Method and system for optimizing formula of low-dielectric glass fiber composition

Abstract

The invention relates to a low-dielectric glass fiber composition formula optimization method and system, and relates to the technical field of glass fiber, and the method comprises the following steps: receiving a formula component type list and a formula component constraint percentage list uploaded by a client, and carrying out formula population initialization to obtain an initial formula set; inputting the initial formula set into a dielectric constant prediction model for analysis to obtain a dielectric constant prediction value set; inputting the initial formula set into a dielectric loss prediction model for analysis to obtain a dielectric loss prediction value set; based on the set of dielectric constant predicted values and the set of dielectric loss predicted values, recipes that are both less than a dielectric constant threshold and a dielectric loss threshold are sorted from the initial recipe set, and returned to the client. The technical problems of low formula optimization efficiency and insufficient precision caused by limitation and subjectivity of manual configuration in the prior art are solved, and the technical effects of improving the formula optimization efficiency and precision and further improving the product performance are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of glass fiber, in particular to a low dielectric glass fiber composition formula optimization method and system. BACKGROUND

[0002] Low dielectric glass fiber composition is a material widely used in the field of high frequency communication, and its dielectric constant and dielectric loss are key factors affecting signal transmission performance. At present, the formula determination method of low dielectric glass fiber composition is mainly based on empirical component adjustment and experimental verification. This method depends on the knowledge reserve and experience of technical personnel, and the proportion of each component is adjusted manually, and the final formula is determined through repeated test verification. On the one hand, due to the limitation of individual knowledge reserve, it is difficult to fully consider the complex interaction between components, resulting in great limitation of formula optimization; on the other hand, the subjective factors of manual configuration have great influence, and different personnel have different judgment standards, resulting in lack of consistency in formula optimization. The limitation of manual knowledge and experience and the instability of subjective judgment make the formula optimization efficiency very low, which is difficult to meet the demand of modern production for high-efficiency formula optimization; at the same time, the quality of formula is unstable, which affects the performance stability of low dielectric glass fiber composition product. SUMMARY

[0003] The present application aims at the technical problem in the prior art that due to the limitation of manual configuration to professional knowledge level and subjective experience judgment, it is difficult to fully and accurately grasp the complex relationship between formula components and dielectric performance, resulting in great limitation of formula optimization, and it is difficult to efficiently develop low dielectric glass fiber composition meeting high performance requirements, and provides a low dielectric glass fiber composition formula optimization method and system to solve the problem.

[0004] The technical scheme of the present application to solve the above technical problem is as follows:

[0005] In a first aspect, the present application provides a low dielectric glass fiber composition formula optimization method, the method comprising: receiving a formula ingredient type list and a formula ingredient constraint percentage list uploaded by a client, performing formula population initialization to obtain an initial formula set; inputting the initial formula set into a dielectric constant prediction model for analysis to obtain a dielectric constant prediction value set, wherein the dielectric constant prediction model is generated by machine learning training using multiple sets of data, each set of data of the multiple sets of data comprising: low dielectric glass fiber composition formula data and a label identifying the dielectric constant of the formula; inputting the initial formula set into a dielectric loss prediction model for analysis to obtain a dielectric loss prediction value set, wherein the dielectric loss prediction model is generated by machine learning training using multiple sets of data, each set of data of the multiple sets of data comprising: low dielectric glass fiber composition formula data and a label identifying the dielectric loss of the formula; based on the dielectric constant prediction value set and the dielectric loss prediction value set, sorting the formula from the initial formula set that is simultaneously less than a dielectric constant threshold and a dielectric loss threshold, and returning to the client.

[0006] In a second aspect, the present application provides a low dielectric glass fiber composition formula optimization system, the system comprising: a formula population initialization module, the formula population initialization module being configured to receive a formula ingredient type list and a formula ingredient constraint percentage list uploaded by a client, perform formula population initialization, and obtain an initial formula set; a dielectric constant prediction module, the dielectric constant prediction module being configured to input the initial formula set into a dielectric constant prediction model for analysis to obtain a dielectric constant prediction value set, wherein the dielectric constant prediction model is generated by machine learning training using multiple sets of data, each set of data of the multiple sets of data comprising: low dielectric glass fiber composition formula data and a label identifying the dielectric constant of the formula; a dielectric loss prediction module, the dielectric loss prediction module being configured to input the initial formula set into a dielectric loss prediction model for analysis to obtain a dielectric loss prediction value set, wherein the dielectric loss prediction model is generated by machine learning training using multiple sets of data, each set of data of the multiple sets of data comprising: low dielectric glass fiber composition formula data and a label identifying the dielectric loss of the formula; a formula screening module, the formula screening module being configured to sort the formula from the initial formula set that is simultaneously less than a dielectric constant threshold and a dielectric loss threshold based on the dielectric constant prediction value set and the dielectric loss prediction value set, and return to the client.

[0007] The beneficial effects of the present application are: by receiving the input of the client, the component type of the formula and the constraint condition of each component are determined, the formula population is initialized, the preliminary formula data is provided for subsequent formula optimization, these data will be used as the basis for model training and optimization, and the starting point of subsequent analysis is ensured to meet the actual production demand. The dielectric constant prediction model trained by machine learning is used to analyze the initial formula, and the dielectric constant prediction value of each formula is obtained. This step objectively and quickly estimates the dielectric performance of each formula in a data-driven manner, thereby providing effective performance indicators for the next step of formula optimization. Similarly, the dielectric loss prediction model trained by machine learning is used to analyze the initial formula, and the dielectric loss prediction value of each formula is obtained, and the performance of the formula on dielectric loss is evaluated. By simultaneously evaluating the performance of the formula on dielectric constant and dielectric loss, it is ensured that the optimization process not only considers a single performance indicator, but also takes into account the overall performance requirement. By setting the threshold values of dielectric constant and dielectric loss, the formula meeting the specific requirements is selected and returned to the client, ensuring that the optimization process not only meets the requirements of dielectric constant, but also ensures that the dielectric loss is within the specified range, achieving a balance of performance indicators and meeting the high performance requirements in actual applications.

