Formula generation method

Abstract

The invention discloses a formula generation method which is applied to the field of chemoinformatics. The method comprises the steps that entity recognition is carried out based on user query to obtain formula demand information, and the formula demand information comprises the type of a target product and description of target performance; a formula is extracted from the at least one data source according to the formula demand information, a reference formula set is obtained, and the reference formula set comprises at least one reference formula related to the formula demand information; and by taking the description of the target performance as a constraint condition, generating a suggested formula of the target product based on a reference formula in the reference formula set by using a formula generation model, the formula generation model comprising a trained artificial intelligence model. Therefore, the formula research and development efficiency can be improved, and the dependence of the research and development process on professionals is reduced.

Description

Technical Field

[0001] This application relates to the field of cheminformatics, specifically to a method for formula generation. Background Technology

[0002] In materials science, formulations involve mixing substances in specific proportions to achieve target properties. Traditional formulation development relies heavily on human experience and trial-and-error experiments. Researchers typically adjust the proportions, concentrations, or order of addition of components based on past experience, literature, or formulation templates to obtain a formulation that meets the target performance. This process is characterized by long development cycles, low efficiency, and a high dependence on the experience of specialized personnel. Summary of the Invention

[0003] To improve the efficiency of formula development and reduce the reliance on professional personnel in the development process, this application provides a formula generation method.

[0004] In a first aspect, embodiments of this application provide a recipe generation method. The method includes: acquiring a user query; performing entity recognition based on the user query to obtain recipe requirement information, the recipe requirement information including a type of target product and a description of target performance; extracting recipes from at least one data source according to the recipe requirement information to obtain a reference recipe set, the reference recipe set including at least one reference recipe related to the recipe requirement information; using the description of the target performance as a constraint, generating a suggested recipe for the target product based on the reference recipes in the reference recipe set using a recipe generation model, the recipe generation model including a trained artificial intelligence model.

[0005] In some embodiments, before generating a suggested formulation for the target product using a formulation generation model based on reference formulations in the reference formulation set, with the description of the target performance as a constraint, the method further includes: generating descriptions of multiple formulation styles based on the description of the target performance; and determining the formulation style to which each reference formulation in the reference formulation set belongs based on the descriptions of the multiple formulation styles and the reference formulation set. Accordingly, generating a suggested formulation for the target product using a formulation generation model based on reference formulations in the reference formulation set, with the description of the target performance as a constraint, includes: for each formulation style, using the formulation generation model to generate a suggested formulation for the target product belonging to that formulation style, with the description of the target performance as a constraint, based on reference formulations in the reference formulation set belonging to that formulation style.

[0006] In some embodiments, determining the formulation style to which each reference recipe in the reference recipe set belongs, based on the descriptions of the plurality of formulation styles, includes: adjusting the mapping relationship between reference recipes and formulation styles in a direction that increases target similarity under preset constraints; the preset constraints are used to restrict each reference recipe to only one formulation style, and there are at least a preset number of reference recipes mapped to each formulation style; the target similarity is the similarity between a relevance score set and the mapping relationship, the relevance score set including the relevance score between each reference recipe in the reference recipe set and the description of each formulation style; and determining the formulation style to which each reference recipe in the reference recipe set belongs based on the adjusted mapping relationship.

[0007] In some embodiments, before generating a suggested formula for the target product using a formula generation model based on reference formulas in the reference formula set, with the description of the target performance as a constraint, the method further includes: selecting a target reference formula from the reference formula set. Correspondingly, generating a suggested formula for the target product using a formula generation model based on reference formulas in the reference formula set, with the description of the target performance as a constraint, includes: adjusting the target reference formula using the formula generation model to generate a suggested formula for the target product, with the description of the target performance as a constraint.

[0008] In some embodiments, extracting a recipe from at least one data source based on the recipe requirement information to obtain a reference recipe set includes: performing a search in the at least one data source based on the recipe requirement information to obtain a reference recipe source set, the reference recipe source set including at least one reference recipe source related to the recipe requirement information; and extracting recipes from the reference recipe source set to obtain the reference recipe set.

[0009] In some embodiments, before generating a suggested formula for the target product using a formula generation model based on reference formulas in the reference formula set, with the description of the target performance as a constraint, the method further includes: aggregating formula information in the reference formula source set to establish a knowledge base. Accordingly, generating a suggested formula for the target product using a formula generation model based on reference formulas in the reference formula set, with the description of the target performance as a constraint, includes: generating a suggested formula for the target product using the formula generation model based on reference formulas in the reference formula set and knowledge in the knowledge base, with the description of the target performance as a constraint.

[0010] In some embodiments, before generating a suggested formula for the target product based on reference formulas in the reference formula set using a formula generation model with the description of the target performance as a constraint, the method further includes: generating an extended query based on the description of the target performance; performing a retrieval in the at least one data source according to the extended query to obtain an extended formula source set, the extended formula source set including at least one extended formula source related to the description of the target performance. Accordingly, the formula information in the reference formula source set is aggregated to establish a knowledge base, including: aggregating the formula information in the reference formula source set and the extended formula source set to establish the knowledge base.

[0011] In some embodiments, the knowledge base includes at least one of the following: the relationship between components and attributes, the mechanism by which components affect target attributes, the relationship between target functions and components, the scientific principles of the formulation, and the co-occurrence relationship of components; wherein the target attributes and the target functions are both related to the target performance.

