Construction of multi-dimensional knowledge graph of chemical industry chain and method for matching low-cost raw material substitution

Abstract

The application discloses a multi-dimensional knowledge graph construction and low-cost raw material replacement matching method for a chemical industry chain, extracts a material-application-product relationship from unstructured scientific literature automatically and constructs a knowledge graph with source tracing, carries out intelligent replacement candidate matching based on a graph structure similarity, and quantifies the uncertainty of a recommended result, so that raw material replacement decision making is upgraded from relying on expert experience to AI driving, data support and credibility quantifiable intelligent decision making, and the intelligent blank of a chemical intelligent factory in research and development design and supply chain optimization links is filled up. The method converts unstructured industry literature knowledge into a structured and calculable knowledge graph, and encapsulates the same into an intelligent decision service that can be called, makes up for the lack of a knowledge management and supply chain optimization module in an intelligent manufacturing system, enables an enterprise to select raw materials based on data driving and intelligent reasoning, and improves the intelligent level and competitiveness of overall manufacturing.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial knowledge automation in the chemical industry, specifically to a method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitution. Background Technology

[0002] The construction of intelligent chemical plants requires not only full automation of the production process but also intelligentization of R&D and supply chain management. The "Reference for Key Performance Indicators for Intelligent Factory Construction" and the "Reference Guide for Typical Scenarios of Intelligent Manufacturing" list data-driven R&D and intelligent supply chain optimization as core scenarios of intelligent manufacturing. Raw material substitution matching, i.e., quickly screening alternatives to expensive raw materials from industrial solid waste, low-cost minerals, and other resources, is a crucial decision-making step for chemical companies to reduce costs and carbon emissions. Currently, this step heavily relies on manual literature review and experience-based judgment, representing a "decision silo" in the overall intelligentization of the chemical process that has not yet been automated.

[0003] In the chemical industry chain, finding alternative low-cost or low-carbon raw materials for specific products (such as concrete, coatings, and plastics) is an important way to reduce costs and increase efficiency. Taking building materials chemicals as an example, using industrial byproducts such as fly ash and blast furnace slag to replace part of the cement can reduce raw material costs and carbon emissions. However, key information such as which materials can be used for which applications, the range of substitution ratios, and the impact on the performance of end products is scattered throughout a large amount of scientific and technological literature, and relying on manual reading and summarization is a large workload and difficult to fully cover.

[0004] While there are existing studies on automatically extracting knowledge from scientific and technological literature using natural language processing technology, the linguistic characteristics of literature in the field of chemical materials—especially in the areas of industrial solid waste resource utilization and alternative raw material screening—pose particular challenges to information extraction: the same material may have multiple names (non-standardized terminology); the material-application-product relationship is often presented as descriptive statements rather than noun phrases (indirect information); related information spans multiple sentences (non-local dependency); and a single document often involves multiple materials and multiple applications (non-injective mapping).

[0005] In existing technologies, named entity recognition-based methods are difficult to handle the above-mentioned complex language phenomena. An attempt has been made to extract information from literature using existing large language models (LLMs), but the computational cost of using a large model with hundreds of billions of parameters to learn context through a small number of samples is high. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitutions, comprising the following steps: S1. Literature Data Acquisition and Preprocessing: Obtain the titles, abstracts and metadata of papers in the target chemical engineering field to construct a literature corpus and record the literature metadata; S2. Establish a knowledge ontology and machine-readable coding system for the chemical industry chain: Define a set of material entities M, a set of application entities A, a set of product entities P, and a set of relation types R; assign a fixed-length machine-readable code to each candidate entity and establish a coding mapping table; S3. Construct instruction templates and generate relevant instructions: Construct instruction templates that include task description, literature context, candidate entity coding list and output format requirements, and generate relevant instructions; S4. Construct a large language model and perform preprocessing, extract relevant literature to supervise the training of the large language model: Select representative literature from the literature corpus and manually annotate relation quadruples, use the annotated data to supervise the training of the pre-trained large language model, so that the large language model outputs the encoding list of relation quadruples and the position index of evidence fragments from the literature text. S5. Three-layer automatic verification: For each extracted result output by the large language model, perform format verification, semantic verification and evidence verification in sequence, and output the relation quadruple that passes all multi-layer verification; S6. Knowledge Graph Construction: Write the relation quadruplets output in step S5 into the knowledge graph. The graph contains material nodes, application nodes, product nodes, and literature source nodes. Each relation edge records the source literature identifier, evidence fragment, extraction confidence, and manual review status, and calculates the composite edge weight between materials and applications. S7. Similarity Calculation: For any two materials, based on their application distribution association in the spectrum, automatically quantify the similarity of the two materials in application functions: When neither material has any known application, the similarity is marked as insufficient evidence. S8. Alternative Candidate Prediction: For material-application pairs that have not yet established effective relationship connections in the knowledge graph, after applying preset industrial exclusion conditions, the alternative prediction score is calculated based on the similarity of known materials to generate potential alternative raw material candidate schemes; S9. Uncertainty Quantification and Recommendation Output: Resample the original literature corpus, apply random perturbation to the composite edge weights based on the extraction error rate, independently perform similarity calculation and alternative prediction score calculation for each graph variant, obtain multiple sets of prediction scores, calculate the mean, standard deviation and ranking stability ratio of the prediction scores, and generate a structured recommendation report containing confidence assessment.

