ternary model can be obtained: {Entity: Pt / H-MOR catalyst and methylcyclohexane system; Relationship: at 300...} o Catalytic dehydrogenation under C conditions; Properties: Metal loading = 1 wt%, Conversion rate = 95%.

[0038] S102. Construct a knowledge base based on the first and second data.

[0039] In one possible implementation, S102 includes: S102a. Construct a unified performance database: Write the first data into structured storage (column storage) and establish an index according to LOHC system-catalyst-condition-performance, supporting condition filtering, statistics and traceability query.

[0040] S102b, Constructing a domain knowledge graph: Perform entity disambiguation and synonym normalization on the entities extracted from the second data, construct a relational network containing catalyst composition, preparation method, process conditions, performance results, and mechanism explanation, and store it in a graph database; establish causal / related edges (e.g., the effect of temperature on hydrogen desorption rate, the effect of support acidity on side reactions) through graph rules or reasoning models.

[0041] S102c, Construct a vectorized retrieval index: Generate vector representations for the first data and the second data respectively, and generate a vector database: Concatenate the fields of the first data into a key-value pair sequence and embed it to obtain a vector representation; For the second data, input the text fragments, table description texts, and figure caption description texts into the embedding model to generate vector representations, wherein the table description texts are obtained by converting table headers, field relationships, values, and units, and the figure caption description texts are obtained by parsing figure captions, experimental conditions, and result information; Establish a mapping relationship between the vector representations and the original content and metadata and write them into the vector database to support nearest neighbor retrieval, mixed retrieval, and reordering.

[0042] S102d. Constructing a Retrieval Enhancement Generation (RAG) Orchestration Layer: Constructing a retrieval enhancement generation orchestration layer based on the LangChain framework, which performs chained orchestration on the vector retrieval module and the large language model inference module, so that when a user query is received, fragments related to the target system are first retrieved from the vector index, and then the fragments and the user input are used together as context input to the large language model for inference.

[0043] For example, in S102a, the LOHC system is an organic hydrogen storage system with reversible hydrogenation / dehydrogenation characteristics, such as methylcyclohexane / toluene; the catalyst refers to the supported or unsupported metal catalyst and its composition and structure information, such as metal type, support type and loading amount; the conditions refer to process parameters such as reaction temperature, pressure, hydrogen partial pressure and space velocity; the performance refers to quantitative indicators characterizing the catalytic dehydrogenation effect, such as conversion rate, target product selectivity and hydrogen release rate.

[0044] Furthermore, in S102b, the graph rules include a set of rules for defining causal or correlational relationships between entities, such as "under similar conditions, increasing the reaction temperature usually increases the hydrogen release rate" and "enhanced carrier acidity usually promotes the generation of light hydrocarbons by side reactions"; the inference model includes a link prediction model based on graph neural networks, rule inference, or embedded representation, used to infer potential structure-performance or condition-performance relationships based on existing nodes and edges.

[0045] Furthermore, in S102c, the embedded model includes models such as all-MiniLM and m3e-base. The original content includes structured records corresponding to the first data and unstructured text such as paragraph text, table content, and figure captions corresponding to the second data; metadata includes literature related to the content, LOHC system name, catalyst information, process conditions, performance indicators, data source type (experimental / simulation), and its specific location identifier in the original literature.

[0046] S103. Vectorize the parameters input by the user and perform similarity retrieval in the knowledge base to obtain the retrieval results.

[0047] In one possible implementation, in S103, the parameters of the original input include: LOHC type, catalyst metal / support / promoter information, metal loading, reaction conditions (temperature / pressure / space velocity / time, etc.), and user-selectable optimization objectives (such as maximizing conversion and selectivity, minimizing temperature or byproducts, etc.).

[0048] In one possible implementation, the original input is in the format of structured key-value pairs or JSON. For example: {LOHC: …, Metal: …, Support: …, Promoter: …, Loading: …, T: …, P: …, Time: …, Flow: …, Target: …}.

