Model generation method, model generation apparatus, model generation program, estimation method, and estimation apparatus

Abstract

To provide novel technologies that enable the provision of information necessary to achieve the required properties of pharmaceutical formulations. [Solution] A model generation method comprising: an acquisition step of acquiring tactile information indicating the feel of a formulation and the blending ratio of raw materials in the formulation for each of several formulations; and a learning step of training a model that uses training data based on the information acquired in the acquisition step to output blending information related to blending ratios, taking tactile information as input.

Description

Technical Field

[0001] The present invention relates to a model generation method, a model generation device, a model generation program, an estimation method, and an estimation device.

Background Art

[0002] Conventionally, there have been numerous technologies for achieving specific characteristics and performance of products by appropriately combining multiple raw materials. In particular, in the manufacturing industries such as cosmetics and food, in order to provide products that meet diverse consumer needs, technologies for combining and formulating various raw materials are widely used. These technologies are important factors that influence the quality and functions of products, and the blending ratio of raw materials has a significant impact on the final texture and appearance of products.

[0003] In conventional product formulation methods, it was common for craftsmen or experts to determine the blending ratio of raw materials based on experience and knowledge. However, in recent years, due to the progress of digital technology, a method of generating a model for optimizing these blending ratios using data has attracted attention. As a result, product formulation can be performed more precisely and efficiently, enabling the rapid provision of customized products.

[0004] For example, the technology described in Patent Document 1 is related to a made-to-order system for cosmetics, and includes a color detection means for detecting the color of a detection object, an input means for a user to input a desired color, a blending ratio setting means for generating an optimal blending ratio from a plurality of cosmetic compositions based on the input color, and a cosmetic formulation device for formulating cosmetics based on the blending ratio.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] The technology described in Patent Document 1 allows for the formulation of cosmetics according to user requests regarding color, moisturizing power, and texture, but it has the drawback of not being able to design products from other perspectives.

[0007] In view of the above circumstances, the object of the present invention is to provide a novel technology that enables the provision of information for realizing the properties required of a pharmaceutical formulation. [Means for solving the problem]

[0008] [1] A model generation method comprising: an acquisition step of acquiring tactile information indicating the feel of a formulation and the blending ratio of raw materials in the formulation for each of several formulations; and a learning step of training a model that uses learning data based on the information acquired in the acquisition step, takes tactile information as input, and outputs blending information relating to the blending ratio.

[0009] By using this configuration, it is possible to generate a model that obtains information about formulations that can achieve the desired feel of a formulation, based on data from multiple formulations.

[0010] [2] The model generation method according to [1], further comprising a learning data creation step between the acquisition step and the learning step, in which the sum of the blending ratios of raw materials belonging to the same classification is calculated as the blending information, wherein in the learning step, the model is trained using the learning data in which the tactile information is an explanatory variable and the blending information obtained in the learning data creation step is an objective variable.

[0011] This configuration allows for the generation of a model capable of proposing the total blending ratio for each classification. Because the raw materials used in pharmaceutical formulations are diverse, data may be scarce for less frequently used materials. However, by training the model based on the total blending ratio for each classification, it is possible to generate a model capable of providing blending information tailored to the classification of the raw materials.

[0012] [3] The model generation method according to [2], wherein raw materials with the same blending purpose are classified as the same category, and the blending information is calculated showing the sum of the blending ratios of raw materials belonging to the same category.

[0013] By using this configuration, it is possible to generate a model that can propose a general configuration for each formulation purpose.

[0014] [4] A model generation method according to any one of [1] to [3], further comprising a learning data creation step between the acquisition step and the learning step, wherein for some of the plurality of formulations, the tactile information is acquired by estimating the tactile information based on the blending ratio, and in the learning step, the model is trained using the learning data including the tactile information obtained in the learning data creation step.

[0015] This configuration allows for the creation of training data by supplementing data even when there are formulations for which tactile information has not been obtained.

[0016] [5] A model generation method according to any one of [1] to [4], wherein the tactile information is obtained by having a subject respond to a subject provide a subjective evaluation value regarding the feel of the preparation, using evaluation terms that represent sensations obtained by human touch as indicators, in the acquisition step.

[0017] This configuration allows for the generation of a model that estimates formulation information based on evaluation values ​​that more directly influence the evaluation of the formulation when it is used.

[0018] [6] A model generation method according to any one of [1] to [5], wherein in the acquisition step, the tactile information is acquired based on the results of multiple subjects evaluating the tactile feel of the target formulation.

[0019] This configuration allows for the acquisition of tactile information regardless of individual differences in perception.

[0020] [7] In the learning process, among the multiple types of formulation information included in the learning data, for the formulation information in which there are a predetermined number or more of data with non-zero values, the model is trained using the formulation information as the target variable, according to the model generation method according to any one of [1] to [6].