[0007] Furthermore, in step S2, the encoding length is specifically detected by using a word segmenter; if the encoding is segmented into fragments that are inconsistent with the preset length by the word segmenter, the encoding is adjusted to a fixed number of segmented fragments by using a dedicated new word list or retaining markers.

[0008] Furthermore, in step S3, each time an instruction is generated, the order of the candidate entity encoding list is independently and randomly rearranged to eliminate the language model's attention bias to the position of the input option.

[0009] Furthermore, the three-layer automatic verification in step S5 specifically includes: Format validation: Validate whether the output code exists in the encoding mapping table, whether the codes within the quadruple are repeated, and whether the evidence fragment position index exceeds the original text length range; Semantic validation: Validate whether the extracted material-application-product-relationship combination violates the predefined knowledge ontology rules. Evidence verification: Extract evidence text from the original text based on the evidence fragment position index, and verify whether the evidence text actually supports the extracted four-tuple relationship in semantics.

[0010] Furthermore, the expression for the relation quadruple in step S4 is: In the formula, m∈M is the material code, a∈A is the application code, p∈P is the product code, r∈R is the relation type, ev is the position index of the evidence fragment supporting the relation in the original text, M represents the set of material entities, A represents the set of application entities, P represents the set of product entities, and R represents the set of relation types. The formula for calculating the composite edge weight w between the material and the application in step S6 is as follows: ; In the formula, I d,m,a A binary variable γ represents whether literature d supports the relationship between material m and application a. d q represents the model's confidence in extracting the relation. d This represents the document quality factor.

[0011] Furthermore, the formula for calculating the similarity of the two materials in terms of application function in step S7 is as follows: In the formula, N(m) i ) represents material m i The set of applications associated in the knowledge graph, N(m) j ) represents material m j A collection of applications linked within a knowledge graph. Material m i Normalized edge weights Material m j Normalized edge weights This represents a small positive constant to prevent the denominator from being zero.

[0012] Furthermore, the specific formula for calculating the substitution prediction score for materials and applications in the knowledge graph for which no effective relational connections have yet been established in step S8 is as follows: In the formula, (m i a k This indicates a material-application pair in the knowledge graph that has not yet established a valid relational connection. Material m j To application a k Normalized edge weights between them; The industrial exclusion conditions in step S8 include at least one of the following: 1) Material-application combinations that are clearly unsuitable as documented in the literature; 2) Combinations where the basic physical or chemical properties of the material conflict with the target application indicators; 3) Combinations where the number of supporting pieces of evidence is below a preset threshold; 4) Combinations that violate safety or environmental regulations.

[0013] Furthermore, the uncertainty quantification in step S9 includes: calculating the mean of the prediction score μ. S and standard deviation σ S Define relative uncertainty The frequency of candidate solutions entering the Top-K recommendation list in repeated experiments is counted as the ranking stability ratio; when the relative uncertainty exceeds the preset risk threshold, the corresponding solution is automatically marked as requiring manual review. in, This represents a small positive constant to prevent the denominator from being zero.

