A raw material formula combination optimization method and system based on a deep calculation model

Abstract

This invention provides a method and system for optimizing raw material formulation combinations based on a deep computing model, relating to the field of machine learning technology. The method includes: establishing a detection task and a reference group for detection conditions; detecting the reference group under target detection conditions; generating benchmark detection results for the reference group; converting the formulation into raw material mass percentage records; using the detection results of the same batch of formulations under target and candidate detection conditions, along with the synchronous detection results of the reference group, as training samples to train a prediction model; obtaining the current detection results based on the target detection conditions and the reference group; iteratively obtaining the formulation to be tested by combining the prediction model; and detecting the tested formulation and updating the training samples, prediction model, or reference group based on the results. Therefore, this invention achieves the differentiation of the same indicator under different detection conditions, maps candidate detection condition data to target detection conditions, avoids model mislearning of condition differences, and improves the optimization stability and iterability of the model.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for optimizing raw material formulation combinations based on deep computing models. Background Technology

[0002] With the continuous growth of personalized skincare demands, the development of customized skincare formulations has become an important technological direction in the medical aesthetics and cosmetics industries. Deep computational models, due to their powerful nonlinear modeling capabilities, are widely used in the performance prediction and combination optimization of raw material formulations.

[0003] In the context of personalized skincare formula development, existing systems typically include a skin testing terminal, a formula database server, a model training module, and a formula optimization module. Specifically, the process begins by collecting historical formulas and their corresponding test results, encoding the types and proportions of raw materials, and training a formula performance prediction model. Then, optimization objectives and formula constraints are set to generate candidate raw material combinations, and the prediction model is used to predict the performance of these combinations. Combinations with unsatisfactory predicted performance are deleted, and qualified combinations are ranked according to performance indicators. Finally, after trial production and verification, the new data is fed back to update the model.

[0004] However, the aforementioned prior art has the following technical problems: In the research and development and trial production of personalized skincare ingredient formulations, existing technologies typically encode historical ingredient types, mass percentages, and test values ​​as training samples, then predict the physicochemical properties of candidate formulations and adjust the ingredient ratios accordingly. However, the same indicator may correspond to different data generation conditions. For example, pH values ​​may originate from the stock solution or aqueous dispersions of different proportions and are affected by temperature and electrode type; particle size may be obtained by laser diffraction or dynamic light scattering; stability may be derived from room temperature, high temperature, freeze-thaw, or centrifugation tests; and color is also affected by the light source, observation angle, and sample thickness. Although the above test values ​​have the same field name, the sample state, testing method, and physical conditions may differ. If data preprocessing only establishes a mapping between formulation parameters and test results without binding sample processing, testing temperature, test duration, instrument parameters, and testing methods to the results, the model may incorrectly attribute differences in testing conditions to changes in ingredient types or ratios. This can lead to incorrect additions, subtractions, or substitutions of ingredients during candidate formulation screening, causing predicted qualified formulations to fail in actual testing or incorrectly excluding originally qualified formulations. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for optimizing raw material formulation combinations based on a deep computing model, so as to solve the technical problem of erroneously attributing differences in detection conditions to changes in the type or proportion of raw materials.

[0006] In a first aspect, the present invention provides a method for optimizing raw material formulation combinations based on a deep computational model, comprising: Establish a reference group for detection tasks and conditions, conduct detection on the reference group under the target detection conditions, and generate benchmark detection results for the reference group.

[0007] The formula is converted into a raw material mass percentage record, and the test results of the same batch of formulas under the target test conditions and candidate test conditions, as well as the synchronous test results of the reference group obtained at the same time as the test results, are used to form a training sample.

[0008] The target detection condition prediction model is trained using the raw material mass percentage record, the synchronous detection results of the reference group, and the benchmark detection results of the reference group as inputs, and the detection results under the target detection conditions as labels.

[0009] The reference group is tested according to the target detection conditions, the current detection results of the reference group are generated, and the candidate formulations are evaluated and screened in combination with the target detection condition prediction model, and the formulation to be tested is determined iteratively.

[0010] Prepare and test the formulation to be tested, obtain the test results, and update the training samples, target detection condition prediction model, or detection condition reference group based on the test results.

[0011] Secondly, the present invention provides a raw material formulation combination optimization system based on a deep computing model, comprising: The reference benchmark establishment module is used to establish a reference group for detection tasks and detection conditions, detect the reference group under the target detection conditions, and generate the reference group benchmark detection results.

[0012] The training sample construction module is used to convert the formula into a raw material mass percentage record and to form a training sample by combining the test results of the same batch of formulas under the target test conditions and candidate test conditions, as well as the synchronous test results of the reference group obtained at the same time as the test results.

[0013] The prediction model training module is used to train a target detection condition prediction model by taking the raw material mass percentage record, the synchronous detection result of the reference group and the benchmark detection result of the reference group as inputs and the detection result under the target detection condition as the label.

[0014] The module for determining the formulation to be tested is used to test the reference group according to the target detection conditions, generate the current detection results of the reference group, and evaluate and screen the candidate formulations in combination with the target detection condition prediction model, and iteratively determine the formulation to be tested.

[0015] The closed-loop correction module is used to prepare and test the formulation to be tested, obtain the test results, and update the training samples, target detection condition prediction model, or detection condition reference group based on the test results.

[0016] The beneficial effects achieved by the technical solutions provided in the embodiments of the present invention include at least the following: (1) This invention divides the testing tasks according to the test object, sample state, test method and test procedure, and establishes a test condition reference group using standard samples and matrix reference formulas. The test results of different batches of test formulas are saved in correspondence with the test results of the reference group, thereby realizing the distinction between the physical meaning of the test results and the test state. This solves the problem of directly merging the same test indicators obtained under different sample states, test methods and test conditions in the prior art.

[0017] (2) This invention forms cross-condition training samples by combining data of the same batch of formulas under target detection conditions and candidate detection conditions, and maps the synchronous detection results of the reference group under both detection conditions to the formula detection results under the target detection conditions, thereby realizing a unified mapping from candidate detection condition data to target detection condition detection results, and solving the problem in the prior art that the deep computing model mistakenly learns the difference in detection conditions as the type of raw materials.

[0018] (3) By comparing the differences in detection results of the reference object and the same batch of trial production formula under the target detection conditions and the candidate detection conditions, the present invention updates the reference group of detection conditions respectively, or supplements cross-condition training samples and updates the target detection condition prediction model, thereby realizing the path closed-loop correction of insufficient identification of the reference group and insufficient model learning, and solving the problem that the source of trial production deviation cannot be distinguished and invalid trial production data is rewritten into the model in the prior art.

