Learning device, control device, training method, and recording medium

Abstract

This learning device: inputs information indicating the composition or structure of a candidate for a raw material for a target object to a raw material characteristics estimation model subjected to machine learning using training data in which information indicating the compositions or structures of substances and information indicating characteristic values pertaining to the substances are combined, thereby estimating a characteristic value pertaining to the candidate for the raw material for the target object; and uses training data in which information indicating conditions for the generation of the target object and including the characteristic value pertaining to the candidate for the raw material for the target object, and information indicating a characteristic value of the target object when the target object is generated, are combined in accordance with said information indicating the conditions, to train a target object characteristics estimation model for outputting an estimate for the characteristic value of the target object when an input of the information indicating the conditions for the generation of the target object has been received.

Description

Learning device, control device, learning method, and recording medium

[0001] The present invention relates to a learning device, a control device, a learning method, and a recording medium.

[0002] Pre-learning may be performed in machine learning. For example, in a feature estimation method described in Patent Literature 1, pre-learning and relearning are performed on a neural network for estimating the physical property values ​​of compounds. In this feature estimation method, the physical property values ​​of each of a plurality of compounds measured under certain measurement conditions are used as known target data, and the physical property values ​​of compounds for which no actual measurement values ​​have been obtained are used as unknown target data, and the unknown target data are used as the estimation target.

[0003] Furthermore, this feature estimation method uses multiple physical property calculation methods to calculate physical properties for all compounds, both those with known target data and those with unknown target data, for each physical property calculation method and for each compound, and generates known auxiliary data. This feature estimation method then pre-trains a neural network using a training dataset including the known auxiliary data, and re-trains the neural network using a training dataset including the known target data. This feature estimation method then estimates unknown target data using the pre-trained and re-trained neural network.

[0004] International Publication No. 2020 / 188971

[0005] Even when sufficient data on the characteristic values ​​of the target object is not available, such as when searching for a method for generating a target object that satisfies certain conditions, it is expected that the characteristic values ​​of the target object can be estimated with a relatively small amount of calculation and with a relatively high degree of accuracy if pre-learning techniques can be used.

[0006] An example of an object of the present disclosure is to provide a learning device, a control device, a learning method, and a recording medium that can solve the above-mentioned problems.

[0007] According to a first aspect of the present disclosure, a learning device includes: a raw material property estimation means for inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of the property value of the substance, and machine learning has been performed using training data that combines information indicating the composition or structure of the substance with information indicating property values ​​related to the substance; and an object property learning means for learning an object property estimation model that receives input of information indicating the composition or structure of a candidate raw material of a target object and outputs an estimated value of the property value of the target object, using training data that combines information indicating the conditions for generating the target object, including the property values ​​of the candidate raw material of the target object, and information indicating the property values ​​of the target object when generated in accordance with that information.

[0008] According to a second aspect of the present disclosure, the control device includes: a raw material property estimation means for inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs estimated values ​​of property values ​​for the substance, and for which machine learning has been performed using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​for the substance; a condition determination means for determining conditions for generating an object whose property values ​​are to be measured, based on the probability distribution of estimated property values ​​of the object output by the object property estimation model that receives information indicating conditions for generating an object, including the property values ​​of the candidate raw material of the target object, and outputs a probability distribution of estimated property values ​​of the object; a control means for controlling the device for generating the object to generate the object in accordance with the determined conditions; and an object property learning means for learning the object property estimation model using training data that combines the determined conditions with measured values ​​of property values ​​of the object generated in accordance with the conditions.

[0009] According to a third aspect of the present disclosure, a learning method includes: inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of a property value of the substance, the raw material property estimation model having undergone machine learning using training data that combines information indicating the composition or structure of a substance and information indicating property values ​​of the substance, estimating property values ​​of the candidate raw material of the target object; and learning an object property estimation model that receives information indicating the conditions for generating the target object, including the property values ​​of the candidate raw material of the target object, and outputs an estimated value of a property value of the target object using training data that combines information indicating the conditions for generating the target object, including the property values ​​of the candidate raw material of the target object, and information indicating the property values ​​of the target object when generated in accordance with the information.

[0010] According to a fourth aspect of the present disclosure, a recording medium is a recording medium storing a program that causes a computer to execute the following steps: inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of a property value of the substance, and machine learning has been performed using training data that combines information indicating the composition or structure of the substance with information indicating property values ​​of the substance; and learning an object property estimation model that receives information indicating the conditions for generating the object, including the property values ​​of the candidate raw material of the target object, and outputs an estimated value of a property value of the object when generated in accordance with the information, using training data that combines information indicating the conditions for generating the object.

[0011] According to one aspect of the present disclosure, it is expected that the pre-learning technique can be used even when sufficient data on the characteristic values ​​of the target object is not available.

[0012] FIG. 1 is a diagram showing an example of the configuration of a learning device according to at least one embodiment; FIG. 2 is a diagram showing an example of data input / output at a processing unit when a learning device according to at least one embodiment learns a raw material property estimation model; FIG. 3 is a diagram showing a first example of data input / output at a processing unit when a learning device according to at least one embodiment searches for conditions for generating an object; FIG. 4 is a diagram showing an example of the configuration of an object property estimation model according to at least one embodiment; FIG. 5 is a diagram showing a second example of data input / output at a processing unit when a learning device according to at least one embodiment searches for conditions for generating an object; FIG. 6 is a diagram showing an example of a processing procedure when a learning device according to at least one embodiment searches for conditions for generating an object; FIG. 7 is a diagram showing an example of the configuration of an object generation system according to at least one embodiment; FIG. 8 is a diagram showing an example of the configuration of a control device according to at least one embodiment; FIG. 9 is a diagram showing an example of the configuration of a learning device according to at least one embodiment; FIG. 10 is a diagram showing an example of the configuration of a control device according to at least one embodiment; FIG. 11 is a diagram showing an example of the processing procedure in a learning method according to at least one embodiment; FIG. 12 is a diagram showing an example of the processing procedure in a control method according to at least one embodiment;

[0013] The following describes embodiments of the present invention, but the following embodiments do not limit the scope of the invention as claimed. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention.

[0014] First Embodiment Fig. 1 is a diagram illustrating an example of the configuration of a learning device according to at least one embodiment. In the configuration illustrated in Fig. 1, the learning device 100 includes a communication unit 110, a display unit 120, an operation input unit 130, a storage unit 170, and a processing unit 180. The processing unit 180 includes a raw material property estimation unit 181, a raw material property learning unit 183, a model selection unit 184, an object property estimation unit 185, an object property learning unit 187, a condition determination unit 188, and a data acquisition unit 189. The raw material property estimation unit 181 includes a raw material property estimation model 182. The object property estimation unit 185 includes an object property estimation model 186.

[0015] The learning device 100 learns an object property estimation model 186 that receives information indicating conditions for generating an object and outputs estimated values ​​of object property values. The conditions for generating an object can also be considered as a method for generating the object. The learning device 100 may be configured using a computer such as a personal computer (PC) or a workstation (WS).

[0016] Here, model (machine learning model) learning refers to adjusting the parameter values ​​of the model so that the accuracy of the model's output becomes higher. In the case of a model that estimates the characteristic values ​​of an object, model learning refers to adjusting the parameter values ​​of the model so that the estimation accuracy of the characteristic values ​​becomes higher (so that the estimation accuracy becomes better). Model learning can also be called model training.

[0017] The following describes an example in which the learning device 100 searches for a method for generating an object having predetermined characteristics. The learning device 100 determines conditions for generating (prototyping) an object using the object characteristic estimation model 186. The learning device 100 then determines whether the characteristics of the object generated according to the determined conditions have the predetermined characteristics. The learning device 100 also trains the object characteristic estimation model 186 using measured values ​​of the characteristic values ​​of the generated object.

[0018] The learning device 100 searches for a method of generating an object (conditions for generating an object) by repeatedly determining the conditions for generating an object, determining whether the generated object has predetermined properties, and learning the object property estimation model 186. However, the use of the learning device 100 is not limited to a specific one. The learning device 100 can be used for various purposes involving learning the object property estimation model 186. For example, a learning device 100 that has progressed to a certain extent in learning the object property estimation model 186 may be used to estimate the properties of a substance.

[0019] The communication unit 110 communicates with other devices. For example, the communication unit 110 may receive measurement values ​​from a device that measures the property values ​​of the generated object. These measurement values ​​are used as correct answer data for training the object property estimation model 186. Furthermore, a device that trains the raw material property estimation model 182 used by the training device 100 may be provided separately from the training device 100. The communication unit 110 may then receive the trained raw material property model from the device that trained the raw material property estimation model 182. The transmission and reception of the model may be performed, for example, by transmitting and receiving parameter values ​​of the model.

[0020] The display unit 120 has a display screen such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images. For example, the display unit 120 may display various information related to the search for a method of generating an object, such as the estimated values ​​of the property values ​​of the object by the object property estimation model 186, the conditions for generating the object, and the measured values ​​of the property values ​​of the generated object.

[0021] Operation input unit 130 includes input devices such as a keyboard and a mouse, and accepts user operations. For example, operation input unit 130 may accept user operations for making various settings related to the processing performed by learning device 100, such as a user operation for setting an acquisition function used to determine conditions for generating an object.

[0022] The storage unit 170 stores various data. For example, the storage unit 170 may store a trained raw material property estimation model 182 and an object property estimation model 186 to be trained by the learning device 100. The storage unit 170 is configured using a storage device included in the learning device 100.

[0023] The processing unit 180 performs various processes by controlling each unit of the learning device 100. The functions of the processing unit 180 are performed, for example, by a CPU (Central Processing Unit) included in the learning device 100 reading and executing a program from the storage unit 170.