[0012] In some embodiments, using the description of the target performance as a constraint, a formula generation model is used to generate a suggested formula for the target product based on reference formulas in the reference formula set. This includes: using the description of the target performance as a constraint, the formula generation model selects a suggested combination of ingredients based on knowledge in the knowledge base; determining a feasible ratio range for each ingredient in the suggested combination based on the reference formulas in the reference formula set; and searching for suggested ratios for each ingredient in the suggested combination within the feasible ratio range to generate a suggested formula for the target product.

[0013] In some embodiments, performing a retrieval in the at least one data source based on the formula requirement information to obtain a reference formula source set includes: performing fragment matching in the at least one data source based on the formula requirement information to obtain a source index set, the source index set including indexes of reference formula sources to which at least one related fragment belongs, the related fragment being related to the formula requirement information; and obtaining the reference formula source set based on the source index set.

[0014] Secondly, embodiments of this application provide an electronic device, including a processor and a memory. The memory stores computer instructions, which, when executed by the processor, implement the method described in the first aspect.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described in the first aspect.

[0016] Fourthly, embodiments of this application provide a computer program product comprising program code, which, when executed by a processor, implements the method described in the first aspect.

[0017] In this embodiment, by identifying formula requirement information and extracting reference formulas, input is provided to the formula generation model, thereby automatically generating a suggested formula for the target product. This improves formula development efficiency and reduces the reliance on specialized personnel in the development process. Attached Figure Description

[0018] Figure 1 This is a first flowchart of a recipe generation method provided in an embodiment of this application.

[0019] Figure 2 This is a second flowchart of the recipe generation method provided in an embodiment of this application.

[0020] Figure 3 The following is an exemplary process for determining the formulation style to which each reference formulation in the reference formulation set belongs.

[0021] Figure 4 This is a third flowchart of the recipe generation method provided in an embodiment of this application.

[0022] Figure 5 The fourth flowchart is provided for the recipe generation method in the embodiments of this application.

[0023] Figure 6 The fifth flowchart is a method for generating a formula provided in an embodiment of this application.

[0024] Figure 7 The diagram illustrates an exemplary process for generating a suggested formula for a target product.

[0025] Figure 8 This is a schematic flowchart of the formula generation method provided in a preferred embodiment of this application.

[0026] Figure 9 A block diagram of the formula generation system provided in the embodiments of this application.

[0027] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] 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.

[0029] Figure 1 This is a first flowchart of a recipe generation method provided in an embodiment of this application. The recipe generation method 100 is executed by a recipe generation system (hereinafter referred to as the system), for example, by... Figure 9 The recipe generation system 900 shown is executed. (As...) Figure 1 As shown, the formula generation method 100 includes steps 110 to 140.

[0030] Step 110: Obtain user query.

[0031] Specifically, users can input at least one type of information, such as text, voice, images, or files, into the user interface for querying.

[0032] In practical applications, users can make one or more rounds of input to provide user queries. During the dialogue, the system can guide users to describe their recipe requirements in order to identify complete recipe requirement information.

[0033] Step 120: Entity recognition is performed based on the user query to obtain formula requirement information.

[0034] Formulation requirements information includes a description of the type of target product and the target performance. Formulation performance refers to the effectiveness of the formulation under specific conditions or applications. Formulation performance can be characterized by quantitative indicators.

[0035] Formula requirements information can be considered as user intent. In this embodiment, user intent is divided into at least two parts: the type of the target product and a description of the target performance. As one possible implementation, user intent can be divided into three parts: the type of the target product, a description of the target performance, and other requirement descriptions. As an example only, if a user inputs "I want to develop a sunscreen with SPF50, waterproof properties, and containing glycerin," the system identifies the target product as "sunscreen" and identifies two target performance properties: SPF50 and waterproof.

[0036] The system can utilize a multi-intent recognition model to process user queries and obtain recipe requirement information. This model can identify multiple concurrent intents from a single user input. Intent recognition models include, but are not limited to, Large Language Models (LLMs), multi-task joint learning models, graph neural networks, and hybrid expert systems.

[0037] In some embodiments, recipe requirement information is stored in the form of structured metadata to provide additional context for models (such as LLM) and to support the model in better understanding user queries in subsequent processes.

[0038] In practical applications, after identifying the formula requirements, this information can be displayed on the user interface. If the user believes the formula requirements need adjustment, they can provide feedback through the user interface so the system can adjust the information accordingly.

[0039] Step 130: Extract the recipe from at least one data source based on the recipe requirement information to obtain a reference recipe set.

[0040] The reference formula set includes at least one reference formula that is related to the formula requirement information.

[0041] The at least one data source includes, but is not limited to, patent databases, academic literature databases, websites, etc. As an example only, the system can extract formulas from patent databases and academic literature databases.

[0042] In some embodiments, the system performs a retrieval in the at least one data source based on the formulation requirement information to obtain a reference formulation source set. The reference formulation source set includes at least one reference formulation source related to the formulation requirement information. Then, the system extracts formulations from the reference formulation source set to obtain a reference formulation set. As an example only, when the at least one data source includes patent databases, academic literature databases, and websites, the reference formulation source can be a patent, academic literature (such as a paper), or a webpage.

[0043] Traditional data retrieval techniques typically focus on the semantic similarity of search results, but this is often not the sole metric for evaluating the quality of those results. For example, users in some industries tend to seek recipes from large, reputable companies or publications. Therefore, the following method provides a way to filter reference recipe sources based on the metadata of candidate recipe sources.