[0019] It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of a raw material formulation combination optimization method based on a deep computing model is shown in an embodiment of this disclosure; Figure 2 This diagram illustrates a raw material formulation combination optimization system structure based on a deep computing model, according to an embodiment of this disclosure. Figure 3 This diagram illustrates the construction of cross-condition training samples for the same batch of formulations in an embodiment of this disclosure. Figure 4 This diagram illustrates the identification of trial production deviation sources and the closed-loop diagram of the sub-paths in the embodiments of this disclosure. Detailed Implementation

[0022] Exemplary embodiments of this disclosure will be described more fully below with reference to the accompanying drawings.

[0023] The accompanying drawings are schematic illustrations of this disclosure and are not necessarily drawn to scale. Some block diagrams shown in the drawings may be functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in hardware modules or integrated circuits, or in networks, processors, or microcontrollers. Implementations can be carried out in various forms and should not be construed as limited to the examples set forth herein. The features, structures, or characteristics described in this disclosure may be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough description of embodiments of this disclosure. However, those skilled in the art will recognize that one or more specific details may be omitted when implementing the technical solutions of this disclosure. Sections, or other methods, components, devices, steps, etc., may be used to replace one or more specific details.

[0024] To make the objectives, technical solutions, and advantages of this disclosure clearer, the embodiments of this disclosure will be further described in detail below with reference to the accompanying drawings. Figure 1 The diagram shown is a flowchart of a raw material formulation combination optimization method based on a deep computing model in an embodiment of this disclosure.

[0025] S1: Establish a testing task and a testing condition reference group. Test the reference group under the target testing conditions and generate benchmark testing results for the reference group. It should be noted that in this embodiment, the testing condition reference group is simply referred to as the reference group. The manufacturer establishes testing tasks based on the final product inspection procedures and the company's testing operation procedures. Each testing task simultaneously records the name of the testing index, the actual measured physical quantity, the sample state, the testing method, the definition of the testing result, and the stability test procedure. The sample state refers to the specific form of the sample at the start of testing, including the original solution, a water dispersion of a specified ratio, a diluted solution, a homogenized sample, a degassed sample, a centrifuged sample, a sample stored at a specified temperature, or a sample after completing a specified freeze-thaw cycle. The definition of the testing result is the physical quantity expressed by the testing value and its expression method. The stability test procedure is the experimental operation process consisting of storage temperature, duration, number of cycles, packaging method, sampling time point, and result reading method. If any content changes and alters the physical meaning of the tested object or result, a new testing task is established. For example, detection tasks were established for the pH of the stock solution and the pH of the 10% aqueous dispersion, the median particle size (D50) by laser diffraction and the intensity-average particle size by dynamic light scattering, and storage at 45°C for four weeks and eight weeks. Instrument serial number, operator, or temperature changes within a specified range that did not alter the physical meaning were used as candidate detection conditions for verification.

[0026] For each testing task, target testing conditions are determined to ultimately determine product qualification. The actual processing parameters corresponding to the determined sample state for that testing task are fixedly recorded, including sample state, dilution ratio and medium, stirring or homogenization method, settling time, testing method, instrument type, measuring accessories, sample and ambient temperature, reading time, number of repetitions, and result retrieval method. The sample state in the testing task is used to determine the physical meaning of the test results, while the sample state in the target testing conditions is used to determine the specific processing method used in the actual testing. When the sample state changes and leads to a change in the object being tested, a new testing task should be established. For particle size testing tasks, refractive index, dispersion medium, and measurement mode are saved; for colorimetry testing tasks, light source, observation angle, and measuring aperture are saved; for stability testing tasks, storage temperature, duration, number of cycles, packaging method, and sampling time are saved. Particle size testing, colorimetry testing, and stability testing are testing task categories, while laser diffraction, dynamic light scattering, and specific colorimetry measurement methods are the testing methods used for the corresponding testing tasks. Candidate test conditions are those that measure the same physical quantities, use the same result definitions, and have the same sample conditions as the target test conditions, but differ only in instrument model, specified temperature range, operator, or sample processing parameters. If no historical test records exist, other instrument models, specified temperature ranges, or sample processing parameters intended for use by the manufacturer are used as candidate test conditions, and paired verification of the same batch of formulations is performed according to this procedure. If no alternative test conditions need to be verified, only the target test conditions are used to generate training data.

[0027] Subsequently, repeatability limits, permissible deviations for retesting, and permissible differences for equivalence of testing conditions are determined. Repeatability conditions are determined by the same laboratory conducting independent, repeated tests on the same sample within a short period, using the same testing method, the same or equivalent testing equipment, and the same sample processing steps. Repeatability limits represent the maximum permissible difference between test results obtained under repeatability conditions. The repeatability limits are preferentially adopted from the repeatability limits explicitly specified in the national, industry, or international standards upon which the testing method is based; if the standard does not specify a repeatability limit, the value confirmed by the manufacturer and incorporated into the testing operating procedures is used; if still not specified, the same laboratory conducts at least ten representative samples twice under repeatability conditions, arranges the absolute differences in ascending order, and takes the minimum upper limit that covers at least nine samples. Permissible deviations for retesting are determined based on the reference object certificate, the assigned uncertainty, and the storage stability verification results. The reference object certificate is a certificate of reference material, a certificate of reference sample, or a certificate of calibration equipment. When the certificate provides an allowable range for the assigned value, half the difference between the upper and lower limits of that range is used as the allowable deviation for retesting. When the certificate only provides the assigned value and expanded uncertainty, the sum of the expanded uncertainty and the maximum change in the test result during the storage stability verification period is used as the allowable deviation for retesting. The assigned uncertainty is preferentially read from the reference object certificate. When the enterprise assigns the value itself, it obtains no fewer than ten independent test results for the same batch of reference objects, uses the arithmetic mean of these results as the assigned value, and adds the squares of the standard uncertainty of repeated testing, the standard uncertainty of the reference testing method, and the standard uncertainty of the instrument resolution, then takes the square root and multiplies it by the coverage factor 2 to form the expanded uncertainty. The coverage factor is a coefficient used to expand the combined standard uncertainty into the expanded uncertainty; typically, 2 corresponds to approximately a 95% confidence level and is derived from the statistical coverage probability under the normal distribution assumption. The storage stability verification results include the initial confirmation average, the average of each subsequent retest after storage, and the absolute difference between the two. The allowable difference in test condition equivalence represents the maximum allowable absolute difference between the average test results of the same reference object or the same formulation under the target test conditions and the candidate test conditions. Equivalent testing conditions refer to two testing conditions measuring the same physical quantity, and the absolute difference of the above average values ​​does not exceed the allowable difference for testing condition equivalence. The allowable difference for testing condition equivalence shall preferentially adopt the cross-instrument or cross-condition allowable difference specified in the testing method; if not specified, the repeatability limit of the testing task shall be adopted.