[0021] By adopting such a configuration, an effect of improving the accuracy of the model can be expected.

[0022] [8] A model generation device including an acquisition unit and a learning unit, wherein the acquisition unit acquires, for each of a plurality of formulations, the tactile information indicating the tactile sensation of the formulation and the blending ratio of the raw materials in the formulation, and the learning unit uses the learning data based on the information acquired by the acquisition unit to train a model that takes the tactile information as an input and outputs the formulation information regarding the blending ratio.

[0023] [9] A model generation program that causes a computer to function as an acquisition unit and a learning unit, wherein the acquisition unit acquires, for each of a plurality of formulations, the tactile information indicating the tactile sensation of the formulation and the blending ratio of the raw materials in the formulation, and the learning unit uses the learning data based on the information acquired by the acquisition unit to train a model that takes the tactile information as an input and outputs the formulation information regarding the blending ratio.

[0024]

[10] An estimation method for estimating based on the relationship between the tactile sensation of a formulation and the blending ratio of raw materials, wherein tactile information indicating the tactile sensation desired for a formulation is input into the model generated by the model generation method according to any one of [1] to [7], and the formulation information regarding the blending ratio of the raw materials of the formulation that realizes the tactile sensation is acquired as the output of the model, thereby estimating the formulation information from the tactile information.

[0025]

[11] The estimation method according to

[10] , wherein the formulation information is estimated to include the blending ratio of the raw materials of the formulation and the total value of the blending ratios of the raw materials belonging to the same classification.

[0026]

[12] An estimation device comprising a reception unit and an estimation unit, wherein the reception unit receives an input of texture information indicating a desired texture for a pharmaceutical preparation, and the estimation unit inputs the texture information received by the reception unit into the model generated by the model generation method according to claim 1, and obtains, as the output of the model, blending information regarding the blending ratio of the raw materials of the pharmaceutical preparation that realizes the texture, thereby estimating the blending information from the texture information.

Advantages of the Invention

[0027] According to the present invention, it is possible to provide a novel technique for estimating blending information for realizing the texture required for a pharmaceutical preparation.

Brief Description of the Drawings

[0028] [Figure 1] Configuration diagram of this embodiment. [Figure 2] Hardware configuration diagram. [Figure 3] Flowchart according to the model generation method of this embodiment. [Figure 4] Flowchart regarding the details of the learning data creation process of this embodiment [Figure 5] An example of the classification of raw materials in this embodiment. [Figure 6] An example of learning data in Example 1. [Figure 7] Accuracy analysis results of Example 1. [Figure 8] Accuracy analysis results of Example 1. [Figure 9] Accuracy analysis results of Example 2. [Figure 10] Accuracy analysis results of Example 2. [Figure 11] Accuracy analysis results of Example 2. [Figure 12] Accuracy analysis results of Example 2.

Modes for Carrying Out the Invention

[0029] This invention relates to a model that generates an output regarding the blending ratio of raw materials based on input regarding the required feel of a product produced by blending multiple raw materials. In particular, this invention relates to a model generation method, a model generation apparatus, and a model generation program. This invention also relates to an estimation method and an estimation apparatus that use the model to estimate the blending ratio.

[0030] In the following embodiment, a model is described as one that estimates the formulation information of a formulation that achieves a given texture from the texture information, using the texture evaluation value as texture information and the sum of the blending ratios for each raw material and / or the blending ratios for each classification of raw materials as formulation information.

[0031] Further details will be provided below with reference to the attached drawings. While preferred embodiments are shown in the drawings, many different forms are possible and the embodiments are not limited to those described herein.

[0032] For example, in this embodiment, the configuration and operation of the model generation device and estimation device are described, but similar effects can be achieved by a system or device having similar functions, a method performed by such a system or device, or a computer program that causes a computer device to perform the method. The program may be provided as a non-transient recording medium that can be read by a computer, or it may be provided so that it can be downloaded from an external server.

[0033] In the following embodiments, "part" may include, for example, hardware resources implemented by a circuit in a broad sense, and information processing of software that can be specifically realized by these hardware resources. In this embodiment, "information" can be represented, for example, by the physical value of a signal value representing voltage or current, the high or low value of a signal value as a set of binary bits composed of 0s or 1s, or by a quantum superposition (so-called qubit), and communication and calculations can be performed on a circuit in a broad sense.

[0034] In a broad sense, a circuit is a set of circuits (Circuitry) that are realized by appropriately combining circuits, processors, and memory. For example, it includes circuits that contain any of the following: CPU (Central Processing Unit), GPU (Graphics Processing Unit), LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), etc.