[0044] Specifically, the system performs a search in at least one data source based on the formula requirement information to obtain a candidate formula source set, which includes multiple candidate formula sources related to the formula requirement information. Then, the system obtains metadata of the multiple candidate formula sources and sorts them according to the metadata. The metadata includes one or more sorting indicators. Finally, the system selects a top-preset number or a top-preset proportion of candidate formula sources as reference formula sources based on the sorting results.

[0045] As an example, for patent databases, the system can use a weighted mechanism to score candidate patents based on metadata such as company (applicant / patentee) size and patent valuation. For academic literature databases, the system can use a weighted mechanism to score candidate academic literature based on metadata such as journal reputation and citation count. The system then ranks the candidate patents and academic literature according to their scores. The weights are adjustable to change the ranking of the formula sources returned to the user.

[0046] In practical applications, reference formulation sources can be displayed in the user interface according to the sorting results. The displayed items for reference formulation sources include, but are not limited to, title, abstract, publication date, and published reference formulation.

[0047] Some document retrieval techniques (such as embedding matching) operate at the fragment level, resulting in a wide variation in the length of the retrieved content. Theoretically, a fragment can contain up to 512 tokens, but log checks show that most fragments are between 4 and 5 sentences long. Since complete recipes typically span long paragraphs, the limited fragment length is often insufficient, leading to misleading model output. Therefore, a method is provided below to obtain the complete recipe source by indexing the recipe source (such as the recipe source ID).

[0048] Specifically, the system performs fragment matching based on the formulation requirement information from at least one data source to obtain a source index set. The source index set includes indexes (such as document IDs) of the reference formulation sources to which each of the relevant fragments belongs, and these relevant fragments are associated with the formulation requirement information. That is, the relevant fragments are fragments of candidate formulation sources. Furthermore, the system obtains the reference formulation sources based on the source index set.

[0049] Step 140: Using the description of the target performance as a constraint, generate a suggested formula for the target product based on the reference formulas in the reference formula set using the formula generation model.

[0050] The recipe generation model includes a trained artificial intelligence model, such as a trained large language model. The recipe generation model and the aforementioned multi-intent recognition model can be different models or the same model.

[0051] In some embodiments, the system can write recipe requirement information and recipes from a reference recipe set into the context memory of a large language model, and process the information in the context memory through the large language model to generate a suggested recipe for the target product.

[0052] Suggested formulations for the target product can include improved formulations and innovative formulations. In practical applications, a label can be displayed for each suggested formulation on the user interface to distinguish between improved and innovative formulations. For details on generating improved and innovative formulations, please refer to [link / reference needed]. Figure 4, Figure 7 The details and related explanations will not be elaborated here.

[0053] In this embodiment, by identifying formulation requirements and extracting reference formulations, input is provided to the formulation generation model, thereby automatically generating a suggested formulation for the target product. This improves formulation development efficiency and reduces the reliance on specialized personnel in the development process.

[0054] Figure 2 This is a second flowchart of a recipe generation method provided in an embodiment of this application. The recipe generation method 200 is executed by a recipe generation system (hereinafter referred to as the system), for example, by... Figure 9 The recipe generation system 900 shown is executed. (As...) Figure 2 As shown, the formula generation method 200 includes steps 210 to 260.

[0055] Step 210: Obtain user query.

[0056] Step 220: Entity recognition is performed based on the user query to obtain formula requirement information.

[0057] The formulation requirements information includes a description of the type of target product and the target performance.

[0058] Step 230: Generate descriptions for multiple formulation styles based on the description of the target performance.

[0059] Different formulation styles imply different focuses during formulation generation. The description of a formulation style is sometimes referred to as a formulation strategy. Specifically, each target performance can correspond to a formulation style; that is, each formulation style emphasizes the achievement of the corresponding target performance. The system can also extend to other formulation styles, for example, by expanding other formulation styles based on technical paths, design directions, constraint identification, and typical path references, or by developing guidelines for specific industries (such as specialty chemicals, consumer products, adhesives, food and beverages, and agricultural chemicals) to expand other formulation styles.

[0060] As an example, when a user inputs "I want to develop a sunscreen with SPF50, waterproof properties, and containing glycerin," the system identifies two target properties: SPF50 and waterproofness. Accordingly, the system can determine two formulation styles: one focusing on sun protection performance and the other on waterproof performance. Furthermore, the system can expand to include more formulation styles, such as those emphasizing environmental friendliness or aesthetic appeal.

[0061] Step 240: Extract the recipe from at least one data source based on the recipe requirement information to obtain a reference recipe set.

[0062] The reference formula set includes at least one reference formula related to the formula requirement information.

[0063] Step 250: Based on the reference formula set and the description of the multiple formula styles, determine the formula style to which each reference formula in the reference formula set belongs.

[0064] In some embodiments, the system calculates a relevance score between each reference recipe in the reference recipe set and the description of each recipe style (e.g., according to a preset scoring rule or using a large language model). Then, the system determines the recipe style to which each reference recipe in the reference recipe set belongs based on the relevance score between each reference recipe in the reference recipe set and the description of each recipe style. As an example, there are a total of 3 recipe styles. For each reference recipe, the system calculates a relevance score between the reference recipe and the description of each recipe style, obtaining a set of relevance scores (s1, s2, s3), which is also called a relevance vector. Then, the system determines the recipe style corresponding to the maximum value among s1, s2, and s3 as the recipe style to which that recipe belongs.

[0065] In some embodiments, it can also be in accordance with Figure 3 The process shown determines the formulation style to which each reference formulation in the reference formulation set belongs.