[0049] In one possible implementation, the similarity retrieval includes: Hybrid retrieval (keyword retrieval for text fields and nearest neighbor retrieval for vector fields); Reordering (perform a secondary relevance assessment on candidate results and output Top-K evidence fragments); Evidence aggregation (merging, deduplicating, and marking conflicts among multiple fragments from the same source or under the same experimental conditions); Evidence structuring (organizing the search results into a structure of "list of experimental cases + mechanism explanation fragments + key data fragments", and attaching similarity scores and source location information).

[0050] S104. The search results and the user's original input are combined to obtain the prompting process; the prompting process is input into the large language model, and the prediction results for the target reaction system are generated through reasoning; the prediction results include product conversion rate, selectivity prediction, credibility assessment and influencing factor analysis.

[0051] In one possible implementation, S104, the method of merging the search results and the user's original input to obtain the prompt engineering includes: using the user's original input as a task input area, organizing the search evidence into an evidence area according to its source and type, and limiting the output fields, units, and ranges in the output constraint area, thereby forming a reusable prompt template.

[0052] For example, in S104, the format of the prompt project is as follows: Task Description: Based on input conditions and evidence, predict conversion rate, selectivity and hydrogen desorption related indicators, and provide the credibility and key influencing factors.

[0053] Input: {user key-value pair}.

[0054] Evidence: Case evidence (including system, catalyst, conditions, performance, and similarity); data evidence (including key table values ​​or ranges); mechanistic evidence (including description of the impact path); mechanistic evidence (including description of the impact path).

[0055] Output: Output in structured format {conversion: …, selectivity: …, confidence: …, factors: [...], reasoning: …}. conversion: conversion rate; selectivity: selectivity; confidence: confidence level; factors: influencing factors; reasoning: sources of evidence.

[0056] In one possible implementation, in S104, the large language model includes an LLaMA model.

[0057] Furthermore, the LLaMA model is a domain-adjusted LLaMA model.

[0058] Specifically, the fine-tuning method of the domain-adjusted LLaMA model includes: constructing a domain instruction dataset (with "input conditions + retrieval evidence" as input and "structured prediction results + explanation" as output); and training using parameter efficient fine-tuning (LoRA) to ensure that the model stably outputs limited fields and units.

[0059] It should be noted that in S104, the selective prediction refers to the prediction of the proportion of the target dehydrogenation product (e.g., the corresponding aromatic / dehydrogenation product) in the product under given reaction conditions; the credibility assessment refers to the quantitative characterization of the reliability of the prediction result, which is obtained by comprehensively considering the retrieval similarity score, evidence coverage, model uncertainty, etc., and outputs a confidence score or prediction interval in the 0-1 range; the impact analysis refers to identifying and ranking the key variables that contribute the most to the performance, and the key variables include, but are not limited to, metal type, support properties, additive ratio, temperature, pressure, space velocity, etc.; the influencing factor analysis is obtained by feature sensitivity analysis, comparative evidence attribution, and path reasoning based on knowledge graph, and provides supporting reasons corresponding to the evidence fragments.

[0060] S105. Based on the prediction results, and according to the historical best record and multi-objective optimization strategy, search for better catalyst composition and reaction conditions under feasibility constraints, and generate recommended solutions.

[0061] In one possible implementation, S105 includes: S105a. Generate a set of candidate solutions: Based on the high-performance historical records and similar system migration rules in the knowledge base, generate candidate catalyst combinations and condition combinations. Furthermore, in S105a, the candidate generation results are generated using methods such as rule enumeration, neighborhood expansion based on knowledge graphs, or parameter suggestions based on generative models.

