Search device, search method, and program

Abstract

A search device (100) comprises: a data set updating unit (110) for generating or updating a data set (121) indicating an objective variable y and an explanatory variable x in association with each other; a reliability calculation unit (130) for calculating the reliability r of the objective variable y on the basis of process data obtained from a material processing device (200) and indicating the state of a raw material processing process; a data set processing unit (140) for processing the data set (121) on the basis of the reliability r; a prediction model generation unit (150) for generating a prediction model (123) indicating the relationship between the explanatory variable x and the objective variable y by using a processed data set (122) which is the data set (121) that has been processed; and a search condition derivation unit (160) for deriving, as a search condition, the value of the explanatory variable x for obtaining the target value of the objective variable y, by performing a search using the prediction model (123).

Description

Searching device, searching method, and program

[0001] The present disclosure relates to a searching device for searching conditions used in processing raw materials.

[0002] Material processing equipment, such as chemical reaction equipment and semiconductor processing equipment, produces products by synthesizing and processing raw materials according to processing conditions. The characteristic values ​​of the products vary depending on the various processing conditions of the material processing equipment. For example, characteristic values ​​of the products are obtained for each of a plurality of different processing conditions. To obtain a product having a target characteristic value, it is necessary to search for input parameters that will produce output parameters that indicate the target characteristic value from a plurality of input parameters that each represent a different value as a processing condition. However, it is not practical to actually measure or evaluate the output parameters for each of the input parameters. Therefore, machine learning techniques, such as neural networks, are used to search for optimal solutions for the input parameters.

[0003] Patent Literature 1 discloses a search device that generates a prediction model showing the relationship between input parameters and output parameters through machine learning and searches for an optimal solution based on this prediction model. Specifically, the search device of Patent Literature 1 calculates, based on the prediction model, predicted values ​​of input parameters that yield output parameters showing target characteristic values, and determines whether the actual measured values ​​corresponding to these predicted values ​​reach the target. The actual measured values ​​are the actual measured characteristic values ​​of a product obtained by processing a semiconductor processing device according to the predicted values. If the actual measured values ​​do not reach the target, the search device adds the predicted and actual measured values ​​to the training data as new input and output parameters, thereby updating the prediction model. This improves search efficiency. Furthermore, if the actual measured values ​​are not close to the target, the search device of Patent Literature 1 excludes from the search area the region where the combination of the actual measured value and the predicted value exists. This improves the accuracy of reaching the optimal solution.

[0004] JP 2019-40984 A

[0005] A search device according to a first aspect of the present disclosure includes a dataset update unit that generates or updates a dataset that associates a dependent variable indicating a characteristic value of a product obtained by processing raw materials according to explanatory variables using a material processing device, and the dependent variable; a reliability calculation unit that calculates the reliability of the dependent variable based on process data obtained from the material processing device and indicating the state of the raw material processing process; a dataset processing unit that processes the dataset based on the reliability; a predictive model generation unit that generates a predictive model indicating the relationship between the explanatory variables and the dependent variable using a processed dataset that is the processed dataset; and a search condition derivation unit that performs a search using the predictive model to derive values ​​of the explanatory variables as search conditions for obtaining a target value of the dependent variable.

[0006] These comprehensive or specific aspects may be realized as a system, a method, an integrated circuit, a computer program, or a computer-readable recording medium such as a CD-ROM, or may be realized as any combination of a system, a method, an integrated circuit, a computer program, and a recording medium. The recording medium may also be a non-transitory recording medium.

[0007] Further advantages and effects of one aspect of the present disclosure will become apparent from the specification and drawings. Such advantages and / or effects are provided by some of the embodiments and configurations described in the specification and drawings, but not necessarily all of the configurations.

[0008] FIG. 1 is a diagram illustrating an example of a search system according to an embodiment. FIG. 2 is a block diagram illustrating an example of the configuration of a material processing apparatus according to an embodiment. FIG. 3 is a block diagram illustrating an example of the configuration of a search device according to an embodiment. FIG. 4A is a diagram illustrating a relationship between process data and a prediction model. FIG. 4B is a diagram illustrating another relationship between process data and a prediction model. FIG. 5 is a diagram illustrating an example of a data set according to an embodiment. FIG. 6 is a diagram illustrating an example of a process abnormality degree according to an embodiment. FIG. 7 is a diagram illustrating an example of a reliability and a reliability ratio according to an embodiment. FIG. 8 is a diagram illustrating weighting of learning amounts according to an embodiment. FIG. 9 is a diagram illustrating a response variable included in a new combination according to an embodiment. FIG. 10 is a diagram illustrating an example of a processed data set according to an embodiment. FIG. 11 is a diagram illustrating an example of processing by a prediction model generation unit according to an embodiment. FIG. 12 is a diagram illustrating search using a prediction model according to an embodiment. FIG. 13 is a flowchart illustrating an example of processing operations of the search device according to an embodiment.

[0009] (Findings that form the basis of the present invention) The present inventor has found that the above-mentioned Patent Document 1, described in the "Background Art" section, has the following problem that an efficient search may not be possible.

[0010] The search device of Patent Document 1 updates or generates a prediction model by using characteristic values ​​of products obtained by processing in a semiconductor processing apparatus as output parameters, regardless of whether the processing is stable or not. Here, output parameters obtained by stable processing are highly reliable, while output parameters obtained by unstable processing are unreliable. If such unreliable output parameters are used as training data to generate a prediction model using machine learning such as a neural network, an erroneous prediction model is generated. As a result, even if such a prediction model is used, it may not be possible to arrive at an optimal solution, or it may take a long time to arrive at the optimal solution. In other words, efficient search becomes difficult.

[0011] For example, a conventional material processing apparatus known as a microreactor, which has a flow channel forming body, is known as a material processing apparatus that contacts and mixes mutually soluble liquids (i.e., reactants) to react with each other to produce a desired product. This microreactor generates a product by flowing multiple types of liquids to be mixed through a microchannel. Therefore, depending on the processing conditions, the products may deposit, adhere, or block the flow channel within the microchannel, making it difficult to achieve a stable reaction. Therefore, when performing a search using the search device of Patent Document 1 on such a material processing apparatus, efficient search may be significantly difficult.

[0012] The present disclosure provides a search device capable of efficient search.

[0013] A search device according to a first aspect of the present disclosure includes a dataset update unit that generates or updates a dataset that associates a dependent variable indicating a characteristic value of a product obtained by processing raw materials according to explanatory variables using a material processing device, and the dependent variable; a reliability calculation unit that calculates the reliability of the dependent variable based on process data obtained from the material processing device and indicating the state of the raw material processing process; a dataset processing unit that processes the dataset based on the reliability; a predictive model generation unit that generates a predictive model indicating the relationship between the explanatory variables and the dependent variable using a processed dataset that is the processed dataset; and a search condition derivation unit that performs a search using the predictive model to derive values ​​of the explanatory variables as search conditions for obtaining a target value of the dependent variable.

[0014] The search device of the present disclosure can perform efficient searches. Specifically, a dataset is processed based on the reliability of the dependent variable, and a predictive model is generated using the processed dataset. Therefore, for example, it is possible to reduce the influence of dependent variable values ​​with low reliability on the generation of the predictive model, and increase the influence of dependent variable values ​​with high reliability on the generation of the predictive model. As a result, even if the processing of raw materials is unstable, i.e., even if unreliable dependent variable values ​​are obtained, a predictive model with high prediction accuracy can be generated. As a result, even in such cases, the predictive model can be used to efficiently search for explanatory variable values. Furthermore, the search speed can be improved.

[0015] In addition, in the search device according to a second aspect, the prediction model generation unit may generate the prediction model by machine learning using the processed dataset, and the dataset processing unit may process the dataset by weighting a learning amount, which is the number of combinations of values ​​of the explanatory variables and values ​​of the target variable, according to the reliability. Note that the second aspect may be dependent on the first aspect.

