Formula optimization method and device, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence and material engineering, and discloses a formula optimization method and device, equipment and a storage medium, and the method comprises the steps: carrying out the knowledge extraction of an obtained structured formula data set and unstructured technical literature data in response to a formula optimization target input by a user, and generating a table literature knowledge set; inputting the table literature knowledge set and the formula optimization target into a large language model to obtain a generated text corresponding to the formula optimization target; performing logic rule screening and risk assessment on the generated text through a knowledge fusion layer to obtain text output conforming to a confidence threshold, and generating a new formula scheme according to the text output; and performing multi-objective optimization on the new formula scheme based on a preset experimental cost constraint condition, and outputting an optimized recommended formula and a support evidence chain. By automatically fusing innovative components and process information in literatures, the output recommended formula has scientific basis and interpretability, and the practicability and innovativeness of an automatic formula are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and material engineering, and particularly relates to a formula optimization method and device, equipment and a storage medium. BACKGROUND

[0002] The existing machine learning method is difficult to use a large amount of natural language formula such as patents and journals to innovate by absorbing industry or general knowledge when automatically designing new formula and new process, which is limited by existing data patterns. SUMMARY

[0003] The main purpose of the present application is to provide a formula optimization method, device, equipment and storage medium, which aims to solve the technical problem that the existing machine learning method cannot absorb industry or general knowledge to innovate when automatically designing new formula and new process.

[0004] To achieve the above purpose, the present application provides a formula optimization method, which comprises:

[0005] In response to a formula optimization target input by a user, knowledge extraction is performed on an obtained structured formula data set and unstructured technical literature data to generate a table literature knowledge set;

[0006] The table literature knowledge set and the formula optimization target are input into a large language model to obtain generated text corresponding to the formula optimization target;

[0007] Logical rule screening and risk assessment are performed on the generated text through a knowledge fusion layer to obtain text output meeting a confidence threshold, and a new formula scheme is generated according to the text output;

[0008] Multi-objective optimization is performed on the new formula scheme based on a preset experimental cost constraint condition to output an optimized recommended formula and a supporting evidence chain.

[0009] Optionally, the step of inputting the table literature knowledge set and the formula optimization target into the large language model to obtain generated text corresponding to the formula optimization target comprises:

[0010] The formula optimization target is analyzed to obtain performance optimization prompts and cost optimization prompts;

[0011] A dynamic prompt template is constructed according to the performance optimization prompts and the cost optimization prompts;

[0012] The dynamic prompt template is input into the large language model, and a task-oriented prompt string is generated through existing patterns in the table literature knowledge set and industry best practices, and the task-oriented prompt string is analyzed to output predicted generated text.

[0013] Optionally, the step of performing multi-objective optimization on the new formulation scheme based on the preset experimental cost constraint condition, outputting the optimized recommended formulation and supporting evidence chain, comprises:

[0014] constructing an experimental operation complexity cost function according to formulation history and external knowledge;

[0015] filtering out high-cost formulations in the new formulation scheme according to the experimental operation complexity cost function, to obtain initial optimized formulations;

[0016] determining risk levels of the initial optimized formulations based on the experimental cost constraint condition, and taking the initial optimized formulation with a low risk level as a recommended formulation of the formulation optimization target;

[0017] generating a supporting evidence chain corresponding to the recommended formulation, the supporting evidence chain being a tree-shaped explanation structure containing parameter decision nodes, literature evidence nodes and risk prompt nodes.

[0018] Optionally, after the step of determining risk levels of the initial optimized formulations based on the experimental cost constraint condition, and taking the initial optimized formulation with a low risk level as a recommended formulation of the formulation optimization target, the method further comprises:

[0019] obtaining a target formulation with a high risk level in the initial optimized formulations;

[0020] when the experimental complexity score and performance improvement score of the target formulation meet a preset optimization threshold, generating a risk prompt corresponding to the new formulation scheme, and triggering a process parameter adjustment instruction;

[0021] modifying ingredient proportion and process time parameters of the new formulation scheme according to the process parameter adjustment instruction and the risk prompt, to generate an adjusted formulation scheme.

[0022] Optionally, after the step of generating a supporting evidence chain corresponding to the recommended formulation, the method further comprises:

[0023] receiving a parameter adjustment instruction of the supporting evidence chain from a user;

[0024] performing dynamic verification on the experimental operation complexity cost function according to the parameter adjustment instruction, to obtain a verification result;

[0025] when the verification result is passed, updating the recommended formulation corresponding to the supporting evidence chain, and adding a user decision branch to the supporting evidence chain, the user decision branch containing a user instruction node, a real-time verification node and an evidence update marker.

[0026] Optionally, the step of knowledge extraction on the obtained structured recipe data set and unstructured technical literature data to generate the table literature knowledge set in response to the user input recipe optimization target comprises:

[0027] In response to the user input recipe optimization target, a structured recipe data set and unstructured technical literature data corresponding to the recipe optimization target are obtained;

[0028] The structured recipe data set is converted into a vector by a semantic encoder to obtain a first relationship vector, the first relationship vector containing the corresponding relationship between the recipe, the process and the performance;

[0029] The unstructured technical literature data set is parsed by a large language model to extract component replacement descriptions and process improvement sentences and converted into structured parameters to obtain a second relationship vector, the second relationship vector filtered by a risk scoring module, and the second relationship vector and the first relationship vector maintaining information complementarity rather than redundancy;

[0030] The table literature knowledge set is generated according to the first relationship vector and the second relationship vector.

[0031] Optionally, the knowledge fusion layer comprises a rule engine, a risk scoring module and a vectorization module; the rule engine checks the generated text of the large language model based on the domain expert rules and the prompt word template; the risk scoring module calculates the credibility score of the checked generated text; and the vectorization module maps the generated text with a credibility score greater than a preset credibility threshold to a vector feature and fuses it with the original table feature.