[0062] Specifically, in S105a, the migration rules for similar systems include: (1) Catalyst migration rules for similar or identical LOHC systems: When the target LOHC system is the same as the historical system, or the molecular skeleton, number of aromatic rings and dehydrogenation pathway are similar, the active metal type, support type and metal loading range of the historical high-performance catalyst will be preferentially migrated. For example, if Pt / Al2O3 exhibits a high conversion rate in the methylcyclohexane / toluene system within a specific temperature range, then when recommending similar dehydrogenation systems such as ethylcyclohexane / ethylbenzene, Pt / Al2O3 and its similar loading combinations can be given priority as candidate solutions.

[0063] (2) Reaction Condition Window Migration Rule: For historical high-performance records similar to the target system, their process parameters such as temperature, pressure, space velocity, and hydrogen partial pressure are used as initial condition windows, and candidate condition combinations are generated within a preset perturbation range. For example, if the optimal dehydrogenation temperature of a similar system is 320°C... o C. Airspeed is 2 h -1 Then it can be between 300 and 340. o Candidate conditions are generated within the range of C and adjacent airspeeds for further screening of the target system.

[0064] (3) Performance-goal-oriented migration rules: When the user's optimization goal differs from the main optimization direction of the historical system, high-performance samples that are consistent with the goal are selected from the historical records for migration. For example, if the user's goal is to maintain high selectivity at lower temperatures, the catalyst composition and process conditions corresponding to the "low temperature high selectivity" samples in the history are migrated first, rather than just the samples with the highest conversion rate.

[0065] Specifically, the rule enumeration includes: Under empirical rules and process constraints, a limited range of alternative values ​​is set for key parameters such as temperature, pressure, space velocity, and metal loading. Parameter combinations that meet safety and feasibility constraints are generated through combined enumeration. Neighborhood expansion based on knowledge graphs involves starting with known high-performance nodes in the knowledge graph and searching for neighboring nodes along the relationship edges of catalyst composition, support properties, additive type, and process conditions within a limited step size and under constraints. New "catalyst-condition" combinations that meet the constraints are selected as candidate solutions. Parameter suggestions based on generative models involve constructing a generative model (such as a conditional generative network or variational autoencoder) with LOHC system and catalyst type as input conditions. The model outputs suggested values ​​for continuous or discrete parameters such as temperature, pressure, space velocity, and metal loading. A set of candidate process condition combinations is obtained through multiple sampling and rule-based constraint screening.

[0066] S105b. Perform multi-objective optimization and constraint filtering on candidate solutions: The candidate solutions are evaluated with conversion rate, selectivity, and hydrogen release-related indicators as the main objective functions, combined with upper temperature limit, pressure range, metal loading range, equipment safety boundary and availability to form feasibility constraints; the Pareto optimal solution set is obtained by screening through the NSGA-II multi-objective optimization method.

[0067] S105c, Output Recommended Schemes and Evidence Chains: Output Top-N schemes from the Pareto optimal solution set, and provide the predictive performance, confidence level, key influencing factors and evidence sources (corresponding case fragments, tabular data fragments or theoretical description fragments) for each scheme to achieve "explainable recommendations".

[0068] Furthermore, in S105c, the Pareto optimal solution set is based on the historical best records and candidate solution set in the unified performance database. It is obtained by performing a multi-objective optimization search with conversion rate, selectivity, etc. as multiple objective functions, under feasibility constraints such as temperature, pressure, hydrogen partial pressure, metal loading, space velocity, and equipment safety boundary. The search is carried out in the model prediction space, while the objective functions and constraints are provided by historical data and domain rules in the knowledge base.

[0069] Specifically, the method for obtaining the historical best records in the unified performance database includes: retrieving historical samples that are the same as or similar to the target LOHC system from the unified performance database; performing field alignment, unit unification, missing value processing, and outlier removal on catalyst composition, support type, additive information, metal loading, reaction temperature, pressure, space velocity, and performance indicators; then, based on the user-defined optimization goals and constraints, sorting, weighting, or Pareto optimization of the historical samples according to conversion rate, selectivity, hydrogen release-related indicators, and stability; and selecting the Top-K high-performance records that meet the conditions as the historical best records. The application scenarios and optimization strategies of the aforementioned method are described below.

