Formulation building system, method, readable storage medium and computer program product

Abstract

The application provides a recipe construction system, method, readable storage medium and computer program product. A dimension reduction module obtains dimension reduction ingredient information according to ingredient information of each historical recipe information and a dimension reduction algorithm; a neural network module obtains a plurality of trained neural network parameters according to the ingredient information and the dimension reduction ingredient information; the neural network module obtains dimension reduction initial ingredient information according to the plurality of trained neural network parameters and initial ingredient information; a search module searches for a plurality of candidate recipe information with a first number from the plurality of historical recipe information; a judgment module judges whether the property information of each candidate recipe information meets the specification, and outputs a recipe solving information meeting the specification.

Description

[0001] This application claims priority to Taiwan Patent Application No. 110134295, filed on September 14, 2021, entitled "Formula Construction System, Formulation Construction Method, Computer-Readable Recording Medium with Internal Program and Non-Transitional Computer Program Product", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of formulation construction, and in particular to a technique for using neural networks to integrate empirical knowledge to construct formulations. Background Technology

[0003] The dose-effect relationship is a form of relationship between input and output variables, exhibiting the same form across different processes. Examples include the relationship between the chemical raw material combination (input) in wire and cable formulation design and the physical properties of the finished coating material (output); the relationship between the controllable machine parameters (input) in the stretching process of soft copper wire in wire and cable and the physical properties of the wire (output); the relationship between the controllable machine parameters (input) in the cold / hot rolling steel process and the quality characteristics of the finished product (output); and the relationship between the controllable parameters (input) in the pickling process of stainless steel finished products (plates, bars, or wire rods) and the weight consumption of the finished product (output). However, traditionally, constructing such dose-effect relationships in practical applications often relies primarily on trial and error based on human experience, failing to effectively utilize the knowledge generated and accumulated from past successful experiences. Summary of the Invention

[0004] In view of the above, some embodiments of this application provide a recipe construction system, method, readable storage medium, and computer program product to improve the problems of the prior art.

[0005] One embodiment of this application provides a formula construction system, which includes a dimensionality reduction module, a neural network module, a search module, and a judgment module. The dimensionality reduction module is configured to receive multiple historical formula information and obtain dimensionality-reduced component information of each historical formula information based on the component information of each historical formula information and a dimensionality reduction algorithm. The neural network module is configured to receive initial component information, wherein the neural network module is configured to train multiple neural network parameters of the neural network module based on the component information of each historical formula information and the dimensionality-reduced component information to obtain multiple trained neural network parameters. The neural network module obtains dimensionality-reduced initial component information based on the multiple trained neural network parameters and the initial component information. The search module is configured to search for multiple candidate formula information with a first number of parameters among the multiple historical formula information based on a first distance metric, the dimensionality-reduced initial component information, and the dimensionality-reduced component information of each historical formula information. The judgment module is configured to determine whether the physical property information of each candidate formula information meets the specifications. In response to the physical property information of the solution formula information among the multiple candidate formula information meeting the specifications, the solution formula information is output.

[0006] One embodiment of this application provides a recipe construction method, executed by a processor. The recipe construction method includes the following steps: receiving multiple historical recipe information and obtaining dimensionality-reduced component information for each historical recipe information based on its component information and a dimensionality reduction algorithm; receiving initial component information; training multiple neural network parameters of a neural network module based on the component information and dimensionality-reduced component information of each historical recipe information to obtain multiple trained neural network parameters, and obtaining dimensionality-reduced initial component information based on the multiple trained neural network parameters and the initial component information; searching for multiple candidate recipe information with a first number of components among the multiple historical recipe information based on a first distance metric, the dimensionality-reduced initial component information, and the dimensionality-reduced component information of each historical recipe information; determining whether the physical property information of each candidate recipe information meets the specifications; and outputting the solution recipe information in response to the physical property information of the solution recipe information among the multiple candidate recipe information meeting the specifications.

[0007] This application provides a computer-readable storage medium containing a program and a non-transitory computer program product that, when loaded and executed by a processor, can complete the aforementioned recipe construction method.

[0008] Based on the above solutions, some embodiments of this application provide a formula construction system, method, computer-readable storage medium with built-in program, and computer program product. A dimensionality reduction module receives multiple historical formula information entries and obtains dimensionality-reduced component information for each historical formula entry based on its component information and a dimensionality reduction algorithm. A neural network module receives initial component information and trains multiple neural network parameters based on the component information and dimensionality-reduced component information of each historical formula entry to obtain multiple trained neural network parameters. The neural network module then obtains dimensionality-reduced initial component information based on the multiple trained neural network parameters and the initial component information. A search module is configured to search for multiple candidate formula entries with a first number of entries among the multiple historical formula entries based on a first distance metric, the dimensionality-reduced initial component information, and the dimensionality-reduced component information of each historical formula entry. A judgment module is configured to determine whether the physical property information of each candidate formula entry meets the specifications. In response to the physical property information of the solution formula entry among the multiple candidate formula entries meeting the specifications, the solution formula entry is output, which can automatically and effectively utilize the knowledge inherent in the historical formula information. Attached Figure Description

[0009] Figure 1A This is a structural block diagram illustrating a formula construction system and historical formula information based on an embodiment of this application.

[0010] Figure 1B This is a structural block diagram illustrating a formula construction system and original historical formula information based on an embodiment of this application.

[0011] Figure 2 This is a structural block diagram illustrating a formula construction system and historical formula information based on an embodiment of this application.

[0012] Figure 3 This is a schematic diagram illustrating the operation of the dimensionality reduction module and the neural network module according to an embodiment of this application.

[0013] Figure 4 This is a schematic diagram illustrating the operation of the search module according to an embodiment of this application.

[0014] Figure 5 This is a schematic diagram illustrating the operation of a clustering module according to an embodiment of this application.

[0015] Figure 6A This is a schematic diagram illustrating a variation operation based on an embodiment of this application.

[0016] Figure 6B This is a schematic diagram of a variation operation illustrated according to another embodiment of this application.

[0017] Figure 7 This is a schematic diagram of an exchange operation illustrated according to an embodiment of this application.

[0018] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application.

[0019] Figure 9 This is a flowchart illustrating a formulation construction method based on an embodiment of this application.