[0032] In addition, to achieve the above-mentioned purposes, the present application also proposes a recipe optimization device, the recipe optimization device comprising:

[0033] A data acquisition module for knowledge extraction on the obtained structured recipe data set and unstructured technical literature data to generate a table literature knowledge set in response to the user input recipe optimization target;

[0034] A text generation module for inputting the table literature knowledge set and the recipe optimization target into a large language model to obtain generated text corresponding to the recipe optimization target;

[0035] A recipe generation module for logical rule screening and risk assessment of the generated text by a knowledge fusion layer to obtain text output conforming to a confidence threshold, and generating a new recipe scheme according to the text output;

[0036] A recipe optimization module for multi-objective optimization of the new recipe scheme based on a preset experimental cost constraint condition to output an optimized recommended recipe and a supporting evidence chain.

[0037] In addition, to achieve the above-mentioned object, the present application also provides a formula optimization device, which comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the formula optimization method as described above.

[0038] In addition, to achieve the above-mentioned object, the present application also provides a storage medium, which is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the formula optimization method as described above.

[0039] In the present application, knowledge extraction is performed on the obtained structured formula data set and unstructured technical literature data in response to a user input formula optimization target to generate a table literature knowledge set; the table literature knowledge set and the formula optimization target are input into a large language model to obtain generated text corresponding to the formula optimization target; logical rule screening and risk assessment are performed on the generated text through a knowledge fusion layer to obtain text output meeting a confidence threshold, and a new formula scheme is generated according to the text output; the new formula scheme is subjected to multi-objective optimization based on a preset experimental cost constraint condition to output an optimized recommended formula and a supporting evidence chain. Through automatic fusion of innovative components and process information in the literature, combined with cost performance dual target optimization, the new formula design boundary is expanded, while ensuring controllable experimental resource consumption, and the output recommended formula has scientific basis and interpretability, effectively improving the practicability and innovativeness of formula optimization. BRIEF DESCRIPTION OF DRAWINGS

[0040] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application together with the specification.

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced here. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0042] Figure 1 Flowchart of the first embodiment of the formula optimization method of the present application;

[0043] Figure 2 Flowchart of the first embodiment of the formula optimization method of the present application;

[0044] Figure 3 Flowchart of the second embodiment of the formula optimization method of the present application;

[0045] Figure 4 The output result self-explaining structure and evidence chain node distribution diagram for the present application;

[0046] Figure 5 The flowchart of the third embodiment of the recipe optimization method of the present application;

[0047] Figure 6 The LLM-driven table text knowledge flow and recipe generation diagram of the present application;

[0048] Figure 7 The module structure diagram of the recipe optimization device of the embodiment of the present application;

[0049] Figure 8 The device structure diagram of the hardware running environment involved in the recipe optimization method of the embodiment of the present application.

[0050] The purpose implementation, functional features and advantages of the present application will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0051] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and do not limit the present application.

[0052] In order to better understand the technical solutions of the present application, the following will be described in detail in conjunction with the drawings and specific embodiments of the specification.

[0053] Traditional recipe optimization methods have exposed many defects in different industries, such as excessive dependence on manual experience, large blind trial, low efficiency, high cost, lack of scientific evaluation model, etc. These problems urgently need an innovative, efficient and intelligent recipe optimization method to solve, in order to meet the needs of various industries for product performance improvement, cost control and research and development efficiency improvement. However, the existing machine learning method is only based on standardized table data, and it is difficult to utilize a large amount of natural language recipe knowledge such as patents and journals. The automatic design of new recipes and new processes is limited by existing data patterns, cannot absorb industry or general knowledge innovation breakthroughs, lacks automated recipe feasibility verification and experimental resource consumption constraints, and the optimization lacks explainability and practicality.

[0054] Therefore, the present application provides a natural language knowledge transfer and experimental cost adaptive recipe optimization technology based on a large language model (LLM), which utilizes a large language model to comprehensively integrate unstructured literature knowledge and structured recipe-process-performance table data, intelligently extrapolates and generates new processes and new recipes, and combines with a cost-performance multi-objective optimization adaptive control method. It is suitable for multiple recipe optimization scenarios such as materials, chemical industry, food, etc.

[0055] It should be noted that the execution subject of the embodiment can be a computing service device with knowledge data processing, model prediction and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions. The adaptive formula optimization system is taken as an example to illustrate the embodiment and the following embodiments.

[0056] Based on this, the application embodiment provides a formula optimization method, referring to Figure 1 , Figure 1 The flowchart of the first embodiment of the formula optimization method of the application is shown.

[0057] In the embodiment, the formula optimization method comprises:

[0058] Step S10, in response to the formula optimization target input by the user, knowledge extraction is performed on the obtained structured formula data set and unstructured technical literature data to generate a table literature knowledge set.

[0059] It should be noted that the formula optimization target is an explicit optimization direction proposed by the user, such as improving product performance (such as strength, durability), reducing cost, shortening production cycle, etc. The structured formula data set can be data in the form of a standardized table, containing formula components, proportions, process parameters and corresponding performance indicators, etc. Information, which is convenient for direct data analysis and processing. Unstructured technical literature data can be text information in natural language form existing in technical literature and research reports, containing component replacement suggestions, process improvement methods, etc. Innovative content, but lack of unified format and structure. The table literature knowledge set is a unified knowledge set formed by integrating the knowledge in the structured data and the unstructured literature, which can be presented in the form of a table, containing the corresponding relationship between the formula, process and performance, and the innovative parameters extracted from the literature.

[0060] It should be understood that in order to ensure that the obtained structured data is accurate and complete, the unstructured literature has reliability and relevance. After obtaining the data, the data needs to be cleaned and preprocessed to remove noise and error information. For example, check the abnormal values in the formula data, verify the authority of the literature source, etc.

[0061] Step S20, inputting the table literature knowledge set and the formula optimization target into a large language model to obtain generated text corresponding to the formula optimization target.

[0062] It should be noted that when inputting the table literature knowledge set and the formula optimization target into the large language model, a suitable semantic encoder and large language model should be selected and necessary tuning should be performed. Different models may perform differently when processing data in a specific field, and selection and optimization need to be made according to actual conditions. For example, in the field of ceramics, a model that understands material science terminology well may need to be selected.