[0028] Each testing task establishes a set of testing condition references, including standard samples for checking the condition of instruments and reagents, and matrix reference formulations to reflect the combined effects of product matrix, sample preparation, and testing conditions. Standard samples utilize standard substances, standard samples, or calibration instruments with assigned values, calibration, or certification basis, such as pH standard buffer solutions, particle size standard microspheres, or standard color charts. Matrix reference formulations are set separately according to product matrix categories, such as oil-in-water emulsions, water-in-oil emulsions, gels, or powder-containing dispersions.

[0029] Before use, the reference material must undergo homogeneity and storage stability verification. Liquid standard samples are taken from at least two independent packaging units within the same batch to compare for differences in filling, concentration, or storage between packaging units; for emulsions, gels, or powder-containing systems, samples are also taken from the top, middle, and bottom of the same packaging unit. Each sample must be tested independently at least twice. The test is valid if the absolute difference between the two results does not exceed the repeatability limit; if more than two tests are conducted, the difference between the maximum and minimum values ​​is used. The reference material can continue to be used if the difference between the periodic retest results and the certificate's assigned range or the initial confirmation result does not exceed the allowable deviation for retesting; otherwise, the batch of reference material should be discontinued. The reference material must be retested according to the retesting frequency specified in the certificate; if the certificate does not specify a retesting frequency, the time interval between two consecutive retests must not exceed one-quarter of the verified stable storage period. The periodic retest results are the arithmetic mean of the test results that meet the repeatability limit in this retest; the certificate's assigned range is the allowable range of the assigned results recorded in the certificate.

[0030] Under the target detection conditions, each reference object is tested. The validity of the current round of results is determined by whether the absolute difference between two test results is not greater than the repeatability limit. If the current round of results is valid, the arithmetic mean of the valid results for each reference object is calculated and arranged in a fixed numerical order as the baseline test results for the reference group. Each reference group version stores the reference object number and order, baseline test results, standard sample value information, matrix reference formulation batch, applicable product matrix category, target detection conditions, three limits, and expiration date. The standard sample value information includes the standard sample number, batch, value result, value uncertainty, value method, applicable temperature, expiration date, and storage conditions. When reference objects are added, replaced, or removed, a new reference group version is generated, without padding or interpolation of missing results from the old training samples.

[0031] When validating candidate detection conditions, homogeneous samples of representative cosmetic formulations from the same production batch are obtained and tested under both target and candidate detection conditions. A reference group of the same version of the detection conditions is also tested simultaneously to determine whether data under candidate detection conditions can be included in the training samples and whether the reference group can identify the differences between the target and candidate detection conditions. Specifically, the absolute difference between the average values ​​of the test results for each reference object and representative cosmetic formulation under the two conditions is calculated and compared with the corresponding allowable difference in detection condition equivalence. For any reference object, the average value of the target detection condition is obtained by summing all valid test results under the target detection condition and dividing by the number of valid tests. Then, the average value of the candidate detection condition is obtained by summing all valid test results under the candidate detection condition and dividing by the number of valid tests. Finally, the absolute value of the difference between the two average values ​​is calculated. The absolute value of the reference object is compared with the allowable difference in detection condition equivalence corresponding to that reference object, and the absolute value of the representative cosmetic formulation is compared with the allowable difference in detection condition equivalence corresponding to the detection task. Different reference objects are judged separately, and data with different dimensions or different physical quantities are not combined.

[0032] When the absolute differences of each reference object and the representative cosmetic formula do not exceed the allowable difference for equivalence of the corresponding detection conditions, the candidate detection conditions are deemed equivalent to the target detection conditions under the product matrix category, and the candidate detection conditions are included in the training data range. When the absolute difference of at least one reference object exceeds the allowable difference for equivalence of the corresponding detection conditions, it indicates that the reference group can identify the differences between the two detection conditions; under the premise that the two conditions measure the same physical quantity, the candidate detection conditions and the synchronous detection results of the reference group are included in the training data range. Among them, measuring the same physical quantity under the two conditions means that the detection results of the two conditions are defined in the same way, such as both being the median particle size D50; when one is D50 and the other is the intensity average particle size, it is not considered that the same physical quantity is measured. When each reference object does not exceed its own allowable difference for equivalence of detection conditions, but the absolute difference of the average detection result of the representative cosmetic formula under the two detection conditions exceeds the allowable difference for equivalence of detection conditions for the detection task, it indicates that the existing reference group cannot reflect the influence of the detection conditions on the product matrix, and is not included in the data under the candidate detection conditions. Instead, a matrix reference formula that can reflect the difference is added, and the scope of application of the reference group is limited; when the physical meanings of the two conditions are different, the detection task is split.

[0033] S2: Convert the formula into a raw material mass percentage record, and establish cross-condition detection units and model training records for the same batch of formulas. For example... Figure 3The diagram illustrates the construction of cross-condition training samples for batch formulations in this embodiment. The system arranges all candidate raw materials according to a unified raw material number. The actual weighed mass of each raw material is divided by the actual total mass of the production batch to form a raw material mass percentage record; unused raw materials are recorded as zero. If the actual weighed mass of a major raw material is missing, the formulation is not used for training. The actual weighed mass of each raw material is recorded as m1, m2, and so on, up to m... n The actual total mass of the production batch is denoted as M. The maximum permissible errors e1, e2, and e3 for the corresponding weighing are read from the verification or calibration certificates of each weighing device. n The maximum permissible error eM for measuring the total mass of the production batch. If the certificate does not specify the maximum permissible error, half of the minimum scale division of the weighing equipment shall be taken as the corresponding maximum permissible error. Calculate the absolute value of the difference between the sum of the actual weighed masses of each raw material and the actual total mass of the production batch. If this absolute value is greater than the sum of the maximum permissible errors for each corresponding weighing, the formula shall be marked as incomplete and not used for training.

[0034] At least two homogeneous samples are obtained from the same production batch. One sample is tested under the target test condition, and the other is tested under the candidate test condition validated by S1. If there is only one candidate test condition, and an even number of homogeneous samples are obtained, half are used for the target test condition and the other half for the candidate test condition; if an odd number of samples are obtained, the extra sample is used as a backup sample. If multiple candidate test conditions exist, at least one independent sample is allocated to each target test condition and each candidate test condition. Each formulation sample undergoes a specified number of independent tests, with the specified number of tests taking precedence over those specified in the testing procedure; if the testing procedure does not specify a number, each sample is tested independently twice. If the absolute difference between two test results exceeds the repeatability limit, the test result for that sample round is invalid, and the average result is not calculated. Reference objects of the same reference group version are tested synchronously within the same test batch. The system correspondingly saves records of raw material mass percentage, production batch, product matrix category, test results under the target test condition, test results under the candidate test condition, synchronous test results of the reference group corresponding to the target test condition and candidate test condition, reference group baseline test results, reference group version, actual sample processing records, and storage time before testing. The reference group baseline test results are the results fixed and saved after testing the reference object according to the target test conditions when the reference group version is established; the reference group synchronous test results are the current results obtained after testing the same reference object within the same test batch of the test formulation sample. The storage time before testing is the time length obtained by subtracting the production completion time stamp from the timestamp of the start of sample processing or testing; the stability test uses the timestamp of the sample entering the specified stability test environment minus the production completion time stamp. Only data from the same production batch with consistent raw material batch, production process, and storage time before testing can form a cross-condition test unit for the same batch formulation. Formulas reproduced according to historical records generate new production batch numbers and are not combined with the original historical samples to form a cross-condition test unit for the same batch formulation.