[0008] In summary, the present application introduces a machine learning prediction model to efficiently and accurately optimize the low dielectric glass fiber composition formula, realizes the automation and intelligentization of data analysis to formula optimization, significantly improves the efficiency and accuracy of low dielectric glass fiber formula optimization, breaks through the limitations of manual optimization, and achieves the technical effect of efficiently developing a formula meeting high performance requirements in a short time. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 A flowchart of a low dielectric glass fiber composition formula optimization method provided by the present application is shown.

[0010] Figure 2 A structure diagram of a low dielectric glass fiber composition formula optimization system provided by the present application is shown.

[0011] The drawings are described as follows:

[0012] The formula population initialization module 10, the dielectric constant prediction module 20, the dielectric loss prediction module 30, and the formula screening module 40. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0014] In the description of the present application, the terms "first", "second", "third" and the like are used only to describe the purpose and are not to be construed as indicating or implying relative importance or a specific number of the technical features indicated. Therefore, the features defined as "first", "second", "third" and the like can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly specified.

[0015] In the description of the present application, the term "for example" is used to indicate "as an example, illustration, or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that a person skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail in order to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope consistent with the principles and features disclosed.

[0016] Embodiment one:

[0017] As shown in the present application, the embodiment provides a low dielectric glass fiber composition formula optimization method, the method comprises: Figure 1

[0018] Step S100: receiving the formula component type list and the formula component constraint percentage list uploaded by the client, initializing the formula population, and obtaining an initial formula set.

[0019] Specifically, the formula component type list and the formula component constraint percentage list are received from the client. The formula component type list contains all component categories in the low dielectric glass fiber composition; the formula component constraint percentage list defines the constraint percentage range of each component. Then, according to the formula component type and the constraint percentage, the value combination is carried out within the constraint percentage range of each component, and a plurality of formula combinations with different proportions are generated, which constitute the initial formula set. For example, the formula component type list and the formula component constraint percentage list can be:

[0020] SiO2 52％～58％ TiO2 1％～3％ B2O3 22％～25％ ZrO2 1％～2％ Al2O3 10％～12％ SnO2 0.1％～0.5％ ZnO 2％～4％ [R2O] ≤0.2％

[0021] ​The generated initial formula set can be as follows: Formula 1: SiO2= 53%, B2O3= 23%, Al2O3= 11%, ZnO = 3%, TiO2= 2%, ZrO2= 1.5%, SnO2= 0.3%, R2O = 0.1%.

[0022] Formula 2: SiO2= 55%, B2O3= 24%, Al2O3= 10.5%, ZnO = 2.5%, TiO2= 1.5%, ZrO2= 1.2%, SnO2= 0.2%, R2O = 0.15%.

[0023] Formula 3: SiO2= 57%, B2O3= 22.5%, Al2O3= 11.5%, ZnO = 3.5%, TiO2= 2.5%, ZrO2= 1.8%, SnO2= 0.4%, R2O = 0.1%.

[0024] By generating a set of multiple possible initial formulas, subsequent optimization and prediction provide a starting point to avoid missing potential good formulas.

[0025] Step S200: inputting the initial formula set into a dielectric constant prediction model for analysis to obtain a dielectric constant prediction value set, wherein the dielectric constant prediction model is generated by machine learning training using multiple sets of data, each set of data including low dielectric glass fiber composition formula data and a label identifying the dielectric constant of the formula.

[0026] Specifically, the dielectric constant prediction model is a model constructed based on machine learning, which is generated by machine learning training using multiple sets of collected data. The machine learning algorithm used here can be a neural network algorithm, a support vector machine, a random forest algorithm, etc. Each set of training data contains low dielectric glass fiber composition formula data and a label identifying the dielectric constant of the formula. For example, one set of data can be a formula (SiO2= 55%, B2O3= 24%, Al2O3= 10.5%, ZnO = 2.5%, TiO2= 1.5%, ZrO2= 1.2%, SnO2= 0.2%, R2O = 0.15%) corresponding to a dielectric constant label of 3.2. Each formula data in the initial formula set obtained in step S100 is input into the dielectric constant prediction model, and the prediction model predicts the corresponding dielectric constant prediction value of the formula according to the previously learned relationship pattern between the formula data and the dielectric constant.

[0027] The dielectric constant is an important physical quantity for measuring the electrical insulation performance of a material, representing the polarization ability of the material in an electric field. For low dielectric glass fiber compositions, the dielectric constant is a very important performance indicator, directly affecting the electrical performance of the material. By introducing a machine learning model, the dielectric constant of each formula in the initial formula set can be efficiently predicted. Without the need for actual dielectric constant testing, the performance of the formula in terms of dielectric constant can be preliminarily judged, providing an important basis for screening suitable formulas.

[0028] Step S300: inputting the initial formula set into a dielectric loss prediction model for analysis to obtain a dielectric loss prediction value set, wherein the dielectric loss prediction model is generated by machine learning training using multiple sets of data, each set of data including low dielectric glass fiber composition formula data and a label identifying the dielectric loss of the formula.

[0029] Specifically, the dielectric loss prediction model is similar to the dielectric constant prediction model and is also a model trained by a machine learning algorithm. It is trained according to multiple sets of data containing low dielectric glass fiber composition formula data and a label identifying the dielectric loss of the formula. For example, one set of data is a formula (SiO2 = 57%, B2O3 = 22.5%, Al2O3 = 11.5%, ZnO = 3.5%, TiO2 = 2.5%, ZrO2 = 1.8%, SnO2 = 0.4%, R2O = 0.1%) corresponding to a dielectric loss label of 0.004. Each formula data in the initial formula set is input into the dielectric loss prediction model, and the prediction model predicts the corresponding dielectric loss prediction value of each formula according to the learned relationship pattern between the formula data and the dielectric loss.