[0035] <1. Configuration of the Model Generation System> Figure 1 is a block diagram showing the configuration of the model generation system of this embodiment. The model generation system shown in Figure 1 comprises a model generation device 1, a database DB, and an estimation device 2. Here, we will first describe the model generation device 1.

[0036] The model generation device 1 comprises an acquisition unit 11, a training data creation unit 12, and a training unit 13. This represents a concrete implementation of software-based information processing by hardware. These configurations do not necessarily need to be implemented by a single computer; they may be implemented through the collaboration of multiple computers. The model generation device 1 is configured to communicate with a database DB via wired or wireless connection, and the generated models are stored in the database DB.

[0037] The acquisition unit 11 performs an acquisition process to acquire tactile information indicating the feel of the formulation and the blending ratio of raw materials in the formulation for each of several formulations. The specific procedure will be described in detail later. In this embodiment, in addition to tactile information, physical property information relating to the physical properties of the formulation is also acquired.

[0038] The learning data creation unit 12 executes a learning data creation process to create learning data based on the tactile information and blending ratio acquired by the acquisition unit 11. In this embodiment, the learning data creation unit 12 further creates learning data based on the physical property information of the formulation.

[0039] The learning data creation unit 12 of this embodiment calculates the sum of the blending ratios of raw materials belonging to the same classification as blending information. In this embodiment, the classification is set according to the blending purpose. Details of the raw material classification will be described later.

[0040] Furthermore, the learning data creation unit 12 of this embodiment obtains estimated tactile information for formulations for which the blending ratio has been obtained but tactile information has not been obtained, based on the blending ratio of the formulation.

[0041] The learning unit 13 uses the learning data created by the learning data creation unit 12 based on the information acquired by the acquisition unit 11 to perform a learning process in which it takes tactile information as input and outputs formulation information related to the formulation ratio. Specifically, the learning data includes pairs of tactile information and formulation information for multiple formulations, and the model learns using tactile information as the explanatory variable and formulation information as the target variable.

[0042] In this embodiment, the learning unit 13 learns not only tactile information but also physical property information as explanatory variables. Details of the tactile information, formulation information, and physical property information in this embodiment, as well as the learning procedure, will be described later.

[0043] <2. Hardware Configuration> Next, the hardware configuration of the model generation system in this embodiment will be described. One or more general-purpose servers or personal computers can be used as the model generation device 1, the estimation device 2, and the database DB. In this embodiment, the model generation device 1 is an information processing device 10 on which a computer program (model generation program) that executes the model generation method is installed. Similarly, in this embodiment, the estimation device 2 is an information processing device 10 on which a computer program (estimation program) that executes the estimation method is installed.

[0044] Figure 2 is a hardware configuration diagram of the information processing device 10. As shown in Figure 2, the information processing device 10 has a control unit 101, a storage unit 102, and a communication unit 103, which are used to perform the functions of each unit and each process.

[0045] The control unit 101 has a processor such as a CPU that can execute instruction sets, and executes the OS and programs. The memory unit 102 includes volatile memory such as RAM capable of storing instruction sets, and non-volatile recording media such as HDDs or SSDs capable of storing the OS, model generation programs, DBMS, etc. The communication unit 103 has an interface for physically connecting to the network and performs communication control with the network NW to input and output information.

[0046] <3. Model Generation Method> Next, the model generation procedure in this embodiment will be described with reference to Figure 3. Figure 3 is a flowchart of the model generation method according to this embodiment. First, in step S11, the acquisition unit 11 acquires tactile information, physical property information, and measured data of the blending ratio for each raw material for each of the multiple formulations.

[0047] Here, tactile information refers to information that represents the sensations obtained by human touch when the target formulation is touched. In this embodiment, subjective evaluation values ​​regarding the feel of the formulation, using evaluation terms that represent tactile sensations as indicators, are used as tactile information.

[0048] Here, "evaluation terms" refer to words that express subjective evaluations. Typical evaluation terms include terms that describe sensations directly evoked by the object itself, such as hard, soft, moist, and smooth (sensory terms). In addition, terms that express emotions, such as like, dislike, fun, sad, and surprised (emotional terms), and terms that express abstract impressions of an object, such as luxurious, attractive, special, and gentle (impressionary terms), may also be used as evaluation terms. Furthermore, multiple types of tactile information may be used, each of these evaluation terms serving as an indicator.