[0016] This allows the amount of learning required to generate a predictive model through machine learning to be weighted according to the reliability. Therefore, for example, it is possible to reduce the influence of a value of a dependent variable with low reliability on the machine learning, and to increase the influence of a value of a dependent variable with high reliability on the machine learning. In other words, it is possible to prevent overlearning for values ​​of dependent variables with low reliability. Therefore, it is possible to appropriately generate a predictive model with high prediction accuracy through machine learning.

[0017] In addition, in the searching device according to a third aspect, the reliability calculation unit may calculate, for each existing combination that is the combination included in the dataset, a reliability of the value of the objective variable based on the process data when the value of the objective variable included in the existing combination is obtained, and the dataset processing unit may weight the learning amount by changing, for each reliability, the number of combinations that include the value of the explanatory variable corresponding to the reliability in accordance with the reliability. Note that the third aspect may be dependent on the second aspect.

[0018] As a result, the processed dataset includes, for each existing combination included in the dataset, a number of new combinations that include the same explanatory variable values ​​as the existing combination, the number of which corresponds to the reliability. As a result, the number of combinations can be adjusted according to the reliability, so that, for example, the influence of a dependent variable value with low reliability on the machine learning can be effectively reduced, and the influence of a dependent variable value with high reliability on the machine learning can be effectively increased. In other words, overfitting for dependent variable values ​​with low reliability can be effectively suppressed. Therefore, a predictive model with high predictive accuracy can be more appropriately generated by machine learning.

[0019] In the search device according to a fourth aspect, when generating a plurality of new combinations from the existing combinations due to an increase in the number of combinations, the dataset processing unit may include, in each of at least one of the plurality of new combinations, a value different from the value of the objective variable included in the existing combination as a new value of the objective variable. Note that the fourth aspect may be dependent on the second or third aspect.

[0020] As a result, values ​​of the objective variables having low reliability, for example, included in the existing combinations, are replaced with values ​​different from the existing values ​​and included in the new combinations, thereby more effectively suppressing overlearning for the values ​​of the objective variables having low reliability.

[0021] In addition, in the searching device according to a fifth aspect, the reliability calculation unit may calculate the reliability to be a smaller value as the fluctuation of the value indicated by the process data is larger, and the data set processing unit may change the number of combinations by increasing the number of combinations as the reliability is larger. Note that the fifth aspect may be dependent on any one of the second to fourth aspects.

[0022] This makes it possible to effectively prevent over-learning for dependent variables with low reliability.

[0023] In a sixth aspect of the present invention, the data set processing unit may select new values ​​of the objective variable included in each of the plurality of new combinations according to a probability distribution having an average value of the objective variable included in the existing combination. Note that the sixth aspect may be dependent on the fourth or fifth aspect.

[0024] This allows the values ​​of the objective variables included in the existing combination to be appropriately replaced with values ​​different from the existing values ​​and included in the new combination.

[0025] In the searching device according to a seventh aspect, the dataset processing unit may use a normal distribution as the probability distribution and randomly select a new value of the objective variable according to the normal distribution. Note that the seventh aspect may be dependent on the sixth aspect.

[0026] This allows the values ​​of the objective variables included in the existing combination to be appropriately replaced with values ​​different from the existing values ​​and included in the new combination.

[0027] In the search device according to an eighth aspect, the data set processing unit may use the normal distribution having a smaller variance as the reliability increases. Note that the eighth aspect may be dependent on the seventh aspect.

[0028] This allows a value of a dependent variable with low reliability included in an existing combination to be replaced with a value that is significantly different from that value and included in a new combination, and a value of a dependent variable with high reliability to be replaced with a value close to that value and included in a new combination, thereby making it possible to more effectively suppress overlearning for values ​​of dependent variables with low reliability.

[0029] Hereinafter, the embodiments will be specifically described with reference to the drawings.

[0030] The embodiments described below are all comprehensive or specific examples. The numerical values, shapes, materials, components, component placement and connection configurations, steps, and step order shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Furthermore, among the components in the following embodiments, components that are not described in the independent claims that represent the highest concepts are described as optional components. Furthermore, each drawing is a schematic diagram and is not necessarily an exact illustration. Furthermore, the same components are designated by the same reference numerals in each drawing.

[0031] (Embodiment) FIG. 1 is a diagram showing an example of a search system according to this embodiment.

[0032] The searching system 1000 in this embodiment includes a searching device 100 and a material processing device 200 .

[0033] The material processing apparatus 200 is an apparatus that processes raw materials to produce a product. For example, the material processing apparatus 200 performs flow synthesis by flowing multiple types of liquids (e.g., a first liquid and a second liquid) that are raw materials through a flow path and sequentially reacting the liquids within the flow path. The liquids flowing within the flow path may be heated or cooled. A product is produced by such flow synthesis. The material processing apparatus 200 in this embodiment may also be referred to as a reactor or a chemical reaction apparatus. The material processing apparatus 200 in this embodiment is not limited to the example shown in FIG. 1 , and may be an apparatus having another configuration, such as a semiconductor processing apparatus, as long as it processes raw materials.

[0034] The searching device 100 searches for processing conditions for the material processing apparatus 200 to obtain target characteristic values ​​(also called physical property values) of the product. In other words, the searching device 100 searches for optimal solutions for the processing conditions. The processing conditions are expressed as explanatory variables or input parameters, and the characteristic values ​​of the product are expressed as response variables or output parameters. The searching device 100 inputs or sets explanatory variables to the material processing apparatus 200 and obtains, from the material processing apparatus 200, response variables obtained by processing raw materials according to the explanatory variables. At this time, the searching device 100 also obtains process data (described below) from the material processing apparatus 200 along with the response variables. The searching device 100 then generates a prediction model using the known explanatory variables, response variables, and process data, and uses the prediction model to derive processing conditions for obtaining the target characteristic values, i.e., the values ​​of the explanatory variables, as optimal solutions. The values ​​of the explanatory variables (processing conditions) are also called search conditions. In a specific example, the search device 100 is configured as a personal computer or a server, and is connected to input devices such as a keyboard and a mouse, and a display device such as a display.

[0035] FIG. 2 is a block diagram showing an example of the configuration of a material processing apparatus 200 according to this embodiment.

[0036] The material processing apparatus 200 in this embodiment is a reaction apparatus that produces a liquid mixture as a product by mixing and synthesizing a first liquid and a second liquid that are soluble in each other. The material processing apparatus 200 includes a first liquid supply unit 210, a second liquid supply unit 220, a mixing channel unit 230, a process data acquisition unit 240, a target variable acquisition unit 250, and a recovery unit 260. In FIG. 2, the flow of the liquid is indicated by solid arrows.

[0037] The first liquid supply unit 210 supplies the first liquid to the mixing channel unit 230 via the first liquid channel 213. Specifically, the first liquid supply unit 210 has a first liquid container 211 and a first liquid pump 212. The first liquid container 211 is a container that stores the first liquid. The first liquid pump 212 pumps out the first liquid from the first liquid container 211 and pumps it to the mixing channel unit 230 via the first liquid channel 213.

[0038] The second liquid supply unit 220 supplies the second liquid to the mixing channel unit 230 via the second liquid channel 223. Specifically, the second liquid supply unit 220 has a second liquid container 221 and a second liquid pump 222. The second liquid container 221 is a container that stores the second liquid. The second liquid pump 222 pumps out the second liquid from the second liquid container 221 and pumps it to the mixing channel unit 230 via the second liquid channel 223.

[0039] The mixing channel section 230 has a mixing channel for sequentially mixing and flowing the first liquid flowing in the first liquid channel 213 and the second liquid flowing in the second liquid channel 223. The first liquid and the second liquid are mixed to produce a mixed liquid, which flows through the mixing channel. The mixed liquid is also called a synthesized product, a product, or a product. The mixing channel section 230 may have a channel configured by, for example, a microreactor and a micromixer. In this case, the diameter of the channel is several μm to several mm, and the material processing apparatus 200 can be said to be a microflow reaction apparatus for performing microflow synthesis.