[0066] Step 260: For each formulation style, with the description of the target performance as a constraint, use the formulation generation model to generate a suggested formulation for the target product belonging to that formulation style based on the reference formulations in the reference formulation set that belong to that formulation style.

[0067] The recipe generation model includes a trained artificial intelligence model.

[0068] In this embodiment, by generating descriptions of multiple formulation styles, the diversity of generated formulations can be enhanced, helping users expand their R&D ideas and product branding opportunities. For example, for sunscreens, by identifying formulation styles that emphasize sun protection performance, water resistance, environmental friendliness, and aesthetics, users can launch sunscreens with different styles.

[0069] In practical applications, suggested recipes for different recipe styles can be grouped and displayed in the user interface, with each group labeled with a description of the corresponding recipe style.

[0070] For more details on recipe generation method 200 and its steps, such as user queries, recipe requirement information, and details related to the recipe generation model, please refer to [link / reference]. Figure 1 The details and related explanations will not be elaborated here.

[0071] Figure 3 The diagram illustrates an exemplary process for determining the formulation style to which each reference formulation in a reference formulation set belongs. As shown in the figure, process 300 includes steps 310-320.

[0072] Step 310: Under preset constraints, adjust the mapping relationship between the reference formula and the formula style in the direction of increasing the similarity of the target.

[0073] The preset constraints restrict each reference recipe to only one recipe style, and the number of reference recipes mapped to each recipe style must be at least a preset number. The target similarity is the similarity between the relevance score set and the mapping relationship, whereby the relevance score set includes the relevance score between each reference recipe in the reference recipe set and the description of each recipe style.

[0074] Specifically, the system can initialize the mapping relationship and calculate an initial value of the target similarity based on the initialized mapping relationship. Then, the system adjusts the mapping relationship in the direction of increasing the target similarity until the target similarity converges or reaches a preset number of adjustments.

[0075] Step 320: Determine the formula style of each reference formula in the reference formula set according to the adjusted mapping relationship.

[0076] As one possible implementation, the system can use a mixed-integer linear programming (MILP) optimization model to classify the multiple existing recipes. The MILP optimization model is as follows: (1) st , (2) , (3) (4) in, R It is an n×3 matrix, consisting of the correlation vectors of n reference formulas; x It is an n×3 matrix, where each row represents the recipe mapping situation (e.g., [0 1 0] means that reference recipe 1 is mapped to recipe style 2). matrix multiplication The trace represents the target similarity, i.e., the optimization objective is to maximize the target similarity (i.e., minimize the negative of the target similarity); the first constraint is used to constrain the matrix. x The elements are binary integers; the second constraint is used to restrict the matrix. x Each row contains only one "1", meaning that each reference recipe can only be mapped to one recipe style; the third constraint requires that there are at least k reference recipes mapped to each recipe style.

[0077] In this embodiment, by adjusting the formula mapping under preset constraints to improve the overall relevance score, the formula distribution can be balanced among multiple formula styles.

[0078] Figure 4 This is a third flowchart of the recipe generation method provided in this application embodiment. The recipe generation method 400 is executed by a recipe generation system (hereinafter referred to as the system), for example, by... Figure 9 The recipe generation system 900 shown is executed. (As...) Figure 4 As shown, the formula generation method 400 includes steps 410 to 450.

[0079] Step 410: Obtain user query.

[0080] Step 420: Entity recognition is performed based on the user query to obtain formula requirement information.

[0081] The formulation requirements information includes a description of the type of target product and the target performance.

[0082] Step 430: Extract the recipe from at least one data source based on the recipe requirement information to obtain a reference recipe set.

[0083] The reference formula set includes at least one reference formula related to the formula requirement information.

[0084] Step 440: Select a target reference formula from the reference formula set.

[0085] The target reference formula can be determined automatically by the system or specified by the user.

[0086] As one possible implementation, the system scores the reference formulas in the reference formula set and sorts them according to the scores. Then, the system identifies the top-scoring reference formulas (either a predetermined number or a predetermined percentage) as target reference formulas. Formula scoring can consider various factors. As an example, primary scoring factors include whether the contained ingredients meet target performance requirements and whether user-specified specific ingredients are included in the formula. Secondary scoring factors include estimated cost-effectiveness, supply chain resilience, compliance, safety, and environmental impact.

[0087] In practical applications, reference recipes can be displayed on the user interface according to the sorting results.

[0088] Step 450: Using the description of the target performance as a constraint, the target reference formula is adjusted using the formula generation model to generate a suggested formula for the target product.

[0089] The recipe generation model includes a trained artificial intelligence model.

[0090] A suggested formula for a target product obtained by adjusting a specific reference formula (target reference formula) can be considered an improved formula for the target product. The system can display the adjustments made to the improved formula of the target product compared to the target reference formula, such as the substitution, addition, or reduction of ingredients, or the adjustment of ingredient ratios.

[0091] In this embodiment, by identifying formula requirements and extracting and selecting reference formulas, input is provided to the formula generation model, thereby automatically generating an improved formula for the target product. Thus, a method for efficiently generating improved formulas for target products is provided.

[0092] For more details on recipe generation method 400 and its steps, such as user queries, recipe requirement information, and recipe generation model details, please refer to [link / reference]. Figure 1 The details and related explanations will not be elaborated here.

