Formula optimization method and device, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a formula optimization method and device, equipment and a storage medium, and the method comprises the steps: carrying out the semantic deconstruction of a target demand in a multi-modal natural language form through a large language model, converting an obtained formula parameter into a differentiable semantic physical quantity, and carrying out the semantic deconstruction of the target demand; according to the method, knowledge rules are extracted from semantic physical quantities through a cognitive distillation algorithm, the semantic physical quantities are verified, the verified semantic physical quantities are optimized through a minimum weight loss function, and then a target formula is obtained through semantic decoding and physical field conversion. According to the method, the formula parameters are accurately extracted through semantic deconstruction, the problem that in the prior art, complex requirements are inaccurately understood is solved, the formula parameters are converted into differentiable semantic physical quantities, mathematical operability is provided for the optimization process, knowledge rules are extracted through a cognitive distillation algorithm for verification, it is ensured that data are reasonable and feasible, and user experience is improved. The defects of lack of an effective verification mechanism and easy optimization deviation in the prior art are overcome.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and storage medium for formula optimization. Background Technology

[0002] In formulation-intensive industries such as chemicals, pharmaceuticals, and food, formulation optimization is crucial. Traditional methods rely on expert experience and trial-and-error, which have limitations such as difficulty in quantification, high costs, and long processing times. While existing formulation optimization methods have improved efficiency, they are insufficient in terms of accuracy, making it difficult to meet the industry's demand for rapid and precise formulation optimization. Summary of the Invention

[0003] The main objective of this application is to provide a method, apparatus, device, and storage medium for formula optimization, aiming to solve the technical problem that the optimization process of existing formula generation methods is not accurate enough.

[0004] To achieve the above objectives, this application proposes a formulation optimization method, the method comprising:

[0005] In response to the target requirements in the form of multimodal natural language, the target requirements are semantically deconstructed through a large language model to obtain the recipe parameters;

[0006] The formula parameters are converted into differentiable semantic physical quantities;

[0007] Knowledge rules are extracted from the semantic physical quantity using a cognitive distillation algorithm, and the semantic physical quantity is verified based on the knowledge rules.

[0008] After successful verification, the verified semantic physical quantities are optimized by minimizing the weighted loss function to obtain intermediate recipe representations.

[0009] The intermediate recipe representation is semantically decoded and physically transformed to obtain the complete target recipe.

[0010] In one embodiment, before the step of extracting knowledge rules from the semantic physical quantity using a cognitive distillation algorithm and verifying the semantic physical quantity based on the knowledge rules, the method further includes:

[0011] Extract domain terminology from the target requirements;

[0012] Establish a terminology parameter mapping table based on the domain terminology and the preset materialization parameter library;

[0013] The terminology parameter mapping table and the semantic physical quantities are verified.

[0014] When verification fails, the failure node in the semantic physical quantity is traced in reverse, and the semantic physical quantity is corrected based on the failure node to obtain the corrected semantic physical quantity.

[0015] Accordingly, the step of extracting knowledge rules from the semantic physical quantity using a cognitive distillation algorithm and verifying the semantic physical quantity based on the knowledge rules includes:

[0016] Knowledge rules are extracted from the semantic physical quantity or the modified semantic physical quantity using a cognitive distillation algorithm, and the semantic physical quantity is verified based on the knowledge rules.

[0017] In one embodiment, the step of reverse tracing the failure node in the semantic physical quantity when verification fails, and correcting the semantic physical quantity based on the failure node to obtain the corrected semantic physical quantity includes:

[0018] When verification fails, counterfactual analysis is used to locate the physical quantity corresponding to the performance deviation index.

[0019] By tracing the generation path of the physical quantity in reverse, conflicting failure nodes are identified;

[0020] Adjust the weight of the failed node or replace the associated data source of the failed node to obtain the corrected semantic physical quantity.

[0021] In one embodiment, the step of obtaining recipe parameters by semantically deconstructing the target requirement using a large language model in response to the target requirement in a multimodal natural language form includes:

[0022] In response to the target requirements of multimodal natural language forms, the term parameter mapping table is matched with the preset text database and preset image database for feature matching, and the recipe parameters with confidence are output.

[0023] If there are contradictions in the data from different modalities during feature matching, the parameter priority of the formula parameters is adjusted through counterfactual analysis to output the final formula parameters.

[0024] In one embodiment, the step of semantically decoding and physically transforming the intermediate recipe representation to obtain the complete target recipe includes:

[0025] The intermediate recipe representation is associated with a preset process database to generate a cost tag set;

[0026] The zero-shot generator is used to compare the matching degree between the text description in the intermediate recipe representation and the image features.

[0027] If the matching degree is lower than a preset threshold, a molecular structure diagram corresponding to the text description is generated, and the molecular structure diagram is mapped to process parameters through physical field transformation.

[0028] The final target formulation is generated based on the cost tag set and the process parameters.