[0040] The process data acquisition unit 240 acquires process data indicating the state of the raw material processing process, i.e., the state of the mixing process of the first liquid and the second liquid, for example, by measurement using a sensor. Specifically, the process data indicates, for example, the pressure of the mixed liquid flowing through the mixing channel unit 230. The process data may also indicate the pressure of the first liquid flowing through the first liquid channel 213 or the pressure of the second liquid flowing through the second liquid channel 223. The process data may also indicate the flow rate, viscosity, pH, temperature, etc. of at least one of the first liquid, the second liquid, and the mixed liquid, or may indicate other numerical values.

[0041] The objective variable acquisition unit 250 has a sensor and acquires objective variables indicating characteristic values ​​of the product, which is a mixed liquid, by using the sensor. For example, the characteristic values ​​may be the particle size, particle size distribution, absorbance, etc. of the product, but are not limited to these and may be other numerical values. The acquired objective variables indicate values ​​actually measured by the above-mentioned sensor.

[0042] The recovery unit 260 recovers the product, which is a mixed liquid. Note that the above-described objective variable acquisition unit 250 may acquire the objective variables of the product recovered by the recovery unit 260.

[0043] Here, the material processing apparatus 200 in this embodiment mixes the first liquid and the second liquid as a process for raw materials in accordance with explanatory variables input or set by the searching device 100. The explanatory variables indicate, for example, the heating temperature, the concentration of the raw materials, the flow rate of the raw materials, etc. Note that the explanatory variables are not limited to these and may indicate other numerical values. The heating temperature is, for example, at least one of the set temperature at which the first liquid supply unit 210 heats the first liquid and the set temperature at which the second liquid supply unit 220 heats the second liquid. In this case, the first liquid supply unit 210 and the second liquid supply unit 220 may be equipped with a thermostatic bath or the like capable of controlling the temperatures of the first liquid and the second liquid in accordance with the explanatory variables. Alternatively, the heating temperature is, for example, the set temperature at which the mixed liquid is heated. In this case, the mixing channel unit 230 may be equipped with a thermostatic bath or the like capable of controlling the temperature of the mixed liquid in accordance with the explanatory variables. Furthermore, the concentration of the raw materials is, for example, the concentration of at least one of the first liquid and the second liquid. In this case, the first liquid supply unit 210 and the second liquid supply unit 220 may have a mechanism capable of adjusting the concentration of the raw material in accordance with the explanatory variable. Furthermore, the flow rate of the raw material is, for example, the flow rate of at least one of the first liquid and the second liquid. In this case, the first liquid supply unit 210 and the second liquid supply unit 220 may adjust the flow rate of the supplied liquid in accordance with the explanatory variable. Furthermore, the mixing ratio of the first liquid and the second liquid may be adjusted by adjusting the concentration or flow rate of the raw material.

[0044] 2 mixes two types of liquids, the number of types of liquids to be mixed is not limited to two and may be three or more. For example, if the number of types of liquids is three or more, the material processing apparatus 200 is provided with liquid supply units equal to the number of types of liquids. These liquid supply units have the same configuration as the first liquid supply unit 210 and the second liquid supply unit 220, and supply the type of liquid corresponding to the liquid supply unit to the mixing channel unit 230.

[0045] In addition, although the main component of the first liquid and the second liquid in this embodiment is water, the first liquid and the second liquid may be either water-soluble or water-insoluble as long as they are soluble in each other. For example, the first liquid and the second liquid may be water or an aqueous solution, or may be an organic solvent or an oil-based liquid.

[0046] Furthermore, the process data acquisition unit 240 may be connected to the first liquid supply unit 210, the second liquid supply unit 220, or the recovery unit 260, and may acquire process data from two or more sources.

[0047] FIG. 3 is a block diagram showing an example of the configuration of the searching device 100 according to this embodiment.

[0048] The search device 100 includes a dataset update unit 110 , a storage unit 120 , a reliability calculation unit 130 , a dataset processing unit 140 , a prediction model generation unit 150 , a search condition derivation unit 160 , and a device control unit 170 .

[0049] The dataset update unit 110 acquires explanatory variables from the apparatus control unit 170 and acquires response variables and process data obtained by mixing according to the explanatory variables from the material processing apparatus 200. That is, the dataset update unit 110 acquires a combination including values ​​indicated by the explanatory variables, values ​​indicated by the response variables, and process data. The values ​​indicated by the explanatory variables are also called explanatory variable values, and the values ​​indicated by the response variables are also called response variable values. The dataset update unit 110 then generates or updates a dataset 121 including the combination. That is, if the dataset 121 is not stored in the storage unit 120, the dataset update unit 110 generates a dataset 121 including the combination and stores it in the storage unit 120. On the other hand, if the dataset 121 is already stored in the storage unit 120, the dataset update unit 110 reads the existing dataset 121 from the storage unit 120 and updates the dataset 121 stored in the storage unit 120 by adding the acquired combination to the existing dataset 121. Such a data set 121 includes one or more of the above-described combinations, thereby indicating the association between the explanatory variables, the response variables, and the process data. Note that, in the present embodiment, the process data is included in the data set 121, but it may also be stored in the storage unit 120 without being included in the data set 121.

[0050] In this way, the dataset update unit 110 in this embodiment generates or updates a dataset 121 that shows, in association with the explanatory variables, the objective variables that indicate the characteristic values ​​of the product obtained by processing the raw materials by the material processing device 200 according to the explanatory variables.

[0051] The storage unit 120 is a recording medium for storing the dataset 121 and the like. The storage unit 120 in this embodiment has a storage capacity for storing not only the dataset 121 but also the processed dataset 122 and the prediction model 123. For example, the storage unit 120 is a hard disk drive, a RAM (Random Access Memory), a ROM (Read Only Memory), or a semiconductor memory. Note that such a storage unit 120 may be volatile or non-volatile.

[0052] The reliability calculation unit 130 reads out the data set 121 from the storage unit 120, and for each combination included in the data set 121, calculates the reliability of the value of the objective variable included in that combination based on the process data included in that combination. The reliability calculation unit 130 then outputs the reliability calculated for each combination to the data set processing unit 140. In this way, the reliability calculation unit 130 in this embodiment calculates the reliability of the objective variable based on the process data indicating the state of the raw material processing process, which is obtained from the material processing apparatus 200.

[0053] The dataset processing unit 140 acquires the reliability for each combination from the reliability calculation unit 130, and processes the dataset 121 based on the reliability. The dataset processing unit 140 generates a processed dataset 122 through the processing, and stores the generated processed dataset 122 in the storage unit 120.

[0054] The prediction model generation unit 150 reads the processed dataset 122, which is the processed dataset 121, from the storage unit 120, and uses the processed dataset 122 to generate a prediction model 123 that indicates the relationship between explanatory variables and dependent variables. This prediction model 123 is, for example, a model generated by machine learning so that, when a value of a dependent variable is input, values ​​of explanatory variables that result in that value are output. In other words, the prediction model generation unit 150 generates the prediction model 123 by machine learning using the processed dataset 122. The prediction model 123 is, for example, a neural network. The prediction model generation unit 150 stores the generated prediction model 123 in the storage unit 120.

[0055] The search condition derivation unit 160 reads the prediction model 123 from the storage unit 120 and performs a search using the prediction model 123 to derive, as search conditions, values ​​of explanatory variables for obtaining a target value of the objective variable. The search condition derivation unit 160 then outputs the derived search conditions, that is, values ​​of explanatory variables, to the device control unit 170. The search conditions are used as conditions for processing by the material processing device 200, i.e., the processing conditions described above.

[0056] The device control unit 170 controls the material processing apparatus 200. Specifically, when the device control unit 170 acquires the values ​​of the explanatory variables as processing conditions from the search condition derivation unit 160, it outputs the values ​​of the explanatory variables to the material processing apparatus 200 and causes the material processing apparatus 200 to perform mixing according to the explanatory variables. The values ​​of the explanatory variables, which are processing conditions, are also output from the device control unit 170 to the dataset update unit 110. As a result, the material processing apparatus 200 outputs new values ​​of the objective variables obtained according to the derived values ​​of the explanatory variables and new process data indicating the state of the mixing process to the dataset update unit 110.