[0014] A knowledge-based intelligent engine system for the chemical industry chain used to implement the method specifically includes: The literature acquisition and preprocessing module is used to retrieve and preprocess literature in the target chemical engineering field from academic literature databases. The knowledge ontology and coding management module is used to define entity and relation sets and assign a fixed-length code to each candidate entity; The information extraction module has a built-in supervised training large language model, which is used to output a list of encoded relation quadruples and an index of the location of evidence fragments from the document text. The three-layer automatic verification module is used to perform format verification, semantic verification and evidence verification on the extracted results respectively. The knowledge graph storage module is used to store material nodes, application nodes, product nodes, and literature source nodes and their relational edges. Each relational edge records the source literature identifier, evidence fragment, extraction confidence level, and verification status. The similarity calculation and substitution prediction module is used to calculate the weighted similarity between materials and generate alternative candidate predictions; The uncertainty quantification module is used to calculate the confidence index of the predicted score through resampling and perturbation analysis; The recommendation output module is used to generate recommendation reports that include scores, confidence levels, and evidence sources.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitution.

[0016] This invention provides a method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitutions, which has the following beneficial effects: (1) By adopting a progressive filtering system consisting of format validation, semantic validation and evidence validation, it is ensured that every relation written into the knowledge graph is verifiable and conforms to ontology constraints. Untraceable extraction results (such as model illusion output) are intercepted at the evidence validation layer.

[0017] (2) By changing “extracting entity lists separately → building edges based on co-occurrence” to “directly extracting relation quadruples”, the problem of incorrect association of materials and applications that do not correspond to each other in the same document is avoided, and the extraction results are auditable by preserving the position of evidence fragments.

[0018] (3) Supervised training is performed using a model with billions of parameters. Compared with the small sample scheme of a large model with hundreds of billions of parameters, the memory usage and inference time are greatly reduced, and it can be deployed on a private server of an enterprise.

[0019] (4) By providing quantifiable reliability indicators for each recommendation through the ratio of relative uncertainty and ranking stability, decision-makers can screen candidate directions with "high score + high stability" for priority verification.

[0020] (5) Fixed-length encoding can adapt to the differentiated word segmentation behavior of different models through the word segmenter detection and adjustment mechanism, without relying on the specific word segmenter's processing method for numbers or symbols.

[0021] (6) This method transforms unstructured industry literature knowledge into structured, computable knowledge graphs and encapsulates them into callable intelligent decision-making services, making up for the lack of knowledge management and supply chain optimization modules in the intelligent manufacturing system. This enables enterprises to select raw materials based on data-driven and intelligent reasoning, thereby improving the overall intelligence level and competitiveness of manufacturing. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 The method flowchart provided by the present invention. Detailed Implementation

[0024] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] The following detailed description of the implementation method of the present invention is in conjunction with the accompanying drawings. The description only covers some embodiments and not all embodiments. For clarity, representations and descriptions unrelated to the present invention have been omitted from the drawings and description.

[0026] To provide a clearer understanding of the technical features, objectives, and beneficial effects of this invention, the technical solutions of this invention are selected and described in detail below. Obviously, the described embodiments are only a portion of the embodiments of this invention, not all of them, and should not be construed as limiting the scope of implementation of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0027] This invention provides a method for constructing a multi-dimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitutions, such as... Figure 1 As shown, it includes the following steps: S1. Literature Data Acquisition and Preprocessing: Obtain the titles, abstracts, and metadata (publication year, DOI, etc.) of papers in the target chemical engineering field to construct a literature corpus and record the literature metadata. This includes standardization processing of the literature text, such as deduplication, sentence segmentation, unified character encoding, and truncation (preserving complete semantic units). Record the retrieval date, database source, and search query for traceability and incremental updates.

[0028] S2. Establish a knowledge ontology and machine-readable coding system for the chemical industry chain: Define a set of material entities M, an application entity set A, a product entity set P, and a relation type set R; assign a fixed-length machine-resolvable code to each candidate entity and establish a coding mapping table. Specifically, a word segmenter is used to detect the code length; if the code is segmented into fragments that are inconsistent with the preset length by the word segmenter, the code is adjusted to a fixed number of segmented fragments using a dedicated new vocabulary or by retaining markers.

[0029] Three types of entity sets are defined by domain experts: Material entity set M: encompasses candidate alternative materials known in the field (including industrial by-products, natural minerals, synthetic materials, biomass ash, etc.), as well as the "unknown material" placeholder option; Application entity set A: covers application scenarios known in the field (such as supplementary cementitious materials, geopolymer raw materials, fine aggregate substitutes, etc.), as well as "unknown applications" placeholder options; Product entity set P: covers end product types in this field (such as cement mortar, structural concrete, etc.), as well as "unknown product" placeholder options.