[0024] 2 is a diagram showing an example of data input / output in the processing unit 180 when the learning device 100 learns the raw material property estimation model 182. The learning device 100 may perform pre-learning of the raw material property estimation model 182. That is, the learning device 100 may perform learning of the raw material property estimation model 182 before estimating the property values ​​of the target object. Furthermore, a device other than the learning device 100 may perform pre-learning of the raw material property estimation model 182.

[0025] In the example of Figure 2, the raw material property estimation model 182 is a model that receives input information indicating the composition or structure of a substance and outputs a property value related to that substance. The properties for which the raw material property estimation model 182 outputs a property value are not limited to specific ones, and can be various properties for which correct values ​​(teaching data) for machine learning can be obtained. The property value related to a certain substance may be a property value of the substance itself, or may be a property value of a substance produced using the substance.

[0026] The number of characteristic values ​​output by the raw material characteristic estimation model 182 may be one or more, and is not limited to a specific number. Hereinafter, it is assumed that the raw material characteristic estimation model 182 outputs a vector indicating the characteristic values ​​of a substance. A vector indicating a characteristic value is also referred to as a feature vector. When the raw material characteristic estimation model 182 outputs a scalar value, that is, when the raw material characteristic estimation model 182 outputs one characteristic value, the scalar value output by the raw material characteristic estimation model 182 may be treated as a feature vector with one element. When the characteristic value of a substance is uniquely determined for each substance and for each characteristic, the feature vector becomes a vector indicating a fixed value or a vector indicating discrete values ​​based on the fixed value for each substance.

[0027] The raw material property estimation unit 181 calculates a feature vector by inputting information indicating the composition or structure of a substance into the raw material property estimation model 182. Information indicating the composition or structure of a substance is also referred to as composition information.

[0028] The representation format of the composition information is not limited to a specific format. Furthermore, the machine learning model used as the raw material property estimation model 182 can be various types of machine learning models depending on the representation format of the composition information. For example, the composition information may be represented as a graph, and the raw material property estimation model 182 may be configured using a graph convolutional neural network (GCN). Alternatively, the composition information may be represented as text (character strings), and the raw material property estimation model 182 may be configured using a language model. The language model here refers to an information processing device that receives input of text, such as a natural language, and outputs the text.

[0029] The raw material property learning unit 183 learns the raw material property estimation model 182. Specifically, the raw material property learning unit 183 acquires a training data set consisting of training data in which the constituent information of the substance to be learned is linked to a feature vector indicating the property value of the substance.

[0030] Then, the raw material property learning unit 183 inputs the composition information from the training data to the raw material property estimation model 182 via the raw material property estimation unit 181. The raw material property learning unit 183 adjusts (updates) the parameter values ​​of the raw material property estimation model 182 so that the value of the feature vector output by the raw material property estimation model 182 approaches the correct value of the feature vector indicated in the training data.

[0031] There is no particular limitation to the method by which the raw material property learning unit 183 learns the raw material property estimation model 182. For example, the raw material property learning unit 183 may learn the raw material property estimation model 182 using a known learning method such as deep learning using backpropagation.

[0032] The following describes an example in which multiple raw material property estimation models 182 are provided. For example, a raw material property estimation model 182 may be provided for each type of substance, for each group into which the properties of substances are grouped, or for each combination thereof. The raw material property learning unit 183 learns each raw material property estimation model 182. This is expected to enable the raw material property estimation model 182 to estimate the properties of substances with relatively high accuracy. In other words, it is expected that the raw material property estimation unit 181 can estimate the properties of substances with relatively high accuracy using the raw material property estimation model 182. However, the number of raw material property estimation models 182 provided in the learning device 100 is not limited to a specific number, and may be one or more.

[0033] 3 is a diagram showing a first example of data input / output in the processing unit 180 when the learning device 100 searches for conditions for generating a target product. In the example of FIG. 3, the model selection unit 184 receives input of information indicating the composition or structure of substances that are candidate raw materials for the target product, and selects a raw material property estimation model 182 corresponding to the composition or structure of the substances. The model selection unit 184 outputs the information indicating the composition or structure of the substances that are candidate raw materials for the target product and information indicating the selected raw material property estimation model 182 to the raw material property estimation unit 181.

[0034] Here, a substance that is a candidate for a raw material of a target product is a substance that has the potential to be used as a raw material of the target product. A substance that is a candidate for a raw material of a target product is also referred to as a raw material candidate. The raw material candidates may include candidates for intermediate products. For example, a substance that is produced using a candidate raw material of a target product may be included in the candidate raw material of the target product. Information indicating the composition or structure of the candidate raw material is also referred to as candidate raw material composition information. Information indicating the raw material property estimation model 182 selected by the model selection unit 184 is also referred to as model selection information.

[0035] The model selection unit 184 may store a correspondence table linking the compositions or structures of raw material candidates with raw material property estimation models 182 corresponding to those compositions or structures. Alternatively, the model selection unit 184 may include a trained machine learning model that receives input raw material candidate composition information and outputs raw material property estimation model 182 selection information. When multiple types of raw materials are used to generate the target product, the model selection unit 184 may select multiple raw material property estimation models 182 according to the raw material candidates.

[0036] The raw material property estimation unit 181 estimates property values ​​of the raw material candidates based on the raw material candidate configuration information. Specifically, the raw material property estimation unit 181 inputs the raw material candidate configuration information to the raw material property estimation model 182 indicated by the model selection information, and calculates a feature vector indicating the property values ​​of the raw material candidates. The raw material property estimation unit 181 outputs the calculated feature vector indicating the property values ​​of the raw material candidates to the target property estimation unit 185.

[0037] The characteristic values ​​of the raw material candidates may be characteristic values ​​of the raw material candidates. Alternatively, the characteristic values ​​of the raw material candidates may be characteristic values ​​of a substance produced using the raw material candidates. In this case, the substance produced using the raw material candidates may be an intermediate product.

[0038] A feature vector indicating the characteristic values ​​of a substance is also referred to as a feature vector for that substance. For example, a feature vector indicating the characteristic values ​​of a raw material candidate is also referred to as a feature vector for the raw material candidate. A feature vector indicating the characteristic values ​​of a substance is also referred to as a feature vector of that substance. For example, a feature vector indicating the characteristic values ​​of an object is also referred to as a feature vector of the object.

[0039] The object characteristic estimation unit 185 estimates the characteristic values ​​of the object. Specifically, the object characteristic estimation unit 185 inputs the feature vectors related to the raw material candidates estimated by the raw material characteristic estimation unit and additional information into the object characteristic estimation model 186 to calculate the characteristic values ​​of the object. The additional information here refers to information indicating the conditions for generating the object other than the raw material candidate configuration information. Therefore, the object characteristic estimation unit 185 inputs information indicating the conditions for generating the object, in which the raw material candidate configuration information has been converted into the feature vectors related to the raw material candidates, into the object characteristic estimation model 186 to calculate the characteristic values ​​of the object.

[0040] The object property estimation model 186 is a model used to estimate the property value of an object. As shown in Equation (1), the object property estimation model 186 receives additional information and a feature vector related to a raw material candidate as input, and outputs an estimate of the property value of the object.

[0041]

[0042] f １ is an expression of the object characteristic estimation model 186 as a function. x represents additional information. z represents a feature vector. y represents an estimated value of the characteristic value of the object. The object characteristic estimation model 186 may be configured to output estimated values ​​of the characteristic value of the object for multiple types of characteristics. In the following, it is assumed that the object characteristic estimation model 186 outputs a vector indicating the characteristic value of the object. In other words, y represents the feature vector of the object. When the object characteristic estimation model 186 outputs a scalar value, that is, when the object characteristic estimation model 186 outputs one characteristic value, the scalar value output by the object characteristic estimation model 186 may be treated as a feature vector with one element.

[0043] The object property estimation model 186 may be configured to include a partial model that outputs an estimated value of the difference between the feature vector of the object and the feature vector of the raw material candidate. Figure 4 is a diagram showing an example of the configuration when the object property estimation model 186 includes a partial model. In the configuration shown in Figure 4, the object property estimation model 186 includes a partial model 186b.

[0044] The partial model 186b receives the additional information as input, and outputs the difference between the feature vector of the target object and the feature vector of the raw material candidate, as shown in equation (2).

[0045]

[0046] f ２ is an expression of the partial model 186b as a function. x, y, and z are the same as in equation (1). In the example of equation (2), the feature vector y of the target object indicates the characteristic value of the same item (same characteristic) as the feature vector z of the raw material candidate output by the raw material characteristic estimation model 182. The difference yz obtained by subtracting the feature vector z of the raw material candidate from the feature vector y of the target object output by the partial model 186b can be considered as the change in characteristic due to the processing indicated by the additional information x.

[0047] When the magnitude of the value y-z obtained by subtracting the feature vector related to the raw material candidate from the feature vector of the object is sufficiently small relative to the magnitude of the value of the feature vector y of the object, it is expected that the value y-z will be able to more easily express small differences in value than the value y. In this respect, it is expected that the partial model 186b will be able to estimate with relatively high accuracy the value of the vector obtained by subtracting the feature vector related to the raw material candidate from the feature vector of the object, and the object property estimation unit 185 will be able to estimate the property value of the object with relatively high accuracy.

[0048] 4, the object property estimation model 186 calculates and outputs the object property vector y by adding the feature vector z related to the raw material candidate to yz, which is obtained by subtracting the feature vector related to the raw material candidate from the feature vector of the object output by the partial model 186b. The calculation by the object property estimation model 186 in this case is shown in equation (3).

[0049]

[0050] The object property estimation model 186 outputs the property values ​​of the object as a probability distribution. For example, the object property estimation model 186 may output the mean value and variance of the property values ​​of the object. A specific type of distribution may be assumed as the probability distribution output by the object property estimation model 186, but is not limited to this. For example, a normal distribution may be assumed as the probability distribution output by the object property estimation model 186, but is not limited to this.

[0051] Various types of models capable of outputting a probability distribution can be used as the object property estimation model 186. For example, the object property estimation model 186 may be configured using a model based on ensemble learning or a model based on a Gaussian process, but is not limited to these.