[0029] In one embodiment, the step of converting the formula parameters into differentiable semantic physical quantities includes:

[0030] By matching similarity, the formula parameters are mapped to a preset physical quantity range to obtain the matched physical quantities;

[0031] The matched physical quantities are converted into continuously differentiable numerical representations using a neural network, thereby obtaining differentiable semantic physical quantities.

[0032] In one embodiment, the step of performing multi-objective optimization on the verified semantic physical quantity by minimizing the weighted loss function after successful verification to obtain an intermediate recipe representation includes:

[0033] After verification, a weighted loss function containing performance and cost items is constructed, wherein the performance items are determined by the performance indicators in the target requirements, and the cost items are determined by the raw material price database and the process energy consumption model.

[0034] Based on the adaptive optimization algorithm, when the performance verification of the verified semantic physical quantity fails to meet the performance target, the weight of the performance item is increased.

[0035] When the cost corresponding to the verified semantic physical quantity exceeds the cost threshold, the weight of the cost item is increased;

[0036] Based on the increased weights, the optimized physical quantities are obtained as intermediate formulation representations.

[0037] Furthermore, to achieve the above objectives, this application also proposes a formula optimization device, which includes:

[0038] The requirement parsing module is used to respond to target requirements in the form of multimodal natural language, and to perform semantic deconstruction of the target requirements through a large language model to obtain the recipe parameters;

[0039] The parameter quantization module is used to convert the formula parameters into differentiable semantic physical quantities.

[0040] The rule verification module is used to extract knowledge rules from the semantic physical quantity through a cognitive distillation algorithm, and to verify the semantic physical quantity based on the knowledge rules;

[0041] The recipe optimization module is used to perform multi-objective optimization on the validated semantic physical quantities by minimizing the weighted loss function after the validation is passed, so as to obtain an intermediate recipe representation.

[0042] The recipe generation module is used to perform semantic decoding and physical field transformation on the intermediate recipe representation to obtain the complete target recipe.

[0043] In addition, to achieve the above objectives, this application also proposes a formulation optimization device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the formulation optimization method as described above.

[0044] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the formula optimization method described above.

[0045] The technical solution proposed in this application responds to the target requirements of multimodal natural language. It semantically deconstructs the target requirements using a large language model to obtain recipe parameters, transforms these parameters into differentiable semantic physical quantities, extracts knowledge rules from these quantities using a cognitive distillation algorithm, and verifies the semantic physical quantities based on these knowledge rules. After successful verification, it performs multi-objective optimization on the verified semantic physical quantities by minimizing a weighted loss function to obtain an intermediate recipe representation. Finally, it performs semantic decoding and physical field transformation on the intermediate recipe representation to obtain the complete target recipe. This application solves the problem of inaccurate understanding of complex requirements in existing technologies by semantically deconstructing multimodal natural language target requirements using a large language model and accurately extracting recipe parameters. Furthermore, it transforms recipe parameters into differentiable semantic physical quantities, providing mathematical operability for the optimization process, making parameter adjustment more flexible and the optimization process easier to control precisely. The use of a cognitive distillation algorithm to extract and verify knowledge rules from the semantic physical quantities not only ensures the reasonableness and feasibility of the parameters but also overcomes the shortcomings of existing technologies, such as the lack of an effective verification mechanism and susceptibility to optimization bias. Attached Figure Description

[0046] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating the formulation optimization method of this application in Embodiment 1.

[0049] Figure 2 This is a flowchart illustrating Example 2 of the formulation optimization method of this application.

[0050] Figure 3 This is a flowchart illustrating the formulation optimization method of this application in Embodiment 3;

[0051] Figure 4 This is a schematic diagram of the module structure of the formula optimization device according to an embodiment of this application;

[0052] Figure 5 This is a schematic diagram of the hardware operating environment involved in the formula optimization method in this application embodiment.

[0053] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0055] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0056] Existing formulation optimization methods are insufficient in terms of accuracy, making it difficult to meet the industry's demand for rapid and precise formulation optimization.

[0057] To overcome the aforementioned shortcomings, this application provides a solution that uses a large language model to semantically deconstruct the target requirements of multimodal natural language, accurately extracts the formula parameters, and further transforms the formula parameters into differentiable semantic physical quantities, providing mathematical operability for the optimization process. The cognitive distillation algorithm is used to extract knowledge rules from the semantic physical quantities and verify them, which not only ensures that the parameters are reasonable and feasible, but also overcomes the shortcomings of existing technologies, such as the lack of an effective verification mechanism and the susceptibility to optimization bias.

[0058] It should be noted that the executing entity of each embodiment of this application can be a computing service system with data processing, network communication, and program execution functions, such as an electronic system or a formula optimization system capable of realizing the above functions. The following description uses a formula optimization system (hereinafter referred to as "the system") as an example to illustrate the following embodiments.

[0059] Based on this, embodiments of this application provide a formula optimization method, referring to... Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of the formulation optimization method of this application.

[0060] In this embodiment, the formula optimization method includes steps S10 to S50:

[0061] Step S10: In response to the target requirement in the form of multimodal natural language, the target requirement is semantically deconstructed through a large language model to obtain the recipe parameters.