[0093] Figure 5 This is a fourth flowchart illustrating the recipe generation method provided in this application embodiment. The recipe generation method 500 is executed by a recipe generation system (hereinafter referred to as the system), for example, by... Figure 9 The recipe generation system 900 shown is executed. (As...) Figure 5 As shown, the formula generation method 500 includes steps 510 to 560.

[0094] Step 510: Obtain user query.

[0095] Step 520: Entity recognition is performed based on the user query to obtain formula requirement information.

[0096] The formulation requirements information includes a description of the type of target product and the target performance.

[0097] Step 530: Perform a search in at least one data source based on the formula requirement information to obtain a reference formula source set.

[0098] The reference formula source set includes at least one reference formula source related to the formula requirement information.

[0099] Step 540: Extract formulas from the reference formula source set to obtain the reference formula set.

[0100] The reference formula set includes at least one reference formula that is related to the formula requirement information.

[0101] Step 550: Aggregate the formula information from the reference formula source set to establish a knowledge base.

[0102] In some embodiments, the knowledge base includes at least one of the following: the relationship between ingredients and properties, the mechanism by which ingredients affect properties, the relationship between functions and ingredients, the scientific principles of the formulation, and ingredient co-occurrence relationships (such as an ingredient co-occurrence matrix, reflecting the compatibility between different ingredients). Here, an attribute (or property) refers to the inherent characteristics of the formulation, such as optical characteristics. A function refers to the design purpose of the formulation, such as sun protection or waterproofing.

[0103] As an example only, a knowledge base may include a property graph dictionary, a summary module, and a functional manual. These three types of knowledge are explained in detail below.

[0104] For each attribute, the attribute graph dictionary can include the related components of that attribute, the influence of those components on the attribute, the mechanism by which those components influence the attribute, and the source of the relevant formulation. The system can easily search for related attributes in structured data and extract the component combinations required to achieve the target effect.

[0105] For each formulation source, the summary module may include the source number, publication date, source type (e.g., patent / academic literature / webpage), scientific principle (e.g., chemical principle), performance information, test data, compliance information, and score.

[0106] The functional manual aims to break down a formulation into multiple functional groups and list the ingredients within each group. For example, sunscreen typically contains solvents, emollients, and UV filters. For each functional group, the manual may include the ingredients, a list of relevant formulation sources, a list of ingredient content ranges, a comprehensive ingredient content range, the maximum overlap range, and the number of occurrences. The ingredient content range list includes the content ranges of the same ingredient from different formulation sources. The comprehensive ingredient content range can be obtained by taking the union of the content ranges of the same ingredient from different formulation sources. The maximum overlap range can be obtained by taking the intersection of the content ranges of the same ingredient from different formulation sources. The number of occurrences indicates how many formulation sources mention the functional group. Recording the proportion range of each ingredient helps the model generate reasonable formulation recommendations and ensures the scientific feasibility of the generated formulation results. For example, when all sources indicate that ingredient X should be within 30% to 50%, a recommendation of "ingredient X at 70%" should be avoided.

[0107] Step 560: Using the description of the target performance as a constraint, generate a suggested formula for the target product based on the reference formulas in the reference formula set and the knowledge in the knowledge base using the formula generation model.

[0108] The recipe generation model includes a trained artificial intelligence model.

[0109] In some embodiments, the system writes formula requirement information, formulas from a reference formula set, and knowledge from a knowledge base into the context memory of a large language model, and processes the information in the context memory through the large language model to generate a suggested formula for the target product.

[0110] The addition of substances and changes in composition exhibit non-linear characteristics, making it difficult to determine whether proposed modifications and specific formulations can guarantee a certain performance level. Conducting experiments to verify the feasibility of specific formulations also disrupts the continuity of iterative cycles. In this embodiment, by retrieving a wide range of formulation sources from data sources and extracting fundamental formulation information (such as chemical principles and compatibility between different components), a solid theoretical foundation is provided for the formulation generation process, facilitating the rapid generation of feasible formulations.

[0111] In some embodiments, the recommended formulation includes basic information and supplementary information. The basic information includes a list of ingredients, a list of ingredient proportions, and a list of functions (each ingredient corresponds to a specific function). Supplementary information includes, but is not limited to, the mechanism of action of each ingredient, the scientific principles behind the formulation, the manufacturing process of the formulation, a risk statement of the formulation, and test instructions for the formulation.

[0112] For improved formulations, additional information may be suggested that the reasons for choosing the target reference formulation as the basis for improvement, the adjustments made compared to the target reference formulation (such as adjustments to ingredients and ingredient ratios), and performance advantages, etc.

[0113] For innovative formulations, it is suggested that the explanatory notes for the formulation may also include introductions to new ingredients, introductions to new functions, and improvements in performance indicators.

[0114] In practical applications, the basic information of the suggested recipe can be displayed first on the user interface, while the additional information of the suggested recipe is hidden by default (such as in a collapsed state). Users can click on preset controls (such as an expand button) to view the additional information of the suggested recipe.

[0115] For more details on recipe generation method 500 and its steps, such as user queries, recipe requirement information, and recipe generation model details, please refer to [link / reference]. Figure 1 The details and related explanations will not be elaborated here.

[0116] Figure 6 This is a fifth flowchart illustrating the recipe generation method provided in this application embodiment. The recipe generation method 600 is executed by a recipe generation system (hereinafter referred to as the system), for example, by... Figure 9 The recipe generation system 900 shown is executed. (As...) Figure 6 As shown, the formula generation method 600 includes steps 610 to 680.

[0117] Step 610: Obtain user query.

[0118] Step 620: Entity recognition is performed based on the user query to obtain formula requirement information.