[0052] The object characteristic learning unit 187 learns the object characteristic estimation model 186. The object characteristic learning unit 187 learns the object characteristic estimation model 186 using training data that is a combination of feature vectors and additional information related to the raw material candidates and measured values ​​of the property values ​​of the object obtained as a result of generating (prototype) the object under the conditions indicated by the information. The feature vectors and additional information related to the raw material candidates correspond to information indicating the conditions for generating the object, in which the raw material candidate configuration information has been converted into feature vectors related to the raw material candidates.

[0053] The method by which the object characteristic learning unit 187 learns the object characteristic estimation model 186 is not limited to the characteristic method, and various methods can be used depending on the type of the object characteristic estimation model 186. For example, the method by which the object characteristic learning unit 187 learns the object characteristic estimation model 186 may be, but is not limited to, ensemble learning or Gaussian process regression.

[0054] The condition determination unit 188 determines conditions for generating the object in accordance with the object property estimation model 186. The condition determination unit 188 determines conditions indicated by the raw material candidate configuration information and the additional information as conditions for generating the object. The determination of the conditions for generating the object by the condition determination unit 188 can be considered to be determining search points in a search for a feature vector of the object as a solution obtained by prototyping the object in the coordinate space of the input data to the object property estimation model 186. In this case, the search for the feature vector of the object can be considered to be a search for conditions for generating the object that will result in a desired object. The method by which the condition determination unit 188 determines the conditions for generating the object may be, but is not limited to, a known method such as Bayesian optimization.

[0055] The condition determination unit 188 determines the conditions for generating the object based on the object property estimation model 186. The condition determination unit 188 determines a point in the coordinate space of the input data to the object property estimation model 186, and converts a feature vector related to the raw material candidate indicated at that point into raw material candidate configuration information.

[0056] To this end, when raw material property estimation unit 181 inputs the raw material candidate composition information into raw material property estimation model 182 and calculates the feature vectors of the raw material candidates, storage unit 170 may associate and store the raw material candidate composition information and the feature vectors of the raw material candidates. Then, condition determination unit 188 may refer to this data and convert the feature vectors of the raw material candidates into raw material candidate composition information.

[0057] The conditions determined by condition determination unit 188 may be presented to a person (a prototype creator of the object) by displaying them on display unit 120, and the person may then generate the object. Alternatively, learning device 100 or another device may automatically generate the object according to the conditions determined by condition determination unit 188.

[0058] The data acquisition unit 189 acquires raw material candidate configuration information, additional information, and actual measured values ​​of characteristic values ​​of the generated target object. The raw material candidate configuration information and additional information correspond to information indicating conditions for generating the target object. The data acquisition unit 189 may acquire the raw material candidate configuration information and additional information from the condition determination unit 188. Alternatively, if the conditions determined by the condition determination unit 188 differ from the conditions under which the target object was actually generated, such as when a prototype maker of the target object generates the target object under conditions different from those determined by the condition determination unit 188, the data acquisition unit 189 may acquire the conditions under which the target object was actually generated. For example, the prototype maker of the target object may input the conditions under which the target object was actually generated via the operation input unit 130, and the data acquisition unit 189 may acquire the input conditions.

[0059] Furthermore, the data acquisition unit 189 may acquire actual measured values ​​of the characteristic values ​​from a device that measured the characteristic values ​​of the object. For example, the device that measured the characteristic values ​​of the object may transmit the measured values, and the transmitted measured values ​​may be received by the communication unit 110. The data acquisition unit 189 may then acquire the measured values ​​received by the communication unit 110. Alternatively, a prototype maker of the object may input the measured values ​​of the characteristic values ​​of the object from the operation input unit 130, and the data acquisition unit 189 may acquire the input actual measured values.

[0060] The distinction between information used to train the raw material property estimation model 182 and additional information may be made depending on the amount of available data. For example, information that is publicly available as open data or semi-open data may be used to train the raw material property estimation model 182, and other information may be treated as additional information. For example, if the amount of available data indicating the property values ​​of raw material candidates is small, information indicating the composition or structure of the raw material candidates may be included in the additional information and input to the target property estimation model 186.

[0061] The additional information may include conditions related to the state of the raw material candidates, such as the temperature of the raw material candidates, conditions related to the amount or ratio of the raw material candidates, such as the ratio when multiple types of raw material candidates are mixed, conditions related to substances used in the process of producing the target product, conditions related to processing time, or two or more of these conditions. The raw material candidates may also include additives or fillers.

[0062] For example, when the target substance is a polymer compound, the additional information may include one or more of the following information: Monomer information Here, the monomer information may include information obtained as a molecular descriptor of the monomer. The monomer information is also referred to as monomer information. The monomer information is an example of information about the raw material candidate. The information obtained as a molecular descriptor of the monomer is an example of information indicating the composition or structure of the raw material candidate, or information indicating the properties of the raw material candidate.

[0063] Polymer information: The polymer information here may include any one or more of the molecular weight, sequence structure, and structural information (configuration or conformation) of the polymer. Polymer information is also referred to as polymer information.

[0064] Polymer information is an example of information about a raw material candidate or information about a target product. The molecular weight, sequence structure, or structural information (configuration or conformation) of a polymer are all examples of information that indicates the composition or structure of a raw material candidate or the composition or structure of a target product to be produced.

[0065] -Additive or filler information The additive or filler information here may include information indicating the type and amount of the additive or filler. The amount of the additive or filler may be indicated as a ratio (volume ratio or weight ratio) to the monomer or polymer. The additive or filler information is also referred to as compound information. Compound information is an example of information indicating conditions regarding the amount or ratio of the raw material candidate.

[0066] Information indicating the conditions for kneading: The information indicating the conditions for kneading here may be information indicating the conditions for kneading multiple types of polymers, or information indicating the conditions for kneading a polymer with an additive or a filler. Information indicating the conditions for processing performed in the process of producing a target product, such as information indicating the conditions for kneading, is also referred to as process information. The process information may include the temperature of the object to be kneaded (polymer, additive, or filler), the kneading time, the rotation speed of the kneading tool, the screw length, diameter, or screw configuration of the kneading equipment, or two or more of these.

[0067] Information indicating the proportion of monomers The information indicating the proportion of monomers here may include information indicating the proportion of each of multiple types of monomers constituting a copolymer. The information indicating the proportion of monomers here is also referred to as monomer ratio information.

[0068] Monomer ratio information is an example of information indicating the composition or structure of a raw material candidate, or an example of information indicating the composition or structure of a target product to be produced. Information indicating the proportions of each of multiple types of monomers constituting a copolymer is also an example of information indicating the composition or structure of a raw material candidate, or an example of information indicating the composition or structure of a target product to be produced.

[0069] FIG. 5 is a diagram showing a second example of data input and output in processing unit 180 when learning device 100 searches for conditions for generating an object.

[0070] The example shown in Fig. 5 corresponds to a specific example of the example shown in Fig. 3. Fig. 5 shows an example in which the target substance is a polymer. Monomer information in SMILES notation is used as raw material candidate composition information. In addition, monomer information in molecular descriptors, polymer information, compound information, process information, and monomer ratio information, or some of these, are used as additional information.

[0071] The raw material candidate composition information and the additional information may contain information on the same monomer. Furthermore, the polymer information contained in the additional information may be information on a polymer obtained by polymerization of the monomer indicated by the monomer information. The feature vector related to the raw material candidate and the feature vector of the target object may indicate characteristic values ​​of the same properties.

[0072] For example, when producing a target polymer using multiple types of monomers, it is possible to use information such as that shown in FIG. 5. The compound information, process information, and monomer information may be information relating to a process of polymerizing multiple types of monomers to produce a copolymer, or information relating to a process of kneading multiple types of polymers (which may be copolymers), or information relating to both of these processes. In this way, the constituent information of the raw material candidate and the additional information may include information about the same substance. In this case, the constituent information of the raw material candidate and the additional information may express the information about the same substance in different formats.

[0073] When the type of monomer used as a raw material for the target object is fixed, the additional information may include at least one of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material (polymer information), information on the kneading of multiple types of polymers (compound information), information on a process for producing the target object using raw materials containing a monomer (process information), and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object (monomer ratio information). As a result, when the type of monomer is fixed, the feature vector from the pre-learning becomes a fixed value. In this respect, it is expected that the target property learning unit 187 can relatively efficiently train the target property estimation model 186.

[0074] When the type of polymer used as the raw material of the target object is fixed, the additional information may include at least one of information regarding the kneading of multiple types of polymers used as the raw material of the target object (compound information) and information regarding the process of producing the target object using raw materials containing a monomer used as the raw material of the target object (process information). As a result, when the type of polymer is fixed, the feature vector from the pre-learning becomes a fixed value. In this respect, it is expected that the target property learning unit 187 can relatively efficiently train the target property estimation model 186.

[0075] The three-dimensional structure of a substance, such as the three-dimensional structure of a raw material candidate, may be expressed using a known expression method, such as a method of expressing the three-dimensional structure of a substance using three-dimensional coordinates. Alternatively, the three-dimensional structure of a substance may be expressed using a more simplified expression method, such as using a one-hot vector. For example, the position of an OH group may be represented by a one-hot vector.

[0076] Regarding the compound information, information on which additives to use may be indicated by a one-hot vector. Also, information on which fillers to use may be indicated by a one-hot vector. Alternatively, information on the type of additive to be used, such as a silica-based additive or a carbon-based additive, or information on the type of filler to be used may be indicated by a one-hot vector. Alternatively, the compound information may be indicated by a character string.

[0077] 6 is a diagram showing an example of a processing procedure in which the learning device 100 searches for conditions for generating an object. In the processing in FIG. 6, the object property learning unit 187 initializes the object property estimation model 186 (step S101). For example, a normal distribution may be assumed as the probability distribution of the property values ​​of the object, and the mean and variance may be indicated. Then, the object property learning unit 187 may set the mean and variance at each point in the coordinate space of the input to the object property estimation model 186, which is a combination of the raw material candidate composition information and the additional information, to predetermined initial values.