[0062] It should be understood that the system first needs to receive and process multimodal target requirements, which can exist in the form of text descriptions, experimental images, or voice commands. The system employs the semantic understanding capabilities of a Large Language Model (LLM) to deeply analyze these inputs. Through word vector space mapping and attention mechanisms, it can transform natural language descriptions into precisely quantified formulation parameter requirements. For example, when the input contains a requirement such as "high-rate fast-charging negative electrode," the model can automatically associate it with "lithium-ion Li..." + The key physical quantity of diffusion coefficient is specified, and a specific threshold requirement is set that it must be greater than 1e. -10 cm 2 / s, while also identifying other implicit performance indicators and process constraints.

[0063] In one implementation, step S10 may include: responding to the target requirements of multimodal natural language form, performing feature matching between the terminology parameter mapping table and a preset text database and a preset image database, and outputting recipe parameters with confidence; during the feature matching process, if there are contradictions in the data of different modalities, adjusting the parameter priority of the recipe parameters through counterfactual analysis, and outputting the final recipe parameters.

[0064] In its implementation, during the formula optimization process, after responding to the target requirements in multimodal natural language and establishing a terminology parameter mapping table, the system performs feature matching operations between the constructed terminology parameter mapping table and a preset text database and a preset image database, respectively. The preset text database stores a large amount of text information related to the formula and is constructed based on domain terms in the target requirements and a preset materialization parameter library; the preset image database contains various image materials related to the formula. Through feature matching, the system can extract information that matches the terminology parameter mapping table from these two databases, and then output formula parameters with confidence levels. The confidence level reflects the reliability of the matching between the output formula parameters and the actual requirements.

[0065] However, due to the differences in the sources and representations of multimodal data (text, images, etc.), there is a probability that contradictory data from different modalities may occur during feature matching. Therefore, if contradictory data exist, counterfactual analysis is used to explore the causes of the contradictions, analyze the degree of influence of different parameters in the formulation, and adjust the parameter priority of the formulation parameters accordingly to obtain the adjusted formulation parameters.

[0066] Step S20: Convert the formula parameters into differentiable semantic physical quantities.

[0067] After obtaining the formulation parameters, a semantic physical field model is established to transform these parameters into differentiable semantic physical quantities. This means using mathematical formulas and physical models to map discrete formulation parameters into continuous, differentiable physical quantities. For example, the physical quantity of "binder flowability" can be represented by a formula... (“binder flowability”) is transformed into a differentiable representation.

[0068] Step S30: Extract knowledge rules from the semantic physical quantity using a cognitive distillation algorithm, and verify the semantic physical quantity based on the knowledge rules.

[0069] It should be noted that cognitive distillation is a technique that extracts key knowledge rules from complex semantic-physical quantities or pre-trained models (teacher models) by constructing a lightweight model (student model), and then uses these rules to validate or optimize the original data. The knowledge rules encompass valuable information such as the model's output probability distribution and intermediate layer feature representations. Then, through a pre-defined knowledge transfer strategy, the knowledge in the teacher model is effectively transferred to the student model. Finally, the knowledge rules mastered by the student model are used to validate or optimize the original semantic-physical quantities. Alternatively, the knowledge rules can also be implicit knowledge rules extracted from existing open-source databases.

[0070] Step S40: After successful verification, the verified semantic physical quantity is optimized by minimizing the weighted loss function to obtain an intermediate recipe representation.

[0071] It should be understood that after verifying the semantic physical quantity and confirming that it meets the preset standards, the optimization process enters the crucial multi-objective optimization stage. At this point, the system uses the verified semantic physical quantity as input and constructs a weighted loss function that includes multiple optimization objectives to collaboratively adjust the key parameters or features of the semantic physical quantity. This loss function is composed of a weighted combination of multiple sub-loss terms, each corresponding to an optimization objective (such as the rationality of the formulation components, the stability of physical properties, the balance of cost and benefit, etc.), and the weight coefficients reflect the relative importance of different objectives in the overall optimization.

[0072] The system can minimize the weighted loss function through optimization algorithms such as gradient descent, and can seek the optimal trade-off among multiple objectives, so that the semantic physical quantities gradually approach the ideal state that satisfies all constraints. Finally, the parameter combination or feature representation obtained after multiple rounds of iterative optimization is regarded as an intermediate recipe representation that can comprehensively balance various indicators.

[0073] Step S50: Semantic decoding and physical field transformation are performed on the intermediate recipe representation to obtain the complete target recipe.

[0074] It's important to note that semantic decoding aims to extract semantically meaningful information from the intermediate formula representation, such as formula components and characteristic descriptions, transforming it from an abstract representation into a language description understandable by humans or systems. Physical field transformation, on the other hand, further converts this semantic information into parameters related to actual physical processes. For example, it converts the description of formula components into specific production process parameters, ensuring that the target formula is not only semantically complete and accurate but also feasible and applicable in real-world physical scenarios. Through the synergistic effect of these two operations, the intermediate formula representation is transformed into a complete target formula with practical guiding significance.