[0030] Define a set of relation types R, which includes at least: "used for" relation (material → application); "formation" relation (material → product); "substitution" relation (material → target raw material being substituted).

[0031] Each candidate entity is assigned a fixed-length machine-resolvable code (such as a fixed-length numeric code or an alphanumeric code). The mapping relationship between the code and the entity name is recorded in the code mapping table. Special codes are set up for "unknown entities" and "no corresponding relationship". For newly added entities, codes are appended according to a unified extended numbering rule. The code length is detected by the word segmenter of the target information extraction model; if the code is segmented into fragments that are inconsistent with the preset length by the word segmenter, it is adjusted by adding a dedicated vocabulary or retaining the marker.

[0032] S3. Construct instruction templates and generate relevant instructions: Construct instruction templates that include task description, literature context, candidate entity encoding list, and output format requirements, and generate relevant instructions. Each time an instruction is generated, the order of the candidate entity encoding list is independently and randomly rearranged to eliminate the language model's attention bias towards the position of input options.

[0033] Taking the material extraction-application-product relationship as an example: The task description specifies that the model must identify specific materials, applications, products, and relationships from the literature and output the corresponding codes and the locations of evidence fragments (supporting the start and end indices of the original sentences or sentence groups in the literature abstract). Candidate entities are presented in coded form; the order of the code list is randomly arranged each time an instruction is generated, allowing the model to learn to focus on the content of the codes rather than their position.

[0034] S4. Construct and preprocess a large language model, and extract relevant literature for supervised training of the large language model: Select representative literature from the literature corpus and manually annotate relation quadruples. Use the annotated data to supervise the training of the pre-trained large language model, enabling the large language model to output a list of encoded relation quadruples and indexes of evidence fragment positions from the literature text. Annotation is performed independently by at least two annotators with backgrounds in chemical materials. After annotation, annotation consistency is calculated, and inconsistent annotations are adjudicated by senior experts. The adjudicated annotations are used as training data to supervise the training of a pre-trained lightweight large language model (with billions of parameters). The model input is the literature text and encoded instructions, and the output is a list of encoded relation quadruples and the positions of evidence fragments.

[0035] The expression for a relation quadruple is: In the formula, m∈M is the material code, a∈A is the application code, p∈P is the product code, r∈R is the relation type, ev is the position index of the evidence fragment supporting the relation in the original text, M represents the set of material entities, A represents the set of application entities, P represents the set of product entities, and R represents the set of relation types.

[0036] S5. Three-layer automatic verification: For each extracted result output by the large language model, perform format verification, semantic verification and evidence verification in sequence, and output the relation quadruple that passes all multi-layer verification.

[0037] The three-layer automatic verification specifically includes: Format validation: Validate whether the output code exists in the encoding mapping table, whether the codes within the quadruple are repeated, and whether the evidence fragment position index exceeds the original text length range; Semantic validation: Validate whether the extracted material-application-product-relationship combination violates the constraints of predefined knowledge ontology rules (for example, if an application is defined in the ontology as "only applicable to mineral materials", but the extraction result associates organic materials with the application, it is marked as a semantic mismatch). Evidence verification: Extract evidence text from the original text based on the evidence fragment position index, and verify whether the evidence text actually supports the extracted four-tuple relationship (manual or automatic judgment to determine whether the text actually supports the extracted relationship; relationships that cannot pass the evidence verification will not be written into the knowledge graph).

[0038] S6. Knowledge Graph Construction: Write the relation quadruples output in step S5 into the knowledge graph. The graph contains material nodes, application nodes, product nodes, and literature source nodes. Each relation edge records the source literature identifier, evidence fragment, extraction confidence, and manual review status, and calculates the composite edge weight between materials and applications.

[0039] The formula for calculating the composite edge weight w between materials and applications is: ; In the formula, I d,m,a A binary variable γ represents whether literature d supports the relationship between material m and application a. d q represents the model's confidence in extracting the relation. d The document quality factor (which can be set based on factors such as the level of the publishing journal and the number of citations, or uniformly set to 1).

[0040] S7. Similarity Calculation: For any two materials, based on the application distribution associated with them in the spectrum, the similarity between the two materials in terms of application function is automatically quantified: when neither material has any known application, the similarity is marked as insufficient evidence.