[0049] Furthermore, evaluation values ​​based on evaluation terms refer to values ​​obtained through the evaluator's subjective evaluation, which considers how well the object fits the impression represented by each evaluation term. The evaluation value is preferably a sensitivity value indicating the strength of the sensation perceived by the evaluator as represented by the evaluation term. The evaluation value can be expressed as a number or integer between 0 and 100, for example, but the minimum and maximum values ​​can be arbitrarily defined. Furthermore, the evaluator's subjective evaluation may be obtained as a binary value of whether or not the evaluation term is appropriate for the object, and this may also be used as the evaluation value. Alternatively, the sensation may simply be numerically evaluated on an axis of good or bad.

[0050] In this embodiment, evaluation values ​​based on the following 18 evaluation terms (A1) to (A18) were used as tactile information. The measured values ​​of these 18 tactile items were obtained using the CATA (Check-All-That-Apply) method.

[0051] (A1) soft, (A2) hard, (A3) smooth, (A4) moist, (A5) smooth, (A6) rough, (A7) warm, (A8) cold, (A9) sticky, (A10) gooey, (A11) firm, (A12) spreadable, (A13) absorbed, (A14) penetrates, (A15) leaves a film, (A16) oily, (A17) fresh, (A18) plump.

[0052] Items (A1) to (A10) above are general tactile sensations based on the tactile dimension, while items (A11) to (A18) are tactile sensations frequently used in the manufacturing of cosmetic formulations.

[0053] In this embodiment, the results of tactile tests conducted by multiple subjects are stored in an arbitrary storage device, and tactile information for each formulation is obtained based on the aggregated results. That is, the tactile information corresponding to each formulation is a compilation of the results of tactile evaluations of that formulation by multiple subjects. Specifically, it is envisioned that multiple subjects will be asked to provide the above evaluation values ​​for each formulation, and the average value of these values ​​will be used as the tactile information for that formulation.

[0054] In this embodiment, in addition to tactile information, physical property information relating to the physical properties of the formulation is also used as an explanatory variable. The physical property information in this embodiment is a quantitative representation of the physical properties of the formulation, including its mechanical and thermal properties. The physical property information is obtained by measuring the formulation with a physical property measuring instrument. Multiple types of physical property information may be used.

[0055] Next, in step S12, the training data creation unit 12 creates training data based on the data acquired by the acquisition unit 11. Specifically, the training data creation unit 12 performs processes to supplement missing training data and processes to calculate the blending information used for training the model. Details of each process will be described later.

[0056] Then, in step S13, the learning unit 13 executes the model learning process based on the learning data created in step S12. Specifically, learning is performed using tactile information and physical property information as explanatory variables and formulation information as the target variable. Further details will be described later.

[0057] As described above, the model generation device 1 executes the acquisition process, the learning data creation process, and the learning process, thereby generating a model that estimates formulation information from tactile information and physical property information. The model thus generated is transmitted to the database DB by the model generation device 1 and stored in the database DB.

[0058] <4. Creating training data> The details of the training data creation process in step S12 of Figure 3 described above will now be explained. In this embodiment, tactile information, physical property information, and formulation information are acquired for multiple products, and these datasets are used as training data.

[0059] Figure 4 is a flowchart showing the various processes included in the learning data creation process in this embodiment. Figure 4(a) shows the process for supplementing the learning data, and Figure 4(b) shows the process for calculating blending information based on the blending ratio of each raw material.

[0060] In the acquisition process described above, the proportions of each raw material in the formulation are often known, but in order to obtain tactile information and physical property information, it is necessary to actually evaluate (measure) the formulation. Therefore, the cost of obtaining tactile and physical property information is high, and there are cases where only the proportions and formulation (mixing conditions, etc.) are obtained in the acquisition process, and there are formulations for which tactile information and physical property information are blank (no data).

[0061] Therefore, in the learning data completion process, data is completed by estimating tactile information and / or physical property information using a pre-generated predictive model. The predictive model can be generated using known techniques, such as the learning method for a predictive model described in Japanese Patent Publication No. 2022-023482 or the learning method for model parameters described in Japanese Patent Publication No. 2023-001429.

[0062] For example, by using the raw material blending ratio as an explanatory variable and physical properties and tactile information as the dependent variable, and training with multiple training datasets, a model can be generated that estimates physical properties and tactile information from the blending ratio. In addition to the blending ratio, the manufacturing method (stirring method, stirring time, stirring speed, and emulsification temperature, etc.) may also be used as an explanatory variable. Furthermore, in estimating tactile information, physical properties may also be added as explanatory variables.

[0063] Any publicly known machine learning algorithm applicable to numerical prediction problems may be used to generate models for estimating tactile information and physical properties. For example, a model may be created using a regression model or a neural network model.

[0064] Figure 4(a) shows the procedure for supplementing data by estimating tactile information based on the blending ratio using such a trained model. Here, the blending ratio obtained in step S11 of Figure 3 before the start of Figure 4(a) is the blending ratio of the raw materials required as input to the pre-generated prediction model.