[0063] It should be understood that before inputting the table literature knowledge set and the formula optimization target into the large language model, the technical field of the current formula can be determined, and the large language model can be fine-tuned to better adapt to the formula optimization task. By fine-tuning on a dataset within the field, the model's understanding of professional terminology and specific tasks can be improved.

[0064] Further, in order to adaptively generate task-oriented prompt strings according to user needs, make the large language model more accurately analyze the needs, and avoid the defects of fixed prompt engineering in traditional methods. The step S20 can include:

[0065] Performing target analysis on the formula optimization target to obtain performance optimization prompts and cost optimization prompts; constructing a dynamic prompt template according to the performance optimization prompts and the cost optimization prompts; inputting the dynamic prompt template into the large language model, generating task-oriented prompt strings through existing patterns in the table literature knowledge set and industry best practices, and analyzing the task-oriented prompt strings to output predicted generated text.

[0066] It should be noted that target analysis can be the process of converting user input fuzzy formula optimization targets into specific and quantifiable performance optimization prompts and cost optimization prompts. For example, "improve product quality" is analyzed into "increase product strength by 20%" and "reduce cost by 15%". Performance optimization prompts are specific performance indicators that the product needs to improve, such as strength, durability, corrosion resistance, thermal conductivity, etc., usually expressed in quantitative form. Cost optimization prompts involve specific requirements for production cost control, including raw material cost, process cost, energy consumption cost, etc., and can include constraints (such as total cost ≤ 100 yuan / unit) or optimization directions (such as reducing raw material cost).

[0067] It should be understood that the dynamic prompt template is a text template automatically generated according to performance and cost optimization prompts, used to guide the large language model to generate task instructions that meet user needs. The template usually contains knowledge set references, constraints and optimization targets. The task-oriented prompt string is a specific operation instruction generated by the large language model according to the dynamic prompt template, for example, trying to replace material B with material A while keeping the cost below X, and adjusting process parameter C to range D.

[0068] Specifically, the natural language processing technique is used to analyze the user input optimization target, identify key performance indicators and cost constraints. For example, the user inputs "develop a low-cost, high-strength ceramic material", the parser identifies "strength" as a performance indicator and "cost" as a constraint condition. Then convert the vague target into quantifiable indicators. For example, convert high strength to compressive strength ≥ 300 MPa, and convert to raw material cost ≤ 50 yuan / kg. At the same time, different optimization targets are assigned weights to reflect their importance. For example, if the user pays more attention to cost, the weight of the cost optimization prompt can be set to 0.7, and the weight of the performance optimization prompt can be set to 0.3. Then generate a dynamic prompt template, for example, "Reference the knowledge set about high-strength ceramic formulations, design a ceramic formulation with compressive strength ≥ 300 MPa under the condition of cost ≤ 50 yuan / kg. Please list the ingredients, proportions and sintering process parameters, and explain the design basis". Then, input the dynamic prompt template into the large language model, and instruct the model to refer to the relevant cases in the table literature knowledge set. The model analyzes the existing patterns and industry best practices in the knowledge set to generate task-based prompt strings, and performs structured analysis on the generated task-based prompt strings to extract ingredient lists, proportions, process parameters, etc. For example, from the prompt string "use 70% alumina, 25% silicon dioxide and 5% magnesium oxide, sinter at 1600°C for 2 hours", extract the ingredient list, proportion and process parameter.

[0069] In step S30, the generated text is subjected to logical rule screening and risk assessment by the knowledge fusion layer to obtain text output that meets the confidence threshold, and a new formulation scheme is generated based on the text output.

[0070] It can be understood that the logical rule screening and risk assessment of the generated text by the knowledge fusion layer can be to apply pre-defined logical rules to filter conflicting or high-risk schemes; then use a quantitative model to calculate the potential credibility of the scheme, and screen out feasible schemes with a credibility higher than the confidence threshold; finally, based on the effective ingredient combination and process parameters in the qualified text, a new formulation scheme that meets the safety, compliance and innovation requirements is generated.

[0071] Further, the knowledge fusion layer includes a rule engine, a risk scoring module and a vectorization module; the rule engine checks the generated text of the large language model based on domain expert rules and prompt word templates; the risk scoring module calculates the credibility score of the checked generated text; the vectorization module maps the generated text with a credibility score greater than a preset credibility threshold to a vector feature and fuses it with the original table feature.

[0072] Specifically, the rule engine first performs a structured check on the text generated by the large language model based on domain expert rules (such as process compatibility, compliance constraints) and predefined prompt templates, filters out logical conflicts or high-risk content, then the risk scoring module quantitatively evaluates the checked text, calculates the credibility score (such as analyzes the potential failure risk through a probability model), and only keeps reliable text whose score exceeds the preset credibility threshold; finally, the vectorization module maps the qualified text into high-dimensional vector features, and fuses them with the original table features to generate enhanced feature sets that can drive formula optimization.

[0073] It can be understood that when processing table data combined with LLM, for features and labels in structured data, LLM can be used to mine the potential relationship between features and labels to supplement feature information. In order to avoid LLM outputting incorrect content, a specific prompt and risk limiting strategy are designed to introduce an agent-like thinking process, and logical rules are constructed at the knowledge fusion layer to filter and constrain the LLM generated content. The token or text content output by LLM is converted into a vector by the knowledge fusion layer, and the vector representation corresponding to each row of the table is distinguished to generate new features. At the same time, a risk assessment and screening mechanism is designed to ensure the quality of new features, and the refined vectors are integrated into the formula prediction model as new features, combined with multi-objective optimization methods such as Bayesian optimization (BO) to improve the accuracy and reliability of the recommended formula results.

[0074] Step S40, multi-objective optimization of the new formula scheme based on a preset experimental cost constraint condition, output of an optimized recommended formula and a supporting evidence chain.

[0075] It should be noted that the experimental cost constraint condition is a pre-set resource limitation parameter, such as a budget upper limit, a time limit, device availability, etc., which is used to filter feasible formula schemes. Multi-objective optimization of the new formula scheme can consider multiple conflicting objectives at the same time, such as performance maximization, cost minimization, and risk controllability, and find the optimal compromise solution through a mathematical model. The supporting evidence chain can be a logical basis presented in a tree structure, including parameter decision nodes, literature evidence nodes, and risk prompt nodes.