[0035] For historical samples, the original production batch number is retained. For formulations re-prepared according to historical production records, the raw material batch, actual weighing value, order of addition, emulsification temperature, homogenization conditions, and cooling process should be re-recorded, and a new production batch number should be generated. For continuous testing such as high-temperature storage and freeze-thaw cycles, multiple independent sealed sample units are obtained from the same production batch. Target and candidate test conditions use separate sample units. Each sample unit, along with the corresponding matrix reference formulation, undergoes the stability test procedure determined by the corresponding test task. Samples that have completed one stability test are not used for another. In one implementation, for stability testing at 45°C for four weeks, three independent sealed sample units are used for both the target and candidate test conditions. The corresponding matrix reference formulation also uses three independent sealed sample units under each test condition.

[0036] Two model training records are generated for each batch of formulations under cross-condition detection units. The first model training record uses the raw material mass percentage record, the synchronous detection result of the reference group under the target detection condition, and the baseline detection result of the reference group as input, and uses the detection result of the same batch of formulations under the target detection condition as the label. The second model training record uses the same raw material mass percentage record, uses the synchronous detection result of the reference group under the candidate detection condition and the baseline detection result of the reference group as input, and still uses the detection result of the same batch of formulations under the target detection condition as the label. The formulation detection results under the candidate detection condition are saved as model validation data to track the difference in detection results between the candidate detection condition and the target detection condition. Historical records with only manually entered temperature, dilution ratio, or instrument parameters, but lacking the detection results of the same batch under the target detection condition or the synchronous detection results of the reference group, are not included in the training samples.

[0037] S3: Training the Target Detection Condition Prediction Model. A separate target detection condition prediction model is established for each detection task and each reference group version. The model inputs are the raw material mass percentage record, the synchronous detection results of the reference group, and the baseline detection results of the reference group. The output is the predicted detection result under the target detection conditions. Different detection result definitions, sample states, detection methods, or stability test procedures do not share the same output layer. The product matrix category is used to select training samples and limit the model's applicability. Different detection result definitions include volume median particle size D50, intensity average particle size, color difference value, or other detection results representing different physical quantities or different statistical methods.

[0038] It should be noted that, in this embodiment, the deep computing model refers to a computing model consisting of an input layer, at least two hidden layers, and an output layer, which establishes the correspondence between input data and detection results under target detection conditions through layer-by-layer nonlinear operations. The input data includes raw material mass percentage records, synchronous detection results of the reference group, and baseline detection results of the reference group. The output is the predicted detection result under target detection conditions. This embodiment employs a multi-layer fully connected neural network. The raw material mass percentage records are input into the formula feature extraction layer, and the synchronous detection results of the reference group and the baseline detection results of the reference group are input into the detection condition feature extraction layer. After connecting the two types of features, the predicted detection result is output by a linear output node. The hidden layers use the ReLU activation function, and the parameters are updated using the Adam optimizer and the mean absolute error loss function.

[0039] The number of network layers and nodes are selected according to the following reproducible rules: Let the total number of input fields be 'd'. The number of nodes in the first hidden layer, h1, is a value that is not less than 16 and not greater than 128, chosen from 2d, 16, and 128. The number of nodes in the second hidden layer, h2, is h1 divided by 2, rounded down, and not less than 8. The number of nodes in the third hidden layer, h3, is h2 divided by 2, rounded down, and not less than 4. Establish a one-layer hidden layer structure containing h1, a two-layer hidden layer structure containing h1 and h2, and a three-layer hidden layer structure containing h1, h2, and h3. Perform five-fold cross-validation according to production batches. If there are fewer than five independent production batches, use one production batch as validation data and the remaining production batches as training data. Select the network structure with the smallest mean absolute error (MAE). If the difference between the MAEs of two structures does not exceed 1% of the smaller value, choose the structure with fewer parameters.

[0040] Training data and model validation data are categorized by formula number and production batch. All model training records from the same production batch are included in the same dataset. After training, validation records are input into the model, and the prediction results are compared with the measured results of the same batch of formulas under target detection conditions. The allowable prediction deviation of the model is predetermined by product inspection requirements or model validation procedures; if not specified, the repeatability limit of the corresponding detection task is used. For each validation record, the absolute value of the difference between the predicted result and the measured result under the target detection conditions is calculated; if the absolute value of any validation record exceeds the model's allowable prediction deviation, the prediction deviation of that validation record is considered to be outside the allowable range. If the above situation occurs in the validation records of a certain product matrix category, the process returns to S2 to supplement the cross-condition detection units of the same batch of formulas for that product category and then retrains.

[0041] The model's applicability is jointly determined by the reference group version, product matrix category, raw material ratio range, and the range of synchronous detection results for the reference group. Specifically, the raw material ratio range is the intersection of the range already covered by the training samples and the allowable range of raw materials; the range of synchronous detection results for the reference group is the minimum to maximum value of each reference object in the training samples. If a product matrix category has corresponding records in both the training and validation data, and the validation records for that product matrix category do not exceed the model's allowable prediction bias, then the product matrix category is considered to be covered.

[0042] The model is applicable only if the following conditions are met: the reference group version is the same as the reference group version saved in the model; the product matrix category is already covered; the mass percentage of each raw material is within the intersection of the minimum to maximum value of the raw material's training samples and the upper and lower limits of the raw material's mass percentage; and the current detection result of the reference group for each reference object is between the minimum and maximum value of the corresponding synchronous detection result of the reference group in the training samples. The model will only be invoked if all of the above conditions are met simultaneously.

[0043] S4: Generate the current test results for the reference group and iteratively determine the formulation to be tested. Before starting the candidate formulation evaluation, test the reference group according to the target test conditions and the version of the target test condition prediction model. Each reference object completes a specified number of independent tests. If the test operation procedure does not specify the number of repeated tests, each reference object is tested independently twice. The repeatability limit is used to determine the validity of the test, the average value of the valid test results is calculated, and the current test results of the reference group are formed according to a fixed numbering order. Valid test results are those that meet the repeatability limit for repeated tests, have complete test records, and are free from instrument malfunctions, sample contamination, or abnormal interruptions. Forming according to a fixed numbering order means that the current test results of the reference group are written into the corresponding fields according to the reference object numbers and arrangement order saved in the reference group version.