[0041] The formula for calculating the similarity of the two materials in terms of application function is: In the formula, N(m) i ) represents material m i The set of applications associated in the knowledge graph, N(m) j ) represents material m j A collection of applications linked within a knowledge graph. Material m i Normalized edge weights Material m j Normalized edge weights This represents a small positive constant to prevent the denominator from being zero. When the denominator is zero (neither of these has any known applications), the similarity is marked as "insufficient evidence" and is not included in subsequent predictions.

[0042] S8. Alternative Candidate Prediction: For material-application pairs that have not yet established effective relationship connections in the knowledge graph, after applying preset industrial exclusion conditions, the alternative prediction score is calculated based on the similarity of known materials to generate potential alternative raw material candidate schemes.

[0043] The industrial exclusion criteria include at least one of the following: 1) Material-application combinations that are clearly unsuitable as documented in the literature; 2) Combinations where the basic physical or chemical properties of a material conflict with the target application indicators (e.g., non-volcanic ash materials used in applications requiring volcanic ash activity). 3) Combinations with a number of supporting evidences below a preset threshold (insufficient predictive basis); 4) Combinations that violate safety or environmental regulations.

[0044] The specific formula for calculating the substitution prediction score for materials in a knowledge graph for which no effective relational connections have yet been established is as follows: In the formula, (m i a k This indicates a material-application pair in the knowledge graph that has not yet established a valid relational connection. Material m j To application a k The normalized edge weights between them.

[0045] S9. Uncertainty Quantification and Recommendation Output: Resample the original literature corpus, apply random perturbation to the composite edge weights based on the extraction error rate, independently perform similarity calculation and alternative prediction score calculation for each graph variant, obtain multiple sets of prediction scores, calculate the mean, standard deviation and ranking stability ratio of the prediction scores, and generate a structured recommendation report containing confidence assessment.

[0046] The sources of uncertainty affecting the reliability of the prediction are handled separately: Uncertainty in document coverage: Bootstrap resampling (with replacement of extracted documents, sample size equal to the original corpus size) is performed on the original document corpus to generate multiple resampled corpora.

[0047] Uncertainty in relation extraction: Based on the model extraction error rate estimate obtained from step S5, a random perturbation of corresponding magnitude is applied to the graph edge weights.

[0048] Link prediction uncertainty: For each map variant after the above processing, steps S7~S8 are performed independently to obtain multiple sets of prediction scores.

[0049] Uncertainty quantification includes: calculating the mean of the predicted score μ S and standard deviation σ S Define relative uncertainty The frequency of candidate solutions entering the Top-K recommendation list in repeated experiments is counted as the ranking stability ratio; when the relative uncertainty exceeds the preset risk threshold, the corresponding solution is automatically marked as requiring manual review. in, This represents a small positive constant to prevent the denominator from being zero.

[0050] The perturbation parameters (number of resampling times, perturbation amplitude, etc.) are determined as follows: when labeled data is available, the false negative rate and false positive rate are estimated through cross-validation to define the perturbation range; when there is no labeled data, the perturbation parameters are used as configurable parameters, and sensitivity analysis is performed on different parameter values ​​to evaluate their impact on the recommendation ranking.

[0051] The uncertainty quantification result can be used as a risk threshold input for the intelligent decision-making system: when the relative uncertainty exceeds the preset threshold, the candidate solution is automatically marked as "requiring manual review" to ensure the safety boundary of fully automated decision-making.

[0052] The generated structured recommendation report includes scores, confidence levels, evidence tracing, and risk warnings. The recommendation list is output in descending order of the mean predicted score, and each entry includes: material name and code, target application name and code, mean and standard deviation of predicted score, relative uncertainty, ranking stability ratio, the Top-N most similar materials supporting the prediction and their contribution, the number of existing evidence documents and representative document identifiers, and a risk warning regarding prohibited or restricted use (if applicable).

[0053] The output results include the following notes: The recommended results are based on computational inference from literature data and are candidate screening in nature; they are not equivalent to proven direct substitutes. Experimental verification of material compatibility, performance, environmental friendliness, and safety should be conducted before actual use. The output results can be integrated into the procurement optimization module or R&D support system of an intelligent manufacturing platform. When production plans change or cost assessment targets are set, the system can automatically invoke this method to quickly generate a set of alternative raw material solutions that have undergone uncertainty assessment, assisting procurement personnel or the system in making automatic decisions and achieving intelligent and automated supply chain processes.