[0065] In step S21, the learning data creation unit 12 inputs the blending ratio into the prediction model and obtains tactile information as the output of the prediction model. Here, we will explain the estimation of tactile information, but physical property information can also be estimated and obtained in the same way.

[0066] Then, in step S22, the learning data creation unit 12 creates learning data by taking the acquired tactile information and physical property information as tactile information and physical property information of the formulation corresponding to the formulation ratio acquired in step S11. In this way, by supplementing the learning data with estimation, the amount of learning data can be secured, and an improvement in the accuracy of the model for estimating formulation information can be expected.

[0067] Furthermore, an upper limit may be set on the ratio of estimated data from the above prediction model to the number of measured data, so that the estimated data does not exceed a certain proportion. In other words, it is preferable to use the measured tactile information and physical property information for learning, and to supplement a portion of it with the above estimated values.

[0068] Next, the formulation information calculation process shown in Figure 4(b) will be explained. In step S31, the formulation ratios for each raw material obtained in step S11 are added together for each classification of raw materials to calculate a total value. Then, in step S31, this total value for each classification is used as formulation information to create training data.

[0069] Here, we will explain the classification of raw materials. In this embodiment, raw materials are classified according to their blending purpose, and the sum of the blending ratios of raw materials belonging to the same classification is used as blending information. In particular, in this embodiment, a classification with multiple levels is defined, and each raw material is classified accordingly. For example, as shown in Figure 5, a classification called "water-soluble components" is defined, and below that, subcategories such as "mono-dihydric alcohols," "tertiary-tetrahydric alcohols," "pentahydric or higher alcohols," "sugar alcohols," and "polymers" are defined. Each raw material is then classified into one of these subcategories. Furthermore, a higher-level classification (major classification) that combines multiple classifications may be defined, and blending information may be calculated for each of these higher-level classifications. If a raw material has multiple blending purposes, the raw material may belong to multiple classifications.

[0070] In this embodiment, the blending information calculation process shown in Figure 4(b) calculates the sum of the blending rates for each classification and uses this as the blending information in the training data. Alternatively, or in addition to this, the sum of the blending rates for each sub-classification or the sum of the blending rates for each major classification may also be used as blending information.

[0071] Because the raw materials used in pharmaceutical formulations are extremely diverse, it may be difficult to obtain sufficient training data for raw materials that are not frequently used. Therefore, as mentioned above, using the blending ratios for each classification according to the purpose of blending can be expected to improve the stability of prediction accuracy.

[0072] <5. Selection of the dependent variable> In this embodiment, the learning process uses both the sum of the blending rates for each classification and the blending rate for each raw material as target variables. However, as mentioned earlier, since the raw materials used in pharmaceutical formulations are diverse, some raw materials may only have data with a blending rate of 0 (no data exists for formulations containing that raw material), or the proportion of data with a blending rate of 0 may be extremely high. For such raw materials, it is considered difficult to achieve accurate predictions even if learning is performed using the individual blending rates as the target variable. The same can be said for the sum of the blending rates for each classification.

[0073] Therefore, in this embodiment, the target variable is an item for which there is a predetermined number or more data points (formulations) where the blending ratio of raw materials and the sum of the blending ratios for each category are not zero, and items for which there are fewer than a predetermined number of non-zero data points are excluded from the target variable. Here, "item" refers to the blending ratio for each raw material or the sum of the blending ratios for each category, which are candidates for the target variable.

[0074] The exclusion criteria for the dependent variable can be set arbitrarily, but for example, one could set a threshold for the proportion of data with a value of 0 in the total data, or the number of non-zero data points, and exclude items that do not meet the threshold from the dependent variable. In this embodiment, items containing 50 or more non-zero data points are adopted as the dependent variable.

[0075] <6. Learning> In this embodiment, a model is generated using machine learning and statistical analysis with the above-mentioned training data. The model configuration and algorithm can be arbitrarily determined, and for example, regression models and neural networks can be used. Alternatively, the distribution of blending ratios for each raw material may be predicted using the Gaussian Mixture Regression method. Furthermore, ensemble learning such as XGboost or Random Forest using a decision tree model may be utilized.

[0076] Furthermore, in this embodiment, tactile information and physical property information are used as explanatory variables to output the blending ratio for each of the multiple raw materials and the sum of the blending ratios for each classification. To output such multiple target variables, a package such as MultiOutputRegressor can be used.

[0077] Here, depending on the raw material, the data may often consist only of the same numerical value. For example, there may be raw materials in which the blending ratio is either 0 or 0.1 in all formulations. For such raw materials, training the model to estimate the "blending ratio" as described above may not yield an accurate estimate. Therefore, for such raw materials, it may be better to generate a model as a classification problem that predicts whether or not the raw material is blended (whether or not it is blended) rather than the blending ratio.