[0119] The formulation requirements information includes a description of the type of target product and the target performance.

[0120] Step 630: Perform a search in at least one data source based on the formula requirement information to obtain a reference formula source set.

[0121] The reference formula source set includes at least one reference formula source related to the formula requirement information.

[0122] Step 640: Extract formulas from the reference formula source set to obtain the reference formula set.

[0123] The reference formula set includes at least one reference formula that is related to the formula requirement information.

[0124] Step 650: Generate extended queries based on the description of the target performance.

[0125] Besides performance targets, query expansion generation can also be considered from the following aspects: technical approach, design direction, constraint identification, and typical path reference. For specific industries, such as specialty chemicals, consumer goods, adhesives, food and beverages, and agricultural chemicals, specialized query expansion strategies can be developed.

[0126] Step 660: Perform a retrieval in the at least one data source based on the extended query to obtain an extended recipe source set.

[0127] The extended formulation source set includes at least one extended formulation source related to the description of the target performance. Obtaining the extended formulation source is similar to obtaining the reference formulation source; that is, for further details regarding step 660, please refer to the relevant description of step 130.

[0128] Query expansion aims to increase the diversity of formula sources in search results, enabling formula generation models to acquire broader formula knowledge, including but not limited to a wider range of considerations, technical information, and ingredient information, thereby improving the quality of formula generation results. For example, query expansion can explore novel ingredients that have not yet been used in the target product but have been proven to achieve the target performance in similar products. As another example, query expansion can obtain knowledge from related industries regarding achieving the target performance.

[0129] Step 670: Aggregate the formula information from the reference formula source set and the extended formula source set to establish a knowledge base.

[0130] Step 680: Using the description of the target performance as a constraint, generate a suggested formula for the target product based on the reference formulas in the reference formula set and the knowledge in the knowledge base using the formula generation model.

[0131] The recipe generation model includes a trained artificial intelligence model.

[0132] For more details regarding recipe generation method 600 and its steps, such as user queries, recipe requirement information, recipe generation models, and knowledge base details, please refer to [link / reference needed]. Figure 1 , Figure 5 The details and related explanations will not be elaborated here.

[0133] Figure 7 The diagram illustrates an exemplary process for generating a suggested formula for a target product. Figure 7 As shown, process 700 includes steps 710 to 730.

[0134] Step 710: Using the description of the target performance as a constraint, the recipe generation model selects suggested ingredient combinations based on knowledge in the knowledge base.

[0135] Referring to the foregoing embodiments, the knowledge base can record knowledge such as the relationship between components and attributes, and the relationship between functions and components, which enables the system to extract the component combinations (i.e., suggested component combinations) required to achieve the target performance from this knowledge.

[0136] In some embodiments, in addition to accessing an internal knowledge base, the system can also access external databases (such as websites containing up-to-date data) and other internal databases (such as customizable chemical databases) to select suggested ingredient combinations.

[0137] Step 720: Determine the feasible ratio range of each component in the recommended ingredient combination based on the reference formulations in the reference formulation set.

[0138] In some embodiments, the system maps multiple reference formulations from a reference formulation set to an n-dimensional space, where n is the number of candidate ingredients, and the value of each dimension represents the proportion of the corresponding ingredient. Furthermore, the system determines a feasibility space in the n-dimensional space based on the multiple reference formulations, the feasibility space representing the feasible proportion range of each ingredient in the proposed ingredient combination.

[0139] In some embodiments, for each formulation style, the system determines the feasible ratio range of each component in the proposed ingredient combination based on multiple reference formulations belonging to that formulation style in the reference formulation set, so as to generate a proposed formulation for the target product belonging to that formulation style. For example, the system can map multiple reference formulations belonging to that formulation style in the reference formulation set to an n-dimensional space, and determine a feasibility space in the n-dimensional space based on the multiple reference formulations belonging to that formulation style.

[0140] Step 730: Within the feasible ratio range, search for the suggested ratios of each component in the suggested ingredient combination to generate a suggested formula for the target product.

[0141] In some embodiments, the system can generate multiple candidate component combinations through multiple iterations and select a suggested component combination from these combinations, for example, selecting the optimal candidate component combination as the suggested component combination. The system can verify the compatibility between different components in the candidate component combinations (e.g., using a component co-occurrence matrix in a knowledge base).

[0142] In some embodiments, the system utilizes the Dirichlet recalibration algorithm to search for suggested proportions of each component in a proposed ingredient combination within a feasible range (such as a feasibility space in n-dimensional space) to generate a suggested formulation for the target product. By controlling the algorithm parameters, the generation of improved or innovative formulations can be controlled. Specifically, by controlling the algorithm parameters, improved formulations can be generated in regions of n-dimensional space close to the reference formulation; by controlling the algorithm parameters, innovative formulations can also be generated in regions of n-dimensional space far from the reference formulation.

[0143] Figure 8 This is a schematic flowchart of the formula generation method provided in a preferred embodiment of this application.

[0144] like Figure 8 As shown, on the one hand, the system identifies formulation requirement information based on user queries and performs document searches in patent databases, academic literature databases, and websites based on this information. On the other hand, the system generates extended queries based on the descriptions of target performance in the formulation requirement information and performs document searches in patent databases, academic literature databases, and websites based on these extended queries.

[0145] The system can score and rank multiple candidate formulation sources (patents / candidate academic literature) according to preset scoring criteria, and then select the top N (i.e. the N with the highest scores) candidate formulation sources as reference formulation sources or extended formulation sources.