[0075] In one implementation, step S50 may include: associating the intermediate recipe representation with a preset process database to generate a cost tag set; comparing the matching degree between the text description and image features in the intermediate recipe representation using a zero-sample generator; if the matching degree is lower than a preset threshold, generating a molecular structure diagram corresponding to the text description, and mapping the molecular structure diagram to process parameters through physical field transformation; and generating the final target recipe based on the cost tag set and the process parameters.

[0076] In practice, the preset process database stores a large amount of process-related data, including the cost of various raw materials and the cost of production processes. By associating the intermediate formula representation with this database, the corresponding cost information can be extracted from the database based on the ingredients and process requirements involved in the intermediate formula, thereby generating a cost tag set.

[0077] A zero-shot generator is a model that can process and analyze data without a large number of labeled samples. It determines the degree of consistency, or matching degree, between textual descriptions (such as textual descriptions of ingredients and properties) and image features (such as features extracted from related images including molecular structure diagrams and process flow diagrams) in intermediate formulation representations.

[0078] If the matching degree between the text description and image features is lower than a preset threshold, it indicates that the current text description is not accurate or complete enough to fully express the key information of the formula. In this case, the system will generate a molecular structure diagram corresponding to the text description to intuitively display the molecular structure and interrelationships of each component in the formula. Then, through physical field conversion technology, the generated molecular structure diagram is mapped to specific process parameters, that is, the information at the molecular level is transformed into process parameters that can be operated in actual production, such as temperature, pressure, and reaction time.

[0079] Finally, based on the cost tag set and process parameters generated above, the system comprehensively considers cost and production process requirements to generate the final target formula. This target formula is not only semantically complete and accurate, meeting the requirements for intermediate formula representation, but also optimized in terms of economy and operability.

[0080] For example, when optimizing the material formulation, upon receiving a request for "high-rate fast charging," it automatically associates it with "lithium-ion Li..." + Diffusion coefficient > 1e -10 cm 2 The physical quantity " / s" generates corresponding core-shell and porous structures. The core-shell structure maps to "TiO2@C", and the porous structure maps to "CNT / graphite".

[0081] This embodiment utilizes LLM to semantically deconstruct multimodal natural language target requirements, accurately extracting formula parameters. This solves the problem of inaccurate understanding of complex requirements in existing technologies. Furthermore, it transforms formula parameters into differentiable semantic physical quantities, providing mathematical operability for the optimization process. This makes parameter adjustment more flexible and the optimization process easier to control precisely. The cognitive distillation algorithm extracts and verifies knowledge rules from the semantic physical quantities, ensuring not only the reasonableness and feasibility of the parameters but also overcoming the shortcomings of existing technologies, such as the lack of effective verification mechanisms and susceptibility to optimization bias. Moreover, LLM is used to match the terminology parameter mapping table with features from a pre-set text and image database, outputting formula parameters with confidence levels. In case of contradictions, counterfactual analysis is used to adjust priorities, improving the comprehensiveness and accuracy of formula parameter acquisition and making formula optimization more aligned with actual needs. Furthermore, by linking intermediate formulation representations with process databases to generate cost tag sets, a zero-sample generator compares the matching degree between text descriptions and image features. If the matching degree is not met, a molecular structure diagram is generated and converted into process parameters. Combined with the cost tag set, the target formulation is generated, realizing the accurate conversion of formulations from abstract representations to specific process parameters. Taking cost factors into account, it ensures that the optimized formulation is feasible and economical in actual production, thereby improving the feasibility and economic benefits of the formulation.

[0082] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. On this basis, refer to Figure 2Steps S21 to S24 may be included before step S30:

[0083] Step S21: Extract domain terminology from the target requirements.

[0084] It should be understood that in formulation optimization scenarios, complex target requirements are mixed with a large number of common words and domain-specific technical terms. To accurately extract key information, natural language processing techniques, such as lexical analysis (identifying word composition and parts of speech), syntactic analysis (analyzing sentence structure and grammatical relationships), and semantic understanding algorithms (mining the deeper meaning of text), are needed to accurately extract domain terms closely related to the formulation from complex target requirements.

[0085] For example, in a pharmaceutical formulation optimization scenario, the target requirement is "to develop a sustained-release tablet for treating hypertension, requiring continuous drug release over 12 hours with a bioavailability of over 70%, while also exhibiting good stability and low side effects." Analyzing this requirement using natural language processing (NLP) reveals domain terms including "treating hypertension," "sustained-release tablet," "12-hour continuous release," "bioavailability of over 70%," "good stability," and "low side effects." "Treating hypertension" clarifies the applicable condition; "sustained-release tablet" specifies the dosage form; "12-hour continuous release" defines the drug's release characteristics; "bioavailability of over 70%" defines the required degree of drug absorption and utilization by the body; "good stability" sets a standard for maintaining the drug's quality during storage and use; and "low side effects" focuses on the drug's safety.

[0086] Step S22: Establish a terminology parameter mapping table based on the domain terminology and the preset materialization parameter library.