[0078] In this case, possible models include, for example, logistic regression, k-nearest neighbors, naive Bayesian modeling, support vector machines (SVM), decision trees, and neural networks. The explanatory variables in this case can also be the same information as described above.

[0079] Whether to predict "presence or absence of combination" (classification problem) or "combination rate" (value estimation) can be determined, for example, by the distribution of values ​​in the training data. For instance, one could check what kinds of combination rate values ​​are included in the training data, and if there are only two types of values ​​(e.g., 0 and 0.1) and they are below a predetermined threshold, then predict "presence or absence of combination" instead of "combination rate." Furthermore, the above judgment may be applied not only to the blending ratio of each raw material, but also to the sum of the blending ratios for each category.

[0080] In other words, the target variable (dependent variable) of the model that estimates the sum of the blending ratios for each raw material and each classification is the item that does not fall under the classification of whether or not it contains the above-mentioned blending, among the items that contain 50 or more non-zero data points. Furthermore, the items to be estimated by the model for estimating the presence or absence of a compound (raw materials or classifications for which presence or absence is estimated) are those items that contain 50 or more non-zero data points and for which the prediction of "presence or absence of compound" is deemed appropriate based on the conditions described above.

[0081] <7. Configuration of the Estimation Device> Next, we will describe the estimation of formulation information using the model generated by the procedure described above. In this embodiment, the estimation device 2 estimates formulation information using the model stored in the database DB. The estimation device 2 comprises a reception unit 21, an estimation unit 22, and an output unit 23. The estimation device 2 is configured to communicate with the database DB via wired or wireless connection and can utilize the model stored in the database DB.

[0082] The reception unit 21 receives input of tactile information indicating the desired feel of the formulation. Specifically, for each type of tactile information, it accepts the range and level of the evaluation value for the desired feel. In this embodiment, the reception unit 21 also accepts the range and level of physical property values ​​for each physical property.

[0083] The estimation unit 22 inputs the tactile information and physical property information received by the reception unit 21 into the model generated by the model generation method described above, which is stored in the database DB. Then, the estimation unit 22 obtains the formulation information as the output of the model.

[0084] It is known that there is a certain relationship between tactile sensation and physical properties. Therefore, based on this relationship, physical property information for input into the model may be obtained by estimating the corresponding physical property information from the tactile information. In this case, the receiving unit 21 only receives tactile information, and the estimation unit 22 estimates the physical property information or identifies it based on a pre-registered correspondence, and inputs it into the model.

[0085] The output unit 23 outputs based on the formulation information estimated by the estimation unit 22. For example, it is expected that the output unit 23 will send the result of display processing of the estimated formulation information to a display device and display it on the screen.

[0086] Here, the tactile information and physical property information input to the model by the estimation unit 22 must have the same items as those used to generate the model. For example, if the evaluation value for the evaluation term "soft" was used in the model generation, the reception unit 21 receives tactile information from the user indicating the desired "softness" of the formulation, and the estimation unit 22 inputs this into the model.

[0087] The model then outputs the formulation information learned as the target variable, and the output unit 23 outputs this information to the user, allowing the user to obtain reference information on the formulation necessary for the desired product.

[0088] The estimation unit 22 may use multiple models. For example, it may use multiple models to predict the blending ratios of different raw materials, or it may use a different classification model to predict whether or not some of the raw materials are included in the blend. [Examples]

[0089] The following describes the learning model generation experiments conducted by the inventors. While the present invention is based on the following experimental methods and results, it is not limited to the following embodiments. In this example, we divided the target formulation into non-emulsified and emulsified types, and attempted to generate two models: one that estimates the blending ratio of each raw material in each category, and another that estimates the total blending ratio for each category.

[0090] (1) Explanatory variables We performed two types of training: one using tactile information and physical property information as explanatory variables, and another using only tactile information.

[0091] (2) Dependent variable For non-emulsified formulations, we predicted the blending ratios for 35 out of 420 candidate raw materials that contained 50 or more non-zero data points. We also divided these raw materials into 18 categories and attempted to predict the total blending ratio for each category. Of these 18 categories, we predicted the total blending ratios for 13 categories that contained 50 or more non-zero data points.

[0092] Figure 6 shows an example of the training data used in this embodiment. As shown, training data was prepared using multiple types of tactile and physical property information as explanatory variables for each formulation, and the blending ratio for each raw material and the sum of the blending ratios for each raw material classification ("Blending Amount by Raw Material Category" in Figure 6) as the objective variables. The "blending ratio" was defined as mass percentage concentration (wt%).