[0044] Calculate the absolute difference between the current test result of each reference object and the baseline test result of the reference group. If none of the absolute differences exceed the allowable difference for the equivalence of the corresponding test conditions, the reference object is within its validity period, and the current test result of the reference group is within the applicable range of the model, then the target test condition prediction model is invoked. If any condition is not met, the candidate formulation evaluation for the corresponding test task is stopped, and the instruments, reagents, sample processing equipment, sample temperature, and reference object status are checked. After eliminating anomalies, the test is repeated. If the invocation conditions are still not met after the first retest, another anomaly check and retest are performed. If the invocation conditions are still not met after two consecutive retests, the candidate formulation evaluation for that test task is stopped, and the reference object number, test result, and corresponding limit value that do not meet the invocation conditions are output. Retesting is then discontinued.

[0045] The actual adjustment step size for each raw material is determined based on the production batch size and the minimum weighing value of the weighing equipment. Specifically, the minimum weighing value of a raw material is divided by the production batch size, and then multiplied by 100% to obtain the minimum mass percentage that the raw material can increase or decrease at one time. The continuous phase carrier raw material is designated as the mass balance raw material, where the continuous phase carrier raw material is the raw material that forms the continuous phase of the product and can be used to supplement the total mass of the formulation, such as purified water in an oil-in-water emulsion. When other raw materials increase, the mass balance raw material decreases by the same mass percentage; when other raw materials decrease, the mass balance raw material increases by the same mass percentage. Multiple raw materials with a fixed ratio are set up as a ratio group, and each raw material in the ratio group is adjusted synchronously according to the fixed ratio relationship.

[0046] At least one formula capable of actual production is used as the starting formula, and this starting formula is set as the current formula. In each round, candidate formulas are generated by increasing the actual adjustment step size by one for one raw material, decreasing the actual adjustment step size by one for one raw material, and increasing or decreasing the actual adjustment step size by one for an entire proportion group. If the amount of a raw material decreases below the minimum actual feed amount, a candidate formula can also be generated to adjust that raw material to zero. A single actual adjustment step size is used to generate candidate formulas, ensuring that the mass difference between adjacent candidate formulas is not less than the minimum change that the weighing equipment can actually perform, and that each candidate formula can be obtained from the current formula through a single weighing adjustment. The upper and lower limits of the mass percentage of candidate raw materials are determined and saved before the candidate formula evaluation begins; their values ​​are determined based on the permitted usage under regulations, product safety assessment results, supplier recommended usage, and production process validation range. The allowable range for mass-balanced raw materials is the upper and lower limits of their corresponding mass percentage. The proportioning relationship is fixed, and the proportioning coefficients and allowable relative deviations of each raw material are saved simultaneously. For any two raw materials within a ratio group, calculate the mass ratio of the two raw materials in the candidate formulation, and then calculate the relative difference between this mass ratio and the ratio of the fixed ratio coefficient. If the relative difference exceeds the allowable relative deviation stored in the fixed ratio relationship, the fixed ratio relationship is considered to be broken. If any raw material in the candidate formulation exceeds the upper or lower limit of its mass percentage, the mass balance raw material exceeds its upper or lower limit of its mass percentage, the fixed ratio relationship is broken, the non-zero raw material is lower than the minimum actual feed amount, the total mass percentage of raw materials is not equal to 100%, or the candidate formulation exceeds the applicable range of the model, then the candidate formulation is excluded. Input the raw material mass percentage records of the retained candidate formulation, the current test results of the reference group, and the benchmark test results of the reference group into the model corresponding to each test task to obtain the predicted test results under the target test conditions, and determine whether they are below, within, or above the qualified range. The qualified range of the test results is determined by the final product inspection procedure, the enterprise testing operation procedure, or the product technical requirements, and is fixed before the candidate formulation evaluation begins.

[0047] If all candidate formulations are qualified (i.e., the predicted test results for all testing tasks fall within the qualified range for their respective test results), calculate the total change in raw materials for each candidate formulation relative to the current formulation. Sort the candidate formulations by their total change in raw materials from smallest to largest, and select the candidate formulation ranked first. If the total change in raw materials is the same, select the candidate formulation with the fewer types of raw materials changed. The total change in raw materials is the sum of the absolute values ​​of the percentage change in the mass of each raw material. When all qualified candidate formulations exist, first exclude those that would render a qualified test task unqualified. A qualified test task is one where the predicted test result of the current formulation is already within the acceptable range. If a candidate formulation causes the predicted test result of that test task to fall outside the acceptable range, then that candidate formulation is deemed to have rendered a qualified test task unqualified. Next, compare the number of test tasks that have entered the acceptable range. Entering the acceptable range means that the predicted test result of the current formulation is outside the acceptable range, while the predicted test result of a candidate formulation has entered the acceptable range. By comparing the number of test tasks that have entered the acceptable range, and without assigning different manual weights to different test tasks, prioritize the candidate formulation that satisfies the requirements for more test tasks.

[0048] If no testing task enters the acceptable range, the lower limit of the acceptable range is denoted as L, the upper limit as U, and the predicted test result as y. When y is less than L, the range distance is L - y; when y is between L and U, the range distance is 0; when y is greater than U, the range distance is y - U. If the range distance of a candidate formula in a certain testing task is less than the range distance corresponding to the current formula, the testing task is considered to be moving closer to the acceptable range. A new current formula is determined sequentially based on the following criteria: a larger number of testing tasks entering the acceptable range, a larger number of testing tasks moving closer to the acceptable range, fewer changes in raw material types, and a smaller total change in raw materials. The above conditions are compared level by level in the order they are recorded; a subsequent condition is compared only if the previous condition is the same. If all conditions are the same, the candidate formula is selected by arranging the smallest changed raw material number in ascending order and choosing the first-ranked candidate formula.

[0049] The new current formulation will continue to be used for the next round of candidate formulation generation and evaluation, and the percentage of raw material mass that has been evaluated will be recorded and saved as visited formulations to avoid duplicate evaluation. When all predicted results for all testing tasks fall within the corresponding acceptable range, or when there are no candidate formulations that can improve unacceptable testing tasks without compromising acceptable testing tasks, or when visited formulations reappear, iteration will stop and the formulation to be tested will be determined. When iteration stops, if there are evaluated formulations with all predicted test results being acceptable, those with fewer changes in the number of raw material types and smaller total changes in raw materials will be selected as the formulation to be tested. If there are no evaluated formulations with all predicted test results being acceptable, those with a higher number of predicted acceptable testing tasks, a higher number of testing tasks closer to the acceptable range of test results relative to the initial formulation direction, fewer changes in the number of raw material types, and smaller total changes in raw materials will be selected as the formulation to be tested.