[0076] It should be understood that multi-objective optimization of the new formula scheme can also adjust the parameters and weights of the cost function for different industries (such as materials science, pharmaceuticals, food processing) to ensure that the evaluation accurately reflects the actual experimental cost. In addition to cost and performance, environmental risk, safety risk, and regulatory compliance should also be considered to avoid missing important risk points.

[0077] In an example, reference is made to Figure 2 , Figure 2To adapt to the task of the present application, the prompt process and the multi-objective optimization interactive flowchart are combined. After the user inputs the formula optimization target, the system decomposes the formula optimization target into two paths through the target analysis module. The performance priority path generates targeted performance optimization prompts, such as increasing the wear resistance of the coating by ≥20%; the cost priority path generates experimental cost constraint prompts, such as the cost of consumables per experiment ≤500; the performance and cost prompts are combined to form a dynamic instruction set, such as maximizing the viscosity within a budget ≤X, and inputting a large language model (LLM). The LLM is based on an internal formula knowledge base, matches industry best practices and literature rules, and outputs candidate formula schemes that balance performance gains and controllable costs; the candidate schemes enter the experimental feasibility analysis module for double verification, and the experimental consumable library (such as the ultrasonic equipment cost table) is used to judge the operability of the schemes. If the verification is passed, the final recommended formula is output, and the explanation module generates an evidence chain, automatically associating supporting literature original text and deducing key parameter adjustment logic; if it is not passed, the risk diagnosis mechanism is triggered to locate the specific conflict point, such as the high-speed stirring equipment exceeding the budget, and generate improvement suggestions, such as using a shear stirring process, and start iterative optimization, automatically adjusting parameters (such as reducing the proportion of ethanol) for high-risk schemes and re-verifying. At the same time, all recommended schemes are accompanied by traceable structured evidence chains, supporting literature citation (automatically associating patent / paper fragments and original text jump paths) and parameter deduction (showing the scientific basis for component replacement and process adjustment through tree-shaped decision nodes).

[0078] In the present embodiment, by automatically fusing innovative components and process information in the literature, combining cost and performance dual target optimization, the design boundary of new formula is expanded, while ensuring controllable consumption of experimental resources, the recommended formula output has scientific basis and explainability, effectively improving the practicality and innovation of formula optimization.

[0079] Reference Figure 3 , Figure 3 The flowchart of the second embodiment of the formula optimization method of the present application is shown in the figure. Based on the first embodiment described above, the second embodiment of the formula optimization method of the present application is proposed.

[0080] In the second embodiment, the step S30 comprises:

[0081] Step S301, constructing an experimental operation complexity cost function according to the formula history and external knowledge.

[0082] It should be noted that the formula history is the recorded data of past formula experiments, including components, proportions, process parameters, and corresponding costs, performance, etc. External knowledge comes from technical literature, industry standards, expert experience, etc. Domain knowledge is used to supplement and verify the deficiencies of formula history data. Experimental operation complexity is an index to measure the difficulty of formula experiment, involving the number of components, process step complexity, equipment requirements, etc.

[0083] It should be understood that the experimental operation complexity can be the number of components of the formula, for example, the complexity increases by 10% for each additional component; it can also be the number of process steps, and it can also be the difficulty of obtaining raw materials, such as the use of rare metals. The experimental operation complexity cost function can be a linear weighted model, or a model obtained by analytic hierarchy process, such as constructing a relative importance matrix between factors by expert judgment, or a machine learning model, such as training a regression model using historical data to predict the cost.

[0084] Step S302, filtering out high-cost formulas in the new formula scheme according to the experimental operation complexity cost function, to obtain initial optimized formulas.

[0085] It should be noted that the high-cost formula can be a formula scheme whose experimental cost exceeds a preset threshold or is ranked at the top in terms of cost after calculation by the cost function. The initial optimized formula is a low-cost or medium-cost formula set after removing the high-cost formula, which serves as the basis for subsequent multi-objective optimization.

[0086] It can be understood that filtering out high-cost formulas in the new formula scheme according to the experimental operation complexity cost function can be sorting the new formula scheme by cost function value from high to low and directly removing the top pre-set number of high-cost formulas; it can also be filtering out inferior formulas with high cost but substandard performance according to the cost-performance Pareto frontier; it can also be sorting by component scarcity cost.

[0087] It should be understood that when filtering out high-cost formulas in the new formula scheme, it cannot rely solely on the single indicator of the cost function, and it needs to be cross-filtered in combination with process feasibility (such as whether the equipment is available) and performance expectation (such as whether the strength meets the standard) to prevent the accidental deletion of potential high-quality formulas. The specific composition of the cost function needs to be displayed during the screening process to facilitate the tracing of the root cause of high cost. The low-cost formulas screened out need to be verified by small-batch experiments to ensure the accuracy of the cost estimation, and to avoid a large deviation between the model prediction and the actual cost. For high-cost but innovative formulas, a whitelist needs to be established for manual intervention and evaluation before deciding whether to remove them.

[0088] Step S303, determining the risk level of the initial optimized formula based on the experimental cost constraint condition, and taking the low-risk initial optimized formula as the recommended formula of the formula optimization target.

[0089] It should be noted that the risk level is a quantitative classification of the potential risks of the formula scheme according to experimental feasibility, performance volatility, and raw material supply, and the risk rating results are also fed back to the knowledge fusion layer for updating the LLM generation prompt in the next round, forming a human-machine collaborative closed loop.

[0090] It can be understood that determining the risk level of the initial optimization formula based on the experimental cost constraint condition can train a random forest model with historical formula cost-risk data to automatically predict the risk level of a new formula; it can also be to divide the formula into risk clusters using clustering algorithm, to determine the level according to the distance of cluster center, and it can also be to construct a risk knowledge graph and infer the level through the association relationship of component-process-risk.