[0093] Furthermore, for emulsified formulations, we predicted the blending ratio for 67 raw materials out of the 808 candidate raw materials that contained 50 or more non-zero data points. In addition, we predicted the total blending ratio for 17 of the 18 categories mentioned above that contained 50 or more non-zero data points.

[0094] Furthermore, we attempted to predict the total blending ratio for each larger category (combining multiple classifications) and the total blending ratio for each smaller category (further subdividing the classifications) for both non-emulsified and emulsified formulations.

[0095] (3) Results Using training data on numerous formulations under the above conditions, test data was input into a model to predict formulation information, and the coefficient of determination R^2 was analyzed for various formulation information. The results of the analysis of indicators related to prediction accuracy are shown in Figures 7 and 8 for non-emulsified and emulsified formulations, respectively. In Figures 7-12, tactile value is a type of tactile information and represents a value indicating tactile sensation. In models using both tactile and physical property information as explanatory variables, R^2 > 0.5 was obtained for the sum of the blending rates of numerous raw materials and the blending rates for each classification, indicating a certain level of prediction accuracy. Furthermore, even in models that used only tactile information as explanatory information, several items (raw material blending ratios and the sum of blending ratios for each category) were found to have R^2 > 0.5, indicating that a certain level of prediction accuracy was observed even with predictions based solely on tactile information.

[0096] In particular, regarding the total blending ratio for each classification, R^2 > 0.5 was predicted for 4 out of 13 classifications in non-emulsified formulations and 7 out of 17 classifications in emulsified formulations, based on tactile and physical property information. Furthermore, regarding the prediction results based solely on tactile information, R^2 > 0.5 was observed for one out of 13 classifications for non-emulsified formulations and for two out of 17 classifications for emulsified formulations.

[0097] Furthermore, for emulsified formulations, the total blending ratio for each major category was also determined to have a certain level of predictive accuracy, as R^2 > 0.5 for some items in both models using both tactile and physical property information as explanatory variables, and models using only tactile information as explanatory variables.

[0098] In predicting the total blending ratio for each subcategory, a model using both tactile and physical property information as explanatory variables resulted in R^2 > 0.5 for some items in both non-emulsified and emulsified formulations. Furthermore, in the case of emulsified formulations, even in models that used only tactile information as an explanatory variable, there were items where R^2 > 0.5 was obtained when predicting the total value of the formulation ratio for each subcategory. [Examples]

[0099] In this example, the target formulations were divided into non-emulsified and emulsified formulations, and an attempt was made to generate a model that estimates (classifies) the presence or absence of each ingredient or classification within each. Two types of algorithms were used for training the models: Balanced RandomForest and RandomForest. RandomForest is an ensemble learning algorithm that uses decision trees as weak learners, while Balanced RandomForest is an algorithm that addresses imbalanced data by adjusting the number of samples for each decision tree.

[0100] (1) Explanatory variables Similar to Example 1, learning was performed using both tactile information and physical property information as explanatory variables, and learning was performed using only tactile information. Examples of explanatory variables are shown in Figure 6.

[0101] (2) Items to be predicted Among the candidate raw materials and classifications, those that contained 50 or more non-zero data points were targeted for prediction of the presence or absence of the ingredient, based on the distribution of the training data, where it was determined that predicting the presence or absence of the ingredient was more appropriate than predicting the blending ratio. In Example 2, instead of the blending ratio itself shown in Figure 6, data indicating whether the blending ratio for each item in each formulation was 0 or non-zero (presence or absence of blending) was prepared and used as training data.

[0102] (3) Results Using training data on numerous formulations under the above conditions, a model was generated and test data was input to predict formulation information. For various formulation information, recall, accuracy, and Kappa coefficient were analyzed. These are all commonly used indicators of estimation accuracy.

[0103] Figure 9 shows a list of recall, accuracy, and Kappa coefficients for the prediction results of the presence or absence of each ingredient in non-emulsified formulations using the Balanced RandomForest model. Figure 10 also shows a list of recall, accuracy, and Kappa coefficients for the prediction results of the presence or absence of each ingredient in emulsified formulations using the Balanced RandomForest model.

[0104] Thus, the Balanced RandomForest model demonstrated a certain level of accuracy across numerous parameters.

[0105] Figure 11 also shows a list of recall, accuracy, and Kappa coefficients for the prediction results of the presence or absence of each ingredient in non-emulsified formulations, based on the RandomForest model.

[0106] In the case of the RandomForest model, a bias in the classes (presence or absence of formulation) in the training data can lead to a decrease in prediction accuracy. Therefore, in this example, the number of data points for formulations with formulations and the number of data points for formulations without formulations were compared, and the classes of the training data were balanced to match the smaller number of data points before training was performed.