[0057] The dataset update unit 110 determines whether the new value of the objective variable has reached the target value, and if it determines that the target value has not been reached, updates the dataset 121 using the new value of the objective variable in the same manner as described above. On the other hand, if the dataset update unit 110 determines that the target value has been reached, it terminates the process of searching for explanatory variables. In other words, as a result of the search, the value of the explanatory variable derived immediately before is determined as the optimal solution.

[0058] Here, if a prediction model is generated from the explanatory variables and the target variable shown in the dataset 121 instead of the processed dataset 122, the prediction model may perform an inappropriate search for the explanatory variables. For example, in the example of Fig. 4A below, an appropriate search is performed, but in the example of Fig. 4B, an inappropriate search is performed.

[0059] 4A is a diagram illustrating the relationship between the process data and the prediction model when the processing of raw materials in the material processing apparatus 200 is stable.

[0060] For example, as shown in (a) of Figure 4A, the mixed liquid flows smoothly through the flow path. That is, even if the mixed liquid contains particles, the particles flow without accumulating or adhering to the inner surface of the flow path. Note that the flow path is, for example, the flow path from the mixing flow path section 230 to the objective variable acquisition section 250 in the material processing apparatus 200.

[0061] In such a case, the data set update unit 110 acquires, for example, process data shown in FIG. 4A (b) from the process data acquisition unit 240 of the material processing apparatus 200. This process data indicates, for example, the pressure of the mixed liquid flowing through the flow path. Note that the horizontal axis of the graph shown in FIG. 4A (b) indicates time, and the vertical axis indicates pressure. In the process data shown in FIG. 4A (b), the pressure is stable without significant fluctuations. In other words, the process of mixing the first liquid and the second liquid by the material processing apparatus 200 is stable.

[0062] Here, when mixing is performed under a plurality of processing conditions by the material processing apparatus 200, if the process data (i.e., pressure) is stable as shown in (b) of Figure 4A, highly reliable values ​​of the objective variable for those processing conditions can be obtained by the material processing apparatus 200. Therefore, by using the explanatory variable x, which is the processing condition, and the objective variable y obtained from that explanatory variable x, a prediction model sufficiently close to the true function can be generated, as shown in (c) of Figure 4A.

[0063] The horizontal axis of the graph shown in (c) of Figure 4A indicates the explanatory variable x, and the vertical axis indicates the response variable y. The solid line in (c) of Figure 4A indicates a prediction model (i.e., a predicted function) generated from multiple actual measurement values. The actual measurement values ​​are a pair of the value of the explanatory variable x used in the processing of the material processing apparatus 200 and the value of the response variable y, which is an actually measured characteristic value of the mixed liquid generated by the processing using the value of the explanatory variable x. The dashed line in (c) of Figure 4A indicates the true relationship between the explanatory variable x and the response variable y, i.e., the true function.

[0064] Therefore, by using such a prediction model, the value of the explanatory variable x for obtaining the target value (e.g., the maximum value) of the objective variable y can be appropriately derived as a candidate for the next experiment (i.e., a search condition).

[0065] 4B is a diagram illustrating another relationship between the process data and the prediction model, specifically, when the processing of the raw material in the material processing apparatus 200 is unstable.

[0066] For example, the mixed liquid does not flow smoothly through the flow path, as shown in (a) of Figure 4B. Specifically, particles contained in the mixed liquid repeatedly flow and stay within the flow path. When the particles stay, they accumulate within the flow path or adhere to the inner surface of the flow path.

[0067] In such a case, the data set update unit 110 acquires, for example, the process data shown in FIG. 4B(b) from the process data acquisition unit 240 of the material processing apparatus 200. This process data also indicates the pressure of the mixed liquid flowing through the flow path, as in the example of FIG. 4A. Note that the horizontal axis of the graph shown in FIG. 4B(b) indicates time, and the vertical axis indicates pressure. In the process data shown in FIG. 4B(b), the pressure fluctuates greatly. In other words, the process of mixing the first liquid and the second liquid by the material processing apparatus 200 is unstable.

[0068] Here, when mixing is performed under a plurality of processing conditions by the material processing apparatus 200, if the process data (i.e., pressure) is unstable as shown in (b) of Figure 4B, unreliable response variable values ​​for those processing conditions are obtained by the material processing apparatus 200. Therefore, if the explanatory variable x, which is the processing condition, and the response variable y obtained from that explanatory variable x are used, a prediction model that is far from the true function or that is significantly different from the true function may be generated, as shown in (c) of Figure 4B.

[0069] The horizontal axis of the graph shown in Figure 4B(c) represents the explanatory variable x, and the vertical axis represents the response variable y. The solid line in Figure 4B(c) represents a prediction model (i.e., a predicted function) generated from multiple actual measurement values. The actual measurement values ​​are a pair of the value of the explanatory variable x used in the processing of the material processing apparatus 200 and the value of the response variable y, which is an actually measured characteristic value of the mixed liquid generated by the processing using the value of the explanatory variable x. The dashed line in Figure 4B(c) represents the true relationship between the explanatory variable x and the response variable y, i.e., the true function.

[0070] Therefore, if such a prediction model is used to derive the value of explanatory variable x as a candidate for the next experiment (i.e., search condition) in order to obtain a target value (e.g., maximum value) of dependent variable y, there is a possibility that the dependent variable y obtained will have a value that is significantly different from the target value. In other words, there is a possibility that an efficient search for dependent variable x will not be possible.

[0071] Therefore, in this embodiment, the dataset processing unit 140 processes the dataset 121 using the reliability to generate a processed dataset 122. Then, the prediction model generation unit 150 generates a prediction model 123 using the processed dataset 122 instead of the dataset 121.

[0072] FIG. 5 is a diagram showing an example of the data set 121. As shown in FIG.

[0073] 5, the data set 121 shows the values ​​of explanatory variable x, the values ​​of response variable y, and process data in association with each other. That is, the data set 121 includes a plurality of combinations, each of which includes the value of explanatory variable x, the value of response variable y, and process data. The values ​​of explanatory variable x and the value of response variable y included in these combinations correspond to the above-mentioned actual measurement values. For example, the data set 121 shows the values ​​of explanatory variable x=x １ and a combination including the objective variable y=0.6 and the process data “Pd1”, and the explanatory variable x=x ２ The combinations include a combination including an explanatory variable x=0.4 and process data "Pd2". Note that these combinations include the values ​​of one type of explanatory variable x and one type of objective variable y, but the number of types of each of the explanatory variables x and objective variables y included in the combination may be two or more. １ ~x ５ are the values ​​of the explanatory variable x, respectively.

[0074] FIG. 6 is a diagram illustrating an example of the process abnormality degree.

[0075] As shown in FIG. 6 , for each combination included in the data set 121, the reliability calculation unit 130 calculates a process anomaly degree σ, which is the degree of anomaly of the value of the objective variable y included in the combination, using the process data included in the combination. The process anomaly degree σ indicates a larger value as the fluctuation of the process data increases, that is, as the process data becomes more unstable. In a specific example, the reliability calculation unit 130 calculates the standard deviation of the process data as the process anomaly degree σ. Alternatively, the reliability calculation unit 130 may calculate the difference between the maximum and minimum values ​​of the process data as the process anomaly degree σ, or may calculate the coefficient of variation, which is the ratio of the standard deviation to the average, as the process anomaly degree σ. Alternatively, the reliability calculation unit 130 may calculate a value indicated by a combination of at least two of the standard deviation, the difference, and the coefficient of variation as the process anomaly degree σ. Furthermore, the reliability calculation unit 130 may calculate the process anomaly degree σ by machine learning, such as a k-nearest neighbor method or a local outlier factor method.

[0076] FIG. 7 is a diagram illustrating an example of reliability and reliability ratio.

[0077] After calculating the process abnormality degree σ for each of the above-described combinations, the reliability calculation unit 130 calculates the reliability r of the value of the objective variable y corresponding to the process abnormality degree σ based on the process abnormality degree σ, as shown in Fig. 7. For example, the reliability calculation unit 130 calculates the reciprocal of the process abnormality degree σ as the reliability r.