[0091] Further, in order to generate a risk prompt and trigger a parameter adjustment instruction for the initial optimization formula with a high risk level when the experimental complexity and performance improvement meet the optimization threshold, so as to modify the component proportion and process time and other parameters, thereby making up for the defects of the traditional method in the processing capacity of high-risk schemes, and improving the feasibility and safety of the formula through dynamic adjustment. After step S303, it further includes:

[0092] Obtaining a target formula of high risk level in the initial optimization formula; when the experimental complexity score and performance improvement score of the target formula meet the preset optimization threshold, generating a risk prompt corresponding to the new formula scheme, and triggering a process parameter adjustment instruction; modifying the component proportion and process time parameters of the new formula scheme according to the process parameter adjustment instruction and the risk prompt, and generating an adjusted formula scheme.

[0093] It should be noted that the high-risk level target formula is a specific formula scheme in the initial optimization formula that is determined to be high risk (such as high experimental failure probability and large performance fluctuation) through risk evaluation. The experimental complexity score is a numerical index for quantitatively evaluating the experimental difficulty of the formula. The performance improvement score is a quantitative value for measuring the improvement amplitude of the formula compared with the benchmark scheme in the performance index (such as strength and durability). The preset optimization threshold is a pre-set judgment standard for judging whether the high-risk formula has optimization value. The process parameter adjustment instruction is a parameter modification scheme automatically generated by the system, such as reducing the proportion of component C from 10% to 5% and extending the sintering time by 30 minutes. The risk prompt is the early warning information generated for the potential problems of the high-risk formula, such as the compatibility risk between components A and B.

[0094] It can be understood that generating the risk prompt corresponding to the new formula scheme and triggering the process parameter adjustment instruction can be based on a preset rule template; it can also be dynamically generated through a large language model; it can also be generated through associated historical cases. Modifying the parameters to generate the adjusted scheme includes component proportion adjustment, process time optimization, and equipment parameter adjustment, etc.

[0095] It should be understood that before obtaining the high-risk formula, the integrity of the initial optimized formula can be checked first, and invalid formulas with missing ingredient data or contradictory process parameters can be removed. After generating the adjusted formula scheme, the formula parameters before and after adjustment, risk level changes, performance prediction comparisons can be automatically visualized for user review.

[0096] Step S304, a support evidence chain corresponding to the recommended formula is generated, and the support evidence chain is a tree-shaped explanatory structure, including parameter decision nodes, literature evidence nodes and risk prompt nodes.

[0097] It should be noted that the support evidence chain is a formula recommendation logic basis presented in a tree structure, which forms a traceable scientific argumentation system through multi-level node association of parameter selection, literature support and risk warning. The parameter decision node records the selection basis of the parameters such as ingredients, proportions, and processes in the formula, including data-driven decision or knowledge-driven decision; the literature evidence node references external knowledge sources such as technical literature, patents, and standards to provide scientific support for parameter decision; and the risk prompt node marks the nodes of potential risks of the formula in experiments, production, performance, etc.

[0098] In an example, referring to Figure 4 , Figure 4 The self-explanatory structure and evidence chain node distribution diagram are output for the application. By constructing a self-explanatory structure of the recommended formula, the output result is presented in a tree form, covering recommended reasons, supporting evidence, formula parameters, risk assessment, and optimization target alignment, etc. The recommended reasons include performance improvement such as viscosity, cost reduction, etc. dimensions; supporting evidence such as ethanol solubility verification, etc.; formula parameters can include ultrasonic process parameters and risks; risk assessment for compatibility risk with or without historical data annotation, optimization target alignment including performance indicators, cost constraints, etc. Each module is associated with each other to form a clear evidence chain node distribution, making the basis, parameters and risks of the recommended formula traceable and easy to understand. The tree structure displays the complete logical chain of the recommended result, including three branches of parameters, reasons, evidence, and risks. The evidence node is directly associated with literature fragments and patent numbers, supporting click jumping to the original text. The risk prompt can also be marked by red / yellow labels for manual review.

[0099] Further, in order to enable the system to flexibly adjust the scheme according to actual needs, while ensuring the rationality of the adjustment through real-time verification. After the step S304, it further includes:

[0100] receive a parameter adjustment instruction of the user on the support evidence chain; call the experimental operation complexity cost function according to the parameter adjustment instruction to perform dynamic verification, and obtain a verification result; when the verification result is passed, update a recommended formula corresponding to the support evidence chain, and add a user decision branch in the support evidence chain, the user decision branch including a user instruction node, a real-time verification node and an evidence update marker.

[0101] It should be noted that the parameter adjustment instruction is a modification instruction proposed by the user for the parameter decision node and the process parameter in the support evidence chain, including component proportion change, process time adjustment and the like. The user decision branch is a newly added branch in the tree structure of the support evidence chain, recording user intervention operation and verification result, including user instruction, verification data and evidence update state. The user instruction node stores the specific parameter adjustment content of the user. The real-time verification node presents dynamic verification process data, and the feasibility of the adjusted scheme can be traced. The evidence update marker is used to identify the node in the support evidence chain that is changed due to user adjustment, and to distinguish the original and manual intervention.

[0102] It can be understood that receiving the parameter adjustment instruction of the user on the support evidence chain can be parsing the text instruction input by the user in the visual interface (such as a formula parameter editing box); can also be capturing the interactive operation of the user on the evidence chain node (such as dragging the slider to modify the parameter after clicking the sintering temperature node); or can be receiving a standardized instruction file to batch parse the adjustment content.

[0103] It can be understood that adding the user decision branch in the support evidence chain can be directly mounting the user instruction node (recording the adjustment content) and the real-time verification node under the parameter node corresponding to the tree structure of the support evidence chain; can also be constructing a three-level sub-branch according to the logic of user instruction, dynamic verification and evidence update, and clearly presenting the intervention process; or can be associated with version control, marking the formula version numbers before and after the adjustment in the branch, and distinguishing the iteration state of the evidence chain.