[0050] In one embodiment, the product is an oil-in-water moisturizing emulsion, with a production batch size of 100 kg. The starting formula includes 73.70% purified water, 5.00% glycerin, 3.00% 1,3-propanediol, 2.00% niacinamide, 0.25% carbomer, 0.15% xanthan gum, 8.00% caprylic / capric triglyceride, 3.00% glyceryl stearate and PEG-100 stearate emulsifier, 2.00% cetearyl alcohol, 1.00% polydimethylsiloxane, 0.20% tocopherol, 0.80% phenoxyethanol, 0.10% ethylhexylglycerin, 0.05% disodium EDTA, 0.70% tromethamine, and 0.05% fragrance, totaling 100.00%. All percentages are the mass percentage of the corresponding raw material to the actual total mass of the formula. Purified water is used as a mass-balancing raw material, and phenoxyethanol and ethylhexylglycerin are mixed in an 8:1 ratio. The target testing tasks include pH of the stock solution at 25℃, D50 of the emulsion droplets by laser diffraction, and appearance stability after four weeks of storage at 45℃. The acceptable ranges are 5.50–6.20 μm, 2.0–6.0 μm, and the layer height ratio after four weeks of storage at 45℃, respectively. The specific stability testing procedure is as follows: k independent sealed sample units from the same production batch are placed into transparent cylindrical containers with millimeter graduations of the same inner diameter, ensuring that the initial sample height of each sample is consistent. After four weeks of storage at 45℃, the samples are transferred to a 25℃ environment and allowed to stand for 24 hours. The total sample height H of each sample is recorded. When a separation layer forms, the height of each separation layer is recorded, and the heights of all separation layers are added together to obtain the total separation layer height h; otherwise, h is recorded as 0. h is divided by H and then multiplied by 100% to obtain the layer height ratio of the sample. The arithmetic mean of the sum of the layer height ratios of the k independent sealed sample units and divided by k is taken as the stability test result of the production batch after four weeks of storage at 45℃. The minimum weighing values ​​for the aqueous and oil phase weighing devices are 50g and 10g, respectively, corresponding to actual adjustment steps of 0.05% and 0.01%. The model predicts the initial formulation to have a pH of 6.34, a D50 of 6.7μm, and a layering height ratio of 2.3%. The system generates candidate formulations by reducing tromethorphan, increasing the compound emulsifier, or adding cetearyl alcohol, and then supplementing or subtracting the same percentage with purified water. Through iteration, a formulation to be tested was determined with 0.60% tromethorphan, 3.10% compound emulsifier, 2.05% cetearyl alcohol, and 73.65% purified water, with other raw materials remaining unchanged; its predicted pH is 6.08, D50 is 5.4μm, and layering height ratio is 0.7%.

[0051] S5: Prepare the test formulation according to the raw material types and mass percentages output in S4, forming a trial production batch, and save the raw material batch, actual weighing value, feeding sequence, emulsification temperature, homogenization speed and time, cooling process, and batch number. Test the test formulation according to the target testing conditions. Within the same testing batch of the test formulation, simultaneously test the same testing condition reference group version as the one used in S4. The reference objects, reference object numbers, and arrangement order in this testing condition reference group are consistent with those used in S4. Obtain the test results of the trial production and the synchronous test results of the reference group. For stability testing, use independent sealed samples, and subject the test sample and the matrix reference formulation to the same stability test procedure. Figure 4 The diagram shown is a sample production deviation source identification and path closed-loop diagram in the embodiments of this disclosure.

[0052] The system determines the validity of the trial production test results based on the repeated test results of the reference group, the validity period of the reference object, the difference between the synchronous test results of the reference group and the baseline test results of the reference group, the actual stability test procedure, and abnormal interruption records. For each reference object, the repeated test of the reference group meets the requirements if the absolute difference between two test results is not greater than the repeatability limit, or the difference between the maximum and minimum values ​​of multiple test results is not greater than the repeatability limit. If the test date does not exceed the validity period of the reference object, the reference object is within its validity period. The absolute value of the difference between the synchronous test results of the reference group and the baseline test results of the reference group is calculated one by one. If each absolute value does not exceed the allowable difference for equivalence of the corresponding test conditions, the test status meets the requirements. If the actual temperature, duration, number of cycles, packaging method, and sampling time are the same as the target test conditions, or the corresponding deviations are within the allowable range specified in the test operation procedure, the actual stability test procedure meets the target test conditions. If the trial production sample and the matrix reference formulation use the same packaging method, start and end the treatment simultaneously in the same equipment, and are tested at the same sampling time, they are considered to have undergone the same treatment.

[0053] If repeated testing fails to meet the repeatability limit, the reference object is expired, the difference between the synchronous test results of the reference group and the baseline test results of the reference group exceeds the allowable difference for equivalence of the test conditions, the stability test procedure does not meet the target test conditions, the trial sample and the matrix reference formulation have not undergone the same treatment, or there is an unrecorded abnormal interruption, the result is deemed invalid, and the test is repeated after the abnormality is eliminated. Invalid trial test results are not used for formula qualification judgment and are not included in the training samples.

[0054] If the trial production test results are valid and all test tasks are qualified, output the raw material formula confirmed by the actual trial production, and save this batch of data as model validation data. If the trial production results are valid but there are unqualified test tasks, and the difference between the predicted results and the trial production results does not exceed the model's allowable prediction deviation, return to S4, and continue screening from other starting formulas or unevaluated candidate formulas; if there are no candidate formulas that can be improved, output the current candidate raw materials and the formulas that have not all qualified under production constraints.

[0055] When the trial production test results are valid, and the difference between the predicted test results and the trial production test results for at least one test task exceeds the model's allowable prediction deviation, independent samples are obtained from the retained samples of the same trial production batch. These samples are then tested according to the target test conditions and the validated candidate test conditions, respectively, while a reference group of the same version of the test conditions is tested simultaneously. The absolute difference between the average test results of each reference object under the target test conditions and the validated candidate test conditions, as well as the absolute difference between the average test results of the same batch of trial production formulations under the target test conditions and the validated candidate test conditions, are calculated and compared with the corresponding allowable difference in test condition equivalence.

[0056] When the absolute difference of each reference object does not exceed the corresponding limit, but the absolute difference of the same batch of trial formulations exceeds the corresponding limit, it is determined that the existing test condition reference group fails to reflect the influence of test conditions on the product matrix. In this case, a matrix reference formulation that can reflect this change is added, or the product matrix is ​​removed from the scope of the existing reference group. If the physical meanings of the two conditions are different or a reliable identification relationship cannot be established, the test task is split. When adding a matrix reference formulation, a candidate matrix reference formulation is selected from formulations belonging to the same product matrix category as the trial formulation and having similar sample processing characteristics. A specified number of tests are performed under both the target test condition and the validated candidate test condition. If the absolute difference of the average test results of the candidate matrix reference formulation under the two test conditions exceeds the allowable difference for test condition equivalence, and passes the homogeneity and storage stability verification, the candidate matrix reference formulation is added to the test condition reference group. When the reference object changes, a new version of the reference group is generated, and training samples and the target test condition prediction model are re-established according to the new version.