[0078] Next, the dataset processing unit 140 obtains the reliability r calculated by the reliability calculation unit 130 for each of all values ​​of the response variable y and calculates the ratio of these reliability r to the total reliability. That is, the dataset processing unit 140 calculates an integrated value by integrating all the reliability r, and then calculates the ratio of the reliability r by dividing the reliability r by the integrated value. For example, the reliability calculation unit 130 calculates reliability r = 3.3, 2.5, 5.0, 2.5, and 1.0 for five values ​​of the response variable y, respectively. In this case, since the reliability r for the response variable y = 0.6 is 3.3, the dataset processing unit 140 calculates the ratio of the reliability r for the response variable y = 0.6 as 3.3 / (3.3 + 2.5 + 5.0 + 2.5 + 1.0) = 0.23. Similarly, the data set processing unit 140 calculates the reliability r ratios "0.17, 0.35, 0.17, 0.07" for the objective variable y = 0.4, 0.3, 0.5, 0.2, respectively.

[0079] The dataset processing unit 140 uses the ratio of the reliability r calculated in this way to weight the number of combinations included in the dataset 121. In other words, the dataset processing unit 140 processes the dataset 121 by weighting the amount of learning, which is the number of combinations of the value of the explanatory variable x and the value of the target variable y, according to the reliability r.

[0080] FIG. 8 is a diagram for explaining weighting of the amount of learning.

[0081] For example, as shown in FIG. 8A, the data set 121 includes first to nth combinations (n ​​is an integer equal to or greater than 4, for example), that is, n combinations. These combinations are existing combinations (also called existing combinations). For example, the first combination is a combination of explanatory variables x = x １ and the objective variable y = y １ and the response variable y=y １ The reliability r of １ Similarly, the second combination is calculated as follows: ２ and the objective variable y = y ２ and the response variable y=y ２The reliability r of ２ Here, the reliability r of n is calculated as r １ , r ２ , ..., r ｎ In １ >... >r ｎ－１ >r ｎ >r ２ It is assumed that there is a relationship between the reliability r and the number of new combinations that can be obtained from an existing combination. The dataset processing unit 140 generates the processed dataset 122 as shown in FIG. 8(b) by increasing the number of new combinations that can be obtained from an existing combination that has a higher reliability r. Specifically, the dataset processing unit 140 increases the number of new combinations that can be obtained from an existing combination that has a higher ratio of reliability r. Note that the width of the blocks corresponding to each of the n existing combinations shown in FIG. 8(b) indicates the ratio of the number of new combinations that can be obtained from that existing combination. In other words, reliability r = r １ The number of new combinations corresponding to the reliability r is greater than the number of new combinations corresponding to any other reliability r, and the reliability r = r ２ The number of new combinations corresponding to the reliability r is smaller than the number of combinations corresponding to any other reliability r. Note that a new combination obtained from an existing combination contains the same value of the explanatory variable x as that existing combination. The new combination is also called a new combination.

[0082] As described above, in this embodiment, the reliability calculation unit 130 calculates, for each existing combination included in the dataset 121, the reliability r of the value of the objective variable y based on the process data when the value of the objective variable y included in that existing combination is obtained. Then, the dataset processing unit 140 weights the amount of learning described above by changing, for each reliability r, the number of combinations including the value of the explanatory variable x corresponding to that reliability r in accordance with that reliability r. Specifically, the reliability calculation unit 130 calculates a smaller reliability r as the fluctuation of the value indicated by the process data increases. Then, the dataset processing unit 140 changes the number of combinations by increasing the number of combinations as the reliability r increases.

[0083] FIG. 9 is a diagram for explaining the objective variable y included in the new combination.

[0084] The data set processing unit 140 generates one or more new combinations from the existing combinations. At this time, the data set processing unit 140 calculates the square of the process abnormality degree σ corresponding to the existing combination as the variance σ ２ and has as its mean the value of the objective variable y included in the existing combination. This probability distribution is, for example, a normal distribution.

[0085] In a specific example, the existing combination is the explanatory variable x = x １ and the objective variable y = y １ The process abnormality degree σ corresponding to the existing combination is σ １ Therefore, the data set processing unit 140 processes the existing combinations as shown in FIG. １ , σ １ ２ ) is used. A normal distribution expressed in this way has a mean y １ and variance σ １ ２ Then, the data set processing unit 140 calculates the objective variable y=y １ The calculated reliability r = r １ The values ​​of the objective variable y included in the new combinations are randomly selected in the number corresponding to the ratio of the explanatory variable x = x １ Includes.

[0086] In addition, other existing combinations are explanatory variables x = x ２ and the objective variable y = y ２ The process abnormality degree σ corresponding to the existing combination is σ ２ Therefore, the data set processing unit 140 processes the existing combinations as shown in FIG. ２ , σ ２ ２ ) is used. A normal distribution expressed in this way has a mean y ２and variance σ ２ ２ Then, the data set processing unit 140 calculates the objective variable y=y ２ The calculated reliability r = r ２ The values ​​of the objective variable y included in the new combinations are randomly selected in the number corresponding to the ratio of the explanatory variable x = x ２ Includes.

[0087] As described above, in this embodiment, the dataset processing unit 140 selects a new value of the objective variable y to be included in each of a plurality of new combinations according to a probability distribution having, as an average, the value of the objective variable y included in the existing combination. Specifically, the dataset processing unit 140 uses a normal distribution as the probability distribution and randomly selects a new value of the objective variable y according to the normal distribution. In other words, when generating a plurality of new combinations from existing combinations due to an increase in the number of combinations, the dataset processing unit 140 includes, in at least one of the plurality of new combinations, a value different from the value of the objective variable y included in the existing combination as a new value of the objective variable y.

[0088] In the example of FIG. 9A, the reliability r=r １ is large, so the variance σ １ ２ Therefore, the distribution width of the normal distribution is small. Therefore, the data set processing unit 140 calculates the mean y １ a value relatively close to y １ In this case, the new combinations are generated by adding the values ​​of the objective variable y (i.e., values ​​similar to the average y) included in the existing combinations. １ ) may be included as a new value for the objective variable y.

[0089] On the other hand, in the example of FIG. 9B, the reliability r=r ２ Since is small, the variance σ ２ ２Therefore, the distribution width of the normal distribution is large. Therefore, the data set processing unit 140 calculates the mean y ２ values ​​relatively far from the mean y ２ In addition, the new combination may include the value of the objective variable y (i.e., a value that is not similar to the average y ２ ) may be included as a new value of the objective variable y. The dataset processing unit 140 generates a smaller number of such new combinations than in the example of FIG. 9(a).

[0090] As described above, in this embodiment, the data set processing unit 140 uses a normal distribution whose variance decreases as the reliability r increases.

[0091] FIG. 10 is a diagram showing an example of the processed data set 122.

[0092] The processed data set 122 has explanatory variables x = x as shown in FIG. １ 23 new combinations including the explanatory variables x = x ２ 17 new combinations including the explanatory variables x = x ３ 35 new combinations including the explanatory variables x = x ４ 17 new combinations including the explanatory variables x = x ５ and seven new combinations including:

[0093] That is, the data set processing unit 140 calculates the explanatory variable x=x １ and from the existing combinations including the objective variable y = 0.6, the explanatory variable x = x １ In this case, since the ratio of the reliability r corresponding to the existing combinations is 0.23, the dataset processing unit 140 determines the number of new combinations to be generated from the existing combinations as 0.23 x 100 = 23. In other words, the dataset processing unit 140 calculates the number of new combinations as 100 times the ratio of the reliability r. Then, the dataset processing unit 140 calculates N(y １, σ １ ２ ) = N (0.6, 0.03), 23 values ​​of the objective variable y' are selected (or extracted). The value of the objective variable y' is the value of the objective variable y included in the new combination. As a result, 23 new combinations are generated. Similarly, the data set processing unit 140 selects (or extracts) 23 values ​​of the objective variable y' according to the normal distribution expressed by ２ and from the existing combinations including the objective variable y = 0.4, the explanatory variable x = x ２ In this case, since the ratio of the reliability r corresponding to the existing combinations is 0.17, the data set processing unit 140 determines the number of new combinations to be generated from the existing combinations as 0.17×100=17. Then, the data set processing unit 140 calculates N(y ２ , σ ２ ２ ) = N (0.4, 0.16), 17 values ​​of the objective variable y' are selected (or extracted). This generates 17 new combinations.