[0078] Next, the condition determination unit 188 determines conditions for generating the object by referring to the object property estimation model 186 (step S102). For example, an acquisition function for Bayesian optimization may be predetermined. The condition determination unit 188 may then search for a point in the coordinate space of the input to the object property estimation model 186 where the evaluation indicated by the acquisition function is the best (e.g., a point where the acquisition function value is maximized), and adopt the conditions indicated by the obtained point as conditions for generating the object. As described above, the storage unit 170 may store raw material candidate configuration information and feature vectors related to the raw material candidates in association with each other. The condition determination unit 188 may then refer to this data and convert the values ​​of the feature vectors related to the raw material candidates into raw material candidate configuration information.

[0079] Furthermore, conditions for generating an object may be determined in advance for each execution of step S102 up to a predetermined number of times, and the condition determination unit 188 may then adopt the predetermined conditions as the conditions for generating an object.

[0080] Next, the data acquisition unit 189 acquires the conditions for generating the target object and the characteristic values ​​of the target object generated according to those conditions (step S103). The data acquisition unit 189 acquires raw material candidate composition information and additional information as the conditions for generating the target object. The data acquisition unit 189 also acquires the measured values ​​of the characteristic values ​​of the target object generated according to those conditions.

[0081] Next, the processing unit 180 determines whether a search termination condition for the conditions for generating the target object is satisfied (step S104). For example, the processing unit 180 determines whether the characteristic values ​​of the target object generated according to the conditions satisfy predetermined conditions as desired conditions. If the processing unit 180 determines that the search termination condition is not satisfied (step S104: NO), the raw material characteristic estimation unit 181 estimates characteristic values ​​of the raw material candidates (step S105). Specifically, the raw material characteristic estimation unit 181 inputs raw material candidate composition information, which is included in the data acquired by the data acquisition unit 189 as the conditions for generating the target object, into the raw material characteristic estimation model 182 to calculate feature vectors for the raw material candidates.

[0082] Next, the object property learning unit 187 updates the object property estimation model 186 (step S106). The object property learning unit 187 uses, as training data for machine learning, data that is a combination of the feature vectors related to the raw material candidates output by the raw material property estimation model 182, additional information among the conditions for generating the object acquired by the data acquisition unit 189, and the feature vectors of the generated object, to update the object property estimation model 186. For example, the object property learning unit 187 may update the object property estimation model 186 using a learning method appropriate for the type of the object property estimation model 186, such as ensemble learning or Gaussian process regression.

[0083] After step S106, the process returns to step S102. On the other hand, if processing unit 180 determines in step S104 that the end condition for the search for the conditions for generating the target object is met (step S104: YES), learning device 100 ends the process in FIG.

[0084] As described above, the object may be generated automatically. In this case, learning device 100 may be configured as a control device having a function of controlling the device that generates the object.

[0085] 7 is a diagram illustrating an example of the configuration of a condition search system according to at least one embodiment. In the configuration illustrated in FIG. 7, the condition search system 1 includes a control device 100b and an object generation device 200. The condition search system 1 searches for conditions for generating an object. In particular, the condition search system 1 automatically searches for an object, measures its characteristic values, and updates an object characteristic estimation model based on the measured characteristic values.

[0086] The object generation device 200 generates an object. In particular, the object generation device 200 automatically generates an object under the control of the control device 100b. The object generation device 200 also measures characteristic values ​​of the generated trees and garden objects. The control device 100b performs the same processing as the learning device 100. Furthermore, when the control device 100b determines conditions for generating an object, it controls the object generation device 200 so that the object is generated in accordance with the determined conditions. For example, the control device 100b transmits data indicating the determined conditions to the object generation device 200. The control device 100b is an example of the learning device 100.

[0087] Fig. 8 is a diagram showing an example of the configuration of a control device 100b. In the configuration shown in Fig. 8, the control device 100b includes a communication unit 110, a display unit 120, an operation input unit 130, a memory unit 170, and a processing unit 180b. The processing unit 180b includes a raw material property estimation unit 181, a raw material property learning unit 183, a model selection unit 184, an object property estimation unit 185, an object property learning unit 187, a condition determination unit 188, a data acquisition unit 189, and a control unit 190. The raw material property estimation unit 181 includes a raw material property estimation model 182. The object property estimation unit 185 includes an object property estimation model 186.

[0088] 8, parts having the same functions as those in FIG. 1 are designated by the same reference numerals (110, 120, 130, 170, 181, 182, 103, 184, 185, 186, 187, 188, 189), and detailed descriptions thereof will be omitted here. In the control device 100b, the processing unit 180b further includes a control unit 190 in addition to the parts included in the processing unit 180 of the learning device 100. In other respects, the control device 100b is similar to the learning device 100.

[0089] The control unit 190 controls the object creation device 200 so as to create an object in accordance with the conditions determined by the condition determination unit 188. For example, the control unit 190 transmits data indicating the conditions determined by the condition determination unit 188 to the object creation device 200 via the communication unit 110. In the condition search system 1, the object creation device 200 transmits the measurement values ​​of the characteristic values ​​of the created object to the control device 100b. In the control device 100b, the communication unit 110 receives the measurement values ​​of the characteristic values ​​of the object and outputs them to the data acquisition unit 189. The condition search system 1 can automatically search for conditions for creating an object.

[0090] As described above, the raw material property estimation unit 181 inputs information indicating the composition or structure of candidate raw materials for the target product into the raw material property estimation model 182, which receives information indicating the composition or structure of a substance and outputs an estimated value of the property value related to that substance, and which has undergone machine learning using training data that combines information indicating the composition or structure of the substance and information indicating the property value related to that substance, and estimates the property value related to the candidate raw material for the target product.

[0091] The object characteristic learning unit 187 uses training data that combines information indicating the conditions for generating an object, including characteristic values ​​related to candidate raw materials for the object, and information indicating the characteristic values ​​of the object when generated in accordance with that information, to learn the object characteristic estimation model 186, which receives input of information indicating the conditions for generating an object and outputs an estimated value of the characteristic values ​​of the object.

[0092] According to the learning device 100, even when sufficient data on the characteristic values ​​of the target object is not available, it is expected that a pre-learning technique can be used to train the raw material characteristic estimation model 182. In this respect, it is expected that the learning device 100 can estimate the characteristic values ​​of the target object with a relatively high degree of accuracy, using a relatively small amount of training data and a relatively small amount of calculation.

[0093] Here, it is conceivable that a sufficient amount of data is not available for the target object whose characteristics are to be estimated by the learning device 100, but a sufficient amount of data is available for the candidate raw materials. For example, consider a case where the learning device 100 is used to develop a target object that satisfies certain conditions regarding its characteristics. In this case, there is no publicly available data for the target object (an object that satisfies certain conditions regarding its characteristics), but open data or semi-open data may be available for objects that could be used as raw materials for the target object.

[0094] When a sufficient amount of data is not available for the target object but a sufficient amount of data is available for the raw material candidates, the learning device 100 is expected to obtain estimates of the property values ​​of the raw material candidates with relatively high estimation accuracy by using the raw material property estimation model 182 that has been trained using a sufficient amount of data.The estimation layer 100 is expected to be able to estimate the property values ​​of the target object with relatively high accuracy by estimating the property values ​​of the raw material candidates with relatively high estimation accuracy.

[0095] Furthermore, the learning device 100 can change a raw material candidate to another candidate, and change the properties of the raw material candidate that are used to estimate the property value of the target product, by replacing the raw material property estimation model 182 with another raw material property estimation model 182. In other words, the learning device 100 can flexibly respond to changes in the feature vector space, which is the coordinate space of the output of the raw material property estimation model 182. The learning device 100 can also flexibly respond to changes in the feature vector space in the additional information.

[0096] Furthermore, by having the learning device 100 estimate the properties of an object using the raw material property estimation model 182, the learning for estimating the properties of the object can be performed by a division of labor. For example, an information device manufacturer may provide the learning device 100 and train the raw material property estimation model 182, and the object developer may perform the learning for estimating the properties of the object other than training the raw material property estimation model 182. This allows the object developer to maintain the confidentiality of information related to the development of the object, which is indicated in the additional information, and to outsource the training of the raw material property estimation model 182, which is a heavy load due to the use of a large amount of data.

[0097] Furthermore, the target property estimation model 186 includes a partial model 186b. The partial model 186b receives input of information indicating the conditions for generating the target object, other than the property values ​​related to the candidate raw materials of the target object, and outputs an estimated value of the difference between the property values ​​related to the candidate raw materials of the target object and the property values ​​related to the candidate raw materials of the target object.

[0098] The object characteristic estimation model 186 inputs information indicating the conditions for generating the object, which is input to the object characteristic estimation model 186, other than the characteristic values ​​related to the candidate raw materials for the object into the partial model 186b, and outputs an estimated value of the characteristic value of the object calculated using the output of the partial model 186b and the characteristic values ​​related to the candidate raw materials for the object.

[0099] When the magnitude of the vector obtained by subtracting the feature vector related to the candidate raw material from the feature vector of the object is sufficiently small relative to the magnitude of the feature vector of the object, it is expected that outputting the vector obtained by subtracting the feature vector related to the candidate raw material from the feature vector of the object will be more likely to represent small differences in the output value than outputting the feature vector of the object from object property estimation model 186. In this respect, it is expected that learning device 100 will be able to estimate the characteristic values ​​of the object with relatively high accuracy.

[0100] Furthermore, the raw material property estimation model 182 receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of the property value of the monomer or an estimated value of the property value of a substance that contains that monomer as a raw material. The target property estimation model 186 receives input of the estimated value output by the raw material property estimation model 182 and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains a monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratios of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product. It is expected that the learning device 100 will be able to perform pre-training of the raw material property estimation model 182 using a large amount of data related to monomers.