[0057] When the absolute difference of at least one reference object exceeds the corresponding limit, it is determined that the reference group of detection conditions has obtained information that can distinguish between the target detection condition and the candidate detection condition. If the model fails to accurately predict the detection result of the same batch of trial production formula under the target detection condition based on the synchronous detection result of the reference group under the candidate detection condition, the raw material mass percentage record of the trial production batch, the formula detection result under the two detection conditions, the corresponding synchronous detection result of the reference group, the reference group benchmark detection result, and the reference group version are added to S2 to form a new cross-condition detection unit for the same batch of formula, and the target detection condition prediction model is retrained according to S3.

[0058] If any of the target detection condition test results, candidate detection condition test results, or reference group synchronous test results are missing, the model or detection condition reference group will not be directly updated based on a single trial production result. The computer system executing this method will mark the record as incomplete data. If there are retained samples from the same trial production batch, an independent sample will be obtained from the retained samples and the missing test results will be supplemented to form a complete cross-condition detection unit for the same batch formulation. If there are no retained samples from the same batch or the test results cannot be supplemented, the update of the current training sample or detection condition reference group will be terminated, and the incomplete record will not be included in the training sample.

[0059] After updating the training samples, target detection condition prediction model, or detection condition reference group, S4 and S5 are re-executed. If all tested qualified formulations cannot be obtained within the limits of candidate raw materials and their mass percentage upper and lower limits, production batch size, minimum weighing value of weighing equipment, applicable range of detection condition reference group, and applicable range of model, the closed loop is terminated and the formulations that do not meet all detection requirements are output. The system also saves and outputs the detection tasks that have not entered the qualified range of detection results, the evaluated formulations, the upper and lower limits of raw material mass percentage, the reason for stopping iteration, and the currently best-ranked formulation to be tested. The formulation developers then re-determine the candidate raw materials, the upper and lower limits of raw material mass percentage, or the production batch size before re-executing S4; the qualified range of detection results is not automatically widened, nor are prediction results outside the applicable range of the model used to replace the actual trial test results.

[0060] like Figure 2 The diagram shown is a structural diagram of a raw material formulation combination optimization system based on a deep computing model, according to an embodiment of this disclosure, including: The reference benchmark establishment module is used to establish a reference group for detection tasks and detection conditions, detect the reference group under the target detection conditions, and generate the reference group benchmark detection results.

[0061] The training sample construction module is used to convert the formula into a raw material mass percentage record and to form a training sample by combining the test results of the same batch of formulas under the target test conditions and candidate test conditions, as well as the synchronous test results of the reference group obtained at the same time as the test results.

[0062] The prediction model training module is used to train a target detection condition prediction model by taking the raw material mass percentage record, the synchronous detection result of the reference group and the benchmark detection result of the reference group as inputs and the detection result under the target detection condition as the label.

[0063] The module for determining the formulation to be tested is used to test the reference group according to the target detection conditions, generate the current detection results of the reference group, and evaluate and screen the candidate formulations in combination with the target detection condition prediction model, and iteratively determine the formulation to be tested.

[0064] The closed-loop correction module is used to prepare and test the formulation to be tested, obtain the test results, and update the training samples, target detection condition prediction model, or detection condition reference group based on the test results.

[0065] As can be seen from the above, the technical solutions disclosed herein can be implemented as methods, apparatus, systems, computer program products, storage media, electronic devices, etc. Those skilled in the art will understand that various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which may be referred to as "circuit," "module," or "system," respectively.

[0066] It should be understood that this disclosure is not limited to the specific methods, steps, or structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. Those skilled in the art will readily conceive of other embodiments based on the specific implementations provided in this disclosure. Therefore, the specific implementations provided in this disclosure are merely exemplary, and the scope and spirit of this disclosure are indicated by the claims, and should cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary technical means in the art not disclosed in this disclosure.

Claims

1. A method for optimizing raw material formulation combinations based on a deep computational model, characterized in that, include: S1: Establish a reference group for detection tasks and detection conditions, detect the reference group under the target detection conditions, and generate the benchmark detection results of the reference group; S2: Convert the formula into a raw material mass percentage record, and use the test results of the same batch of formulas under the target test conditions and candidate test conditions, as well as the synchronous test results of the reference group obtained synchronously with the test results, to form a training sample; S3: Using the raw material mass percentage record, the synchronous detection result of the reference group, and the benchmark detection result of the reference group as input, train the target detection condition prediction model with the detection result under the target detection condition as the label; S4: Detect the reference group according to the target detection conditions, generate the current detection results of the reference group, and evaluate and screen the candidate formulas in combination with the target detection condition prediction model, and iteratively determine the formula to be tested; S5: Prepare and test the formulation to be tested, obtain the test results, and update the training samples, the target detection condition prediction model, or the detection condition reference group based on the test results.

2. The raw material formulation combination optimization method based on a deep computing model according to claim 1, characterized in that, S1 includes: For candidate detection conditions, samples of representative cosmetic formulas were obtained from the same production batch and tested under both target and candidate detection conditions, while a reference group under the same detection conditions was tested simultaneously. Calculate the absolute value of the difference between the average detection result of each reference object under the target detection conditions and the average detection result of the reference object under the candidate detection conditions, and calculate the absolute value of the difference between the average detection result of the representative cosmetic formula under the target detection conditions and the average detection result of the representative cosmetic formula under the candidate detection conditions. The absolute value corresponding to each reference object is compared with the allowable difference of the equivalence of the detection conditions corresponding to that reference object, and the absolute value corresponding to the representative cosmetic formula is compared with the allowable difference of the equivalence of the detection conditions corresponding to the detection task. When the absolute value corresponding to each reference object does not exceed the allowable difference of the equivalence of the detection conditions corresponding to that reference object, and the absolute value corresponding to the representative cosmetic formula does not exceed the allowable difference of the equivalence of the detection conditions corresponding to the detection task, the candidate detection conditions are included in the training data range. When the absolute difference of at least one reference object exceeds the corresponding limit, the candidate detection conditions and the synchronous detection results of the reference group will be included in the training data range. When the absolute difference of each reference object does not exceed the corresponding limit, but the absolute difference of the representative cosmetic formula exceeds the corresponding limit, the data under the candidate detection condition will not be written into the training sample, and a matrix reference formula will be added, the applicable scope of the detection condition reference group will be adjusted, or the detection task will be split.