[0070] I. Typical Application Scenarios The method is used for new catalyst screening, reaction condition optimization, and cross-system migration prediction. In new catalyst screening, users input candidate formulations in batches, and the system predicts their performance sequentially and provides a priority ranking. In condition optimization, users fix the catalyst formulation to obtain recommended optimal combinations of parameters such as temperature, pressure, and gas flow rate. When migrating across systems, the method can infer appropriate catalyst adjustment and process modification strategies based on the similarity of different LOHC systems.

[0071] II. Active Learning and Model Update During the prediction process, if the identified results have high uncertainty or the recommended solution has significant potential for improvement, a priority experimental verification list will be generated.

[0072] After the experiment is completed, the results are fed back into the database and model to update the vector index and fine-tune the parameters, thereby continuously improving the prediction and recommendation capabilities.

[0073] III. Long-term operation and performance improvement Through continuous data accumulation and iterative optimization, the method's cross-system generalization ability and prediction accuracy will gradually improve. Simultaneously, the method retains the evidence chain for each prediction and recommendation, facilitating subsequent tracing and analysis, and providing a long-term reliable reference for scientific research decision-making.

[0074] The following describes the device for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model provided in this application. The device for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model described below can be referred to in correspondence with the method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model described above.

[0075] Figure 2 This is a schematic diagram of the structure of the organic hydrogen storage liquid catalytic dehydrogenation performance prediction device based on a large language model provided in the embodiments of this application, as shown below. Figure 2As shown, it includes: a preprocessing module 21, a knowledge base construction module 22, a retrieval module 23, a prediction module 24, and a generation module 25, wherein: Preprocessing module 21 is used to acquire structured data, preprocess it to obtain first data, and acquire unstructured data, preprocess it to obtain second data. Knowledge base construction module 22 is used to construct a knowledge base based on the first data and the second data; The retrieval module 23 is used to vectorize the parameters input by the user and perform similarity retrieval in the knowledge base to obtain the retrieval results. The prediction module 24 is used to merge the search results and the user's original input to obtain a prompting process; the prompting process is input into a large language model, and a prediction result for the target reaction system is generated through reasoning; the prediction result includes product conversion rate, selectivity prediction, credibility assessment and influencing factor analysis; The generation module 25 is used to search for better catalyst composition and reaction conditions under feasibility constraints based on the prediction results, historical best records and multi-objective optimization strategies, and generate recommended solutions.

[0076] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communications bus 340. The processor 310, communications interface 320, and memory 330 communicate with each other via the communications bus 340. The processor 310 can call logic instructions from the memory 330 to execute a method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model.

[0077] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model provided by the above methods.

[0079] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the above-described methods for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on large language models.

[0080] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0081] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model, characterized in that, include: Obtain structured data and preprocess it to obtain the first data; Unstructured data is acquired and preprocessed to obtain the second data; A knowledge base is built based on the first and second data; The parameters input by the user are vectorized, and similarity is searched in the knowledge base to obtain the search results. The search results and the user's original input are combined to obtain the prompt engineering; the prompt engineering is input into the large language model, and the prediction results for the target reaction system are generated through reasoning; the prediction results include product conversion rate, selectivity prediction, credibility assessment and influencing factor analysis; Based on the prediction results, and using historical best records and multi-objective optimization strategies, we search for better catalyst composition and reaction conditions under feasibility constraints, and generate recommended solutions.

2. The prediction method according to claim 1, characterized in that, The structured data includes LOHC type, catalyst information, process conditions, and performance indicators; The unstructured data includes: the main text of the experimental section of academic papers, supplementary information, screenshots of tables, graphs, and spectra.