[0101] Furthermore, the object property estimation model 186 outputs a probability distribution of the estimated values ​​of the object property values. The learning device 100 can reflect the uncertainty of the estimated values ​​of the object property values ​​in the estimation results. Furthermore, the learning device 100 can use known methods such as Bayesian optimization to search for conditions for generating the object.

[0102] Furthermore, the condition determination unit 188 determines conditions for generating an object whose characteristic values ​​are to be measured, based on the probability distribution of the estimated values ​​of the object's characteristic values. The object characteristic learning unit 187 uses training data that combines the determined conditions with the measured values ​​of the object's characteristic values ​​generated according to the conditions to train the object characteristic estimation model 186. It is expected that the learning device 100 will be able to estimate the object's characteristic values ​​with a relatively high degree of accuracy, even at a stage where the number of searches for conditions for generating the object is relatively small.

[0103] The type of monomer used as a raw material for the target object is fixed. The condition determination unit 188 determines, as a condition for generating the target object whose characteristic values ​​are to be measured, at least one of information on a polymer containing a monomer that is a candidate raw material for the target object, information on the kneading of multiple types of polymers, information on a process for generating the target object using raw materials containing a monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object. The learning device 100 is expected to enable relatively efficient training of the target object property estimation model 186.

[0104] The type of polymer used as the raw material for the target object is fixed. The condition determination unit 188 determines, as a condition for generating the target object whose characteristic values ​​are to be measured, at least one of information regarding the kneading of multiple types of polymers used as the raw material for the target object and information regarding a process for generating the target object using raw materials containing the monomer used as the raw material for the target object. The learning device 100 is expected to enable relatively efficient training of the target object property estimation model 186.

[0105] Furthermore, the raw material property estimation unit 181 inputs information indicating the composition or structure of candidate raw materials for the target product into a raw material property estimation model 182 that receives information indicating the composition or structure of a substance and outputs an estimated value of the property value of the substance, and has undergone machine learning using training data that combines information indicating the composition or structure of the substance with information indicating the property value of the substance, thereby estimating the property value of the candidate raw material for the target product.

[0106] The condition determination unit 188 determines the conditions for generating an object whose characteristic values ​​are to be measured based on the probability distribution of estimated values ​​of the characteristic values ​​of the object output by the object characteristic estimation model 186, which receives input of information indicating the conditions for generating an object, including characteristic values ​​related to candidate raw materials for the object, and outputs a probability distribution of estimated values ​​of the characteristic values ​​of the object.

[0107] The control unit 190 controls the object generating device to generate the object according to the determined conditions. The object characteristic learning unit 187 learns the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated according to the conditions.

[0108] According to the control device 100b, even when sufficient data on the characteristic values ​​of the target object is not available, it is expected that a pre-learning technique can be used to train the raw material characteristic estimation model 182. In this respect, it is expected that the control device 100b can estimate the characteristic values ​​of the target object with a relatively high degree of accuracy, using a relatively small amount of training data and a relatively small amount of calculation.

[0109] Furthermore, the control device 100b can change a raw material candidate to another candidate, and change the properties of the raw material candidate that are used to estimate the property value of the target product, by replacing the raw material property estimation model 182 with another raw material property estimation model 182. In other words, the control device 100b can flexibly respond to changes in the feature vector space, which is the coordinate space of the output of the raw material property estimation model 182.

[0110] Furthermore, by having the control device 100b estimate the properties of an object using the raw material property estimation model 182, the learning for estimating the properties of the object can be performed by a division of labor. For example, an information device manufacturer may provide the control device 100b and train the raw material property estimation model 182, while the learning for estimating the properties of the object, except for training the raw material property estimation model 182, may be performed by the object developer. This allows the object developer to maintain the confidentiality of information related to the development of the object, which is indicated in the additional information, and to outsource the training of the raw material property estimation model 182, which is a heavy load due to the use of large amounts of data. Furthermore, the control device 100b can automatically search for conditions for generating an object.

[0111] Second Embodiment Fig. 9 is a diagram showing an example of the configuration of a learning device according to at least one embodiment. In the configuration shown in Fig. 9, a learning device 610 includes a raw material property estimation unit 611 and a target property learning unit 612.

[0112] In this configuration, the raw material property estimation unit 611 inputs information indicating the composition or structure of a candidate raw material of the target product into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of a property value related to that substance, and that has been subjected to machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​related to that substance, and estimates the property values ​​related to the candidate raw material of the target product.

[0113] The object characteristic learning unit 612 uses training data that combines information indicating the conditions for generating an object, including characteristic values ​​related to candidate raw materials for the object, and information indicating the characteristic values ​​of the object when generated in accordance with that information, to learn an object characteristic estimation model that receives input of information indicating the conditions for generating an object and outputs estimated values ​​of the characteristic values ​​of the object. The raw material characteristic estimation unit 611 is an example of a raw material characteristic estimation means. The object characteristic learning unit 612 is an example of an object characteristic learning means.

[0114] The learning device 610 is expected to be able to utilize a pre-learning technique for training a raw material property estimation model even when sufficient data on the property values ​​of the target object is not available. In this respect, the learning device 610 is expected to be able to estimate the property values ​​of the target object with a relatively high degree of accuracy, using a relatively small amount of training data and a relatively small amount of calculation.

[0115] Furthermore, the learning device 610 can change a raw material candidate to another candidate, or change the properties of the raw material candidate that are used to estimate the property value of the target product, by replacing the raw material property estimation model with another raw material property estimation model. In other words, the learning device 610 can flexibly respond to changes in the feature vector space, which is the coordinate space of the output of the raw material property estimation model.

[0116] Furthermore, by having the learning device 610 estimate the properties of an object using the raw material property estimation model, the learning for estimating the properties of the object can be performed by a division of labor. For example, an information device manufacturer may provide the learning device 610 and train the raw material property estimation model, while the object developer may perform the learning for estimating the properties of the object other than training the raw material property estimation model. This allows the object developer to maintain the confidentiality of information related to the development of the object, which is indicated in the additional information, and to outsource the training of the raw material property estimation model, which is a heavy load due to the use of a large amount of data.

[0117] The raw material property estimation unit 611 can be realized, for example, by using the functions of the raw material property estimation unit 181 in Fig. 1. The target property learning unit 612 can be realized, for example, by using the functions of the target property estimation unit 185 in Fig. 1.

[0118] Third Embodiment Fig. 10 is a diagram showing an example of the configuration of a control device according to at least one embodiment. In the configuration shown in Fig. 10, a control device 620 includes a raw material property estimation unit 621, a condition determination unit 622, a control unit 623, and a target property learning unit 624.

[0119] In this configuration, the raw material property estimation unit 621 inputs information indicating the composition or structure of a candidate raw material of the target product into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of the property value of that substance, and that has been subjected to machine learning using training data that combines information indicating the composition or structure of the substance and information indicating the property value of that substance, and estimates the property value of the candidate raw material of the target product.

[0120] The condition determination unit 622 determines the conditions for generating an object whose characteristic values ​​are to be measured based on the probability distribution of estimated values ​​of the characteristic values ​​of the object output by an object characteristic estimation model that receives input of information indicating the conditions for generating the object, including characteristic values ​​related to candidate raw materials for the object, and outputs a probability distribution of estimated values ​​of the characteristic values ​​of the object.

[0121] The control unit 623 controls the device for generating the object so as to generate the object in accordance with the determined conditions. The object characteristic learning unit 624 learns the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated in accordance with those conditions. The raw material characteristic estimation unit 621 is an example of raw material characteristic estimation means. The condition determination unit 622 is an example of condition determination means. The control unit 623 is an example of control means. The object characteristic learning unit 624 is an example of object characteristic learning means.

[0122] It is expected that the control device 620 can utilize a pre-learning technique for training a raw material property estimation model even when sufficient data on the property values ​​of the target object is not available. In this respect, it is expected that the control device 620 can estimate the property values ​​of the target object with a relatively high degree of accuracy, using a relatively small amount of training data and a relatively small amount of calculation.

[0123] Furthermore, the control device 620 can change a raw material candidate to another candidate, or change the properties of the raw material candidate that are used to estimate the property value of the target product, by replacing the raw material property estimation model with another raw material property estimation model. In other words, the control device 620 can flexibly respond to changes in the feature vector space, which is the coordinate space of the output of the raw material property estimation model.

[0124] Furthermore, by having the control device 620 estimate the properties of the object using the raw material property estimation model, the learning for estimating the properties of the object can be performed by a division of labor. For example, an information device manufacturer may provide the control device 620 and train the raw material property estimation model, while the object developer may perform the learning for estimating the properties of the object other than training the raw material property estimation model. This allows the object developer to maintain the confidentiality of information related to the development of the object, which is indicated in the additional information, and to outsource the training of the raw material property estimation model, which is a heavy load due to the use of large amounts of data. Furthermore, the control device 620 can automatically search for conditions for generating the object.

[0125] 11 is a diagram showing an example of a processing procedure in a learning method according to at least one embodiment. The learning method shown in FIG. 11 includes estimating properties of candidate raw materials (step S611) and estimating property values ​​of a target object (step S612).

[0126] In estimating the properties of candidate raw materials (step S611), a computer inputs information indicating the composition or structure of a candidate raw material of a target product into a raw material property estimation model that receives input of information indicating the composition or structure of a substance and outputs an estimated value of a property value related to that substance, and that has undergone machine learning using training data that combines information indicating the composition or structure of the substance with information indicating property values ​​related to that substance, and estimates the property values ​​related to the candidate raw material of the target product.