[0094] Here, the explanatory variable x = x ３ The ratio of reliability r corresponding to the existing combination including ５ The ratio of the reliability r corresponding to the existing combination including the explanatory variable x = x is the smallest. ３ The largest number of new combinations are generated from existing combinations containing explanatory variable x = x, and the values ​​of the objective variable y' contained in these new combinations are relatively close to the values ​​of the objective variable y contained in the existing combinations. ５ The smallest number of new combinations are generated from the existing combinations including the above, and the values ​​of the objective variable y′ included in these new combinations include values ​​that are relatively far from the values ​​of the objective variable y included in the existing combinations.

[0095] As described above, in this embodiment, the dataset processing unit 140 generates new combinations from existing combinations included in the dataset 121, the number of which corresponds to the reliability r of the existing combinations. That is, the dataset processing unit 140 generates more new combinations as the reliability r increases. In other words, the dataset processing unit 140 generates a larger number of new combinations than the existing combinations by weighting the number of existing combinations included in the dataset 121 according to the reliability r. This increases the number of combinations. Because many new combinations are generated from existing combinations with high reliability r, the amount of machine learning learning by the predictive model generation unit 150 based on those existing combinations can be increased. On the other hand, because a smaller number of new combinations are generated from existing combinations with low reliability r, overlearning in machine learning by the predictive model generation unit 150 based on those existing combinations can be suppressed.

[0096] In the above example, the number of new combinations is determined by multiplying the ratio of reliability r by 100, but the ratio of reliability r may also be multiplied by a value greater than 100. This generates more new combinations, making it possible to further suppress the above-mentioned overlearning. Alternatively, the number of new combinations may be determined by multiplying the ratio of reliability r by a value smaller than 100. In this case, the number of new combinations can be reduced compared to when multiplying by 100, thereby reducing the processing load on the searching device 100. Furthermore, the time required for searching by the searching device 100 can be reduced.

[0097] FIG. 11 is a diagram illustrating an example of processing by the prediction model generating unit 150.

[0098] To generate the prediction model 123 by ensemble learning, the prediction model generation unit 150 divides the processed dataset 122 into, for example, m subsets (m is an integer equal to or greater than 2). Then, for each subset, the prediction model generation unit 150 generates a sub-prediction model, which is a prediction model corresponding to that subset. For example, when the explanatory variable x = x１ Each of the 23 new combinations including the first to mth subsets is classified into one of the first to mth subsets, and the explanatory variable x = x ２ Each of the 17 new combinations, including the first sub-prediction model and the mth sub-prediction model, is classified into one of the first to mth subsets. The prediction model generation unit 150 then generates a first sub-prediction model corresponding to the first subset from the first subset, and generates a second sub-prediction model corresponding to the second subset from the second subset. Similarly, sub-prediction models corresponding to the other subsets are generated. This generates m sub-prediction models, i.e., the first to mth sub-prediction models. Each sub-prediction model represents the relationship between the explanatory variable x and the objective variable y included in the subset corresponding to that sub-prediction model. The prediction model generation unit 150 generates such sub-prediction models using machine learning. Specific machine learning techniques include, for example, support vector machines, neural networks, and random forests. Machine learning is not limited to a single technique, and may be performed by combining two or more techniques.

[0099] Next, the prediction model generation unit 150 generates the prediction model 123 by integrating the first to m-th sub-prediction models.

[0100] Fig. 12 is a diagram for explaining a search using the prediction model 123. Note that the horizontal axis of each graph in Fig. 12(a) and (b) indicates the explanatory variable x, and the vertical axis indicates the objective variable y. Fig. 12(a) shows a comparative example of a search, and Fig. 12(b) shows an example of a search using the prediction model 123.

[0101] For example, explanatory variable x = x １ , x ２ , x ３ , x ４ , x ５ By mixing using the material processing apparatus 200 under the processing conditions, five values ​​of the objective variable y indicating the characteristic values ​​of the mixed liquid are obtained. １ , y １ ), (x２ , y ２ ), (x ３ , y ３ ), (x ４ , y ４ ), (x ５ , y ５ ) are obtained as actual measured values. In the example of FIG. 12(a), a prediction model is generated from these actual measured values. Here, as in the examples of FIGS. 4A and 4B, if the process data is unstable, the prediction model will show a function that is significantly different from the true function. In other words, in the machine learning for generating the prediction model, equal weights are used for combinations with high reliability r and combinations with low reliability r. Therefore, if the target value of the objective variable y is, for example, the maximum value, and the search range of the explanatory variable x is x = x １ ~x ５ In this case, the predictive model shown in FIG. 12(a) derives a value for the explanatory variable x, which is an incorrect search condition that differs from the optimal solution. In other words, in the example of FIG. 12(a), the reliability r is not taken into consideration, making it difficult to derive appropriate search conditions. The example of FIG. 12(a) can also be said to be an example of conventional search.

[0102] On the other hand, in this embodiment, as shown in (b) of Fig. 12, it is possible to appropriately derive the value of the explanatory variable x, which is a search condition. That is, when m = 5, for example, the prediction model generation unit 150 generates five sub-prediction models and integrates the five sub-prediction models to generate the prediction model 123. This prediction model 123 is a prediction model generated based on the processed data set 122, i.e., based on the reliability r. In generating the processed data set 122, the explanatory variable x = x １ and the objective variable y = y １ From the existing combinations including １ Using a probability distribution (specifically, a normal distribution) based on the reliability r of the explanatory variable x, the number of new combinations corresponding to the reliability r is generated. ２ and the objective variable y = y ２ From the existing combinations including ２Using a probability distribution (specifically, a normal distribution) based on the reliability r of each of the variables, a number of new combinations corresponding to the reliability r are generated. By weighting the existing combinations in this way, multiple new combinations are generated. In other words, a higher weight is assigned to combinations with a higher reliability r than to combinations with a lower reliability r. Then, a prediction model 123 is generated by machine learning using the new combinations generated by such weighting.

[0103] Therefore, even if the process data is unstable, the prediction model 123 shows a function closer to the true function than the prediction model in the example of FIG. 12A. Therefore, when the target value of the objective variable y is, for example, the maximum value and the search range of the explanatory variable x is, for example, x=x １ ~x ５ In this case, the search condition derivation unit 160 can use the prediction model 123 to derive an appropriate value of the explanatory variable x that is close to the optimal solution as the search condition.

[0104] In this way, in this embodiment, the normal distribution N(y, σ ２ ), the value of the objective variable y is processed into the value of the objective variable y′. This allows a value different from the value of the objective variable y to be used in machine learning when generating the prediction model 123, making it possible to generate a more flexible prediction model 123. In addition, ２ ), the larger the value of the objective variable y is for the process abnormality degree σ, the larger the value of the objective variable y' that differs from that value is included in the new combination and can be used for machine learning. Therefore, it is possible to suppress overlearning for values ​​of the objective variable y with low reliability r.

[0105] In the above example, the variance of the normal distribution used to select the value of the response variable y′ is the square of the process abnormality σ (i.e., σ ２ ) is used in the present example, but the variance is not limited to this and may be other values. For example, if the magnitude of the influence of a process abnormality (i.e., the process abnormality degree σ) on the objective variable y is known in advance, σ weighted according to the magnitude of the influence can be used. ２may be used for the variance of the normal distribution mentioned above.

[0106] In the above example, the value of the objective variable y′ is selected using the normal distribution N(y, σ ２ ) is used, but random numbers in the range of y±σ may also be used. In this case, values ​​smaller than y-σ and values ​​larger than y+σ are not selected as the value of the objective variable y'. Also, in the above example, a normal distribution is used, but other probability distributions may be used instead of the normal distribution.