3. The raw material formulation combination optimization method based on a deep computing model according to claim 2, characterized in that, The S2 includes: At least two uniform samples were obtained from the same production batch and tested according to the target test conditions and the validated candidate test conditions, respectively. During the testing of each sample, the reference objects of the same reference group version were tested simultaneously. Records of raw material mass percentage, production batch, product matrix type, test results under two test conditions, corresponding synchronous test results of the reference group, baseline test results of the reference group, and reference group version are kept. The above data, which are consistent in terms of raw material batch, production process, and storage time before testing, are combined into a cross-condition testing unit for the same batch of formula.

4. The raw material formulation combination optimization method based on a deep computing model according to claim 3, characterized in that, Also includes: Two model training records are generated for each cross-condition detection unit of the same batch of formulas; The first model training record uses the raw material mass percentage record, the synchronous detection result of the reference group under the target detection condition, and the baseline detection result of the reference group as input, and the detection result under the target detection condition as the label; The second model training record uses the same raw material mass percentage record, the synchronous detection results of the reference group under the candidate detection conditions, and the baseline detection results of the reference group as input, and uses the detection results of the same batch of formula under the target detection conditions as labels; The detection results under the candidate detection conditions are saved as model validation data.

5. The raw material formulation combination optimization method based on a deep computing model according to claim 4, characterized in that, The S3 includes: For each testing task and each reference group version, a target testing condition prediction model is established. The raw material mass percentage record, the synchronous testing results of the reference group, and the baseline testing results of the reference group are input into the target testing condition prediction model, and the predicted testing results under the target testing conditions are output. Training data and model validation data are divided according to formula number and production batch, so that all model training records formed in the same production batch are included in the same data set. The applicability of the target detection condition prediction model is determined based on the reference group version, product matrix category, raw material ratio range, and the range of synchronous detection results of the reference group.

6. The raw material formulation combination optimization method based on a deep computing model according to claim 5, characterized in that, The S4 includes: The current detection result of the reference group is compared with the baseline detection result of the reference group. When the difference between each reference object does not exceed the allowable difference of the corresponding detection condition equivalence, and the current detection result of the reference group is within the applicable range of the target detection condition prediction model, the target detection condition prediction model is invoked. The actual adjustment step size of the raw materials is determined based on the production batch and the minimum weighing value of the weighing equipment. The raw materials for mass balance are specified, and the raw materials with fixed proportions are set as proportion groups. Using the initial recipe as the current recipe, increase or decrease one actual adjustment step to generate candidate recipes, and exclude candidate recipes that do not meet the candidate recipe constraints. The candidate formulation constraints are jointly defined by the upper and lower limits of the mass percentage of the candidate raw materials, the fixed ratio relationship, the production batch, the minimum weighing value of the weighing equipment, the mass balance raw materials, and the applicable scope of the target detection condition prediction model.

7. The raw material formulation combination optimization method based on a deep computing model according to claim 6, characterized in that, Also includes: The raw material mass percentage records of the candidate formulation, the current test results of the reference group, and the baseline test results of the reference group are input into the target test condition prediction model corresponding to each test task to obtain the predicted test results; The predicted test result is determined to be below the acceptable range of the test result, within the acceptable range of the test result, or above the acceptable range of the test result. Prioritize candidate formulations where all predicted test results fall within the corresponding acceptable range and the total change in raw materials is relatively small. If no candidate formulations are found to be entirely qualified, exclude candidate formulations that would render a qualified test task unqualified. The new current formula is determined based on the number of testing tasks that have entered the acceptable range of test results, the number of testing tasks that are moving closer to the acceptable range of test results, the number of raw material types that have been changed, and the total amount of raw material changes. Candidate formulas are then generated based on the new current formula until the formula to be tested is determined.

8. The raw material formulation combination optimization method based on a deep computing model according to claim 7, characterized in that, The S5 includes: Prepare the formulation to be tested to form a trial production batch, test the formulation to be tested according to the target testing conditions, and simultaneously test the testing conditions reference group with the same reference object and arrangement order as the testing conditions reference group used in S4 within the same testing batch, and obtain the trial production test results and the synchronous test results of the reference group. The validity of the trial production test results is determined based on the repeated test results of the test conditions reference group, the validity period of the reference object, the difference between the synchronous test results of the reference group and the baseline test results of the reference group, the actual stability test procedure, and the abnormal interruption records. The model is set to allow prediction bias. When the trial production test results are valid, there are unqualified test tasks, and the difference between the predicted test results and the trial production test results does not exceed the model's allowed prediction bias, the process returns to S4 to continue determining the test production formula from other starting formulas or unevaluated candidate formulas.

9. The raw material formulation combination optimization method based on a deep computing model according to claim 8, characterized in that, Also includes: When the trial production test results are valid, and the difference between the predicted test results of at least one test task and the trial production test results exceeds the prediction deviation allowed by the model, independent samples are obtained from the same trial production batch and tested according to the target test conditions and candidate test conditions respectively, and the same test condition reference group is tested simultaneously. Calculate the absolute difference between the average values ​​of the test results of each reference object and the same batch of trial-produced formula under the two test conditions, and compare them with the allowable difference of the equivalence of the corresponding test conditions. When the absolute difference of each reference object does not exceed the corresponding limit, but the absolute difference of the same batch of trial formulations exceeds the corresponding limit, update the matrix reference formulation or scope of application in the test condition reference group and generate a new reference group version. When the absolute difference of at least one reference object exceeds the corresponding limit, the raw material mass percentage record of the trial production batch, the test results under two test conditions, the corresponding synchronous test results of the reference group, and the reference group benchmark test results are added to the training sample, and the target test condition prediction model is updated.

10. A raw material formulation combination optimization system based on a deep computing model, characterized in that, include: The reference benchmark establishment module is used to establish a reference group for detection tasks and detection conditions, detect the reference group under the target detection conditions, and generate the reference group benchmark detection results. The training sample construction module is used to convert the formula into a raw material mass percentage record and to form a training sample by combining the test results of the same batch of formulas under the target test conditions and candidate test conditions, as well as the synchronous test results of the reference group obtained synchronously with the test results. The prediction model training module is used to train a target detection condition prediction model with the raw material mass percentage record, the synchronous detection result of the reference group and the benchmark detection result of the reference group as input and the detection result under the target detection condition as the label. The module for determining the formulation to be tested is used to test the reference group according to the target detection conditions, generate the current detection results of the reference group, and evaluate and screen the candidate formulations in combination with the target detection condition prediction model, and iteratively determine the formulation to be tested. The closed-loop correction module is used to prepare and test the formulation to be tested, obtain the test test results, and update the training samples, the target detection condition prediction model, or the detection condition reference group based on the test test results.

Images