[0146] After obtaining recipe sources from multiple data sources, the system aggregates the recipe information from these sources to build a knowledge base.

[0147] Regarding formula extraction, the system can use the N highest-scoring candidate formula sources retrieved from patent databases and academic literature databases as reference formula sources, and extract formulas from these sources to obtain a reference formula set. This reference formula set is used in the subsequent formula generation stage. Furthermore, the reference formula set can also be used to build a knowledge base.

[0148] The system can generate descriptions of multiple formulation styles based on the description of the target performance. Furthermore, the system can classify the reference formulations in the reference formulation set, i.e., determine the formulation style to which each reference formulation belongs.

[0149] For each formulation style, the system can select M target reference formulations belonging to that formulation style based on the knowledge in the knowledge base. These M target reference formulations are used to generate a suggested formulation for the target product.

[0150] The system can invoke iterative search tools and deterministic functions to write knowledge from the knowledge base, web search results, and data from the chemical database into the context memory of the large language model, and then process the information in the context memory through the large language model to generate a suggested formula for the target product.

[0151] During the formula generation stage, the system first uses a large language model to select ingredients and obtain suggested ingredient combinations. Then, based on a reference formula set, the system determines the suggested proportions of each ingredient in the suggested ingredient combinations. Finally, the system outputs a suggested formula for the target product, which includes the suggested ingredient combinations and the suggested proportions of each ingredient in the suggested ingredient combinations.

[0152] It's worth noting that the recipe generation method can call the large language model multiple times. For example, it can be used to identify recipe requirements, generate descriptions of multiple recipe styles, and generate suggested recipes for the target product. These steps based on the large language model can all utilize few-shot prompting techniques to encourage the model to generate output in the expected structured format. Few-shot prompting enables subsequent steps to functionally parse and process the model output.

[0153] In the recipe generation stage, Retrieval-Augmented Generation (RAG) technology can be utilized. Specifically, the recipe generation process involves multiple iterations. In each iteration, the system processes information from the context memory (such as knowledge in the knowledge base and recipe requirement information) through a large language model to obtain the generation result for the current round. Then, the system generates reliability feedback for the current round based on the generated content, i.e., determining whether the generated content for the current round is reliable. If the generated content for the current round is determined to be reliable, the system writes the reliability feedback for the current round into the context memory of the large language model to proceed to the next round of generation. If the generated content for the current round is determined to be reliable, the system directly outputs the generated content for the current round as the final generated content (such as a suggested recipe for the target product). This mechanism helps to iteratively select the best result at each step and gradually converge to the final answer.

[0154] Figure 9This is an exemplary block diagram of the recipe generation system provided in the embodiments of this application. Figure 9 As shown, the recipe generation system 900 includes a query acquisition module 910, a demand identification module 920, a recipe extraction module 940, and a recipe generation module 960. In practical applications, the recipe generation system 900 can be implemented based on an intelligent agent. The intelligent agent utilizes artificial intelligence technologies (such as reasoning, planning, memory, and learning abilities) to perceive the environment, make autonomous decisions, and execute actions to achieve specific goals or complete tasks assigned on behalf of the user.

[0155] The query retrieval module 910 is used to retrieve user queries.

[0156] The demand identification module 920 is used to: perform entity identification based on the user query to obtain formula demand information. The formula demand information includes the type of the target product and a description of the target performance.

[0157] The formula extraction module 940 is used to: extract formulas from at least one data source based on the formula requirement information to obtain a reference formula set. The reference formula set includes at least one reference formula related to the formula requirement information.

[0158] The formula generation module 960 is used to: generate a suggested formula for the target product based on reference formulas in the reference formula set, using the description of the target performance as a constraint, and employing a formula generation model. The formula generation model includes a trained artificial intelligence model.

[0159] In some embodiments, the formula generation system 900 further includes a style determination module 930. The style determination module 930 is configured to: generate descriptions of multiple formula styles based on the description of the target performance; and determine the formula style to which each reference formula in the reference formula set belongs, based on the reference formula set and the descriptions of the multiple formula styles. Correspondingly, the formula generation module 960 is configured to: for each formula style, using the description of the target performance as a constraint, and utilizing the formula generation model to generate a suggested formula for the target product belonging to that formula style based on the reference formulas in the reference formula set belonging to that formula style.

[0160] In some embodiments, the formula extraction module 940 is configured to: perform a retrieval in the at least one data source based on the formula requirement information to obtain a reference formula source set, the reference formula source set including at least one reference formula source related to the formula requirement information; and extract formulas from the reference formula source set to obtain the reference formula set. Based on this, the formula generation system 900 further includes a knowledge base construction module 950. The knowledge base construction module 950 is configured to: aggregate the formula information in the reference formula source set to establish a knowledge base. The formula generation module 960 is configured to: use the description of the target performance as a constraint, and utilize the formula generation model to generate a suggested formula for the target product based on the reference formulas in the reference formula set and the knowledge in the knowledge base.

[0161] For more details about the Recipe Generation System 900 and its modules, please refer to [link / reference]. Figures 1-8 The details and related explanations will not be elaborated here.

[0162] refer to Figure 10 This application also provides an electronic device 1000. The electronic device 1000 includes a processor 1010 and a memory 1020. The memory 1020 stores computer instructions, which, when executed by the processor 1010, implement the recipe generation method provided in this application.

[0163] This application also provides a computer-readable storage medium. The storage medium stores computer instructions, which, when executed by a processor, implement the recipe generation method provided in this application.