[0020] Figure 10 This is a flowchart illustrating a formulation construction method according to another embodiment of this application.

[0021] Figure 11 This is an evolutionary calculation flowchart illustrated according to an embodiment of this application.

[0022] Figure 12 This is a flowchart illustrating the exchange operation based on an embodiment of this application.

[0023] Figure 13 This is a flowchart illustrating the variation operation based on an embodiment of this application.

[0024] Explanation of reference numerals in the attached figures:

[0025] 100, 100', 200: Recipe construction system;

[0026] 101: Dimensionality Reduction Module;

[0027] 102: Neural Network Module;

[0028] 103: Search module;

[0029] 104: Judgment module;

[0030] 105: Historical formula information;

[0031] 106: Standardized Modules;

[0032] 107: Original historical recipe information;

[0033] 201: Clustering module;

[0034] 202: Model Building Module;

[0035] 203: Prediction Model Selection Module;

[0036] 204: Evolutionary computation module;

[0037] 301: Component information vector;

[0038] 302: Dimensionally reduced component information vector;

[0039] 401: Initial component information vector for dimensionality reduction;

[0040] 402: Dimensionally reduced component information vector;

[0041] 501, 502, 503, 504: Data groups;

[0042] 600, 600', 603, 603', 603”, 701, 701', 702, 702': Component information vector;

[0043] 601: Components of the component information vector;

[0044] 602, 602', 604, 604', 604”: numerical values;

[0045] 703, 704: Cut-off points;

[0046] 800: Electronic devices;

[0047] 801: Processor;

[0048] 802: Internal memory;

[0049] 803: Non-volatile memory;

[0050] S901~S904, S1001~S1007, S1101~S1105, S1201~S1202, S1301: Steps. Detailed Implementation

[0051] The foregoing and other technical contents, features, and effects of this application will be clearly presented in the following detailed description of the embodiments with reference to the accompanying drawings. The proportions or dimensions of the elements in the drawings are exaggerated, omitted, or generalized for the understanding and reading of those skilled in the art, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. The same reference numerals will be used to denote the same or similar elements in all the drawings. The terms "coupled" or "connected" as used in the following embodiments can refer to any direct or indirect, wired or wireless connection means.

[0052] Figure 1A This is a structural block diagram of the formula construction system 100 and historical formula information 105 drawn according to an embodiment of the present invention. Figure 1B This is a structural block diagram illustrating the formula construction system 100 and the original historical formula information 107 according to an embodiment of the present invention. Please also refer to... Figure 1A and Figure 1B .like Figure 1AAs shown, in one software implementation, the recipe construction system 100 includes a dimensionality reduction module 101, a neural network module 102, a search module 103, and a judgment module 104. The dimensionality reduction module 101 is configured to receive multiple historical recipe information 105.

[0053] Each formula in the historical formula information 105 contains one component information and one physical property information. In this embodiment, the component information contains the quantity of each chemical raw material. The physical property information contains the physical properties of the finished product. The component information and physical property information are presented in vector form. For example, the component information is presented in the form of the following component information vector:

[0054] (1.0, 2.0, 1.5, 4, 2.3, 1.7, 10).

[0055] Each component of the composition information vector represents a different chemical raw material component, and its value represents the amount of each chemical raw material component used in manufacturing. For example, the first component represents the amount of the first type of main adhesive, the second component represents the amount of the second type of main adhesive, the third component represents the amount of the first type of filler, and so on. The information displayed by the composition information vector is the composition information.

[0056] Physical property information is presented in the form of the following physical property information vectors:

[0057] (1.12, 341, 33.67, 5.52).

[0058] Each component of the physical property information vector represents a different physical property. These physical properties are the physical properties of the finished product manufactured based on the corresponding component information. For example, the first component of the physical property information vector represents tensile strength, the second component represents elongation percentage, the third component represents the viscosity of the compound, the fourth component represents compression set, and so on. The information displayed by the physical property information vector is the physical property information.

[0059] It should be noted that, although in this embodiment, the composition information includes the quantity of each chemical raw material, and the physical property information includes the physical properties of the finished product. In one embodiment of this application, the composition information includes the machine controllable parameters of the stretching process for soft copper wire in wire and cable, and the physical property information includes the physical properties of the wire. In one embodiment of this application, the composition information includes the machine controllable parameters of the cold / hot rolling steel process, and the physical property information includes the quality characteristics of the finished product. In one embodiment of this application, the composition information includes the controllable parameters of the pickling process for stainless steel finished products (plates, bars, or wire rods), and the physical property information includes the weight consumption of the finished product. As long as the composition information and physical property information contain a relationship of change, this application is not limited to them. Furthermore, although the composition information and physical property information are presented in vector form in this embodiment, they can also be presented in matrix form, and this application is not limited to them.

[0060] The formula construction system 100 receives a specification from an external source. The neural network module 102 of the formula construction system 100 receives initial ingredient information from an external source. The goal of the formula construction system 100 is to output a component information vector that can produce a finished product with physical properties that meet the specification, based on the initial ingredient information received from the external source and historical formula information 105. That is, the component information represented by the output component information vector.

[0061] In this embodiment, the specification is a set of physical property limits. Taking the aforementioned physical property information vector as an example, the specifications are: tensile strength > 13, elongation percentage > 300, compound viscosity > 20, and compression set < 10.

[0062] In this embodiment, the component information vector of the historical formula information 105 received by the dimensionality reduction module 101 is a standardized value. That is, the value of each component of the component information vector is the standardized value of the vector of the original component information of the original historical formula information. In this embodiment, standardization is performed by subtracting the corresponding average value from the vector of the original component information of each original historical formula information, and then dividing by the corresponding standard deviation. For example, if the original historical formula information consists of the following three sets of vectors:

[0063] Group 1: Original component information vector (273, 82, 105, 210, 9, 904, 680), physical property information vector (23, 62, 34.99).

[0064] The second group consists of the original component information vector (163, 149, 191, 180, 12, 843, 746) and the physical property information vector (0, 20, 41.14).

[0065] The third group: original component information vector (162, 148, 191, 179, 16, 840, 743), physical property information vector (1, 20, 41.81).