[0030] Dielectric loss refers to the energy dissipation part of a material under the action of an electric field. Dielectric loss is another important performance indicator of low dielectric glass fiber compositions. The lower the dielectric loss, the better the material performance. Through the dielectric loss prediction model, the dielectric loss of each formula in the initial formula set can be predicted without actual testing, providing data on the key performance indicator of dielectric loss for comprehensive evaluation of formula performance, which helps to screen formulas that meet the dielectric loss requirements.

[0031] Step S400: based on the dielectric constant prediction value set and the dielectric loss prediction value set, sorting the formulas from the initial formula set that are both less than the dielectric constant threshold and the dielectric loss threshold, and returning to the client.

[0032] Specifically, the dielectric constant threshold is a pre-set upper limit of a dielectric constant. For example, in the production of certain electronic devices, the dielectric constant threshold is set to 3.0, so only the formula whose dielectric constant prediction value is less than 3.0 can be selected. The dielectric loss threshold is a pre-set upper limit of dielectric loss. For example, in the manufacture of certain communication devices, the dielectric loss threshold is set to 0.003, and only the formula whose dielectric loss prediction value is less than 0.003 meets the requirements. First, the dielectric constant prediction value set obtained in step S200 and the dielectric loss prediction value set obtained in step S300 are obtained. Then, for each formula in the initial formula set, it is checked whether the dielectric constant prediction value is less than the dielectric constant threshold and whether the dielectric loss prediction value is less than the dielectric loss threshold. For example, for a formula, if its dielectric constant prediction value is 2.8 (less than the dielectric constant threshold 3.0) and its dielectric loss prediction value is 0.002 (less than the dielectric loss threshold 0.003), then the formula meets the conditions. Finally, the formulas that meet the conditions are screened out and returned to the client.

[0033] By comprehensively considering the two key performance indicators of dielectric constant and dielectric loss, the formula that meets the requirements is screened out and returned to the client, ensuring that the formula obtained by the client reaches a good level in low dielectric performance and meets the requirements of actual production on the performance of low dielectric glass fiber composition, improving the quality and performance of the product.

[0034] Further, step S100 includes:

[0035] Step S110: based on the formula ingredient type list and the formula ingredient constraint percentage list, constructing a population initialization constraint condition: wherein F(x) represents the formula initialization fitness, x i represents the assignment percentage of the i-th formula ingredient type, l i represents the upper limit of the constraint percentage of the i-th formula ingredient type, u i represents the lower limit of the constraint percentage of the i-th formula ingredient type, N represents the total number of formula ingredient types, w i represents the weight of the i-th formula ingredient type, w total represents the constraint percentage fitness weight.

[0036] Step S120: receiving the micro percentage threshold uploaded by the client; wherein when the assignment percentage of the i-th formula ingredient type is less than or equal to the micro percentage threshold, the weight of the i-th formula ingredient type belongs to 500-1000, and when the assignment percentage of the i-th formula ingredient type is greater than the micro percentage threshold, the weight of the i-th formula ingredient type belongs to 10-50; wherein the constraint percentage fitness weight is equal to 1000.

[0037] Step S130: according to the population initialization constraint condition, sorting the initial formula set with a preset number of fitness threshold values.

[0038] Specifically, according to the formula ingredient type list and the constraint percentage list, the population initialization constraint condition is constructed as above. The constraint condition is an initialization fitness function, which can quantitatively evaluate the compliance between each possible formula combination and the formula ingredient constraint condition, and provide a clear numerical basis for subsequent screening of the initial formula set that meets the requirements, so as to find a more reasonable initial formula.

[0039] In determining the initial formula set, first, according to the formula ingredient type list and the formula ingredient constraint percentage list uploaded by the client, the constraint percentage upper limit l i and the constraint percentage lower limit u i of each formula ingredient type are determined. i Then, for N formula ingredient types, different percentage values x

[0040] The micro percentage threshold uploaded by the client is received, which is a set limit value for distinguishing the weight value range of the formula ingredient type. In calculating the weight of the formula ingredient type, according to the size relationship between the assigned percentage x i of the i-th formula ingredient type and the micro percentage threshold, the value range of the weight w i corresponding to the formula ingredient type is determined. If the assigned percentage of a formula ingredient type is less than or equal to the micro percentage threshold, its weight will be between 500 and 1000; if the assigned percentage of a formula ingredient type is greater than the threshold, the weight will be between 10 and 50. By introducing the micro percentage threshold to distinguish the weight value range, the importance of different formula ingredient types can be more reasonably considered in calculating the formula initialization fitness. For ingredients with smaller assigned percentages (which may be trace components but have important influence on performance), higher weights are given, so that the rationality of the formula can be more accurately evaluated.

[0041] The percentage value x i of each formula ingredient type in the above obtained multiple possible formulas, the constraint percentage upper limit l i , the constraint percentage lower limit u i , and the weight w iThe constraint condition formula is substituted to calculate the initial fitness F(x) of each formula. Then, from all possible formulas, a preset number of formulas with fitness less than or equal to the fitness threshold are sorted to form an initial formula set. The preset number is the total number of formulas in the initial formula set set in advance. For example, if the preset number is 10 and the fitness threshold is 100, 10 formulas with fitness less than or equal to 100 are selected to form the initial formula set. By setting the fitness threshold, the initial formula set that is relatively more reasonable in terms of formula ingredient proportion can be screened out, and the formulas in this set all meet the formula ingredient constraint condition, providing an optimized basis for further analysis of the formula (prediction of dielectric constant and dielectric loss).

[0042] Further, step S200 includes:

[0043] Step S210: obtaining a first initial formula, wherein the first initial formula includes a first component type set and a first component percentage set.

[0044] Step S220: counting the number of component types in the first component percentage set greater than 0.

[0045] Step S230: based on the number of component types, obtaining the dielectric constant prediction model by integrating the dielectric constant prediction unit configuration, wherein the number of integrated dielectric constant prediction units is greater than twice the number of component types, and the dielectric constant prediction unit is generated by machine learning training using multiple sets of data, each set of data including low dielectric glass fiber composition formula data and a label indicating the dielectric constant of the formula.

[0046] Step S240: inputting the first component type set and the first component percentage set into the dielectric constant prediction model to output a first formula dielectric constant prediction value and add it to the dielectric constant prediction value set.