[0087] After parsing out the domain terminology, each term needs to be matched and associated with the information in the physical and chemical parameter library to establish a term-parameter mapping table, that is, to transform the abstract domain terminology into quantifiable and operable physical and chemical parameters.

[0088] Step S23: Verify the terminology parameter mapping table and the semantic physical quantity.

[0089] It should be understood that the verification process mainly includes two aspects: First, checking whether the correspondence between terms and parameters in the terminology parameter mapping table is accurate and reasonable, that is, judging whether the physicochemical parameters corresponding to each domain term meet the actual meaning and requirements of the term in the formulation; Second, verifying whether the semantic physical quantities can accurately reflect the characteristics and target requirements of the formulation. This can be done by substituting the semantic physical quantities into a preset model or algorithm for simulation analysis to verify whether it can produce results that meet the target requirements.

[0090] Step S24: When verification fails, reverse trace the failure node in the semantic physical quantity, and correct the semantic physical quantity based on the failure node to obtain the corrected semantic physical quantity.

[0091] When verification fails, it indicates a problem with the terminology parameter mapping table or semantic physical quantities, resulting in an inaccurate reflection of formulation characteristics or failure to meet target requirements. Reverse tracing involves analyzing the generation process of semantic physical quantities from the outset, starting with the final verification failure, and systematically examining each step and parameter that generates the semantic physical quantities to identify the points of deviation or error—the failure nodes. Once the failure nodes are found, they are corrected according to the specific circumstances, such as adjusting parameter values ​​or re-establishing the relationships between parameters. After correction, the corrected semantic physical quantities are obtained, and then verification is performed again to ensure that the corrected semantic physical quantities accurately reflect the formulation characteristics and meet the target requirements.

[0092] In one implementation, step S24 in this embodiment may include: when verification fails, locating the physical quantity corresponding to the performance deviation index through counterfactual analysis; identifying the conflicting failure node by tracing the generation path of the physical quantity in reverse; adjusting the weight of the failure node or replacing the associated data source of the failure node to obtain the corrected semantic physical quantity.

[0093] It should be noted that counterfactual analysis is a hypothetical reasoning method that imagines how the outcome would change if certain conditions or parameters were altered.

[0094] In this step, counterfactual analysis is used to hypothesize different values ​​for physical quantities and compare the differences between the actual and target results, thereby accurately identifying the key physical quantities causing performance deviations. For example, in the formulation optimization of a high-performance battery anode material, if verification shows that the battery's cycle life does not meet the expected target, counterfactual analysis can be used to hypothesize changes in physical quantities such as the porosity and conductive agent content of the anode material, observe the impact on cycle life, and then determine which physical quantity's value deviation caused the insufficient cycle life.

[0095] After determining the physical quantity corresponding to the performance deviation index, the generation path of the physical quantity is traced in reverse. Reverse tracing involves starting from the physical quantity result that ultimately failed verification and searching backwards along its generation path, analyzing each link that generated that physical quantity. During this process, if problems are found such as unreasonable data processing methods, incorrect parameter settings, or unreliable data sources in certain links, then these problematic links are the conflicting failure nodes.

[0096] For identified failure nodes, if the failure node is due to an unreasonable weighting of a parameter in the calculation, its weight can be adjusted to make its impact on the overall calculation result more reasonable. Another approach is to replace the associated data source of the failure node. When a failure node is found to be caused by an unreliable or erroneous data source, a more accurate and reliable data source is found and replaced. For example, insufficient precision equipment detection data in the granularity detection stage can be replaced with high-precision equipment detection data. After correcting the failure node using these two methods, the semantic physical quantity is recalculated to obtain the corrected semantic physical quantity.

[0097] Accordingly, step S30 includes: extracting knowledge rules from the semantic physical quantity or the modified semantic physical quantity using a cognitive distillation algorithm, and verifying the semantic physical quantity based on the knowledge rules.

[0098] This embodiment parses domain terms to establish a terminology parameter mapping table before extracting knowledge rules, and verifies it together with semantic physical quantities. In case of failure, it performs reverse tracking and correction, enhancing the accuracy of formulation parameters and ensuring that knowledge rules are extracted based on reliable parameters, avoiding optimization failures due to parameter errors. Furthermore, when verification fails, counterfactual analysis is used to locate the deviating physical quantities, reverse-track and identify the failure nodes, and adjust weights or replace the data source for correction. This allows for precise problem location, effective error correction, and ensures the correctness of semantic physical quantities, further improving the accuracy and stability of formulation optimization.

[0099] Based on the first embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Step S40 may include steps S401 to S404:

[0100] Step S401: After verification, construct a weighted loss function that includes performance and cost items, wherein the performance items are determined by the performance indicators in the target requirements, and the cost items are determined by the raw material price database and the process energy consumption model.

[0101] It should be understood that once the semantic physical quantities are validated, it means that the currently constructed semantic physical quantity system can accurately reflect the characteristics of the formulation. In order to further optimize the formulation in terms of performance and cost, a weighted loss function is constructed to quantitatively integrate the two key factors of performance and cost.