[0066] The mean of the first component of the original component information vector is (273 + 163 + 162) / 3 = 199.33, and the standard deviation is 52.09. Using the mean of the first component of the original component information vector (199.3) and the standard deviation (52.09), along with the formula... Normalize the value of the first component of all original ingredient information vectors to obtain (273-199.3) / 52.09 = 1.41, (163-199.3) / 52.09 = -0.69, and (162-199.3) / 52.09 = -0.71. This completes the standardization of the first component of the original ingredient information vectors. Repeat the above steps to standardize each component of each original ingredient information vector to obtain the ingredient information vector of historical formula information 105. The mean and standard deviation of each component of the original ingredient information vectors are called the standardization parameters.

[0067] It should be noted that the aforementioned standardization method is not the only method. Appropriate standardization parameters can be selected to standardize the original component information based on the characteristics of the actual data.

[0068] Please refer to Figure 1B In one embodiment of this application, the formula construction system 100' further includes a standardization module 106. The standardization module 106 is configured to receive original historical formula information 107 and standardize the original historical formula information 107 according to the method described above to obtain the component information of historical formula information 105. The historical formula information 105 is then transmitted to the dimensionality reduction module 101.

[0069] The following is a detailed description, with reference to the accompanying drawings, of the formulation construction method of this application and how the modules of the formulation construction system 100 and 100' work together in accordance with each other.

[0070] Figure 3 This is a schematic diagram illustrating the operation of the dimensionality reduction module 101 and the neural network module 102 according to an embodiment of this application. Figure 9 This is a flowchart illustrating a formulation construction method according to an embodiment of this application. Please refer to it as well. Figure 1A , Figure 1B , Figure 3 and Figure 9 In step S901, the dimensionality reduction module 101 receives multiple historical formula information 105 and obtains the dimensionality-reduced component information of each historical formula information 105 based on the component information of each historical formula information 105 and the dimensionality reduction algorithm.

[0071] In this embodiment, as Figure 3 As shown, the dimensionality reduction module 101 uses the t-SNE (t-distributed stochastic neighbor embedding) algorithm as the dimensionality reduction algorithm to reduce the dimensionality of the component information vector 301 of all historical formula information 105 to obtain the dimensionality-reduced component information vector 302, which is located in a two-dimensional vector space.

[0072] In step S902, the neural network module 102 uses the component information vectors 301 and corresponding dimensionality-reduced component information vectors 302 of each historical formula information 105 as training samples to train multiple neural network parameters of the neural network module 102 to obtain multiple trained neural network parameters. That is, in this embodiment, the internal process of learning the dimensionality reduction algorithm is utilized using multiple neural network parameters of the neural network module 102. After obtaining the trained neural network parameters, the neural network module 102 receives initial component information from the outside, which is represented by an initial component information vector in this embodiment. Dimensionality-reduced initial component information is obtained based on the trained neural network parameters and the initial component information. This initial component information vector has been standardized using the mean and standard deviation (i.e., standardization parameters) of each component of the aforementioned original component information vector. Figure 1B The illustrated embodiment first receives the original initial component information by the standardization module 106, then standardizes it using the mean and standard deviation of each component of the original component information vector to obtain the initial component information, which is then passed to the neural network module 102 to obtain the dimensionality-reduced initial component information.

[0073] The aforementioned internal process of learning the dimensionality reduction algorithm using the neural network module 102 can obtain the corresponding dimensionality-reduced component information vector 302 without affecting the dimensionality reduction process of the component information vector 301 of the historical formula information 105 with the initial component information, while maintaining the independence of the dimensionality-reduced component information vector 302, so that the dimensionality-reduced component information vector 302 is not affected by the initial component information.

[0074] In some embodiments of this application, the dimensionality reduction module 101 uses other nonlinear algorithms (e.g., LocallyLinear Embedding (LLE) or Isometric Mapping algorithm) as dimensionality reduction algorithms. This application does not limit the type of nonlinear algorithm.

[0075] In some embodiments of this application, the dimensionality reduction module 101 uses a linear algorithm (e.g., principal component analysis (PCA)) as the dimensionality reduction algorithm.

[0076] Figure 4 This is a schematic diagram illustrating the operation of the search module 103 according to an embodiment of this application. Please also refer to... Figure 4 .

[0077] In step S903, the search module 103 searches for multiple candidate formula information with a first number in multiple historical formula information 105 based on the first distance metric, the initial component information of dimensionality reduction, and the component information of each historical formula information 105.

[0078] In this embodiment, the search module 103, based on a first distance metric (Euclidean distance in a two-dimensional vector space in this embodiment), searches among multiple historical formula information vectors for a first number of vectors that are closest to the initial component information vector 401 after dimensionality reduction. Figure 4 In the illustrated example, there are 5) dimension-reduced component information vectors 402. The search module 103 then searches for a first number of historical formula information corresponding to these dimension-reduced component information vectors 402. For convenience, these first number of historical formula information are referred to as candidate formula information below.

[0079] In step S904, after selecting a first number of candidate formulation information, the judgment module 104 further judges whether there is any physical property information among these first number of candidate formulation information that meets the specifications. If there is physical property information among these first number of candidate formulation information that meets the specifications, for ease of explanation, the candidate formulation information corresponding to the physical property information that meets the specifications is called the solution formulation information. The judgment module 104 then outputs this solution formulation information.

[0080] It should be noted that, although in the aforementioned embodiments, the first distance metric is the Euclidean distance in a two-dimensional vector space, and the dimensionality reduction module 101 uses the t-SNE algorithm as the dimensionality reduction algorithm to reduce the dimensionality of the component information vectors 301 of all historical formula information 105 to obtain the dimensionality-reduced component information vectors 302 in a two-dimensional vector space, other metric methods can also be selected for the first distance metric (such as Manhattan distance, Chebyshev distance, etc.), as long as they can assign distances to points in a two-dimensional vector space. The dimensionality reduction module 101 can also use the t-SNE algorithm to reduce the dimensionality of the component information vectors 301 of all historical formula information 105 to 3 dimensions, or use other dimensionality reduction algorithms to reduce the dimensionality of the component information vectors 301 of all historical formula information 105 to other dimensions. This application is not limited to the aforementioned embodiments.