[0104] In the embodiment, the experimental operation complexity cost function is constructed according to the formula history and external knowledge; the high-cost formula in the new formula scheme is screened out according to the experimental operation complexity cost function, and the initial optimization formula is obtained; the risk level of the initial optimization formula is determined based on the experimental cost constraint condition, and the initial optimization formula with low risk is taken as the recommended formula of the formula optimization target; the support evidence chain corresponding to the recommended formula is generated, and the support evidence chain is a tree-shaped explanation structure, including a parameter decision node, a literature evidence node and a risk prompt node. When the multi-objective optimization is carried out based on the experimental cost constraint condition, the high-cost formula is automatically screened and the risk level is determined, and the tree-shaped support evidence chain containing parameters, literature evidence and risk prompt is generated, so that the quantitative management of formula cost and risk is realized, and not only the high-cost unfeasible scheme can be filtered out, but also the complete logical chain and traceable literature basis can be provided for the recommended formula.

[0105] Reference Figure 5 , Figure 5 The flowchart of the third embodiment of the formula optimization method of the application is shown in FIG. 3. Based on the above embodiments, the third embodiment of the formula optimization method of the application is proposed.

[0106] In the third embodiment, the step S10 comprises:

[0107] In the third embodiment, the step S10 comprises:

[0108] It can be understood that the structured formula data set corresponding to the formula optimization target can be connected to the enterprise formula database, and the matched formula table is screened according to the optimization target; or the industry data platform API can be called, and the standardized formula data set is retrieved according to the target keyword; or the historical experimental record file is parsed to extract the formula parameters and result data related to the optimization target, and the structured data set is regularized.

[0109] In an example, referring to Table 1, Table 1 is a data sample table of the structured formula data set of the embodiment. When the structured formula data set corresponding to the formula optimization target and the unstructured technical literature data are obtained, the model automatically combines the English patent or journal translation and summarizes the key innovation points. The formula suggestion containing the ethanol-new stirring scheme is output, the required experimental complexity or consumable budget is automatically evaluated, and the note “no adverse cases in the industry for this ratio scheme, low batch test is recommended” can be added.

[0110] Table 1 Data sample table of structured formula data set

[0111] id Formulation processing Product performance 001 PAA dissolution - water bath stirring for 30 min Viscosity 122 mPa-s 002 PAA + NaCl - ultrasonic for 10 min Viscosity 146 mPa-s 003 New process: ethanol as solvent, shear -

[0112] Step S102, converting the structured formula data set into a vector through a semantic encoder to obtain a first relationship vector, the first relationship vector containing the corresponding relationship between the formula, process and performance.

[0113] It should be noted that the first relationship vector: after conversion by the semantic encoder, the corresponding relationship between the formula, process and performance is presented in the form of a vector, and the vector dimension reflects the associated features. The semantic encoder can be a pre-trained BERT model or a traditional machine learning encoder.

[0114] It should be understood that converting the structured formula data set into a vector through the semantic encoder can be to first sort the triple relationship of formula, process and performance, and then use the encoder to convert each triple into a vector to retain the association of the three; or the structured data can be taken as a sample by row, each row containing formula, process and performance fields, and the entire row sample can be encoded by the encoder to make the generated vector fuse the corresponding relationship of the three types of information; or the formula, process and performance can be encoded respectively, and then a composite relationship vector containing the association of the three can be constructed through vector splicing operation.

[0115] Step S103, analyzing the unstructured technical literature data set through a large language model to extract component replacement description and process improvement sentences, and converting them into structured parameters to obtain a second relationship vector.

[0116] It should be noted that the replacement description is a description in the literature or patent about replacing the original component with a new component, such as replacing the traditional solvent methanol with ethanol to improve safety. The process improvement sentence is a text describing the optimization method of the process, such as shortening the stirring time to 10 minutes by ultrasonic treatment. The second relationship vector is a vector representation obtained by encoding the structured parameters, used to associate the semantic relationship between component replacement, process improvement and performance influence.

[0117] It can be understood that when the unstructured technical literature data set is analyzed through the large language model, the API interface of the LLM is first called, the literature fragment is input and the output format is specified, the key sentences are automatically parsed, the LLM generated prompt is generated, and the text description output by the LLM is mapped to the preset field, and then converted into a vector through a word vector model. After the whole process is completed, the consistency of the LLM output and the expert annotation can be compared through manual sampling of the literature analysis results to evaluate the accuracy of the system.

[0118] In an example, referring to Table 2, Table 2 is a data sample table of the unstructured formula data set of the present embodiment. It is disclosed through existing literature that: "ethanol can replace traditional solvents to reduce the viscosity of PAA by 18±2% (implementation conditions: 40% ethanol concentration + 200 rpm stirring). Note: the pH needs to be maintained at 6-8 to avoid phase separation". The second relationship vector obtained by analyzing the above unstructured technical literature through the large language model can be as follows:

[0119] Table 2 Data sample table of unstructured formulation dataset

[0120]

[0121] Step S104, generating a table literature knowledge set according to the first relationship vector and the second relationship vector.

[0122] It can be understood that generating a table literature knowledge set according to the first relationship vector and the second relationship vector can be splicing the first relationship vector and the second relationship vector by dimension to generate a composite vector and then mapping into a table row; or can be weighting and fusing the same type of feature vectors to convert into a parameter column in the table; or can be filtering key dimensions of the vectors through an attention mechanism to generate an optimization basis column in the table.