[0107] As a result, as shown in Figures 11 and 12, the RandomForest model also demonstrated a certain level of accuracy in many items.

[0108] As described above, it was suggested that the formulation information that realizes a given texture can be estimated with a certain degree of accuracy from tactile information. In particular, it was possible to predict with a certain degree of accuracy the formulation ratio for each raw material and the sum of the formulation ratios for each classification of raw materials. Furthermore, the prediction accuracy of the model may be further improved by adding training data, modifying the model configuration, and adjusting parameters. Moreover, it was shown that accuracy can be further improved by using physical property information as the target variable in addition to tactile information.

[0109] Thus, by estimating the blending ratio of each raw material, or the sum of the blending ratios for each classification of raw materials, based on information indicating the desired texture, it is possible to provide useful information for formulation design. In particular, it has been shown that the sum of the blending ratios for each classification according to the purpose of formulation can be predicted with relatively high accuracy, and that this can provide information about the overall framework related to formulation design. [Explanation of symbols]

[0110] 1: Model Generator 11: Acquisition part 12: Training Data Creation Section 13: Learning Department 2: Estimation device 21: Reception Department 22:Estimation part 23: Output section 10: Information Processing Device 101: Control Unit 102: Storage section 103: Communications Department NW: Network

Claims

1. A process for obtaining tactile information indicating the feel of the formulation and the proportion of raw materials in the formulation for each of several formulations, A model generation method comprising: a learning step of training a model that uses learning data based on the information acquired in the acquisition step, takes tactile information as input, and outputs blending information relating to the blending ratio.

2. The process further includes a learning data creation step between the acquisition step and the learning step, in which the sum of the blending ratios of raw materials belonging to the same classification is calculated as the blending information. The model generation method according to claim 1, wherein in the learning step, the model is trained using the learning data in which the tactile information is used as an explanatory variable and the formulation information obtained in the learning data creation step is used as an objective variable.

3. The model generation method according to claim 2, wherein in the learning data creation step, raw materials with the same blending purpose are classified as the same category, and the blending information is calculated to show the sum of the blending ratios of raw materials belonging to the same category.

4. The process further includes a learning data creation step, in which, between the acquisition step and the learning step, the tactile information for a portion of the plurality of formulations is acquired by estimating the tactile information based on the blending ratio, The model generation method according to claim 1, wherein the learning step involves training the model using the learning data, which includes the tactile information obtained in the learning data creation step.

5. The model generation method according to claim 1, wherein in the acquisition step, the tactile information is acquired by having a subject respond to a subject provide a subjective evaluation value regarding the feel of the formulation, using evaluation terms that represent sensations obtained by human touch as indicators.

6. The model generation method according to claim 5, wherein in the acquisition step, the tactile information is acquired based on the results of evaluations of the tactile feel of the target formulation by multiple subjects.

7. The model generation method according to claim 1, wherein in the learning step, the model is trained using, as the target variable, a combination of combination information in which a predetermined number of non-zero values ​​exist among the multiple types of combination information included in the learning data.

8. A model generation device comprising an acquisition unit and a learning unit, The acquisition unit acquires tactile information indicating the feel of the formulation and the blending ratio of raw materials in the formulation for each of several formulations. The learning unit is a model generation device that uses learning data based on the information acquired by the acquisition unit, takes tactile information as input, and outputs blending information related to the blending ratio.

9. A model generation program that causes a computer to function as an acquisition unit and a learning unit, The acquisition unit acquires tactile information indicating the feel of the formulation and the blending ratio of raw materials in the formulation for each of several formulations. The learning unit is a model generation program that uses learning data based on the information acquired by the acquisition unit, takes tactile information as input, and outputs blending information related to the blending ratio.

10. An estimation method that makes estimations based on the relationship between the feel of a formulation and the mixing ratio of raw materials, An estimation method for estimating formulation information from the tactile information, wherein tactile information indicating the desired feel for a formulation is input to a model generated by the model generation method described in claim 1, and formulation information relating to the blending ratio of raw materials for a formulation that realizes the said tactile information is obtained as the output of the model.

11. The estimation method according to claim 10, wherein the aforementioned formulation information includes estimating the blending ratio of the raw materials of the formulation and the sum of the blending ratios of raw materials belonging to the same classification.

12. An estimation device comprising a reception unit and an estimation unit, The aforementioned reception unit receives input of tactile information indicating the desired feel for the formulation. The estimation device comprises an estimation unit which inputs tactile information received by the reception unit into a model generated by the model generation method described in claim 1, and obtains formulation information relating to the blending ratio of raw materials for a formulation that realizes the tactile sensation as an output of the model, thereby estimating formulation information from the tactile information.

Images