[0107] Note that overlearning in this embodiment refers to, for example, learning that generates a predictive model that makes highly accurate predictions for data used in learning but cannot make similarly accurate predictions for unknown data, resulting in reduced search efficiency. In the above example, overlearning can be suppressed by using ensemble learning on the processed data set 122. Note that ensemble learning does not necessarily have to be used to generate a predictive model.

[0108] 13 is a flowchart showing an example of the processing operation of the searching device 100 according to this embodiment. Note that, in the example shown in FIG. 13, there is one type of explanatory variable x and one type of objective variable y, but there may be two or more types.

[0109] First, the search condition derivation unit 160 of the search device 100 acquires a target value of the objective variable y and a search range of the explanatory variable x (step S1). For example, the search condition derivation unit 160 may acquire the target value and the search range in response to an input operation on an input device such as a keyboard, or, if the target value and the search range are stored in the storage unit 120, may acquire them by reading them from the storage unit 120. Alternatively, if the search device 100 has a communication interface, the search condition derivation unit 160 may acquire the target value and the search range from outside the search device 100 via the communication interface.

[0110] Next, the dataset update unit 110 acquires an explanatory variable x from the apparatus control unit 170 through an initial experiment using the material processing apparatus 200, and acquires an objective variable y and process data obtained by mixing according to the explanatory variable x from the material processing apparatus 200. That is, the dataset update unit 110 acquires a combination including the value of the explanatory variable x and the value of the objective variable y. Then, the dataset update unit 110 generates a dataset 121 including the combination and stores it in the storage unit 120 (step S2). Note that the processing of step S2 may be performed before the processing of step S1 or may be performed simultaneously with the processing of step S1.

[0111] In the initial experiment, the apparatus control unit 170 may randomly select a value for the explanatory variable x from the search range acquired in step S1 and set that value in the material processing apparatus 200. Alternatively, the apparatus control unit 170 may select a value for the explanatory variable x using a statistical method such as experimental design. In step S2, the dataset update unit 110 may acquire two or more of the above-described combinations and generate a dataset 121 including those combinations. The number of combinations is not particularly limited and may be one or two. However, the greater the number of combinations, the higher the prediction accuracy of the prediction model 123 generated from the dataset 121.

[0112] Next, the reliability calculation unit 130 calculates the reliability r of the value of the objective variable y included in each combination shown in the dataset 121, i.e., for each existing combination (step S3). Then, the dataset processing unit 140 processes the dataset 121 using the reliability r calculated in step S3 to generate a processed dataset 122 (step S4).

[0113] Next, the prediction model generation unit 150 generates a prediction model 123 through machine learning using the processed dataset 122 generated in step S4 (step S5).Then, the search condition derivation unit 160 uses the prediction model 123 generated in step S5 to derive, as search conditions, values ​​of the explanatory variable x for obtaining a target value of the objective variable y from the search range acquired in step S1 (step S6).For example, the search condition derivation unit 160 derives the value of the explanatory variable x that maximizes or minimizes the value of the objective variable y.

[0114] Next, the apparatus control unit 170 sets the value of the explanatory variable x, which is the search condition derived in step S6, in the material processing apparatus 200, and causes the material processing apparatus 200 to perform processing according to the value of the explanatory variable x, i.e., mixing of the first liquid and the second liquid. As a result, the dataset update unit 110 obtains the value of the explanatory variable x from the apparatus control unit 170, and obtains the value of the response variable y and process data obtained by mixing according to the value of the explanatory variable x from the material processing apparatus 200. In other words, the dataset update unit 110 obtains a combination including the value of the explanatory variable x and the value of the response variable y as an actual measurement value (step S7).

[0115] Next, the dataset update unit 110 determines whether the value of the objective variable y included in the actual measurement value acquired in step S7 reaches the target value acquired in step S1 (step S8). For example, if the target value is a lower limit, the dataset update unit 110 determines that the value of the objective variable y has reached the target value if the value of the objective variable y is equal to or greater than the lower limit. Alternatively, if the target value is an upper limit, the dataset update unit 110 determines that the value of the objective variable y has reached the target value if the value of the objective variable y is equal to or less than the upper limit. Here, if the dataset update unit 110 determines that the value of the objective variable y has not reached the target value (No in step S8), it updates the dataset 121 stored in the storage unit 120 (step S9). That is, the dataset update unit 110 adds the combination including the value of the explanatory variable x and the value of the objective variable y acquired in step S7 to the dataset 121. Then, after the processing of step S9, the searching device 100 repeatedly executes the processing from step S3.

[0116] On the other hand, if the dataset update unit 110 determines in step S8 that the value of the objective variable y has reached the target value (Yes in step S8), it outputs the value of the explanatory variable x acquired in step S7 and ends the search for the explanatory variable x. When outputting the value of the explanatory variable x, the dataset update unit 110 may display the value on a display device connected to the searching device 100, or may store the value in the storage unit 120. Alternatively, the dataset update unit 110 may transmit the value to a device external to the searching device 100 via the above-mentioned communication interface.

[0117] As described above, in this embodiment, the data set 121 is processed based on the reliability r of the dependent variable y, and the prediction model 123 is generated using the processed data set 121. Therefore, for example, it is possible to reduce the influence of a value of the dependent variable y having a low reliability r on the generation of the prediction model 123, and to increase the influence of a value of the dependent variable y having a high reliability r on the generation of the prediction model 123. As a result, even if the processing of raw materials is unstable, that is, even if an unreliable value of the dependent variable y is obtained, it is possible to generate a prediction model 123 with high prediction accuracy. As a result, even in such cases, the value of the explanatory variable x can be efficiently searched for by using the prediction model 123. Furthermore, the search speed can be improved.

[0118] Furthermore, in this embodiment, the amount of learning required to generate the prediction model 123 by machine learning is weighted according to the reliability r. Therefore, for example, it is possible to reduce the influence of a value of the dependent variable y with a low reliability r on the machine learning, and to increase the influence of a value of the dependent variable y with a high reliability r on the machine learning. In other words, it is possible to prevent overlearning for values ​​of the dependent variable y with a low reliability r. Therefore, it is possible to appropriately generate a prediction model 123 with high prediction accuracy by machine learning.

[0119] Furthermore, in this embodiment, by using the weighting described above, the processed data set 122 includes, for each existing combination included in the data set 121, a number of new combinations corresponding to the reliability r, each of which includes the same value of the explanatory variable x as that existing combination. As a result, the number of combinations can be adjusted according to the reliability r. This makes it possible, for example, to effectively reduce the influence of a value of the dependent variable y with a low reliability r on the machine learning, and effectively increase the influence of a value of the dependent variable y with a high reliability r on the machine learning. In other words, it is possible to effectively prevent overlearning for values ​​of the dependent variable y with a low reliability r. Therefore, a prediction model 123 with high prediction accuracy can be more appropriately generated by machine learning.

[0120] Furthermore, in this embodiment, a value of the objective variable y having a small reliability r, for example, included in an existing combination, is replaced with a value different from that value and included in a new combination, thereby more effectively suppressing overlearning for the value of the objective variable y having a small reliability r.

[0121] Furthermore, in this embodiment, the greater the fluctuation in the value indicated by the process data, the smaller the reliability r is calculated, and the greater the reliability r, the greater the number of combinations is changed to. This makes it possible to effectively suppress overlearning for the objective variable y having a small reliability r.

[0122] Furthermore, in this embodiment, a new value of the objective variable y to be included in each of a plurality of new combinations is selected according to a probability distribution having the value of the objective variable y included in the existing combination as the mean. This allows the value of the objective variable y included in the existing combination to be appropriately replaced with a value different from that value and included in the new combination.

[0123] In this embodiment, a normal distribution is used as the probability distribution, and a new value of the objective variable y is selected randomly according to the normal distribution. This allows the value of the objective variable y included in an existing combination to be appropriately replaced with a value different from the original value and included in the new combination.