[0081] Figure 2 This is a structural block diagram illustrating a formula construction system 200 and historical formula information 105 according to an embodiment of this application. Please refer to... Figure 2 , Figure 2 The illustrated formula construction system 200 and Figure 1A In comparison, it also includes a clustering module 201, a model building module 202, a prediction model selection module 203, and an evolutionary calculation module 204. Figure 5 This is a schematic diagram illustrating the operation of the clustering module 201 according to an embodiment of this application. Figure 10 This is a flowchart illustrating a formulation construction method according to an embodiment of this application. Please refer to it as well. Figure 2 , Figure 5 and Figure 10 .

[0082] In response to the absence of physical property information that meets the specifications among the aforementioned first number of candidate formulation information, in step S1001, the clustering module 201 divides the multiple historical formulation information 105 into multiple data clusters based on the clustering algorithm and the dimensionality reduction component information of each historical formulation information 105.

[0083] In this embodiment, the clustering module 201 uses hierarchical clustering as the clustering algorithm and divides the dimensionality-reduced component information into multiple data clusters based on the dimensionality-reduced component information of each historical formula information 105. Figure 5 The illustrated example shows that the clustering module 201 divides the dimensionality reduction component information into multiple data clusters 501, 502, 503, and 504. Correspondingly, the historical formula information 105 also corresponds to data clusters 501, 502, 503, and 504, respectively.

[0084] In step S1002, the model building module 202 builds candidate prediction models for each data group based on the component information and physical property information of each member in each data group.

[0085] exist Figure 5 In the illustrated example, the model building module 202 establishes candidate prediction models for each data group 501, 502, 503, and 504 based on the component information and physical property information of each member. In this embodiment, the model building module 202 uses multivariable linear regression to establish linear models of the component information vectors and physical property information vectors in each data group 501, 502, 503, and 504. These linear models serve as candidate prediction models for their respective data groups.

[0086] It should be noted that, in this embodiment, the clustering module 201 uses hierarchical clustering as the clustering algorithm, and the model building module 202 uses multivariable linear regression to build candidate prediction models for each data cluster 501, 502, 503, and 504. In some embodiments of this application, the clustering module 201 uses k-means as the clustering algorithm, and the model building module 202 uses a fully connected neural network to build candidate prediction models for each data cluster 501, 502, 503, and 504. As long as the clustering module 201 achieves the clustering objective, the model building module 202 can build candidate prediction models for each data cluster 501, 502, 503, and 504; this application is not limited to the aforementioned embodiments.

[0087] In some embodiments of this application, the clustering module 201 uses density-based spatial clustering of applications with noise (DBSCAN) as the clustering algorithm. In some embodiments of this application, the clustering module 201 uses the expectation-maximization algorithm (EM) as the clustering algorithm.

[0088] In step S1003, the prediction model selection module 203 selects an approximate group from multiple data groups based on multiple candidate formulation information, and sets the candidate prediction model of the selected approximate group as the prediction model.

[0089] In this embodiment, the prediction model selection module 203 selects the data group containing the most dimensionality reduction component information vectors of the candidate formulation information as an approximation group based on the number of such vectors in each data group. The prediction model selection module 203 then sets the candidate prediction model of the selected approximation group as the prediction model.

[0090] by Figure 5 The illustrated example shows that after the prediction model selection module 203 counts, it obtains a dimensionality-reduced component information vector of candidate formulation information (in...). Figure 5 In the illustrated example, the reduced-dimensional component information vector 402 is located in the data group 501. Therefore, the prediction model selection module 203 selects the data group 501 as the approximate group. The prediction model selection module 203 then sets the candidate prediction model of the data group 501 as the prediction model.

[0091] In step S1004, the evolutionary calculation module 204 executes steps S1005 to S1007. In step S1005, the evolutionary calculation module 204 sets the fitness function according to the specification. In this embodiment, the evolutionary calculation module 204 selects closed intervals that conform to all physical property constraints included in the specification, and calculates the midpoints of these closed intervals. Then, the Manhattan distance between the physical property information vector and the vector formed by the midpoints of the multiple closed intervals is used as the fitness function.

[0092] Taking the aforementioned specifications—tensile strength > 13, elongation percentage > 300%, compound viscosity > 20, and compression set < 10—as an example, the evolution calculation module 204 selects closed intervals that meet the property limitations as follows: tensile strength: [13, 15], elongation percentage: [300, 310], compound viscosity: [20, 30], and compression set: [9, 10]. The vector formed by the midpoints of these closed intervals is (14, 305, 25, 9.5). The next goal of the evolution calculation module 204 is to calculate a set of component information vectors such that the Manhattan distance between the property information vectors predicted by the prediction model based on these component information vectors and the vector (14, 305, 25, 9.5) formed by the midpoints of these closed intervals is as small as possible. Therefore, the evolutionary computation module 204 selects the Manhattan distance (14, 305, 25, 9.5) formed by the property vector to the midpoint of multiple closed intervals as the fitness function.

[0093] In other words, the fitness function can be expressed as:

[0094]

[0095] Where N represents the dimension of the material property information vector, m i x represents the midpoint of each closed interval that conforms to the physical property constraints. i This represents the i-th component of the property information vector predicted based on the calculated component information vector. In this example, m1 = 14, m2 = 305, m3 = 25, and m4 = 9.5.

[0096] Since the ingredients used are recorded in the historical formula information 105, they may not be available at the current time. Alternatively, due to cost considerations, these ingredients may not be used. Therefore, before calculating the new ingredient information vector, the evolutionary calculation module 204 receives an adjustment parameter from an external source in step S1006. This adjustment parameter indicates which ingredients in the historical formula information 105 and the ingredient information to be calculated are considered and adjustable in the evolutionary calculation, and which are not considered. In other words, the adjustment parameter indicates multiple adjustable pieces of information in the ingredient information.

[0097] Taking the aforementioned component information vector (1.0, 2.0, 1.5, 4, 2.3, 1.7, 10) as an example, the first component represents the amount of the first main adhesive, the second component represents the amount of the second main adhesive, and the third component represents the amount of the first filler. If the first main adhesive is not considered in the evolution calculation, the parameter adjustment indicates that the first main adhesive is not considered. In the following step S1007, the vector component corresponding to the first main adhesive will be set to 0 so that this vector component does not participate in the evolution calculation. This part will be explained in detail later.