[0047] Specifically, when analyzing the initial formula set using the dielectric constant prediction model, the first formula is extracted from the initial formula set, denoted as the first initial formula. The selection process can be random extraction or extraction according to the storage order of the formulas in the initial formula set, etc. The first initial formula includes a first component type set (the types of components selected in the first initial formula) and a first component percentage set (the percentages of components in the first initial formula). For example, the first component type set in the first initial formula is {SiO2, B2O3, Al2O3}, and the corresponding first component percentage set is {55%, 23%, 12%}.

[0048] For the first initial formula, the first component percentage set is checked one by one, and the number of component types greater than 0 is counted. For example, the first component percentage set is {55%, 23%, 12%}, and here the three component percentages are all greater than 0, so the number of component types counted is 3.

[0049] The dielectric constant prediction model is a model integrating multiple dielectric constant prediction units, each of which is an independent machine learning algorithm model trained based on multiple sets of data (including low dielectric glass fiber composition formula data and labels identifying the dielectric constant of the formula) and can predict the corresponding dielectric constant according to the input initial formula data. The number of dielectric constant prediction units in the dielectric constant prediction model is not fixed, but is dynamically adjusted according to the input formula data. According to the number of component types counted, the number of dielectric constant prediction units in the dielectric constant prediction model is configured. The configuration rule is that the number of dielectric constant prediction units integrated is greater than twice the number of component types with component percentages greater than 0. According to the number of dielectric constant prediction units determined, dielectric constant prediction units are selected and combined together to form a dielectric constant prediction model. For example, if the number of component types is 3, the number of dielectric constant prediction units integrated can be set to 7 (greater than 2x3), and 7 dielectric constant prediction units that have been trained are combined to form a dielectric constant prediction model. By reasonably configuring the number of dielectric constant prediction units integrated according to the number of component types, a dielectric constant prediction model suitable for the current first initial formula can be constructed. Such a model has higher accuracy and adaptability in predicting dielectric constant.

[0050] The first component type set and the first component percentage set are input into the dielectric constant prediction model constructed above, and the model uses multiple prediction units inside to analyze the first component type set and the first component percentage set, and outputs the first formula dielectric constant prediction value. Then, this prediction value is added to the dielectric constant prediction value set. The same method is used to predict the dielectric constant of the remaining initial formulas in the initial formula set, and the dielectric constant prediction value set is gradually constructed, providing a data basis for subsequent evaluation of the entire initial formula set.

[0051] Further, the dielectric constant prediction unit is obtained, including:

[0052] Step A: According to the pre-set low dielectric glass fiber composition component type attribute set, the low dielectric glass fiber composition formula data and the label identifying the dielectric constant of the formula are collected.

[0053] Step B: With the label of the identified formulation dielectric constant as supervision and the formulation data of the low dielectric glass fiber composition as input, an original training dataset is constructed to generate a first dielectric constant prediction unit by machine learning training, wherein the first dielectric constant prediction unit has a first loss training dataset with an output accuracy less than or equal to an output accuracy threshold.

[0054] Step C: When the data amount of the first loss training dataset is less than or equal to a data amount threshold, the first dielectric constant prediction unit is set as the dielectric constant prediction unit.

[0055] Specifically, the pre-set low dielectric glass fiber composition ingredient type attribute set is a pre-set set containing various attributes related to ingredient types in low dielectric glass fiber compositions. For example, attributes such as chemical properties and physical properties of ingredients, which are helpful to determine which formulation data needs to be collected. According to the pre-set low dielectric glass fiber composition ingredient type attribute set, the corresponding low dielectric glass fiber composition formulation data and the label of the identified formulation dielectric constant are collected.

[0056] With the collected label of the identified formulation dielectric constant as a supervision signal and the low dielectric glass fiber composition formulation data as input data, an original training dataset is constructed. Then, a machine learning algorithm (such as a neural network algorithm, a support vector machine, etc.) is used to train this original training dataset to obtain a first dielectric constant prediction unit. During the training process, the difference (loss) between the predicted result and the label is constantly calculated through a loss function (such as a mean square error loss function, a mean absolute error loss function, etc.), and the parameters of the prediction unit are adjusted according to this loss. For example, a gradient descent algorithm is used to minimize the loss function, and when the training reaches a certain convergence condition (such as the loss no longer obviously decreases), the first dielectric constant prediction unit is obtained, and a first loss training dataset is also obtained. This first loss training dataset is a dataset generated during the training of the first dielectric constant prediction unit, containing training data with an output accuracy less than or equal to a pre-set output accuracy threshold during the training process, reflecting the error of the first dielectric constant prediction unit during the training process.

[0057] The data amount of the first loss training data set is compared with a data amount threshold. The data amount threshold is a preset maximum data amount of the loss data set, which is used to determine whether the model needs to be further optimized. If the data amount of the first loss training data set is less than or equal to the threshold, it is considered that the accuracy of the first dielectric constant prediction unit meets the requirements, and the first dielectric constant prediction unit can be set as the dielectric constant prediction unit. By evaluating the output accuracy of the model and the data amount of the loss data set, it is ensured that the generated dielectric constant prediction unit has high accuracy and stability, thereby effectively improving the accuracy of dielectric constant prediction and providing a reliable basis for subsequent formula optimization.

[0058] Further, obtaining the dielectric constant prediction unit further includes:

[0059] Step D: When the data amount of the first loss training data set is greater than the data amount threshold, increasing the loss calculation weight of the first loss training data set in the original training data set, and training a second dielectric constant prediction unit.

[0060] Step E: Until the data amount of the Mth loss training data set is less than or equal to the data amount threshold, the first dielectric constant prediction unit, the second dielectric constant prediction unit, and the Mth dielectric constant prediction unit are outputted and integrated to obtain the dielectric constant prediction unit.