[0127] In estimating the characteristic values ​​of the target object (step S612), the computer uses training data that combines information indicating the conditions for generating the target object, including characteristic values ​​related to candidate raw materials for the target object, and information indicating the characteristic values ​​of the target object when generated in accordance with that information, to learn an target object characteristic estimation model that receives input information indicating the conditions for generating the target object and outputs an estimated value of the characteristic value of the target object.

[0128] The learning method shown in Fig. 11 is expected to enable the use of a pre-learning technique for training a raw material property estimation model even when sufficient data on the property values ​​of the target object is not available. In this respect, the learning method shown in Fig. 11 is expected to enable the property values ​​of the target object to be estimated with a relatively high degree of accuracy, using a relatively small amount of training data and a relatively small amount of calculation.

[0129] 11, changing a raw material candidate to another candidate, or changing the properties of the raw material candidate used to estimate the property value of the target product, can be handled by replacing the raw material property estimation model with another raw material property estimation model. In other words, the learning method shown in FIG. 11 can flexibly handle changes in the feature vector space, which is the coordinate space of the output of the raw material property estimation model.

[0130] Furthermore, since the learning method shown in Figure 11 estimates the properties of an object using a raw material property estimation model, learning for estimating the properties of the object can be done through a division of labor. For example, an information device manufacturer may provide the device that executes the learning method shown in Figure 11 and train the raw material property estimation model, while a developer of the object may perform all of the learning for estimating the properties of the object except for training the raw material property estimation model. This allows the developer of the object to maintain the confidentiality of information related to the development of the object, which is shown in the additional information, and to outsource the training of the raw material property estimation model, which is a heavy load due to the use of a large amount of data.

[0131] Fifth Embodiment Fig. 12 is a diagram showing an example of a processing procedure in a control method according to at least one embodiment. The control method shown in Fig. 12 includes estimating properties of candidate raw materials (step S621), determining conditions (step S622), performing control (step S623), and performing learning (step S624).

[0132] The control method shown in Fig. 12 is expected to be able to utilize a pre-learning technique for training a raw material property estimation model even when sufficient data on the property values ​​of the target object is not available. In this respect, the control method shown in Fig. 12 is expected to be able to estimate the property values ​​of the target object with relatively high accuracy using a relatively small amount of training data and a relatively small amount of calculation.

[0133] 12, the raw material candidate can be changed to another candidate, and the properties of the raw material candidate used to estimate the property value of the target product can be changed by replacing the raw material property estimation model with another raw material property estimation model. In other words, the control method shown in FIG. 12 can flexibly respond to changes in the feature vector space, which is the coordinate space of the output of the raw material property estimation model.

[0134] Furthermore, since the control method shown in FIG. 12 estimates the properties of an object using a raw material property estimation model, the learning for estimating the properties of the object can be performed by a division of labor. For example, an information device manufacturer may provide the device that executes the control method shown in FIG. 12 and train the raw material property estimation model, while the object developer may perform all of the learning for estimating the properties of the object except for training the raw material property estimation model. This allows the object developer to maintain the confidentiality of information related to the development of the object, which is indicated in the additional information, and to outsource the training of the raw material property estimation model, which is a heavy load due to the use of large amounts of data. Furthermore, the control method shown in FIG. 12 allows the search for conditions for generating an object to be performed automatically.

[0135] 13 is a diagram illustrating an example of a computer configuration according to at least one embodiment. In the configuration shown in FIG. 8, a computer 700 includes a CPU 710, a main memory device 720, an auxiliary memory device 730, an interface 740, and a non-volatile recording medium 750.

[0136] One or more of the learning device 100, control device 100b, learning device 610, and control device 620, or a portion thereof, may be implemented in a computer 700. In this case, the operation of each of the above-described processing units is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-described processing in accordance with the program. The CPU 710 also allocates storage areas in the main storage device 720 corresponding to each of the above-described storage units in accordance with the program. Communication between each device and other devices is performed by an interface 740 having a communication function and communicating under the control of the CPU 710. The interface 740 also has a port for a non-volatile recording medium 750, and reads information from the non-volatile recording medium 750 and writes information to the non-volatile recording medium 750.

[0137] When learning device 100 is implemented in computer 700, the operations of processing unit 180 and each of its components are stored in the form of a program in auxiliary storage device 730. CPU 710 reads the program from auxiliary storage device 730, loads it into main storage device 720, and executes the above-described processing in accordance with the program.

[0138] Furthermore, the CPU 710 allocates a storage area for the storage unit 170 in the main storage device 720 in accordance with the program. Communication with other devices by the communication unit 110 is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Display of images by the display unit 120 is performed by the interface 740 having a display device and displaying various images under the control of the CPU 710. Reception of user operations by the operation input unit 130 is performed by the interface 740 having an input device and receiving user operations under the control of the CPU 710.

[0139] When the control device 100b is implemented in the computer 700, the operations of the processing unit 180b and each of its units are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-described processing in accordance with the program.

[0140] Furthermore, the CPU 710 allocates a storage area for the storage unit 170 in the main storage device 720 in accordance with the program. Communication with other devices by the communication unit 110 is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Display of images by the display unit 120 is performed by the interface 740 having a display device and displaying various images under the control of the CPU 710. Reception of user operations by the operation input unit 130 is performed by the interface 740 having an input device and receiving user operations under the control of the CPU 710.

[0141] When the learning device 610 is implemented in the computer 700, the operations of the raw material property estimation unit 611 and the target property learning unit 612 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-described processing in accordance with the program.

[0142] Furthermore, CPU 710 allocates a storage area in main memory 720 for learning device 610 to perform processing in accordance with the program. Communication between learning device 610 and other devices is performed by interface 740, which has a communication function and operates under the control of CPU 710. Interaction between learning device 610 and a user is performed by interface 740, which has an input device and an output device, presenting information to the user via the output device under the control of CPU 710 and accepting user operations via the input device.

[0143] When the control device 620 is implemented in the computer 700, the operations of the raw material property estimation unit 621, the condition determination unit 622, the control unit 623, and the target property learning unit 624 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing in accordance with the program.

[0144] Furthermore, the CPU 710 allocates a storage area in the main storage device 720 for the control device 620 to perform processing in accordance with the program. Communication between the control device 620 and other devices is performed by the interface 740, which has a communication function and operates under the control of the CPU 710. Interaction between the control device 620 and a user is performed by the interface 740, which has an input device and an output device, presenting information to the user via the output device under the control of the CPU 710 and accepting user operations via the input device.

[0145] One or more of the above-described programs may be recorded on nonvolatile recording medium 750. In this case, interface 740 may read the programs from nonvolatile recording medium 750. Then, CPU 710 may directly execute the programs read by interface 740, or may temporarily store the programs in main storage device 720 or auxiliary storage device 730 and then execute them.

[0146] Alternatively, a program for executing all or part of the processing performed by learning device 100, control device 100b, learning device 610, and control device 620 may be recorded on a computer-readable recording medium, and the program may be loaded into a computer system and executed to perform the processing of each component. Note that the term "computer system" herein includes hardware such as an operating system (OS) and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, read-only memories (ROMs), and compact disc read-only memories (CD-ROMs), as well as storage devices such as hard disks built into computer systems. The program may be designed to implement part of the aforementioned functions, or may be capable of implementing the aforementioned functions in combination with a program already stored in the computer system.

[0147] Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments and includes designs within the scope of the present invention. Furthermore, the above-described embodiments may be combined with other embodiments as appropriate.

[0148] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes.

[0149] (Supplementary Note 1) A learning device comprising: a raw material property estimation means for inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of the property value of the substance, and that has been subjected to machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​related to the substance; and an object property learning means for training an object property estimation model that receives input of information indicating the conditions for generating the target object, including the property values ​​of the candidate raw material of the target object, and outputs an estimated value of the property value of the target object when generated in accordance with the information, using training data that combines information indicating the conditions for generating the target object.

[0150] (Supplementary Note 2) The target object property estimation model comprises a partial model that receives input of information indicating the conditions for generating the target object other than property values ​​related to candidate raw materials of the target object, and outputs an estimated value of the difference between the property values ​​related to the candidate raw materials of the target object and the property values ​​related to the candidate raw materials of the target object; and the learning device described in Supplementary Note 1 inputs information indicating the conditions for generating the target object other than the property values ​​related to the candidate raw materials of the target object, which is input to the target object property estimation model, into the partial model, and outputs an estimated value of the property value of the target object, calculated using the output of the partial model and the property values ​​related to the candidate raw materials of the target object.

[0151] (Supplementary Note 3) The raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratios of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product. (Supplementary Note 3) The learning device described in Supplementary Note 1 or Supplementary Note 2.

[0152] (Supplementary Note 4) The learning device according to any one of Supplementary Notes 1 to 3, wherein the object property estimation model outputs a probability distribution of estimated values ​​of the property values ​​of the object.

[0153] (Supplementary Note 5) A learning device as described in Supplementary Note 4, further comprising a condition determination means for determining conditions for generating an object whose characteristic values ​​are to be measured based on a probability distribution of estimated values ​​of the characteristic values ​​of the object, and the object characteristic learning means for learning the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated according to the conditions.

[0154] (Appendix 6) The type of monomer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for generating the target object whose characteristic value is to be measured, at least one of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for generating the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

[0155] (Appendix 7) A learning device as described in Appendix 5, wherein the type of polymer used as a raw material for the target object is fixed, and the condition determination means determines at least one of information regarding the mixing of multiple types of polymers used as raw materials for the target object and information regarding a process for producing the target object using raw materials including a monomer used as a raw material for the target object as a condition for producing the target object whose characteristic value is to be measured.

[0156] (Supplementary Note 8) A control device comprising: a raw material property estimation means that estimates property values ​​for a candidate raw material of a target object by inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of a property value for the substance, the raw material property estimation model having undergone machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​for the substance; a condition determination means that determines conditions for generating an object that is the subject of property value measurement, based on a probability distribution of estimated property values ​​of the object output by the object property estimation model that receives information indicating conditions for generating an object, including the property values ​​of the candidate raw material of the target object; a control means that controls an apparatus for generating the object so as to generate the object in accordance with the determined conditions; and an object property learning means that learns the object property estimation model using training data that combines the determined conditions with measured values ​​of property values ​​of the object generated in accordance with the conditions.