[0102] The determination of performance items revolves around the performance indicators in the target requirements. For example, when developing a new battery electrode material formulation, the target requirements clearly stipulate that the battery's charge-discharge efficiency must reach over 90% and its cycle life must exceed 1000 cycles. These specific performance indicators constitute the core content of the performance items. The determination of cost items relies on a raw material price database and a process energy consumption model. The raw material price database records the market price fluctuations of various raw materials, and the process energy consumption model can accurately calculate the energy consumption cost of the process required to produce the formulation. By comprehensively considering these two factors, the cost of the formulation can be accurately quantified.

[0103] Step S402: Based on the adaptive optimization algorithm, when the performance verification of the verified semantic physical quantity fails to meet the performance target, the weight of the performance item is increased.

[0104] It should be noted that the adaptive optimization algorithm in this embodiment is an intelligent optimization algorithm that can automatically adjust the weights of the performance and cost terms in the weighted loss function based on the performance verification results and cost situation during the formulation optimization process, so as to guide the optimization direction towards meeting the performance target and controlling the cost.

[0105] When the performance verification results corresponding to the verified semantic physical quantities fail to meet the performance target, the weight of the performance term in the weighted loss function is increased through an adaptive optimization algorithm. That is, in the subsequent optimization process, more emphasis will be placed on improving the performance of the formulation, guiding the optimization direction towards meeting the performance target, so that the optimized formulation can better meet the performance requirements of actual applications.

[0106] Step S403: When the cost corresponding to the verified semantic physical quantity exceeds the cost threshold, increase the weight of the cost item.

[0107] When the cost corresponding to the verified semantic physical quantity exceeds a pre-set cost threshold, it means that the cost of the current formula is too high. An adaptive optimization algorithm will increase the weight of the cost term in the weighted loss function, making the algorithm focus more on cost reduction in subsequent optimization processes. For example, this can be achieved using the formula minL = α·MSE. 性能 +β·Cost achieves the goal of minimizing the weighted loss function in this embodiment, where α and β are the weights of the performance term and the cost term, respectively.

[0108] Step S404: Based on the increased weights, obtain the optimized physical quantities as an intermediate formula representation.

[0109] After the adaptive optimization algorithm increases the weights of performance and cost terms based on performance and cost considerations respectively, it recalculates the formulation based on the new weights. During this process, the algorithm comprehensively considers the balance between performance and cost, continuously adjusting various parameters in the formulation to minimize the weighted loss function. Through the above calculations and optimizations, the optimized physical quantities are finally obtained. These physical quantities represent the content of each component in the formulation, process parameters, and other information, collectively forming the intermediate formulation representation.

[0110] This embodiment constructs a weighted loss function containing performance and cost terms. The adaptive optimization algorithm dynamically adjusts the weights based on the verification results to ensure that the optimized formulation meets the performance target while keeping the cost under control, thus achieving the optimal balance between performance and cost.

[0111] In one implementation, step S20 may include: mapping the formula parameters to a preset physical quantity range through similarity matching to obtain the matched physical quantity; and converting the matched physical quantity into a continuously differentiable numerical representation through a neural network to obtain a differentiable semantic physical quantity.

[0112] In practice, formulation parameters often exist in diverse forms, such as discrete numerical values, textual descriptions, or data with different dimensions. To enable unified optimization analysis and calculations later, these parameters need to be standardized to a specific range. Similarity matching is used to compare and analyze the formulation parameters with standard values ​​within a preset physical quantity range.

[0113] For example, when developing a novel lithium-ion battery anode material formulation, the formulation parameters include a "binder content" parameter, whose raw data is given as discrete values ​​in percentage form, such as 5%, 10%, etc. The preset physical quantity range is a reasonable range set based on materials science knowledge and experience, such as 3%-12%. Through a similarity matching algorithm, the system calculates the similarity between the actual "binder content" parameter and the values ​​within this range, finding the closest standard value, thus mapping the "binder content" to the 3%-12% preset physical quantity range, obtaining the matched physical quantity. This process not only unifies formulation parameters of different forms and dimensions to a suitable range but also eliminates some interference caused by differences in data format.

[0114] It should be understood that neural networks possess powerful nonlinear mapping capabilities, enabling them to transform input data into a specific form of output. When converting matched physical quantities into continuously differentiable numerical representations, multi-layered neural networks are employed, such as fully connected neural networks or convolutional neural networks (selected based on data characteristics). Taking a fully connected neural network as an example, it consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the matched physical quantities as input data, which may be discrete or not continuously differentiable. The hidden layers contain multiple neurons, each of which performs a nonlinear transformation on the input data through an activation function (such as the sigmoid function, hyperbolic tangent function, or linear rectified function).

[0115] During neural network training, learning is achieved through a large amount of sample data. This sample data includes the matched physical quantity of the input and the corresponding expected output (a numerical representation with a certain degree of continuous differentiability). Based on the error between the input data and the expected output, the neural network adjusts the connection weights between neurons in each layer using the backpropagation algorithm. The backpropagation algorithm calculates the gradient of the error with respect to each weight based on the chain rule, iteratively updating the weights to make the neural network's output gradually approach the expected output.