3. The prediction method according to claim 1, characterized in that, Building a knowledge base based on the first and second data includes: Construct a unified performance database: Write the first data into structured storage and create an index by LOHC system-catalyst-condition-performance, supporting condition filtering, statistics and traceability queries; Constructing a domain knowledge graph: Entity disambiguation and synonym normalization are performed on the entities extracted from the second data to construct a relational network containing catalyst composition, preparation method, process conditions, performance results, and mechanism explanation, and store it in a graph database; causal relationship edges are established. Construct a vectorized retrieval index: Generate vector representations for the first and second data respectively, and generate a vector database; concatenate the fields of the first data into a key-value pair sequence and embed it to obtain a vector representation; for the second data, input the text fragments, table description text, and figure caption description text into the embedding model to generate vector representations, where the table description text is obtained by converting table headers, field relationships, values, and units, and the figure caption description text is obtained by parsing figure captions, experimental conditions, and result information; establish a mapping relationship between the vector representations and the original content and metadata, and write them into the vector database to support nearest neighbor retrieval, mixed retrieval, and reordering; Constructing a retrieval enhancement generation orchestration layer: Constructing a retrieval enhancement generation orchestration layer based on the LangChain framework, which performs chained orchestration on the vector retrieval module and the large language model inference module, so that when a user query is received, the fragments related to the target system are first retrieved from the vector index, and then the fragments and the user input are used together as context input to the large language model for inference.

4. The prediction method according to claim 1, characterized in that, The parameters initially input by the user include: LOHC type, catalyst metal, support, additive information, metal loading, reaction conditions, and user-selectable optimization objectives.

5. The prediction method according to claim 1, characterized in that, Methods for combining search results and user input to obtain suggestions include: The user's original input is used as the task input area, the retrieved evidence is organized into evidence areas according to source and type, and the output fields, units and ranges are limited in the output constraint area to form a reusable prompt template.

6. The prediction method according to claim 1, characterized in that, The method for constructing the large language model includes: The model takes a domain instruction dataset as input and outputs structured prediction results and explanations; it is trained using efficient parameter fine-tuning to build a large language model.

7. The prediction method of claim 1, wherein, Based on the prediction results, and according to historical best records and a multi-objective optimization strategy, a better catalyst composition and reaction conditions are searched under feasibility constraints, and a recommended scheme is generated, including: Generate a set of candidate solutions: Based on the high-performance historical records and similar system migration rules in the knowledge base, generate candidate catalyst combinations and condition combinations; Multi-objective optimization and constraint filtering of candidate solutions: The candidate solutions are selected with conversion rate, selectivity and hydrogen release-related indicators as the main objective functions, and feasibility constraints are formed by combining the upper limit of temperature, pressure range, metal loading range, equipment safety boundary and availability; the Pareto optimal solution set is obtained by screening through the NSGA-II multi-objective optimization method. Output recommended solutions and evidence chains: Output the Top-N solutions from the Pareto optimal solution set, and provide the prediction performance, confidence level, key influencing factors and evidence sources for each solution to generate recommended solutions.

8. A device for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model, characterized in that, include: The preprocessing module is used to acquire structured data and obtain the first data after preprocessing. Unstructured data is acquired and preprocessed to obtain the second data; The knowledge base construction module is used to build a knowledge base based on the first and second data. The retrieval module is used to vectorize the parameters input by the user, perform similarity retrieval in the knowledge base, and obtain the retrieval results. The prediction module is used to merge the search results and the user's original input to obtain a prompting process; the prompting process is input into a large language model, and the model generates a prediction result for the target reaction system through reasoning; the prediction result includes product conversion rate, selectivity prediction, credibility assessment and influencing factor analysis; The generation module is used to search for better catalyst composition and reaction conditions under feasibility constraints based on prediction results, historical best records, and multi-objective optimization strategies, and to generate recommended solutions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model as described in any one of claims 1 to 7. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for predicting the catalytic dehydrogenation performance of organic hydrogen storage liquids based on a large language model as described in any one of claims 1 to 7.

Images