[0061] Specifically, when the data amount of the first loss training data set is greater than the preset data amount threshold, it indicates that the prediction accuracy on these data needs to be further optimized in the model training process. Therefore, according to the training condition, the weight of the first loss training data set in the overall loss calculation is increased, and more attention is paid to these data points in the subsequent training process, thereby enhancing the prediction ability of these data points. Then, based on the updated loss weight and the original training data set, the second dielectric constant prediction unit is trained by a machine learning algorithm, and the training process is similar to that of the first dielectric constant prediction unit. By increasing the loss calculation weight of the first loss training data set, the second dielectric constant prediction unit can pay more attention to the error situation embodied by the first loss training data set in the training process, thereby improving the accuracy of the second dielectric constant prediction unit and making it more suitable.

[0062] During the model training process, the plurality of dielectric constant prediction units (from the second dielectric constant prediction unit to the Mth dielectric constant prediction unit) are continuously adjusted and trained. Each time of training checks whether the data amount of the corresponding loss training data set is less than or equal to the data amount threshold. When the data amount of the Mth loss training data set is less than or equal to the data amount threshold, the training is stopped. Then, the output results of these trained dielectric constant prediction units (the first dielectric constant prediction unit, the second dielectric constant prediction unit to the Mth dielectric constant prediction unit) are mean integrated to obtain the dielectric constant prediction unit.

[0063] For example, the data amount of the first loss training data set is 120, which is greater than the set data amount threshold of 100. The loss calculation weight of the first loss training data set is adjusted from 1 to 2. The model is retrained using the data set after the weight adjustment to generate the second dielectric constant prediction unit. The data amount of the second loss training data set is 110, which is still greater than 100. The weight continues to be adjusted and trained until the data amount of the loss data set generated by the Mth training is less than or equal to 100. The output results of the first, second to Mth dielectric constant prediction units are mean integrated to generate the final dielectric constant prediction unit. For example, M = 3, that is, there are three prediction units, and the dielectric constant prediction values of these prediction units for a certain formula are 3.2, 3.3 and 3.1 respectively. Then, the dielectric constant prediction value obtained by the output mean integration is (3.2 + 3.3 + 3.1) / 3 = 3.2.

[0064] By increasing the weight of the key training data set, the model can pay more attention to the data that is more important or more difficult to fit in the training, thereby improving the prediction ability and accuracy of the model for these key data and reducing the risk of overfitting. By output mean integrating a plurality of dielectric constant prediction units, the advantages of multiple prediction units can be integrated to improve the accuracy and stability of the final obtained dielectric constant prediction unit. This integration method can reduce the errors that may exist in a single prediction unit to some extent, thereby obtaining more reliable dielectric constant prediction results.

[0065] Further, the method described in the embodiments of the present application further comprises:

[0066] Step S510: when the number of formulas less than the dielectric constant threshold and the dielectric loss threshold is equal to 0 at the same time, a fitness function for optimization is constructed: wherein f(x) represents the fitness of any one formula, k represents the iteration number, ε(x) represents the dielectric constant prediction value, ε0 represents the dielectric constant threshold, δ(x) represents the dielectric loss prediction value, δ0 represents the dielectric loss threshold, w ε represents the dielectric loss fitness weight, w δ represents the dielectric loss fitness weight.

[0067] Step S520: According to the optimization fitness function, perform iterative optimization on the initial formula set, extract the candidate formula set whose output value of the optimization fitness function is less than or equal to the optimization fitness threshold value.

[0068] Step S530: Sort the formulas from the candidate formula set that are simultaneously less than the dielectric constant threshold value and the dielectric loss threshold value, and return to the client.

[0069] Specifically, when there is no formula whose dielectric constant and dielectric loss are simultaneously less than the respective threshold values (i.e., the number of formulas that are simultaneously less than the dielectric constant threshold value and the dielectric loss threshold value is equal to 0) in the formula screening process, an optimization algorithm is guided by constructing an optimization fitness function. This optimization fitness function takes into account the difference between the predicted values of the dielectric constant and the dielectric loss of the formula and the corresponding threshold values, and calculates these data differences by weighting with a natural exponential function to produce an fitness value for each formula. If the dielectric constant and the dielectric loss of the formula are close to their respective threshold values, the corresponding fitness value will be lower, and vice versa.

[0070] According to the constructed optimization fitness function, the fitness value of each formula in the initial formula set is calculated. Then, using an optimization algorithm (such as a genetic algorithm, a particle swarm optimization algorithm, etc.), the formulas in the initial formula set are iteratively optimized according to their fitness values until the iteration number k is met. The fitness values of the formulas after iteration are calculated and compared with the optimization fitness threshold value, and those whose fitness values are less than or equal to the optimization fitness threshold value are selected to form the candidate formula set. The optimization fitness threshold value is a pre-set standard used to determine whether the formula has met the optimization goal. If the fitness value of the formula is less than or equal to this threshold value, it means that it meets the optimization goal.

[0071] For each formula in the obtained candidate formula set, check whether its predicted value of dielectric constant is less than the dielectric constant threshold value and whether its predicted value of dielectric loss is less than the dielectric loss threshold value. Then, sort out the formulas that simultaneously satisfy these two conditions and return them to the client.

[0072] By introducing the optimization fitness function, the optimization direction can be adjusted when there is a lack of formulas that meet the requirements. This function will motivate the optimization algorithm to search for formulas that are closer to the target values of dielectric constant and dielectric loss, avoiding the limitations brought by relying solely on preliminary prediction results. By further screening the candidate formula set, the final formula that simultaneously meets the threshold values of dielectric constant and dielectric loss is obtained and returned to the client, which can ensure that the formula obtained by the client meets the expected requirements in terms of dielectric performance and meets the needs of actual applications (such as electronic device manufacturing) for low-dielectric glass fiber compositions.

[0073] Further, step S520 comprises:

[0074] Step S521: mutating the initial recipe set to obtain a first expanded recipe set.

[0075] Step S522: analyzing a first minimum fitness of the first expanded recipe set based on the fitness function.

[0076] Step S523: analyzing a second minimum fitness of the initial recipe set based on the fitness function.

[0077] Step S524: when the first minimum fitness is greater than or equal to the second minimum fitness, performing a second iteration based on the initial recipe set.