[0116] Furthermore, the step of converting the matched physical quantity into a continuously differentiable numerical representation through a neural network to obtain a differentiable semantic physical quantity may include: automatically routing to a dedicated processing channel based on the structural parameters / process parameters of the matched physical quantity; the structural parameters enter a topology-aware encoder, which extracts the interaction network features between material components through a neighborhood aggregation algorithm; the process parameters flow to a time-series causal encoder, which uses gated convolution to capture the dynamic dependencies in the process path; and obtaining the branched output based on the interaction network features and dynamic dependencies.

[0117] Embedding prior knowledge of materials science at the output of the neural network: By transforming physical rules such as Gibbs free energy and phase transition critical point into differentiable penalty terms, a boundary constraint matrix is ​​constructed; by using gradient redirection technology, when the output value exceeds the feasible domain of the material, backpropagation automatically adjusts the gradient direction along the physical constraint surface to obtain the final differentiable semantic physical quantity.

[0118] This embodiment transforms formulation parameters into differentiable semantic physical quantities through similarity matching and neural networks. Similarity matching accurately maps formulation parameters to a preset range of physical quantities, ensuring an accurate correspondence between parameters and actual physical quantities and avoiding optimization deviations caused by inaccurate parameters. The use of neural networks converts the matched physical quantities into continuously differentiable numerical representations, making these parameters mathematically differentiable and allowing for precise adjustment of parameters during the optimization process using methods such as gradient descent. Furthermore, the nonlinear fitting capability of neural networks can capture complex relationships between parameters, further improving optimization accuracy.

[0119] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the formulation optimization method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0120] This application also provides a formula optimization device, please refer to... Figure 4 The formula optimization device includes:

[0121] The requirement parsing module 10 is used to respond to target requirements in the form of multimodal natural language, and to perform semantic deconstruction of the target requirements through a large language model to obtain recipe parameters;

[0122] The parameter quantization module 20 is used to convert the formula parameters into differentiable semantic physical quantities.

[0123] The rule verification module 30 is used to extract knowledge rules from the semantic physical quantity through the cognitive distillation algorithm, and to verify the semantic physical quantity based on the knowledge rules;

[0124] The recipe optimization module 40 is used to perform multi-objective optimization on the validated semantic physical quantity by minimizing the weighted loss function after the validation is passed, so as to obtain an intermediate recipe representation.

[0125] The recipe generation module 50 is used to perform semantic decoding and physical field transformation on the intermediate recipe representation to obtain the complete target recipe.

[0126] The formula optimization apparatus provided in this application, employing the formula optimization method described in the above embodiments, can solve the technical problem of insufficient accuracy in the optimization process of existing formula generation methods. Compared with the prior art, the beneficial effects of the formula optimization apparatus provided in this application are the same as those of the formula optimization method provided in the above embodiments, and other technical features in the formula optimization apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0127] This application provides a formula optimization device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; 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 1.

[0128] The following is for reference. Figure 5 The diagram illustrates a structural schematic of a formula optimization device suitable for implementing embodiments of this application. The formula optimization device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Devices), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The illustrated formula optimization device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0129] like Figure 5 As shown, the recipe optimization device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the recipe optimization device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the formulation optimization device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows formulation optimization devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0130] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0131] The formula optimization device provided in this application, employing the formula optimization method described in the above embodiments, can solve the technical problem of insufficient accuracy in the optimization process of existing formula generation methods. Compared with the prior art, the beneficial effects of the formula optimization device provided in this application are the same as those of the formula optimization method provided in the above embodiments, and other technical features of the formula optimization device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0132] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0134] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the recipe optimization method in the above embodiments.

[0135] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0136] The aforementioned computer-readable storage medium may be included in the formulation optimization device; or it may exist independently and not be assembled into the formulation optimization device.

[0137] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the recipe optimization device, the recipe optimization device: responds to the target requirement in the form of multimodal natural language, performs semantic deconstruction of the target requirement through a large language model to obtain recipe parameters, converts the recipe parameters into differentiable semantic physical quantities, extracts knowledge rules from the semantic physical quantities through a cognitive distillation algorithm, and verifies the semantic physical quantities based on the knowledge rules. After verification, it performs multi-objective optimization on the verified semantic physical quantities by minimizing a weighted loss function to obtain an intermediate recipe representation, and performs semantic decoding and physical field transformation on the intermediate recipe representation to obtain the complete target recipe.

[0138] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, entirely on a remote computer or server, or on an ARM (Advanced RISC Machines) development board. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0139] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0140] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0141] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described formula optimization method, which can solve the technical problem that the optimization process of existing formula generation methods is not accurate enough. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the formula optimization method provided in the above embodiments, and will not be repeated here.

[0142] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for optimizing a formulation, characterized in that, The method includes the following steps: In response to the target requirements in the form of multimodal natural language, the target requirements are semantically deconstructed through a large language model to obtain the recipe parameters; The formula parameters are converted into differentiable semantic physical quantities; Knowledge rules are extracted from the semantic physical quantity using a cognitive distillation algorithm, and the semantic physical quantity is verified based on the knowledge rules. After successful verification, the verified semantic physical quantities are optimized by minimizing the weighted loss function to obtain intermediate recipe representations. The intermediate recipe representation is semantically decoded and physically transformed to obtain the complete target recipe.