[0098] In this embodiment, parameter adjustment is implemented using a list mask, where MASK is the list name. The values ​​in the list mask consist of 0s and 1s. If the value at position i is 0, it means that the component represented by the component information vector at that position is not considered in the evolution calculation. For example, with the list mask = [0, 1, 1, 1, 1, 1, 1], the first element of the list mask is 0, indicating that the component corresponding to the first component in the component information vector (in this example, the first main adhesive) will not be considered in the evolution calculation. The second to seventh elements of the list mask are 1, indicating that the components corresponding to the second to seventh components in the component information vector will be considered in the evolution calculation.

[0099] In step S1007, the evolution calculation module 204 outputs evolution recipe information based on the fitness function, adjustment parameters, evolution algorithm, and approximate group.

[0100] There are many types of evolutionary algorithms. Figure 11 This is a flowchart illustrating an evolutionary computation based on an embodiment of this application. The aforementioned step S1007 includes... Figure 11 The steps S1101 to S1105 are illustrated. Please also refer to [the relevant documentation / reference]. Figure 2 , Figure 5 as well as Figure 11 Since in step S1003, the prediction model selection module 203 selects the data group containing the most dimensionality-reduced component information vectors of the candidate formulation information as the approximate group, in step S1101, the evolution calculation module 204 selects the candidate formulation information located in the approximate group as the approximate formulation information. Based on the adjustment parameters, the component information of each approximate formulation information is adjusted to obtain multiple adjusted approximate formulation information.

[0101] The evolution calculation module 204 sets an initial set based on multiple adjusted approximate recipe information, and then sets the initial set as the current evolution set.

[0102] Here, with Figure 5The example shown illustrates this. In step S1003, the prediction model selection module 203 selects the data group 501 containing the most dimensionality-reduced component information vectors of the candidate formulation information as the approximation group. The evolution calculation module 204 selects the candidate formulation information located in the approximation group. In this example, the candidate formulation information is the historical formulation information corresponding to the five dimensionality-reduced component information vectors 402. If the historical formulation information vectors corresponding to these five dimensionality-reduced component information vectors 402 are respectively:

[0103] First: Component information vector (1, 2, 3, 4, 5, 6, 7, 8), property information vector (23, 62, 34.99);

[0104] The second one; component information vector (1.1, 2, 3, 4.1, 5, 6, 7.1, 8), property information vector (0, 20, 41.14);

[0105] The third: component information vector (1.1, 2, 3, 4.1, 5, 6, 7.1, 8), property information vector (1, 25, 42.89);

[0106] The fourth: component information vector (1.2, 2.2, 3, 4.3, 5, 6, 7.1, 8), property information vector (2, 19, 48.80);

[0107] Fifth: Component information vector (1.4, 2, 3.3, 4.1, 5, 6, 7.1, 8.1), property information vector (3, 20, 41.81);

[0108] The adjustment parameter is a string Mask = [1, 1, 1, 0, 0, 0, 1, 1]. Based on this adjustment parameter, five adjusted approximate recipe information vectors are obtained as follows:

[0109] First: Adjusted component information vector (1, 2, 3, 0, 0, 0, 7, 8), physical property information vector (23, 62, 34.99);

[0110] The second one; the adjusted component information vector (1.1, 2, 3, 0, 0, 0, 7.1, 8), and the property information vector (0, 20, 41.14);

[0111] The third: Adjusted component information vector (1.1, 2, 3, 0.0, 0, 7.1, 8), physical property information vector (1, 25, 42.89);

[0112] Fourth: Adjusted component information vector (1.2, 2.2, 3, 0, 0, 0, 7.1, 8), physical property information vector (2, 19, 48.80);

[0113] Fifth: Adjusted component information vector (1.4, 2, 3.3, 0, 0, 0, 7.1, 8.1), physical property information vector (3, 20, 41.81).

[0114] The evolutionary calculation module 204 then adds the adjusted component information vectors from these five adjusted approximate recipe information vectors to the initial set, and then expands the initial set to a fixed number of offspring. There are many methods for expansion. In some embodiments, the evolutionary calculation module 204 directly copies the existing adjusted component information vectors in the initial set to expand the initial set to a fixed number of offspring. In some embodiments, the evolutionary calculation module 204 randomly selects a component information vector from the initial set, then adds a random real number to the values ​​of the components to be considered in the evolutionary calculation (indicated by the serial Mask) within a certain range to create a new component information vector, and calculates the property information vector of the new component information vector using a prediction model. The new component information vector and its property information vector are then added to the initial set until the initial set is expanded to the required number of offspring.

[0115] In step S1102, the evolutionary calculation module 204 adds the current evolutionary set to the candidate class to record the current evolutionary set calculated each time. The evolutionary calculation module 204 then selects multiple reproductive individuals from the current evolutionary set based on the fitness function and the prediction model.

[0116] In this embodiment, the evolutionary calculation module 204 calculates the fitness function values ​​of all members in the current evolutionary set, and then sorts all members in the current evolutionary set according to their fitness function values ​​from smallest to largest. The minimum fitness function value among all members is defined in this application as the representative fitness function value of the current evolutionary set. The evolutionary calculation module 204 then selects a specified number of members from the current evolutionary set as reproducing individuals based on this sorting.

[0117] In step S1103, multiple reproductive individuals of the same number as the evolutionary number are sequentially exchanged and mutated to obtain the next evolutionary set. As explained in step S1003 above, the next evolutionary set is adjusted according to the adjustment parameters and then set as the current evolutionary set.

[0118] In step S1104, steps S1102 and S1103 are repeated until the number of repetitions reaches a preset number of executions.

[0119] Finally, in step S1105, the judgment module 104 selects an evolutionary set from the candidate classes based on the fitness function, wherein the representative fitness function value of this evolutionary set is the smallest among the representative fitness function values ​​of all evolutionary sets in the candidate classes. The judgment module 104 outputs this selected evolutionary set as evolutionary recipe information.