[0054] The method also includes step S10: real-time acquisition of the purchase price, pre-processing cost, and transportation cost of candidate materials, weighted calculation based on the substitution prediction score, and outputting a comprehensive recommendation score that combines technical feasibility and economic optimization. If cost data for each candidate material (including purchase price, pre-processing cost, transportation cost, and disposal cost) is available, a comprehensive cost index can be added to each recommendation. The comprehensive cost index and the predicted technical score can be weighted and combined to form the comprehensive recommendation score. In fully automated industrial intelligent manufacturing scenarios, cost data is usually automatically provided by the ERP system. This method can connect with the enterprise cost database in real time to achieve joint optimization decisions based on technical feasibility and economic efficiency.

[0055] It also includes a chemical industry chain knowledge intelligent engine system for implementing the method, specifically including: The literature acquisition and preprocessing module is used to retrieve and preprocess literature in the target chemical engineering field from academic literature databases. The knowledge ontology and coding management module is used to define entity and relation sets and assign a fixed-length code to each candidate entity; The information extraction module has a built-in supervised training large language model, which is used to output a list of encoded relation quadruples and an index of the location of evidence fragments from the document text. The three-layer automatic verification module is used to perform format verification, semantic verification and evidence verification on the extracted results respectively. The knowledge graph storage module is used to store material nodes, application nodes, product nodes, and literature source nodes and their relational edges. Each relational edge records the source literature identifier, evidence fragment, extraction confidence level, and verification status. The similarity calculation and substitution prediction module is used to calculate the weighted similarity between materials and generate alternative candidate predictions; The uncertainty quantification module is used to calculate the confidence index of the predicted score through resampling and perturbation analysis; The recommendation output module is used to generate recommendation reports that include scores, confidence levels, and evidence sources.

[0056] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitution.

[0057] Example: Taking the field of concrete and cement-based materials as an example.

[0058] (1) Document Corpus A corpus was retrieved using databases such as Scopus using the following method: TITLE-ABS-KEY(("waste" OR "by-product" OR "slag" OR "ash" OR "alternative" OR "substitution") AND ("concrete" OR "cement" OR "mortar")), limiting the document type to journal articles and conference papers, and the time range to 1971 to 2023. After deduplication, a corpus of documents was obtained. The title and abstract of each document were standardized, and the DOI was retained as a unique identifier.

[0059] (2) Definition of Entity and Relationship There are approximately 76 types of material entities: fly ash (Class F), fly ash (Class C), blast furnace slag, silica fume, metakaolin, waste glass powder, limestone powder, rice husk ash, red mud, mine tailings, calcined clay, etc., plus "unknown materials".

[0060] There are approximately 13 types of applications: supplementary cementitious materials, geopolymer raw materials, fine aggregate substitutes, coarse aggregate substitutes, clinker raw material substitutes, fillers / micro-aggregates, fiber-reinforced materials, alkali activators, lightweight aggregates, lime-volcanic ash cement systems, etc., plus "unknown applications".

[0061] There are approximately 16 types of products: cement mortar, ordinary concrete, high-strength concrete, self-compacting concrete, lightweight concrete, road concrete, precast components, blocks / bricks, etc., plus "unknown products".

[0062] Relationship types: Used for (material → application), forming (material → product), and substituting (material → replaced material).

[0063] Each entity is assigned a fixed-length code (e.g., M01~M76 represent materials, A01~A13 represent applications, P01~P16 represent products, and U00 represents unknown). The code mapping table records the correspondence between codes and entity names.

[0064] (3) Training data and model Approximately 100 representative papers were selected from the literature corpus. Two annotators with materials science backgrounds independently annotated the relation quadruples in each paper. Annotation consistency indices (such as Cohen's Kappa coefficient) were calculated, and inconsistencies were resolved by senior experts. After annotation, the training data was split into training and test sets according to the document ID (ensuring that samples from the same document did not cross sets).

[0065] A pre-trained large language model with approximately 2.8 billion parameters was selected as the base model for the information extraction model. Supervised training of this model was performed using labeled data. After training, the relation-level extraction performance was evaluated on a test set.

[0066] (4) Full sampling and verification The trained model is used to extract information from the full corpus of literature. Each output contains a list of encoded relation quadruples and an index of the location of the evidence fragment.

[0067] Perform three-layer validation on each item. Format validation filters out abnormal outputs with non-existent or duplicate codes. Semantic validation filters out unreasonable relationships based on predefined ontology constraints. Evidence validation locates evidence fragments from the original text and determines whether they actually support the relationship. Relationships that pass all validations are then entered into the knowledge graph.