2. The formulation optimization method as described in claim 1, characterized in that, Before the step of extracting knowledge rules from the semantic physical quantity using a cognitive distillation algorithm and verifying the semantic physical quantity based on the knowledge rules, the method further includes: Extract domain terminology from the target requirements; Establish a terminology parameter mapping table based on the domain terminology and the preset materialization parameter library; The terminology parameter mapping table and the semantic physical quantities are verified. When verification fails, the failure node in the semantic physical quantity is traced in reverse, and the semantic physical quantity is corrected based on the failure node to obtain the corrected semantic physical quantity. Accordingly, the step of extracting knowledge rules from the semantic physical quantity using a cognitive distillation algorithm and verifying the semantic physical quantity based on the knowledge rules includes: Knowledge rules are extracted from the semantic physical quantity or the modified semantic physical quantity using a cognitive distillation algorithm, and the semantic physical quantity is verified based on the knowledge rules.

3. The formulation optimization method as described in claim 2, characterized in that, The step of reverse tracing the failure node in the semantic physical quantity when verification fails, and correcting the semantic physical quantity based on the failure node to obtain the corrected semantic physical quantity includes: When verification fails, counterfactual analysis is used to locate the physical quantity corresponding to the performance deviation index. By tracing the generation path of the physical quantity in reverse, conflicting failure nodes are identified; Adjust the weight of the failed node or replace the associated data source of the failed node to obtain the corrected semantic physical quantity.

4. The formulation optimization method according to any one of claims 1 to 3, characterized in that, The step of responding to the target requirement in a multimodal natural language form and obtaining the recipe parameters by semantically deconstructing the target requirement through a large language model includes: In response to the target requirements of multimodal natural language forms, the term parameter mapping table is matched with the preset text database and preset image database for feature matching, and the recipe parameters with confidence are output. If there are contradictions in the data from different modalities during feature matching, the parameter priority of the formula parameters is adjusted through counterfactual analysis to output the final formula parameters.

5. The formulation optimization method according to any one of claims 1 to 3, characterized in that, The step of semantically decoding and physically transforming the intermediate recipe representation to obtain the complete target recipe includes: The intermediate recipe representation is associated with a preset process database to generate a cost tag set; The zero-shot generator is used to compare the matching degree between the text description in the intermediate recipe representation and the image features. If the matching degree is lower than a preset threshold, a molecular structure diagram corresponding to the text description is generated, and the molecular structure diagram is mapped to process parameters through physical field transformation. The final target formulation is generated based on the cost tag set and the process parameters.

6. The formulation optimization method according to any one of claims 1 to 3, characterized in that, The step of converting the formula parameters into differentiable semantic physical quantities includes: By matching similarity, the formula parameters are mapped to a preset physical quantity range to obtain the matched physical quantities; The matched physical quantities are converted into continuously differentiable numerical representations using a neural network, thereby obtaining differentiable semantic physical quantities.

7. The formulation optimization method according to any one of claims 1 to 3, characterized in that, The step of performing multi-objective optimization on the verified semantic physical quantity by minimizing the weighted loss function after successful verification to obtain an intermediate recipe representation includes: After verification, a weighted loss function containing performance and cost items is constructed, wherein the performance items are determined by the performance indicators in the target requirements, and the cost items are determined by the raw material price database and the process energy consumption model. Based on the adaptive optimization algorithm, when the performance verification of the verified semantic physical quantity fails to meet the performance target, the weight of the performance item is increased. When the cost corresponding to the verified semantic physical quantity exceeds the cost threshold, the weight of the cost item is increased; Based on the increased weights, the optimized physical quantities are obtained as intermediate formulation representations.

8. A formula optimization device, characterized in that, The formula optimization device includes: The requirement parsing module is used to respond to target requirements in the form of multimodal natural language, and to perform semantic deconstruction of the target requirements through a large language model to obtain the recipe parameters; The parameter quantization module is used to convert the formula parameters into differentiable semantic physical quantities. The rule verification module is used to extract knowledge rules from the semantic physical quantity through a cognitive distillation algorithm, and to verify the semantic physical quantity based on the knowledge rules; The recipe optimization module is used to perform multi-objective optimization on the validated semantic physical quantities by minimizing the weighted loss function after the validation is passed, so as to obtain an intermediate recipe representation. The recipe generation module is used to perform semantic decoding and physical field transformation on the intermediate recipe representation to obtain the complete target recipe.

9. A formula optimization device, characterized in that, The formulation optimization device includes: a memory, a processor, and a formulation optimization program stored in the memory and executable on the processor, wherein the formulation optimization program, when executed by the processor, implements the formulation optimization method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a recipe optimization program, which, when executed by a processor, implements the recipe optimization method as described in any one of claims 1 to 7.

Images