[0078] Step S525: when the first minimum fitness is less than the second minimum fitness, integrating a predetermined number of better solutions of the first expanded recipe set and the initial recipe set to obtain an updated recipe set, and performing a second iteration based on the updated recipe set.

[0079] Specifically, each recipe in the initial recipe set is randomly mutated, i.e., the percentage of ingredients is randomly adjusted within the constraint range to generate a new recipe, forming a first expanded recipe set. Through the mutation operation, the diversity of the recipe is increased, the search space is expanded, and it is helpful to find a better recipe. The newly generated first expanded recipe set may contain some recipes that are better in dielectric performance, providing more possibilities for the subsequent optimization process.

[0080] For example, the initial recipe set is:

[0081] Recipe 1: SiO2=53%, B2O3=23%, Al2O3=11%, ZnO=3%, TiO2=2%, ZrO2=1.5%, SnO2=0.3%, R2O=0.1%.

[0082] Recipe 2: SiO2=55%, B2O3=24%, Al2O3=10.5%, ZnO=2.5%, TiO2=1.5%, ZrO2=1.2%, SnO2=0.2%, R2O=0.15%.

[0083] Recipe 3: SiO2=57%, B2O3=22.5%, Al2O3=11.5%, ZnO=3.5%, TiO2=2.5%, ZrO2=1.8%, SnO2=0.4%, R2O=0.1%.

[0084] Mutate the initial recipe set to generate a first expanded recipe set:

[0085] Formula 4: SiO2= 54%, B2O3= 22.5%, Al2O3= 11.5%, ZnO = 3.5%, TiO2= 2.5%, ZrO2= 1.8%, SnO2= 0.4%, R2O = 0.1%.

[0086] Formula 5: SiO2= 56%, B2O3= 23.5%, Al2O3= 10%, ZnO = 2%, TiO2= 2.5%, ZrO2= 1%, SnO2= 0.2%, R2O = 0.1%.

[0087] The fitness of each formula in the first expanded formula set is calculated using the fitness optimization function, and the minimum value is extracted, denoted as the first minimum fitness.

[0088] The fitness of each formula in the initial formula set is calculated using the fitness optimization function, and the minimum value is extracted from the initial formula set, denoted as the second minimum fitness.

[0089] The first minimum fitness and the second minimum fitness are compared, and if the first minimum fitness is greater than or equal to the second minimum fitness, it means that the expanded formula set does not bring optimization effect, so the second iteration is continued based on the initial formula set.

[0090] When the first minimum fitness is less than the second minimum fitness, it means that the first expanded formula set is better in terms of fitness optimization. At this time, a preset number of optimal solutions are selected from the first expanded formula set and the initial formula set according to the fitness value. The preset number of optimal solutions is a certain number of formulas with better fitness selected from the first expanded formula set and the initial formula set. For example, 10 formulas with the smallest fitness can be selected as the optimal solutions. Then, the optimal solutions are integrated to form an updated formula set. Finally, the second iteration is performed based on the updated formula set, including mutation again, recalculation of fitness, etc. By integrating the optimal solutions to form an updated formula set and performing the second iteration based on it, the optimization can be continued on the basis of a more advantageous formula set, and the efficiency of finding a formula that meets the dielectric constant and dielectric loss threshold requirements at the same time can be improved.

[0091] The low dielectric glass fiber composition formula optimization method provided by the embodiment of the application has at least the following technical effects:

[0092] The embodiment of the present application provides the initial formula set by receiving the component information and constraint conditions provided by the client, provides basic data for the whole optimization process, and ensures that the starting point of optimization meets the actual production demand. Subsequently, the dielectric constant prediction model and the dielectric loss prediction model generated by machine learning training are used for prediction and analysis of the initial formula set respectively, and the prediction value set of the dielectric constant and the dielectric loss is obtained, so that the performance of different formulas is quickly and objectively evaluated. Finally, by setting the threshold conditions of the dielectric constant and the dielectric loss, the formula meeting the threshold requirements of the dielectric constant and the dielectric loss is selected from the initial formula set, and the inefficient process of manual trial and error and multiple experiments is avoided.

[0093] Overall, the embodiment of the present application realizes automation and intelligentization from data analysis to formula optimization, significantly improves the efficiency and precision of low-dielectric glass fiber formula optimization, breaks through the limitation of manual optimization, and achieves the technical effect of efficiently developing a formula meeting high-performance requirements in a short time, thereby providing higher-quality and more stable material selection for the field of high-frequency communication.

[0094] Embodiment two:

[0095] As shown in Figure 2 based on the same inventive concept of the low-dielectric glass fiber composition formula optimization method provided in embodiment one, the embodiment of the present application further provides a low-dielectric glass fiber composition formula optimization system, which comprises:

[0096] The formula population initialization module 10 is used for receiving the formula component type list and the formula component constraint percentage list uploaded by the client, performing formula population initialization, and obtaining an initial formula set.

[0097] The dielectric constant prediction module 20 is used for inputting the initial formula set into a dielectric constant prediction model for analysis to obtain a dielectric constant prediction value set, wherein the dielectric constant prediction model is generated by machine learning training using multiple sets of data, and each set of data of the multiple sets of data comprises low-dielectric glass fiber composition formula data and a label identifying the dielectric constant of the formula.

[0098] The dielectric loss prediction module 30 is used for inputting the initial formula set into a dielectric loss prediction model for analysis to obtain a dielectric loss prediction value set, wherein the dielectric loss prediction model is generated by machine learning training using multiple sets of data, and each set of data of the multiple sets of data comprises low-dielectric glass fiber composition formula data and a label identifying the dielectric loss of the formula.

[0099] A recipe screening module 40 for sorting recipes from the initial recipe set that are both less than a dielectric constant threshold and a dielectric loss threshold based on the set of dielectric constant prediction values and the set of dielectric loss prediction values, back to the client.