[0068] (5) Map construction The graph is stored in a graph database or graph computing framework. Nodes are materials, applications, products, and literature. Each relation edge records the source literature, evidence fragments, extraction confidence level, and verification status. The weight of the material-application edge is determined by the literature count weighted by the overall confidence level.

[0069] The knowledge graph can be deployed on an enterprise's private cloud or edge computing nodes and linked in real time with the production execution system. When production plans change or raw material supply fluctuates, the system automatically triggers alternative candidate re-matching, enabling autonomous response and optimization of the supply chain.

[0070] (6) Alternative candidate prediction Taking the "lime-volcanic ash cement system" application as an example, steps S7-S9 are performed to predict alternative candidates. The Bootstrap resampling is performed 100 times, and the perturbation amplitude of the relationship extraction is set with reference to the error rate estimated in the verification stage. The predicted scores are sorted in descending order of mean, and the top 20 material-application pairs with the highest scores are included in the candidate recommendation list. Pairs with relative uncertainty below a preset threshold are marked as "high confidence candidates".

[0071] Output example format: |Fly ash (Class F)| 0.68| 0.03| 1.00| ~12| |Zeolite|0.66|0.03|0.98|~8| (7) R&D Design of Intelligent Chemical Plants and Adaptation to Intelligent Supply Chain Scenarios This system can be directly embedded into two typical smart manufacturing scenarios in chemical intelligent factories: intelligent R&D design and intelligent supply chain optimization. Intelligent R&D design: R&D personnel input the target product, and the system automatically searches literature, constructs a graph, matches candidate alternative raw materials, and includes a credibility assessment, reducing the traditional weeks of manual literature research to minutes. Intelligent supply chain optimization: By connecting to raw material price data in the enterprise's ERP (an optional extension of step S10), the system can automatically output the most cost-effective alternative raw material solutions under the current market conditions, enabling data-driven autonomous decision-making in raw material procurement. Intelligent carbon emission management: If carbon footprint data is embedded in the graph, the system can automatically recommend low-carbon alternatives to support the enterprise's ESG goals.

[0072] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitution, characterized in that, Includes the following steps: S1. Literature Data Acquisition and Preprocessing: Obtain the titles, abstracts and metadata of papers in the target chemical engineering field to construct a literature corpus and record the literature metadata; S2. Establish a knowledge ontology and machine-readable coding system for the chemical industry chain: Define a set of material entities M, a set of application entities A, a set of product entities P, and a set of relation types R; assign a fixed-length machine-readable code to each candidate entity and establish a coding mapping table; S3. Construct instruction templates and generate relevant instructions: Construct instruction templates that include task description, literature context, candidate entity coding list and output format requirements, and generate relevant instructions; S4. Construct a large language model and perform preprocessing, extract relevant literature to supervise the training of the large language model: Select representative literature from the literature corpus and manually annotate relation quadruples, use the annotated data to supervise the training of the pre-trained large language model, so that the large language model outputs the encoding list of relation quadruples and the position index of evidence fragments from the literature text. S5. Three-layer automatic verification: For each extracted result output by the large language model, perform format verification, semantic verification and evidence verification in sequence, and output the relation quadruple that passes all multi-layer verification; S6. Knowledge Graph Construction: Write the relation quadruplets output in step S5 into the knowledge graph. The graph contains material nodes, application nodes, product nodes, and literature source nodes. Each relation edge records the source literature identifier, evidence fragment, extraction confidence, and manual review status, and calculates the composite edge weight between materials and applications. S7. Similarity Calculation: For any two materials, based on their application distribution association in the spectrum, automatically quantify the similarity of the two materials in application functions: When neither material has any known application, the similarity is marked as insufficient evidence. S8. Alternative Candidate Prediction: For material-application pairs that have not yet established effective relationship connections in the knowledge graph, after applying preset industrial exclusion conditions, the alternative prediction score is calculated based on the similarity of known materials to generate potential alternative raw material candidate schemes; S9. Uncertainty Quantification and Recommendation Output: Resample the original literature corpus, apply random perturbation to the composite edge weights based on the extraction error rate, independently perform similarity calculation and alternative prediction score calculation for each graph variant, obtain multiple sets of prediction scores, calculate the mean, standard deviation and ranking stability ratio of the prediction scores, and generate a structured recommendation report containing confidence assessment.