[0157] (Supplementary Note 9) The control device according to Supplementary Note 8, wherein the object property estimation model comprises a partial model that receives input of information other than characteristic values ​​related to candidate raw materials of the object from among information indicating conditions for generating the object, and outputs an estimated value of a difference between the characteristic values ​​related to the candidate raw materials of the object and the characteristic values ​​related to the candidate raw materials of the object; and the control device inputs information other than characteristic values ​​related to the candidate raw materials of the object from among information indicating conditions for generating the object input to the object property estimation model into the partial model, and outputs an estimated value of the characteristic value of the object calculated using an output of the partial model and the characteristic values ​​related to the candidate raw materials of the object.

[0158] (Supplementary Note 10) The control device according to Supplementary Note 8 or Supplementary Note 9, wherein the raw material property estimation model receives input of information indicating a composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer, or an estimated value of a property value of a substance that contains the monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

[0159] (Appendix 11) The control device according to any one of Appendices 8 to 10, wherein the type of monomer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for producing the target object whose characteristic value is to be measured, at least one of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

[0160] (Supplementary Note 12) The control device according to any one of Supplementary Notes 8 to 10, wherein the type of polymer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for producing the target object whose characteristic value is to be measured, at least one of information regarding the kneading of multiple types of polymers used as raw materials for the target object and information regarding a process for producing the target object using raw materials including a monomer used as a raw material for the target object.

[0161] (Supplementary Note 13) A learning method comprising: a computer inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of a property value of the substance, the raw material property estimation model having undergone machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​of the substance; and learning an object property estimation model that receives information indicating the conditions for generating the object, including the property values ​​of the candidate raw material of the target object, and outputs an estimated value of a property value of the object, using training data that combines information indicating the conditions for generating the object, including the property values ​​of the candidate raw material of the target object, and information indicating the property values ​​of the object when generated in accordance with that information.

[0162] (Supplementary Note 14) The object property estimation model comprises a partial model that receives input of information indicating the conditions for generating the object other than the property values ​​related to the candidate raw materials of the object, and outputs an estimated value of the difference between the property values ​​related to the candidate raw materials of the object and the property values ​​related to the candidate raw materials of the object; and the learning method described in Supplementary Note 13, wherein information indicating the conditions for generating the object input to the object property estimation model other than the property values ​​related to the candidate raw materials of the object is input to the partial model, and an estimated value of the property value of the object is output, calculated using the output of the partial model and the property values ​​related to the candidate raw materials of the object.

[0163] (Supplementary Note 15) The raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratios of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product. (Supplementary Note 15) The learning method described in Supplementary Note 13 or Supplementary Note 14.

[0164] (Supplementary Note 16) The learning method according to any one of Supplementary Notes 13 to 15, wherein the object property estimation model outputs a probability distribution of estimated values ​​of the property values ​​of the object.

[0165] (Supplementary Note 17) The learning method described in Supplementary Note 16, including the computer determining conditions for generating an object whose characteristic values ​​are to be measured based on a probability distribution of estimated values ​​of the characteristic values ​​of the object, and learning the object characteristic estimation model includes the computer learning the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated in accordance with the conditions.

[0166] (Appendix 18) The learning method described in Appendix 17, wherein the type of monomer used as a raw material for the target object is fixed, and determining the conditions includes the computer determining, as conditions for generating the target object whose characteristic values ​​are to be measured, at least any of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for generating the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

[0167] (Appendix 19) The learning method described in Appendix 17, wherein the type of polymer used as a raw material for the target object is fixed, and determining the conditions includes the computer determining, as conditions for generating the target object whose characteristic values ​​are to be measured, at least one of information regarding the mixing of multiple types of polymers used as raw materials for the target object and information regarding a process for generating the target object using raw materials including a monomer used as a raw material for the target object.

[0168] (Supplementary Note 20) A control method comprising: a computer inputting information indicating the composition or structure of candidate raw materials of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs estimated values ​​of property values ​​related to the substance, and that has been subjected to machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​related to the substance, thereby estimating property values ​​related to the candidate raw materials of the target object; determining conditions for generating a target object whose property values ​​are to be measured, based on the probability distribution of estimated values ​​of property values ​​of the target object output by the target property estimation model that receives information indicating conditions for generating a target object, including the property values ​​related to the candidate raw materials of the target object; controlling a device for generating the target object so that the device generates the target object according to the determined conditions; and training the target object property estimation model using training data that combines the determined conditions with measured values ​​of property values ​​of the target object generated according to the conditions.

[0169] (Supplementary Note 21) The target object property estimation model comprises a partial model that receives input of information indicating the conditions for generating the target object other than property values ​​related to candidate raw materials of the target object, and outputs an estimated value of a difference between the property values ​​related to the candidate raw materials of the target object and the property values ​​related to the candidate raw materials of the target object; and the control method described in Supplementary Note 20, wherein information indicating the conditions for generating the target object input to the target object property estimation model other than the property values ​​related to the candidate raw materials of the target object is input to the partial model, and an estimated value of the property value of the target object is calculated using an output of the partial model and the property values ​​related to the candidate raw materials of the target object.

[0170] (Supplementary Note 22) The control method according to Supplementary Note 20 or Supplementary Note 21, wherein the raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer, or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

[0171] (Appendix 23) The control method according to any one of Appendices 20 to 22, wherein the type of monomer used as a raw material for the target object is fixed, and determining the conditions includes the computer determining, as conditions for producing the target object, the target object whose characteristic value is to be measured, at least any of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

[0172] (Appendix 24) The control method according to any one of Appendices 20 to 22, wherein the type of polymer used as a raw material for the target object is fixed, and determining the conditions includes the computer determining, as conditions for producing the target object whose characteristic values ​​are to be measured, at least one of information regarding the kneading of multiple types of polymers used as raw materials for the target object and information regarding a process for producing the target object using raw materials including a monomer used as a raw material for the target object.

[0173] (Supplementary Note 25) A program causing a computer to execute the following steps: inputting information indicating the composition or structure of candidate raw materials of a target object into a raw material property estimation model that has been subjected to machine learning using training data that combines information indicating the composition or structure of a substance with information indicating property values ​​related to the substance, to estimate property values ​​related to the candidate raw materials; and learning an object property estimation model that receives input of information indicating the conditions for generating the target object, including property values ​​related to the candidate raw materials of the target object, and outputs estimated property values ​​of the target object when generated in accordance with the information, using training data that combines information indicating the conditions for generating the target object.

[0174] (Appendix 26) The object characteristic estimation model comprises a partial model that receives input of information indicating the conditions for generating the object other than characteristic values ​​related to candidate raw materials of the object, and outputs an estimated value of the difference between the characteristic values ​​related to the candidate raw materials of the object and the characteristic values ​​related to the candidate raw materials of the object; and the program described in Appendix 25, wherein information indicating the conditions for generating the object input to the object characteristic estimation model other than the characteristic values ​​related to the candidate raw materials of the object is input to the partial model, and an estimated value of the characteristic value of the object is output, calculated using the output of the partial model and the characteristic values ​​related to the candidate raw materials of the object.

[0175] (Appendix 27) The program described in Appendix 25 or Appendix 26, wherein the raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

[0176] (Supplementary Note 28) The program according to any one of Supplementary Notes 25 to 27, wherein the object property estimation model outputs a probability distribution of estimated values ​​of the property values ​​of the object.

[0177] (Appendix 29) The program described in Appendix 28, which causes the computer to execute the steps of: determining conditions for generating an object whose characteristic values ​​are to be measured based on a probability distribution of estimated values ​​of the characteristic values ​​of the object; and, in training the object characteristic estimation model, causing the computer to execute training of the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated according to the conditions.

[0178] (Appendix 30) The program according to Appendix 29, wherein the type of monomer used as a raw material for the target object is fixed, and determining the conditions causes the computer to determine, as conditions for producing the target object whose characteristic values ​​are to be measured, at least any of information on polymers containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object, by determining the conditions for producing the target object whose characteristic values ​​are to be measured.

[0179] (Appendix 31) The program according to Appendix 29, wherein the type of polymer used as a raw material for the target object is fixed, and determining the conditions causes the computer to determine, as conditions for producing the target object whose characteristic values ​​are to be measured, at least one of information regarding the kneading of multiple types of polymers used as raw materials for the target object and information regarding a process for producing the target object using raw materials including a monomer used as a raw material for the target object.

[0180] (Supplementary Note 32) A program causing a computer to execute the following steps: inputting information indicating the composition or structure of candidate raw materials of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs estimated values ​​of property values ​​related to the substance, and that has been subjected to machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​related to the substance, thereby estimating property values ​​related to the candidate raw materials of the target object; determining conditions for generating a target object whose property values ​​are to be measured, based on the probability distribution of estimated property values ​​of the target object output by the target object property estimation model that receives information indicating conditions for generating a target object, including the property values ​​related to the candidate raw materials of the target object, and outputs a probability distribution of estimated property values ​​of the target object; controlling a device for generating the target object so as to generate the target object according to the determined conditions; and training the target object property estimation model using training data that combines the determined conditions with measured values ​​of property values ​​of the target object generated according to the conditions.

[0181] (Appendix 33) The object characteristic estimation model comprises a partial model that receives input of information indicating the conditions for generating the object other than characteristic values ​​related to candidate raw materials of the object, and outputs an estimated value of the difference between the characteristic values ​​related to the candidate raw materials of the object and the characteristic values ​​related to the candidate raw materials of the object; and the program described in Appendix 32, wherein information indicating the conditions for generating the object input to the object characteristic estimation model other than the characteristic values ​​related to the candidate raw materials of the object is input to the partial model, and an estimated value of the characteristic value of the object is output, calculated using the output of the partial model and the characteristic values ​​related to the candidate raw materials of the object.