[0120] Figure 6A , Figure 6B This is a schematic diagram illustrating a variation operation based on an embodiment of this application. Figure 7 This is a schematic diagram of an exchange operation illustrated according to an embodiment of this application. Figure 12 This is a flowchart illustrating the exchange operation based on an embodiment of this application. Figure 13 This is a flowchart illustrating a variation operation based on an embodiment of this application. Please also refer to... Figure 6A , Figure 6B , Figure 7 , Figure 12 as well as Figure 13 .

[0121] In step S1102, the evolutionary calculation module 204 selects multiple breedable individuals from the current evolutionary set based on the fitness function and the prediction model. In step S1201, the evolutionary calculation module 204 randomly performs the following two steps based on the exchange rate: (i) randomly selects the first and second individuals from the multiple breedable individuals, and exchanges some of the component information of the first individual with some of the component information of the second individual to generate a third and a fourth individual. The evolutionary calculation module 204 adds the third and fourth individuals to the multiple breedable individuals. (ii) randomly generates a fifth and a sixth individual based on the first range restriction and adjustment parameters, and adds the fifth and sixth individuals to the multiple breedable individuals.

[0122] In this embodiment, the commutation rate is a real number between [0, 1], representing the probability of performing the aforementioned step (i). The evolutionary computation module 204 utilizes a random function common to software simulations, such as the random() function of the random module in Python. The random() function randomly generates a real number between [0, 1). The evolutionary computation module 204 utilizes the random() function of the random module in Python and the following program structure:

[0123] If random.random() < commutative property;

[0124] Perform the aforementioned step (i).

[0125] else:

[0126] Perform the aforementioned step (ii).

[0127] The procedure structure executes the steps described in (i) with a probability of commutation and the steps described in (ii) with a probability of (1 - commutation). For example, if the commutation is 0.7, the procedure structure will execute step (i) with a probability of 0.7 and step (ii) with a probability of 0.3.

[0128] In step (i), the evolutionary computation module 204 randomly selects the first and second individuals from the reproductive population (using the sample() function of the random module in Python), then randomly selects two cutoff points (using the randint() function of the random module in Python to select integers within the length of the component information vector), and swaps the contents of the components between the two cutoff points in the component information vector of the first individual. Figure 7 The illustrated example shows that component 601 of the component information vector is the part that the adjustment parameter indicates should not be considered, and therefore is adjusted to 0. The evolutionary calculation module 204 randomly selects the first and second individuals from the breedable population. The component information vector of the first individual is 701, and the component information vector of the second individual is 702. The evolutionary calculation module 204 randomly selects two cutoff points 703 and 704, and then swaps the contents of the components between the two cutoff points 703 and 704 to generate component information vectors 701' and 702'. The evolutionary calculation module 204 uses a prediction model to calculate the property information vector of component information vector 701' to generate the third individual. The evolutionary calculation module 204 uses a prediction model to calculate the property information vector of component information vector 702' to generate the fourth individual. The evolutionary calculation module 204 adds the third and fourth individuals to the breedable population.

[0129] In step (ii), the evolutionary calculation module 204, based on the first range limit [-2, 2] and the adjustment parameters, randomly generates real numbers between [-2, 2] for the components of the component information vector to be considered, as indicated by the adjustment parameters, to generate two component information vectors. The evolutionary calculation module 204 uses a prediction model to calculate the physical property information vectors corresponding to these two component information vectors to generate a fifth and a sixth individual. The fifth and sixth individuals are added to the breedable individuals. Taking Mask = [0, 1, 1, 1, 1, 1, 1] as an example, the two randomly generated component information vectors are (0, 0.1, -0.3, 1, 1.5, 1, 1) and (0, 1, 1, 1.5, 0.2, -1.7, 0.8). The prediction model is then used to calculate the corresponding physical property information vectors, for example, (2, 33, 7) and (3, 44, 5). Then, the following can be set:

[0130] Fifth individual: Composition information vector (0, 0.1, -0.3, 1, 1.5, 1, 1), property information vector (2, 33, 7);

[0131] The sixth individual: composition information vector (0, 1, 1, 1.5, 0.2, -1.7, 0.8), property information vector (3, 44, 5).

[0132] In step S1202, the evolutionary calculation module 204 determines whether the number of fertile individuals has reached a predetermined number. In this embodiment, the predetermined number is the fixed number of offspring mentioned above.

[0133] It should be noted that the predetermined number does not necessarily have to be the number of offspring, and can be greater than the number of offspring. After the evolutionary calculation module 204 generates the predetermined number of fertile individuals, other processing is performed, and this application does not impose any restrictions.

[0134] After the exchange operation is performed, the evolutionary calculation module 204 performs a mutation operation. In step S1301, the component information of each reproducing individual is adjusted within a second range limit according to the adjustment parameters.

[0135] In this embodiment, the evolutionary calculation module 204, based on the second range constraints (0.01, 0.1) and (-1, -0.01), adds a real number randomly generated in (0.01, 0.1) and (-1, -0.01) to the components of the component information vector to be considered by each adjustment parameter for the component information information of each breeding individual (using the uniform() function of the random module in Python) to generate a mutated component information vector. The evolutionary calculation module 204 then uses a prediction model to calculate the physical property information vector corresponding to the mutated component information vector to generate mutated breeding individuals.

[0136] In one embodiment, after adding a real number randomly generated in (0.01, 0.1) and (-1, -0.01), the evolution calculation module 204 further checks whether the values ​​of the components of the mutated component information vector are reasonable. If the evolution calculation module 204 restores the values ​​of the components of the mutated component information vector back to the actual manufacturing quantity and finds that the restored quantity is negative, it indicates that the values ​​of the components of the mutated component information vector are unreasonable. Therefore, the values ​​of the components of the mutated component information vector are set to 0.

[0137] by Figure 6A To illustrate with the example shown, in this example, Mask = [1, 1, 1, 0, 0, 0, 1, 1], component 601 of the component information vector 600 is the part that is not considered according to the adjustment parameter. The evolution calculation module 204 randomly generates a real number 0.09 based on the second range constraints (0.01, 0.1) and (-1, -0.01), and adds the value 602 (actual value is 2) of the second component of the component information vector 600 to obtain a new value 602' (actual value is 2.09) of the second component, thus generating a new component information vector 600'.