[0124] Furthermore, in this embodiment, a normal distribution with a smaller variance is used as the reliability r increases. As a result, a value of the objective variable y with a small reliability r, for example, included in an existing combination, can be replaced with a value that is significantly different from that value and included in a new combination, and a value of the objective variable y with a large reliability r can be replaced with a value close to that value and included in a new combination. Therefore, overlearning for values ​​of the objective variable y with a small reliability r can be more effectively suppressed.

[0125] (Other Modifications) While the search device 100 according to one aspect of the present disclosure has been described above based on the above embodiment, the present disclosure is not limited to the above embodiment. As long as the modifications do not deviate from the spirit of the present disclosure, various modifications that would occur to those skilled in the art may also be included in the present disclosure.

[0126] For example, in the above embodiment, the ratio of the reliability r is used to calculate the number of new combinations, but the number of new combinations may be calculated without using that ratio. For example, the number of new combinations may be calculated by multiplying the reliability r by a predetermined coefficient. Alternatively, the number of new combinations may be calculated by inputting the reliability r into a predetermined function. With this function, the larger the input reliability r, the larger the value output as the number of new combinations. Furthermore, in the above embodiment, the square of the process abnormality degree σ is used as the variance of the normal distribution, but the variance of the normal distribution does not have to be the square of the process abnormality degree σ. The variance of the normal distribution may be any value as long as it increases as the process abnormality degree σ increases.

[0127] Furthermore, the prediction model 123 in the above embodiment may be any model as long as it is a model that indicates the relationship between the explanatory variables and the objective variables, that is, the correlation.

[0128] In the above-described embodiments, each component may be configured with dedicated hardware, or may be realized by executing a software program suitable for each component. Each component may be realized by a program execution unit such as a CPU or processor reading and executing a software program recorded on a recording medium such as a hard disk or semiconductor memory. Here, the software that realizes the search device 100 of the above-described embodiments is, for example, a program that causes a computer to execute each step of the flowchart shown in FIG. 13.

[0129] The following cases are also included in this disclosure:

[0130] (1) The at least one device is specifically a computer system comprising a microprocessor, ROM (Read Only Memory), RAM (Random Access Memory), a hard disk unit, a display unit, a keyboard, a mouse, etc. A computer program is stored in the RAM or hard disk unit. The at least one device achieves its function when the microprocessor operates in accordance with the computer program. Here, the computer program is composed of a combination of multiple instruction codes that indicate commands to the computer to achieve a predetermined function.

[0131] (2) Some or all of the components constituting at least one of the above devices may be configured as a single system LSI (Large Scale Integration). A system LSI is an ultra-multifunctional LSI manufactured by integrating multiple components on a single chip, and specifically, is a computer system configured to include a microprocessor, ROM, RAM, etc. A computer program is stored in the RAM. The system LSI achieves its functions by the microprocessor operating in accordance with the computer program.

[0132] (3) Some or all of the components constituting at least one of the above devices may be configured as an IC card or a standalone module that can be attached to or detached from the device. The IC card or module is a computer system configured with a microprocessor, ROM, RAM, etc. The IC card or module may include the above-mentioned ultra-multifunctional LSI. The IC card or module achieves its functions when the microprocessor operates in accordance with a computer program. This IC card or module may be tamper-resistant.

[0133] (4) The present disclosure may be embodied as the methods described above, a computer program that implements these methods on a computer, or a digital signal that includes the computer program.

[0134] The present disclosure may also be a computer program or a digital signal recorded on a computer-readable recording medium, such as a flexible disk, a hard disk, a CD (Compact Disc)-ROM, a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray (registered trademark) Disc), a semiconductor memory, etc. Alternatively, the present disclosure may be a digital signal recorded on such a recording medium.

[0135] The present disclosure may also be applied to transmitting a computer program or digital signal via a telecommunications line, a wireless or wired communication line, a network such as the Internet, data broadcasting, or the like.

[0136] Furthermore, the program or digital signal may be recorded on a recording medium and transferred, or the program or digital signal may be transferred via a network or the like, so that the program or digital signal may be implemented by another independent computer system.

[0137] The searching device of the present disclosure can be applied to, for example, devices and systems that search for processing conditions for producing a product having desired characteristic values ​​as search conditions.

[0138] 100 Searching device 110 Data set update unit 120 Storage unit 121 Data set 122 Processed data set 123 Prediction model 130 Reliability calculation unit 140 Data set processing unit 150 Prediction model generation unit 160 Search condition derivation unit 170 Device control unit 200 Material processing device 210 First liquid supply unit 211 First liquid container 212 First liquid pump 213 First liquid flow path 220 Second liquid supply unit 221 Second liquid container 222 Second liquid pump 223 Second liquid flow path 230 Mixing flow path unit 240 Process data acquisition unit 250 Objective variable acquisition unit 260 Recovery unit 1000 Searching system

Claims

1. A searching device comprising: a dataset updating unit that generates or updates a dataset indicating a relationship between an explanatory variable and a response variable indicating a characteristic value of a product obtained by processing raw materials by a material processing device in accordance with an explanatory variable; a reliability calculation unit that calculates the reliability of the response variable based on process data obtained from the material processing device and indicating a state of the processing of the raw materials; a dataset processing unit that processes the dataset based on the reliability; a predictive model generation unit that generates a predictive model indicating a relationship between the explanatory variable and the response variable using a processed dataset that is the processed dataset; and a search condition derivation unit that performs a search using the predictive model to derive a value of the explanatory variable for obtaining a target value of the response variable as a search condition.

2. The search device described in claim 1, wherein the predictive model generation unit generates the predictive model by machine learning using the processed dataset, and the dataset processing unit processes the dataset by weighting the amount of learning, which is the number of combinations of the values ​​of the explanatory variables and the values ​​of the target variable, according to the reliability.

3. The search device described in claim 2, wherein the reliability calculation unit calculates, for each existing combination that is a combination included in the dataset, a reliability of the value of the objective variable based on the process data when the value of the objective variable included in the existing combination is obtained, and the dataset processing unit weights the amount of learning by changing, for each reliability, the number of combinations that include the value of the explanatory variable corresponding to that reliability in accordance with that reliability.

4. The search device described in claim 3, wherein when generating a plurality of new combinations from the existing combinations by increasing the number of combinations, the dataset processing unit includes, in at least one of the plurality of new combinations, a value different from the value of the objective variable as a new value of the objective variable, instead of the value of the objective variable included in the existing combination.

5. The search device described in claim 4, wherein the reliability calculation unit calculates a smaller reliability value as the fluctuation of the value indicated by the process data increases, and the data set processing unit changes the number of combinations by increasing the number of combinations as the reliability increases.

6. The search device described in claim 5, wherein the dataset processing unit selects new values ​​of the objective variable included in each of the plurality of new combinations according to a probability distribution having as its average the value of the objective variable included in the existing combinations.

7. The search device according to claim 6, wherein the data set processing unit uses a normal distribution as the probability distribution and randomly selects a new value of the objective variable in accordance with the normal distribution.

8. The search device according to claim 7, wherein the data set processing unit uses a normal distribution having a smaller variance as the reliability increases.

9. A searching method using a computer, comprising: generating or updating a dataset showing an association between an explanatory variable and a response variable indicating a characteristic value of a product obtained by processing raw materials by a material processing device in accordance with the explanatory variables; calculating a reliability of the response variable based on process data obtained from the material processing device indicating a state of the processing of the raw materials; processing the dataset based on the reliability; generating a predictive model indicating the relationship between the explanatory variables and the response variable using a processed dataset, which is the processed dataset; and performing a search using the predictive model to derive a value of the explanatory variable for obtaining a target value of the response variable as a search condition.

10. A program that causes a computer to execute the following steps: generate or update a dataset that correlates explanatory variables with a response variable that indicates a characteristic value of a product obtained by processing raw materials by a material processing device in accordance with explanatory variables; calculate the reliability of the response variable based on process data obtained from the material processing device that indicates the state of the processing of the raw materials; process the dataset based on the reliability; generate a prediction model that indicates the relationship between the explanatory variables and the response variable using a processed dataset that is the processed dataset; and derive, as a search condition, the value of the explanatory variable for obtaining a target value of the response variable by performing a search using the prediction model.

Images