2. The alternative matching method according to claim 1, characterized in that, In step S2, the encoding length is specifically detected by a word segmenter. If the encoding is segmented into fragments that are inconsistent with the preset length by the word segmenter, the encoding is adjusted to a fixed number of segmented fragments by using a dedicated new word list or retaining the marker.

3. The substitution matching method according to claim 1, characterized in that, In step S3, each time an instruction is generated, the order of the candidate entity encoding list is independently and randomly rearranged to eliminate the language model's attention bias to the position of the input option.

4. The substitution matching method according to claim 1, characterized in that, The three-layer automatic verification in step S5 specifically includes: Format validation: Validate whether the output code exists in the encoding mapping table, whether the codes within the quadruple are repeated, and whether the evidence fragment position index exceeds the original text length range; Semantic validation: Validate whether the extracted material-application-product-relationship combination violates the predefined knowledge ontology rules. Evidence verification: Extract evidence text from the original text based on the evidence fragment position index, and verify whether the evidence text actually supports the extracted four-tuple relationship in semantics.

5. The substitution matching method according to claim 1, characterized in that, The expression for the relation quadruple in step S4 is: In the formula, m∈M is the material code, a∈A is the application code, p∈P is the product code, r∈R is the relation type, ev is the position index of the evidence fragment supporting the relation in the original text, M represents the set of material entities, A represents the set of application entities, P represents the set of product entities, and R represents the set of relation types. The formula for calculating the composite edge weight w between the material and the application in step S6 is as follows: ; In the formula, I d,m,a A binary variable γ represents whether literature d supports the relationship between material m and application a. d q represents the model's confidence in extracting the relation. d This represents the document quality factor.

6. The alternative matching method according to claim 5, characterized in that, The formula for calculating the similarity of the two materials in terms of application function in step S7 is as follows: In the formula, N(m) i ) represents material m i The set of applications associated in the knowledge graph, N(m) j ) represents material m j A collection of applications linked within a knowledge graph. Material m i Normalized edge weights Material m j Normalized edge weights This represents a small positive constant to prevent the denominator from being zero.

7. The substitution matching method according to claim 6, characterized in that, The specific formula for calculating the substitution prediction score for materials and applications in the knowledge graph that have not yet established effective relational connections in step S8 is as follows: In the formula, (m i a k This indicates a material-application pair in the knowledge graph that has not yet established a valid relational connection. Material m j To application a k Normalized edge weights between them; The industrial exclusion conditions in step S8 include at least one of the following: 1) Material-application combinations that are clearly unsuitable as documented in the literature; 2) Combinations where the basic physical or chemical properties of the material conflict with the target application indicators; 3) Combinations where the number of supporting pieces of evidence is below a preset threshold; 4) Combinations that violate safety or environmental regulations.

8. The alternative matching method according to claim 1, characterized in that, The uncertainty quantification in step S9 includes: calculating the mean of the prediction score μ. S and standard deviation σ S Define relative uncertainty The frequency of candidate solutions entering the Top-K recommendation list in repeated experiments is counted as the ranking stability ratio; when the relative uncertainty exceeds the preset risk threshold, the corresponding solution is automatically marked as requiring manual review. in, This represents a small positive constant to prevent the denominator from being zero.

9. A chemical industry chain knowledge intelligent engine system for implementing the method according to any one of claims 1 to 8, characterized in that, Specifically, it includes: The literature acquisition and preprocessing module is used to retrieve and preprocess literature in the target chemical engineering field from academic literature databases. The knowledge ontology and coding management module is used to define entity and relation sets and assign a fixed-length code to each candidate entity; The information extraction module has a built-in supervised training large language model, which is used to output a list of encoded relation quadruples and an index of the location of evidence fragments from the document text. The three-layer automatic verification module is used to perform format verification, semantic verification and evidence verification on the extracted results respectively. The knowledge graph storage module is used to store material nodes, application nodes, product nodes, and literature source nodes and their relational edges. Each relational edge records the source literature identifier, evidence fragment, extraction confidence level, and verification status. The similarity calculation and substitution prediction module is used to calculate the weighted similarity between materials and generate alternative candidate predictions; The uncertainty quantification module is used to calculate the confidence index of the predicted score through resampling and perturbation analysis; The recommendation output module is used to generate recommendation reports that include scores, confidence levels, and evidence sources.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for constructing a multidimensional knowledge graph of the chemical industry chain and matching low-cost raw material substitution as described in any one of claims 1 to 8.

Images