[0123] In an example, referring to Figure 6 , Figure 6 is a schematic diagram of LLM-driven table-text knowledge flow and formulation generation of the present application. In the left branch, the structured table data is converted into a vector by a semantic encoder, retaining the formulation-process-performance relationship. First, the formulation history table is constructed to sort out the information of ingredients, process, performance, etc. of past formulations, and then the table semantic encoder is used to convert it into a computable vector or coded form containing semantic association, to mine the potential relationship among formulation-process-performance. For unstructured literature or patents, in the right branch, relying on LLM knowledge extraction technology, key knowledge such as ingredient replacement and process improvement is identified and extracted from the text, and a natural language understanding module is used to analyze the semantics and regularize the information, and convert it into structured parameters. In the fusion layer, the knowledge fusion layer internally includes a planning rule engine and a reliability evaluation module. The former performs one-time logical constraint on LLM based on explicit prompt and domain rule (such as ingredient compatibility matrix), and the latter performs second screening with a score threshold (trust degree ≥ τ) to ensure that only high-quality tokens are vectorized. Combine the two to generate candidate formulations, and filter the final results through cost-performance optimization. Subsequently, the new formulation generator generates a new formulation based on the fused knowledge, combines the formulation optimization target, and constructs an initial new formulation scheme through an algorithm (such as a generative model, rule deduction, etc.). Then, the multi-objective optimization section screens, adjusts and optimizes the initial formulation according to cost, performance, risk and other multi-dimensional constraints, balances the conflicts between targets, for example, for the "performance > 150 mPa·s and experimental material budget does not increase" target, the new scheme of "PAA: NaCl: ethanol ratio x:1, ultrasonic 12 min" is automatically recommended, and the cost change is listed. Finally, the output section outputs the recommended formulation and explanation content.

[0124] In the embodiment, a structured recipe dataset and unstructured technical literature data corresponding to a user-inputted recipe optimization target are obtained; the structured recipe dataset is converted into a vector by a semantic encoder to obtain a first relationship vector containing the corresponding relationship among recipes, processes and performances; the unstructured technical literature dataset is parsed by a large language model to extract ingredient replacement descriptions and process improvement sentences and convert them into structured parameters to obtain a second relationship vector; and a table literature knowledge set is generated according to the first relationship vector and the second relationship vector. By using the vector containing the recipe-process-performance relationship and the innovative points extracted from the unstructured literature and converted into structured parameters, the system can simultaneously absorb the regularity of the table data and the innovative experience in the literature to form a more comprehensive knowledge system, provide rich knowledge support for the large language model to generate high-quality new recipes, and further improve the innovativeness and accuracy of the recipe design.

[0125] It should be noted that the above examples are only used for understanding the present application and do not constitute a limitation on the recipe optimization method of the present application. More forms of simple transformation based on the technical concept are within the protection scope of the present application.

[0126] The present application also provides a recipe optimization device, which refers to Figure 7 The recipe optimization device comprises:

[0127] A data acquisition module 10 is configured to perform knowledge extraction on the acquired structured recipe dataset and unstructured technical literature data in response to a user-inputted recipe optimization target, and generate a table literature knowledge set.

[0128] A text generation module 20 is configured to input the table literature knowledge set and the recipe optimization target into a large language model to obtain generated text corresponding to the recipe optimization target.

[0129] A recipe generation module 30 is configured to perform logical rule screening and risk assessment on the generated text by a knowledge fusion layer to obtain text output meeting a confidence threshold, and generate a new recipe scheme according to the text output.

[0130] A recipe optimization module 40 is configured to perform multi-objective optimization on the new recipe scheme based on a pre-set experimental cost constraint condition, and output an optimized recommended recipe and a supporting evidence chain.

[0131] The formula optimization device provided in the present application adopts the formula optimization method in the above embodiment, and can solve the technical problem that the existing machine learning method cannot absorb industry or general knowledge for innovation when automatically designing new formula and new process. Compared with the prior art, the formula optimization device provided in the present application has the same beneficial effects as the formula optimization method provided in the above embodiment, and other technical features in the formula optimization device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0132] The present application provides a formula optimization device, which comprises at least one processor and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the formula optimization method in the above embodiment one.

[0133] Reference will be made to the following description of the embodiments of the present application, taken in conjunction with the accompanying drawings, in which Figure 8 which shows a structural diagram of a formula optimization device suitable for implementing the embodiments of the present application. The formula optimization device in the embodiments of the present application can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 8 The formula optimization device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the present application.

[0134] As Figure 8As shown, the recipe optimization device can include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to programs stored in a read-only memory 1002 or loaded from a storage device 1003 into a random access memory 1004. Various programs and data required for the recipe optimization device to operate are also stored in the random access memory 1004. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other by a bus 1005. An input / output interface 1006 is also connected to the bus. Generally, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; the storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the recipe optimization device to communicate with other devices wirelessly or by wire to exchange data. Although the recipe optimization device with various systems is shown in the figure, it should be understood that all the shown systems are not required to be implemented or possessed. More or fewer systems can be alternatively implemented or possessed.

[0135] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program codes for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by a communication device, or installed from the storage device 1003, or installed from the read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the methods of the embodiments disclosed in the present application are performed.

[0136] The recipe optimization device provided in the present application adopts the recipe optimization method in the above-mentioned embodiments, and can solve the technical problem that the existing machine learning method cannot absorb industry or general knowledge for innovation in the automatic design of new recipes and new processes. Compared with the prior art, the recipe optimization device provided in the present application has the same beneficial effects as the recipe optimization method provided in the above-mentioned embodiments, and other technical features in the recipe optimization device are the same as the features disclosed in the previous embodiment method, which will not be described here.

[0137] It should be understood that portions of the application disclosed can be implemented in hardware, software, firmware, or combinations thereof. In the description of the embodiments above, specific features, structures, materials or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0138] The above description is merely illustrative of the application and is not intended to limit the scope of the application. Any changes and modifications that can be made to the application in light of the teachings described herein are contemplated in the broad scope of the application. Accordingly, the scope of the application should be determined not with reference to the above description but with reference to the claims appended hereto.

[0139] The application provides a computer readable storage medium having stored thereon computer readable program instructions (i.e., a computer program) for performing the recipe optimization method in the above embodiments.

[0140] The computer readable storage medium provided by the application may, for example, be a U disk, but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more conductive wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer readable storage medium can be transmitted in any suitable medium, including but not limited to electrical wire, optical cable, RF (Radio Frequency), etc., or any suitable combination of the above.

[0141] The above computer readable storage medium can be included in the recipe optimization device; or can exist separately and not be assembled into the recipe optimization device.

[0142] The above computer readable storage medium carries one or more programs, which, when executed by the recipe optimization device, cause the recipe optimization device to perform the recipe optimization method described above.

[0143] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0144] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0145] The modules involved in the embodiments of the present application can be implemented in the form of software or in the form of hardware. In some cases, the name of the module does not constitute a limitation on the module itself.