[0164] This application also provides a computer program product comprising program code, which, when executed by a processor, implements the recipe generation method provided in this application.

[0165] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications or equivalent substitutions made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating a formula, characterized in that, include: Get user queries; Entity recognition is performed based on the user query to obtain formula requirement information, which includes the type of target product and a description of the target performance. Based on the formula requirement information, a formula is extracted from at least one data source to obtain a reference formula set, the reference formula set including at least one reference formula related to the formula requirement information; Using the description of the target performance as a constraint, a recommended formula for the target product is generated based on reference formulas in the reference formula set using a formula generation model, wherein the formula generation model includes a trained artificial intelligence model. Based on the formula requirement information, formulas are extracted from at least one data source to obtain a reference formula set, including: Based on the formula requirement information, a retrieval is performed in the at least one data source to obtain a reference formula source set, the reference formula source set including at least one reference formula source related to the formula requirement information; Formulas are extracted from the reference formula source set to obtain the reference formula set.

2. The method according to claim 1, characterized in that, Before generating a suggested formulation for the target product based on reference formulations in the reference formulation set using a formulation generation model with the description of the target performance as a constraint, the method further includes: Multiple formulation style descriptions are generated based on the description of the target performance; Based on the reference formula set and the descriptions of the multiple formula styles, determine the formula style to which each reference formula in the reference formula set belongs; Using the description of the target performance as a constraint, a suggested formula for the target product is generated based on reference formulas in the reference formula set using a formula generation model, including: For each formulation style, with the description of the target performance as a constraint, the formulation generation model generates a suggested formulation for the target product belonging to that formulation style based on reference formulations belonging to that formulation style in the reference formulation set.

3. The method according to claim 2, characterized in that, Based on the reference recipe set and the descriptions of the multiple recipe styles, the recipe style to which each reference recipe in the reference recipe set belongs is determined, including: Under preset constraints, the mapping relationship between reference recipes and recipe styles is adjusted in the direction of increasing target similarity; the preset constraints are used to restrict each reference recipe to only one recipe style, and there are at least a preset number of reference recipes mapped to each recipe style; the target similarity is the similarity between the relevance score set and the mapping relationship, and the relevance score set includes the relevance score between each reference recipe in the reference recipe set and the description of each recipe style; The formulation style to which each reference formulation in the reference formulation set belongs is determined based on the adjusted mapping relationship.

4. The method according to claim 1, characterized in that, Before generating a suggested formulation for the target product based on reference formulations in the reference formulation set using a formulation generation model with the description of the target performance as a constraint, the method further includes: Select a target reference formula from the reference formula set; Using the description of the target performance as a constraint, a suggested formula for the target product is generated based on reference formulas in the reference formula set using a formula generation model, including: Using the description of the target performance as a constraint, the target reference formula is adjusted using the formula generation model to generate a suggested formula for the target product.

5. The method according to claim 1, characterized in that, Before generating a suggested formulation for the target product based on reference formulations in the reference formulation set using a formulation generation model with the description of the target performance as a constraint, the method further includes: The formula information from the reference formula source set is aggregated to establish a knowledge base; Using the description of the target performance as a constraint, a suggested formula for the target product is generated based on reference formulas in the reference formula set using a formula generation model, including: Using the description of the target performance as a constraint, the formula generation model generates a suggested formula for the target product based on reference formulas in the reference formula set and knowledge in the knowledge base.

6. The method according to claim 5, characterized in that, Before generating a suggested formulation for the target product based on reference formulations in the reference formulation set using a formulation generation model with the description of the target performance as a constraint, the method further includes: Generate extended queries based on the description of the target performance; The extended query is used to perform a retrieval in the at least one data source to obtain an extended recipe source set, which includes at least one extended recipe source related to the description of the target performance. The formula information from the reference formula source set is aggregated to establish a knowledge base, including: The formula information in the reference formula source set and the extended formula source set is aggregated to establish the knowledge base.

7. The method according to claim 5, characterized in that, The knowledge base includes at least one of the following: the relationship between components and attributes, the mechanism by which components affect target attributes, the relationship between target functions and components, the scientific principles of formulation, and component co-occurrence relationships; wherein the target attributes and the target functions are both related to the target performance.

8. The method according to claim 5, characterized in that, Using the description of the target performance as a constraint, a suggested formula for the target product is generated based on reference formulas in the reference formula set using a formula generation model, including: Using the description of the target performance as a constraint, the recipe generation model selects suggested ingredient combinations based on knowledge in the knowledge base; Based on the reference formulations in the reference formulation set, determine the feasible ratio range of each component in the proposed ingredient combination; Within the feasible ratio range, the suggested ratios of each component in the suggested ingredient combination are searched to generate a suggested formulation for the target product.

9. The method according to claim 1, characterized in that, Based on the formula requirement information, a retrieval is performed in the at least one data source to obtain a reference formula source set, including: Based on the formula requirement information, fragment matching is performed in at least one data source to obtain a source index set, the source index set including the index of the reference formula source to which each of the relevant fragments belongs, the relevant fragments being related to the formula requirement information; The reference formula source set is obtained based on the source index set.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer instructions, which, when executed by the processor, implement the method as described in any one of claims 1 to 9.

11. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions, which, when executed by the processor, implement the method as described in any one of claims 1 to 9.

12. A computer program product, characterized in that, It includes program code that, when executed by a processor, implements the method as described in any one of claims 1 to 9.

Images