[0100] Further, the recipe population initialization module 10 of the embodiment of the present application is further configured to perform the following steps:

[0101] Based on the recipe ingredient type list and the recipe ingredient constraint percentage list, a population initialization constraint condition is constructed. Wherein, F(x) represents the recipe initialization fitness, x i represents the assigned percentage of the i-th recipe ingredient type, l i represents the upper limit of the constraint percentage of the i-th recipe ingredient type, u i represents the lower limit of the constraint percentage of the i-th recipe ingredient type, N represents the total number of recipe ingredient types, w i represents the weight of the i-th recipe ingredient type, w total represents the constraint percentage fitness weight.

[0102] A trace percentage threshold is received uploaded by the client; wherein when the assigned percentage of the i-th recipe ingredient type is less than or equal to the trace percentage threshold, the weight of the i-th recipe ingredient type belongs to 500-1000, and when the assigned percentage of the i-th recipe ingredient type is greater than the trace percentage threshold, the weight of the i-th recipe ingredient type belongs to 10-50; wherein the constraint percentage fitness weight is equal to 1000; based on the population initialization constraint condition, the initial recipe set of a preset number less than or equal to the fitness threshold is sorted.

[0103] Further, the dielectric constant prediction module 20 of the embodiment of the present application is further configured to perform the following steps:

[0104] A first initial recipe is obtained, wherein the first initial recipe includes a first ingredient type set and a first ingredient percentage set; the number of ingredient types with a first ingredient percentage greater than 0 in the first ingredient percentage set is counted; based on the number of ingredient types, a dielectric constant prediction unit integration configuration is obtained, and the dielectric constant prediction model is obtained, wherein the number of dielectric constant prediction unit integrations is greater than 2 times the number of ingredient types, the dielectric constant prediction unit is generated by machine learning training using multiple sets of data, and each set of data of the multiple sets of data includes low dielectric glass fiber composition recipe data and a label identifying the dielectric constant of the recipe; the first ingredient type set and the first ingredient percentage set are input into the dielectric constant prediction model, and a first recipe dielectric constant prediction value is output, which is added to the set of dielectric constant prediction values.

[0105] Further, the system of the embodiment of the present application is also used to execute the following steps:

[0106] According to the preset low dielectric glass fiber composition component type attribute set, the low dielectric glass fiber composition formula data and the label identifying the formula dielectric constant are collected; the label identifying the formula dielectric constant is taken as supervision, the low dielectric glass fiber composition formula data is taken as input, and a first dielectric constant prediction unit is generated by machine learning training of a raw training data set, wherein the first dielectric constant prediction unit has a first loss training data set with an output accuracy less than or equal to an output accuracy threshold; when the data amount of the first loss training data set is less than or equal to a data amount threshold, the first dielectric constant prediction unit is set as the dielectric constant prediction unit.

[0107] Further, the system of the embodiment of the present application is also used to execute the following steps:

[0108] When the data amount of the first loss training data set is greater than the data amount threshold, the loss calculation weight of the first loss training data set in the raw training data set is increased, and a second dielectric constant prediction unit is trained; until when the data amount of the Mth loss training data set is less than or equal to the data amount threshold, the first dielectric constant prediction unit, the second dielectric constant prediction unit, and the Mth dielectric constant prediction unit are integrated for output mean value set to obtain the dielectric constant prediction unit.

[0109] Further, the system of the embodiment of the present application is also used to execute the following steps:

[0110] When the number of formulas less than the dielectric constant threshold and the dielectric loss threshold at the same time is equal to 0, an optimization fitness function is constructed: Wherein, f(x) represents the fitness of any one formula, k represents the iteration number, ε(x) represents the dielectric constant prediction value, ε0 represents the dielectric constant threshold, δ(x) represents the dielectric loss prediction value, δ0 represents the dielectric loss threshold, w ε represents the dielectric loss fitness weight, w δ represents the dielectric loss fitness weight; according to the optimization fitness function, iteration optimization is performed on the initial formula set, and a candidate formula set with an output value of the optimization fitness function less than or equal to an optimization fitness threshold is extracted; from the candidate formula set, formulas less than the dielectric constant threshold and the dielectric loss threshold at the same time are sorted, and returned to the client.

[0111] Further, the system of the embodiment of the present application is also used to execute the following steps:

[0112] The initial formula set is mutated to obtain a first expanded formula set; a first minimum fitness of the first expanded formula set is analyzed based on the optimization fitness function; a second minimum fitness of the initial formula set is analyzed based on the optimization fitness function; when the first minimum fitness is greater than or equal to the second minimum fitness, second iteration is performed based on the initial formula set; when the first minimum fitness is less than the second minimum fitness, a preset number of optimal solutions of the first expanded formula set and the initial formula set are integrated to obtain an updated formula set, and second iteration is performed based on the updated formula set.

[0113] It should be noted that in the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0114] Although the preferred embodiments of the present application have been described, those skilled in the art who have the basic inventive concept can make further changes and modifications to these embodiments.

[0115] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for optimizing the formulation of a low-dielectric glass fiber composition, characterized in that, include: Receive the list of formula ingredient types and the list of formula ingredient constraint percentages uploaded by the client, initialize the formula population, and obtain the initial formula set; The initial set of formulations is input into the dielectric constant prediction model for analysis to obtain a set of predicted dielectric constant values. The dielectric constant prediction model is generated by training multiple sets of data through machine learning. Each set of data includes: low dielectric glass fiber composition formulation data and a label identifying the dielectric constant of the formulation. The initial set of formulations is input into the dielectric loss prediction model for analysis to obtain a set of dielectric loss prediction values. The dielectric loss prediction model is generated by training multiple sets of data through machine learning. Each set of data includes: low dielectric glass fiber composition formulation data and a label identifying the dielectric loss of the formulation. Based on the set of predicted dielectric constant values ​​and the set of predicted dielectric loss values, recipes that are simultaneously less than the dielectric constant threshold and the dielectric loss threshold are selected from the initial recipe set and returned to the client.