[0146] The readable storage medium provided by the present application is a computer readable storage medium, which stores computer readable program instructions (i.e., computer programs) for executing the above-mentioned formula optimization method, and can solve the technical problem that the existing machine learning method cannot absorb industry or general knowledge for innovation when automatically designing new formulas and new processes. Compared with the prior art, the computer readable storage medium provided by the present application has the same beneficial effects as the formula optimization method provided by the above-mentioned embodiments, and will not be described here.

[0147] The above merely describes some embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation, direct / indirect application in other related technical fields, or the like, which is made based on the technical concept of the present application and the content of the specification and drawings, is included in the patent protection scope of the present application.

Claims

1. A formulation optimization method, characterized by, The formula optimization method comprises: In response to a user input formula optimization target, knowledge extraction is performed on an acquired structured formula data set and unstructured technical literature data to generate a table literature knowledge set; The table literature knowledge set and the formula optimization target are input into a large language model to obtain generated text corresponding to the formula optimization target; Logical rule screening and risk assessment are performed on the generated text through a knowledge fusion layer to obtain text output meeting a confidence threshold, and a new formula scheme is generated according to the text output; The new formula scheme is subjected to multi-objective optimization based on a preset experimental cost constraint condition, and an optimized recommended formula and a supporting evidence chain are output.

2. The formulation optimization method of claim 1, wherein, The step of inputting the table literature knowledge set and the formula optimization target into the large language model to obtain generated text corresponding to the formula optimization target comprises: The formula optimization target is subjected to target analysis to obtain performance optimization prompts and cost optimization prompts; A dynamic prompt template is constructed according to the performance optimization prompts and the cost optimization prompts; The dynamic prompt template is input into the large language model, and a task-type prompt string is generated through existing modes in the table literature knowledge set and industry best practices, and the task-type prompt string is analyzed to output predicted generated text.

3. The formulation optimization method of claim 1, wherein, The step of subjecting the new formula scheme to multi-objective optimization based on a preset experimental cost constraint condition to output an optimized recommended formula and a supporting evidence chain comprises: An experimental operation complexity cost function is constructed according to formula history and external knowledge; High-cost formulas in the new formula scheme are screened out according to the experimental operation complexity cost function to obtain initial optimization formulas; The risk levels of the initial optimization formulas are determined based on the experimental cost constraint condition, and an initial optimization formula with a low risk level is taken as a recommended formula of the formula optimization target; A supporting evidence chain corresponding to the recommended formula is generated, the supporting evidence chain is a tree-shaped explanation structure, and comprises parameter decision nodes, literature evidence nodes and risk prompt nodes.

4. The formulation optimization method of claim 3, wherein, After the step of determining the risk levels of the initial optimization formulas based on the experimental cost constraint condition and taking an initial optimization formula with a low risk level as a recommended formula of the formula optimization target, the method further comprises: An objective formula with a high risk level in the initial optimization formulas is acquired; When an experimental complexity score and a performance improvement score of the objective formula meet a preset optimization threshold, a risk prompt corresponding to the new formula scheme is generated, and a process parameter adjustment instruction is triggered; Composition proportion and process time parameters of the new formula scheme are modified according to the process parameter adjustment instruction and the risk prompt to generate an adjusted formula scheme.

5. The formulation optimization method of claim 3, wherein, After the step of generating a supporting evidence chain corresponding to the recommended formula, the method further comprises: A parameter adjustment instruction of the supporting evidence chain is received from a user; A dynamic verification is performed on the experimental operation complexity cost function according to the parameter adjustment instruction to obtain a verification result; When the verification result is passed, the corresponding recommended formula of the support evidence chain is updated, and a user decision branch is added in the support evidence chain, the user decision branch comprising a user instruction node, a real-time verification node and an evidence update mark.

6. The formulation optimization method of claim 1, wherein, The step of performing knowledge extraction on the obtained structured formula data set and unstructured technical literature data in response to the user input formula optimization target to generate the table literature knowledge set comprises: In response to the user input formula optimization target, a structured formula data set and unstructured technical literature data corresponding to the formula optimization target are obtained; The structured formula data set is converted into a vector through a semantic encoder to obtain a first relationship vector, the first relationship vector comprising corresponding relationships among formulas, processes and performances; The unstructured technical literature data set is analyzed through a large language model to extract component replacement descriptions and process improvement sentences and is converted into structured parameters to obtain a second relationship vector, the second relationship vector being filtered through a risk scoring module and the second relationship vector and the first relationship vector being information complementary rather than redundant; The table literature knowledge set is generated according to the first relationship vector and the second relationship vector.

7. The formulation optimization method of any one of claims 1 to 6, wherein, The knowledge fusion layer comprises a rule engine, a risk scoring module and a vectorization module; the rule engine checks generated text of the large language model based on a domain expert rule and a prompt word template; the risk scoring module calculates a credibility score of the checked generated text; and the vectorization module maps generated text with a credibility score greater than a preset credibility threshold to a vector feature and fuses the vector feature with original table features.

8. A formulation optimization apparatus, characterized by, The device comprises: A data acquisition module configured to perform knowledge extraction on an obtained structured formula data set and unstructured technical literature data in response to a user input formula optimization target to generate a table literature knowledge set; A text generation module configured to input the table literature knowledge set and the formula optimization target into a large language model to obtain generated text corresponding to the formula optimization target; A formula generation module configured to perform logical rule screening and risk assessment on the generated text through a knowledge fusion layer to obtain text output meeting a confidence threshold and generate a new formula scheme according to the text output; A formula optimization module configured to perform multi-objective optimization on the new formula scheme based on a preset experimental cost constraint condition to output an optimized recommended formula and a support evidence chain.

9. A recipe optimization device, characterized by, The device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement steps of the formula optimization method according to any one of claims 1 to 7.

10. A storage medium, characterized by The storage medium is a computer readable storage medium, and the storage medium stores a computer program, the computer program being executable by a processor to implement steps of the formula optimization method according to any one of claims 1 to 7.

Images