[0138] by Figure 6BTo illustrate with the example shown, in this example, Mask = [1, 1, 1, 0, 0, 0, 1, 1], component 601 of component information vector 603 is the part that the adjustment parameter indicates should not be considered. Evolutionary calculation module 204, based on the second range constraints (0.01, 0.1) and (-1, -0.01), randomly generates a real number -0.08 and adds it to the value 604 of the eighth component of component information vector 603 (actual value is 0.01) to obtain a new value 604' (actual value is -0.07) of the eighth component, thus generating a new component information vector 603'. However, evolutionary calculation module 204 determines that the value of -0.07 is unreasonable (in this example, it is assumed that the average value of the eighth component is 0), therefore, it sets the value of the eighth component to 0 to obtain a new value 604" of the eighth component, thus generating a new component information vector 603.

[0139] It should be noted that the initial component information vectors of the preceding embodiments can be obtained through prior evaluation. In this case, the formulation construction systems 100, 100', and 200 disclosed in this application can fully integrate this initial component information vector with the knowledge inherent in the historical formulation information 105. Of course, the initial component information vectors of the preceding embodiments can also be generated randomly. The formulation construction systems 100, 100', and 200 disclosed in this application can utilize the knowledge inherent in the historical formulation information 105 to find component information that meets the specifications. In one embodiment of this application, the formulation construction system 200 uses a component information vector from the output evolved formulation information as the initial component information vector.

[0140] It should also be noted that while the foregoing explanations mentioned using internal functions from Python's `random` module to implement random selection, and using these functions in conjunction with Python's `if` statement to achieve random execution steps, one could also utilize functions from the C++ Standard Library. <random>The function library's internal functions implement random selection functionality, and utilize C++ language features. <random>The internal function of the function library is combined with the structure of if in the C++ syntax to realize the random execution step, and other program languages having the function of generating random numbers can be used to realize the random selection function, and other program languages having the function of generating random numbers are combined with the structure of the conditional branch instruction in the program language syntax to realize the random execution step, and the present application is not limited thereto.

[0141] Figure 8 FIG. 8 is a structural schematic diagram of an electronic device 800 according to an embodiment of the present application. As shown in FIG. 8, at the hardware level, the electronic device 800 includes a processor 801, an internal memory 802 and a non-volatile memory 803. The internal memory 802 is, for example, a random-access memory (RAM). The non-volatile memory is, for example, at least one disk memory. Of course, the electronic device 800 can also include other hardware required by functions. Figure 8

[0142] The internal memory 802 and the non-volatile memory 803 are used to store programs, which can include program codes including computer operation instructions. The internal memory 802 and the non-volatile memory 803 provide instructions and data for the processor 801. The processor 801 reads the corresponding computer programs from the non-volatile memory 803 to the internal memory 802 and then runs, and forms the recipe construction system 100, 100' or 200 at the logical level. The processor 801 is specifically used to execute the steps described in the specification. Figures 9 to 13

[0143] The processor 801 can be an integrated circuit chip having a signal processing capability. In the implementation process, the methods and steps disclosed in the foregoing embodiments can be completed by the integrated logic circuits or the instructions in the software form in the processor 801. The processor 801 can be a general-purpose processor, including a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, which can realize or execute the methods and steps disclosed in the foregoing embodiments.

[0144] ​​The embodiments of the present specification further provide a computer-readable storage medium storing at least one instruction, which, when executed by the processor 801 of the electronic device 800, can enable the processor 801 of the electronic device 800 to perform the methods and steps disclosed in the foregoing embodiments.

[0145] Examples of storage media of a computer include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other internal memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disks storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, the computer-readable storage medium does not include transitory media, such as modulated data signals and carriers.< / random> < / random>

Claims

1. A formulation construction system, characterized in that, Include: The dimensionality reduction module is configured to receive multiple historical formula information and obtain the dimensionality-reduced component information of each historical formula information based on the component information of each historical formula information and the dimensionality reduction algorithm; A neural network module is configured to receive initial component information. The neural network module is configured to train multiple neural network parameters based on the component information of each historical formula and the dimensionality-reduced component information to obtain multiple trained neural network parameters, and to obtain dimensionality-reduced initial component information based on the multiple trained neural network parameters and the initial component information. The search module is configured to search for multiple candidate formula information with a first number among the multiple historical formula information based on a first distance metric, the initial dimensionality reduction component information, and the dimensionality reduction component information of each of the historical formula information; The judgment module is configured to determine whether the physical property information of each candidate formulation information meets the specifications; in response to the physical property information of the solution formulation information among the multiple candidate formulation information meeting the specifications, the module outputs the solution formulation information; and A standardization module is configured to receive multiple original historical formula information and original initial ingredient information, wherein each original historical formula information includes the physical property information and the original ingredient information. The standardization module obtains multiple standardization parameters based on the original ingredient information of each original historical formula information. The standardization module obtains the multiple historical formula information and the ingredient information of each historical formula information based on the multiple standardization parameters and the multiple original historical formula information. The standardization module obtains the initial ingredient information based on the multiple standardization parameters and the original initial ingredient information.

2. The formulation construction system as described in claim 1, characterized in that, The formulation construction system also includes: The clustering module is configured to divide the multiple historical formula information into multiple data clusters based on the clustering algorithm and the dimensionality reduction component information of each of the historical formula information. The model building module is configured to build candidate prediction models for each data group based on the component information and physical property information of each member in each data group. The prediction model selection module is configured to select an approximate group from the multiple data groups based on the multiple candidate formulation information, and set the candidate prediction model of the approximate group as the prediction model; and The evolutionary computation module is configured to perform the following steps: (a) Set the fitness function according to the specifications; (b) Receive adjustment parameters; (c) Output evolution recipe information based on the fitness function, the adjustment parameters, the evolution algorithm, and the approximate group; The adjustment parameter indicates multiple adjustable information in the component information of each member in the approximate group.