2. The method as described in claim 1, characterized in that, Receive the list of formula ingredient types and the list of formula ingredient constraint percentages uploaded by the client, initialize the formula population, and obtain an initial formula set, including: Based on the list of formulation ingredient types and the list of formulation ingredient constraint percentages, construct the population initialization constraints: Where F(x) represents the fitness of the formulation initialization, x i The percentage of the assigned value representing the type of ingredient in the i-th formulation, l i The upper limit of the constrained percentage representing the type of ingredient in the i-th formulation, u i The lower limit of the constraint percentage representing the i-th type of formulation ingredient, N represents the total number of formulation ingredient types, w i The weight representing the type of ingredient in the i-th formulation, w total Characterize the percentage of constraint-adaptive weights; The threshold for receiving minute percentages uploaded by the client; Wherein, when the assigned percentage of the i-th formulation component type is less than or equal to the trace percentage threshold, the weight of the i-th formulation component type is between 500 and 1000; when the assigned percentage of the i-th formulation component type is greater than the trace percentage threshold, the weight of the i-th formulation component type is between 10 and 50. Wherein, the constraint percentage adaptation weight is equal to 1000; Based on the population initialization constraints, the initial recipe set with a preset number less than or equal to the fitness threshold is selected.

3. The method as described in claim 1, characterized in that, The initial formula set is input into the dielectric constant prediction model for analysis to obtain a set of predicted dielectric constant values, including: Obtain a first initial formula, wherein the first initial formula includes a first set of ingredient types and a first set of ingredient percentages; Count the number of component types with a percentage greater than 0 in the first component set; Based on the number of component types, the dielectric constant prediction unit is integrated and configured to obtain the dielectric constant prediction model. The number of dielectric constant prediction units integrated is greater than twice the number of component types. The dielectric constant prediction unit is generated by machine learning training using multiple sets of data. Each set of data includes: low dielectric glass fiber composition formulation data and a label identifying the dielectric constant of the formulation. The first set of component types and the first set of component percentages are input into the dielectric constant prediction model, and the first formulation dielectric constant prediction value is output and added to the dielectric constant prediction value set.

4. The method as described in claim 3, characterized in that, The dielectric constant prediction unit is generated through machine learning training using multiple sets of data. Each set of data includes: low-dielectric glass fiber composition formulation data and labels identifying the dielectric constant of the formulation, including: Based on a preset set of low-dielectric glass fiber composition component type attributes, collect the low-dielectric glass fiber composition formulation data and the label identifying the dielectric constant of the formulation; Using the label identifying the dielectric constant of the formulation as supervision and the low dielectric glass fiber composition formulation data as input, an original training dataset is constructed and a first dielectric constant prediction unit is generated through machine learning training. The first dielectric constant prediction unit has a first loss training dataset with an output accuracy less than or equal to an output accuracy threshold. When the amount of data in the first loss training dataset is less than or equal to the data amount threshold, the first dielectric constant prediction unit is set as the dielectric constant prediction unit.

5. The method as described in claim 4, characterized in that, Also includes: When the amount of data in the first loss training dataset is greater than the data amount threshold, the loss calculation weight of the first loss training dataset in the original training dataset is increased, and the second dielectric constant prediction unit is trained. Until the amount of data in the Mth loss training dataset is less than or equal to the data amount threshold, the output mean of the first dielectric constant prediction unit, the second dielectric constant prediction unit, and up to the Mth dielectric constant prediction unit is integrated to obtain the dielectric constant prediction unit.

6. The method as described in claim 1, characterized in that, Also includes: When the number of recipes that are simultaneously less than the dielectric constant threshold and the dielectric loss threshold is equal to 0, construct the fitness function for optimization: Where f(c) represents the fitness of any formulation, k represents the number of iterations, ε(x) represents the predicted dielectric constant, ε0 represents the dielectric constant threshold, δ(x) represents the predicted dielectric loss, δ0 represents the dielectric loss threshold, and w ε The fitness weights representing dielectric loss, w δ Weights characterizing dielectric loss fitness; According to the optimization fitness function, iterative optimization is performed on the initial formula set to extract the candidate formula set whose output value of the optimization fitness function is less than or equal to the optimization fitness threshold; Formulas that are simultaneously less than the dielectric constant threshold and dielectric loss threshold are selected from the set of candidate formulas and returned to the client.

7. The method as described in claim 6, characterized in that, Based on the optimization fitness function, iterative optimization is performed on the initial formula set to extract a set of candidate formulas whose output value of the optimization fitness function is less than or equal to the optimization fitness threshold, including: The initial recipe set is mutated to obtain an expanded recipe set; The first minimum fitness of the first expansion recipe set is analyzed based on the optimization fitness function. The second minimum fitness of the initial recipe set is analyzed based on the optimization fitness function. When the first minimum fitness is greater than or equal to the second minimum fitness, a second iteration is performed based on the initial recipe set; When the first minimum fitness is less than the second minimum fitness, the optimal solution of the first expansion formula set and the initial formula set is integrated to obtain an updated formula set, and a second iteration is performed based on the updated formula set.

8. A system for optimizing the formulation of low-dielectric glass fiber compositions, characterized in that, The system is used to perform a method for optimizing the formulation of a low-dielectric glass fiber composition according to any one of claims 1-7, comprising: The formula population initialization module is used to receive the list of formula component types and the list of formula component constraint percentages uploaded by the client, perform formula population initialization, and obtain an initial formula set. The dielectric constant prediction module is used to input the initial formula set into the dielectric constant prediction model for analysis to obtain a set of predicted dielectric constant values. The dielectric constant prediction model is generated by training multiple sets of data through machine learning. Each set of data includes: low dielectric glass fiber composition formula data and a label identifying the dielectric constant of the formula. The dielectric loss prediction module is used to input the initial formula set into the dielectric loss prediction model for analysis to obtain a set of dielectric loss prediction values. The dielectric loss prediction model is generated by training multiple sets of data through machine learning. Each set of data includes: low dielectric glass fiber composition formula data and a label identifying the dielectric loss of the formula. The recipe screening module is used to select recipes from the initial recipe set that are simultaneously less than the dielectric constant threshold and the dielectric loss threshold based on the dielectric constant prediction value set and the dielectric loss prediction value set, and return them to the client.

Images