[0182] (Appendix 34) The program described in Appendix 32 or Appendix 33, wherein the raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

[0183] (Appendix 35) The program according to any one of Appendices 32 to 34, wherein the type of monomer used as a raw material for the target object is fixed, and determining the conditions causes the computer to determine, as conditions for producing the target object whose characteristic values ​​are to be measured, at least any of information on polymers containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object, by determining the conditions for producing the target object whose characteristic values ​​are to be measured.

[0184] (Appendix 36) The program according to any one of Appendices 32 to 34, wherein the type of polymer used as a raw material for the target object is fixed, and determining the conditions causes the computer to determine, as conditions for producing the target object whose characteristic values ​​are to be measured, at least one of information regarding the kneading of multiple types of polymers used as raw materials for the target object and information regarding a process for producing the target object using raw materials including a monomer used as a raw material for the target object.

[0185] This application claims priority based on Japanese Patent Application No. 2024-103986, filed on June 27, 2024, the disclosure of which is incorporated herein by reference in its entirety.

[0186] The present disclosure may be applied to a learning device, a control device, a learning method, and a recording medium.

[0187] 1 Condition search system 100, 610 Learning device 100b, 620 Control device 110 Communication unit 120 Display unit 130 Operation input unit 170 Memory unit 180, 180b Processing unit 181, 611, 621 Raw material property estimation unit 182 Raw material property estimation model 183 Raw material property learning unit 184 Model selection unit 185 Object property estimation unit 186 Object property estimation model 187, 612, 624 Object property learning unit 188, 622 Condition determination unit 189 Data acquisition unit 190, 623 Control unit

Claims

1. A learning device comprising: a raw material property estimation means for inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of the property value of the substance, and for which machine learning has been performed using training data that combines information indicating the composition or structure of the substance and information indicating the property values ​​of the substance; and an object property learning means for training an object property estimation model that receives input of information indicating the conditions for generating the target object, including the property values ​​of the candidate raw material of the target object, and outputs an estimated value of the property value of the target object when generated in accordance with that information.

2. The object characteristic estimation model according to claim 1, wherein the object characteristic estimation model comprises a partial model that receives input of information indicating the conditions for generating the object other than characteristic values ​​related to candidate raw materials of the object, and outputs an estimated value of the difference between the characteristic values ​​related to the candidate raw materials of the object and the characteristic values ​​related to the candidate raw materials of the object; and the learning device according to claim 1, wherein information indicating the conditions for generating the object that is input to the object characteristic estimation model other than characteristic values ​​related to the candidate raw materials of the object is input to the partial model, and outputs an estimated value of the characteristic value of the object calculated using the output of the partial model and the characteristic values ​​related to the candidate raw materials of the object.

3. The learning device according to claim 1 or claim 2, wherein the raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains as a raw material the monomer that is a candidate raw material for the target product, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratios of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

4. The learning device according to any one of claims 1 to 3, wherein the object property estimation model outputs a probability distribution of estimated values ​​of the object property values.

5. A learning device as described in claim 4, further comprising a condition determination means for determining conditions for generating an object whose characteristic values ​​are to be measured based on a probability distribution of estimated values ​​of the characteristic values ​​of the object, and the object characteristic learning means learns the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated in accordance with those conditions.

6. The learning device described in claim 5, wherein the type of monomer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for generating the target object whose characteristic value is to be measured, at least one of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for generating the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

7. The learning device described in claim 5, wherein the type of polymer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for generating the target object whose characteristic value is to be measured, at least one of information regarding the mixing of multiple types of polymers used as raw materials for the target object and information regarding a process for generating the target object using raw materials containing a monomer used as a raw material for the target object.

8. A control device comprising: a raw material property estimation means that estimates property values ​​for a candidate raw material of a target object by inputting information indicating the composition or structure of a candidate raw material of a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs estimated property values ​​for the substance, the raw material property estimation model having undergone machine learning using training data that combines information indicating the composition or structure of the substance and information indicating property values ​​related to the substance; a condition determination means that determines conditions for generating an object whose property values ​​are to be measured based on the probability distribution of estimated property values ​​of the object output by the object property estimation model that receives information indicating conditions for generating an object, including property values ​​related to the candidate raw material of the target object, and outputs a probability distribution of estimated property values ​​of the object; a control means that controls an apparatus for generating the object so as to generate the object in accordance with the determined conditions; and an object property learning means that learns the object property estimation model using training data that combines the determined conditions with measured property values ​​of the object generated in accordance with the conditions.

9. The control device according to claim 8, wherein the object characteristic estimation model comprises a partial model that receives input of information other than characteristic values ​​related to candidate raw materials of the object from among information indicating the conditions for generating the object, and outputs an estimated value of the difference between the characteristic values ​​related to the candidate raw materials of the object and the characteristic values ​​related to the candidate raw materials of the object; and the control device according to claim 8, wherein information other than characteristic values ​​related to the candidate raw materials of the object from among the information indicating the conditions for generating the object input to the object characteristic estimation model is input to the partial model, and outputs an estimated value of the characteristic value of the object calculated using the output of the partial model and the characteristic values ​​related to the candidate raw materials of the object.

10. The control device according to claim 8 or claim 9, wherein the raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains the monomer that is a candidate raw material for the target product as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

11. A control device as described in any one of claims 8 to 10, wherein the type of monomer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for producing the target object whose characteristic value is to be measured, at least one of information on a polymer containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for producing the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

12. A control device as described in any one of claims 8 to 10, wherein the type of polymer used as a raw material for the target object is fixed, and the condition determination means determines, as a condition for producing the target object whose characteristic value is to be measured, at least one of information regarding the kneading of multiple types of polymers used as raw materials for the target object and information regarding a process for producing the target object using raw materials containing a monomer used as a raw material for the target object.

13. A learning method comprising: a computer inputting information indicating the composition or structure of candidate raw materials for a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of a property value for the substance, the raw material property estimation model having undergone machine learning using training data that combines information indicating the composition or structure of the substance with information indicating property values ​​for the substance; and training an object property estimation model that receives information indicating the conditions for generating the target object, including the property values ​​for the candidate raw materials for the target object, and outputs an estimated value of a property value for the target object, using training data that combines information indicating the conditions for generating the target object, including the property values ​​for the candidate raw materials for the target object, and information indicating the property values ​​of the target object when generated in accordance with that information.

14. The learning method described in claim 13, wherein the object property estimation model comprises a partial model that receives information other than characteristic values ​​related to candidate raw materials of the object from among information indicating the conditions for generating the object, and outputs an estimated value of the difference between the characteristic values ​​related to the candidate raw materials of the object and the characteristic values ​​related to the candidate raw materials of the object; and the learning method described in claim 13, wherein information other than characteristic values ​​related to the candidate raw materials of the object from among the information indicating the conditions for generating the object input to the object property estimation model is input to the partial model, and outputs an estimated value of the characteristic value of the object calculated using the output of the partial model and the characteristic values ​​related to the candidate raw materials of the object.

15. The learning method according to claim 13 or 14, wherein the raw material property estimation model receives input of information indicating the composition or structure of a monomer that is a candidate raw material for the target product, and outputs at least one of an estimated value of a property value of the monomer or an estimated value of a property value of a substance that contains that monomer as a raw material; and the target property estimation model receives input of the estimated value output by the raw material property estimation model and at least one of a molecular descriptor of the monomer that is a candidate raw material for the target product, information on a polymer that contains as a raw material the monomer that is a candidate raw material for the target product, information on the kneading of multiple types of polymers, information on a process for producing the target product using raw materials that contain the monomer, and information on the monomer ratios of multiple types of monomers that are candidate raw materials for the target product, and outputs an estimated value of the property value of the target product.

16. The learning method according to any one of claims 13 to 15, wherein the object property estimation model outputs a probability distribution of estimated values ​​of the object property values.

17. The learning method according to claim 16, further comprising the step of: determining conditions for generating an object whose characteristic values ​​are to be measured based on a probability distribution of estimated values ​​of the characteristic values ​​of the object; and learning the object characteristic estimation model includes the step of the computer learning the object characteristic estimation model using training data that combines the determined conditions with measured values ​​of the characteristic values ​​of the object generated in accordance with the conditions.

18. The learning method described in claim 17, wherein the type of monomer used as a raw material for the target object is fixed, and determining the conditions includes the computer determining, as conditions for generating the target object whose characteristic values ​​are to be measured, at least any of information on polymers containing a monomer that is a candidate raw material for the target object as a raw material, information on the kneading of multiple types of polymers, information on a process for generating the target object using raw materials containing the monomer, and information on the monomer ratio of multiple types of monomers that are candidate raw materials for the target object.

19. The learning method described in claim 17, wherein the type of polymer used as a raw material for the target object is fixed, and determining the conditions includes the computer determining, as conditions for generating the target object whose characteristic values ​​are to be measured, at least one of information regarding the mixing of multiple types of polymers used as raw materials for the target object and information regarding a process for generating the target object using raw materials including a monomer used as a raw material for the target object.

20. A recording medium storing a program that causes a computer to execute the following steps: inputting information indicating the composition or structure of candidate raw materials for a target object into a raw material property estimation model that receives information indicating the composition or structure of a substance and outputs an estimated value of the property value of the substance, and that has undergone machine learning using training data that combines information indicating the composition or structure of the substance with information indicating the property values ​​of the substance; and learning an object property estimation model that receives information indicating the conditions for generating the target object, including the property values ​​of the candidate raw materials for the target object, and outputs an estimated value of the property value of the target object, using training data that combines information indicating the conditions for generating the target object, including the property values ​​of the candidate raw materials for the target object, with information indicating the property values ​​of the target object when generated in accordance with that information.

Images