3. The formulation construction system as described in claim 2, characterized in that, Step (c) includes the following steps: (c1) Select multiple approximate formulation information that are located in the approximation group from the multiple candidate formulation information, and adjust the component information of each approximate formulation information according to the adjustment parameter to obtain multiple adjusted approximate formulation information, set an initial set according to the multiple adjusted approximate formulation information, and set the initial set as the current evolution set; (c2) Add the current evolutionary set to the candidate class, and select multiple reproductive individuals from the current evolutionary set according to the fitness function and the prediction model; (c3) Perform exchange and mutation operations on the plurality of reproductive individuals in sequence to obtain the next evolutionary set, and adjust the next evolutionary set according to the adjustment parameters and set it as the current evolutionary set; as well as (c4) Repeat steps (c2) and (c3) a number of times; The judgment module is configured to select an evolutionary set from the candidate classes based on the fitness function as the output of the evolutionary recipe information.

4. The formulation construction system as described in claim 3, characterized in that, The exchange operation includes the following steps: (c31) Based on the exchange rate, perform the following two steps randomly: (i) Randomly select a first individual and a second individual from the plurality of fertile individuals, exchange partially or entirely of the component information of the first individual and partially or entirely of the component information of the second individual to generate a third individual and a fourth individual, and add the third individual and the fourth individual to the plurality of fertile individuals; and (ii) Randomly generate a fifth individual and a sixth individual according to the first range limitation and the adjustment parameters, and add the fifth individual and the sixth individual to the plurality of fertile individuals; and (c32) Repeat step (c31) until the number of the plurality of reproductive individuals reaches a predetermined number.

5. The formulation construction system as described in claim 3, characterized in that, The mutation operation includes the following steps: Based on the adjustment parameters, the component information of each of the reproductive individuals is randomly adjusted within the second range limit.

6. The formulation construction system as described in claim 1, characterized in that, The dimensionality reduction algorithm is a nonlinear algorithm.

7. The formulation construction system as described in claim 6, characterized in that, The dimensionality reduction algorithm is the t-distribution random neighbor embedding algorithm.

8. A method for formula construction, characterized in that, Executed by a processor, the recipe construction method includes: Receive multiple historical formula information, and obtain the dimensionality-reduced component information of each historical formula information based on the component information of each historical formula information and the dimensionality reduction algorithm; Receive initial component information, train multiple neural network parameters of the neural network module based on the component information of each of the historical formula information and the dimensionality-reduced component information to obtain multiple trained neural network parameters, and obtain dimensionality-reduced initial component information based on the multiple trained neural network parameters and the initial component information; Based on the first distance metric, the initial dimensionality reduction component information, and the dimensionality reduction component information of each of the historical formula information, multiple candidate formula information with a first number are searched among the multiple historical formula information; as well as Determine whether the physical property information of each candidate formulation information meets the specifications, and in response to the physical property information of the solution formulation information among the multiple candidate formulation information meeting the specifications, output the solution formulation information; The formulation construction method further includes: Receive multiple original historical formula information and original initial ingredient information, wherein each of the original historical formula information includes the physical property information and the original ingredient information; Multiple standardized parameters are obtained based on the original ingredient information of each of the original historical formulas. Based on the multiple standardized parameters and the multiple original historical formula information, the multiple historical formula information and the component information of each of the historical formula information are obtained; and, The initial component information is obtained based on the multiple standardized parameters and the original initial component information.

9. The formulation construction method as described in claim 8, characterized in that, The formulation construction method further includes: Based on the clustering algorithm and the dimensionality reduction component information of each historical formula information, the multiple historical formula information is divided into multiple data clusters; Based on the component information and physical property information of each member in each data group, a candidate prediction model is established for each data group; Based on the multiple candidate formulation information, an approximate group is selected from the multiple data groups, and the candidate prediction model of the approximate group is set as the prediction model; and Perform the following evolutionary calculation steps: (a) Set the fitness function according to the specifications; (b) Receive adjustment parameters; (c) Output evolution recipe information based on the fitness function, the adjustment parameters, the evolution algorithm, and the approximate group; The adjustment parameter indicates multiple adjustable information in the component information of each member in the approximate group.

10. The formulation construction method as described in claim 9, characterized in that, The evolutionary computation step (c) includes the following steps: (c1) Adjust the multiple approximate recipe information located in the approximation group among the multiple candidate recipe information according to the adjustment parameters to obtain multiple adjusted approximate recipe information, set an initial set according to the multiple adjusted approximate recipe information, and set the initial set as the current evolution set; (c2) Add the current evolutionary set to the candidate class, and select multiple reproductive individuals from the current evolutionary set according to the fitness function and the prediction model; (c3) Perform exchange and mutation operations on the plurality of reproductive individuals to obtain the next evolutionary set, and adjust the next evolutionary set according to the adjustment parameters and set it as the current evolutionary set; (c4) Repeat steps (c2) and (c3) a number of times; and (c5) Select an evolutionary set from the candidate classes based on the fitness function as the output of the evolutionary recipe information.

11. The formulation construction method as described in claim 10, characterized in that, The exchange operation includes: (c31) Based on the exchange rate, perform the following two steps randomly: (i) Randomly select a first individual and a second individual from the plurality of reproductive individuals, exchange partially or completely of the component information of the first individual with partially or completely of the component information of the second individual to generate a third individual and a fourth individual, and add the third individual and the fourth individual to the plurality of reproductive individuals; and (ii) Based on the first range limitation and the adjustment parameters, a fifth individual and a sixth individual are randomly generated, and the fifth individual and the sixth individual are added to the plurality of breedable individuals; and (c32) Repeat step (c31) until the number of the plurality of reproductive individuals reaches a predetermined number.

12. The formulation construction method as described in claim 10, characterized in that, The mutation operation includes: Based on the adjustment parameters, the component information of each of the reproductive individuals is randomly adjusted within the second range limit.

13. The formulation construction method as described in claim 8, characterized in that, The dimensionality reduction algorithm is nonlinear.

14. The formulation construction method as described in claim 13, characterized in that, The dimensionality reduction algorithm is the t-distribution random neighbor embedding algorithm.

15. A computer-readable storage medium containing a stored program, characterized in that, When the processor loads and executes the memory program, it completes the method as described in any one of claims 8 to 14.

16. A non-transitory computer program product, characterized in that, The system stores at least one instruction that, when executed by a processor, causes the processor to perform the method as described in any one of claims 8 to 14.

Images