JP7610108B2 - Physical property prediction device and program - Google Patents

Description

Application of Article 30, paragraph 2 of the Patent Act Published on the website on February 15, 2021, https://www.y-yokohama.com, https://www.y-yokohama.com/release/, https://www.y-yokohama.com/release/?id=3517, https://www.y-yokohama.com/release/pdf/2021021515tr001.pdf

The present invention relates to a property prediction device and program.

The following Patent Document 1 describes a method for predicting the physical properties (safety such as toxicity) of an unknown sample, which is a compound, by searching for samples (multiple known samples) whose similarity to an input unknown sample is equal to or greater than a specified value, creating a subsample set from the searched samples, and performing data analysis on this subsample set. In addition, the following Patent Document 2 describes a method for searching for a base model and a correction model (a model that predicts the inverse of the residual of the base model) for a representative vector close to an unknown input vector, which is data for the unknown sample, calculating a predicted physical property value based on the predicted values of the base model and the correction model, and calculating a risk value for the predicted physical property value based on the predicted value of the correction model.

Patent No. 5083320 JP 2020-187417 A

The accuracy of physical property predictions varies from prediction to prediction. Knowing the degree of accuracy for each prediction can help users decide which prediction to rely on for prototyping.

The objective of this disclosure is to provide a property prediction device and program that allows users to know the degree of certainty of each property prediction.

The physical property prediction device according to the present disclosure includes a learning condition data storage means for storing a plurality of learning condition data each indicating a generation condition of a substance, a parameter storage means for storing parameters of a machine learning model that is learned using the plurality of learning condition data and outputs predicted physical property data indicating the properties of a substance generated under the generation conditions indicated by the input condition data when given input condition data is input, a similarity determination means for determining the similarity between each of the plurality of learning condition data and the input condition data, and an output means for outputting the predicted physical property data and outputting information based on any of the similarities as the accuracy of the predicted physical property data. This makes it possible to know the degree of certainty of a physical property prediction for each prediction.

The program according to the present disclosure is a program for causing a computer to function as: a learning condition data storage means for storing a plurality of learning condition data each indicating a generation condition of a substance; a parameter storage means for storing parameters of a machine learning model that is learned using the plurality of learning condition data and outputs predicted physical property data indicating the properties of a substance generated under the generation conditions indicated by the input condition data when given input condition data is input; a similarity determination means for determining the similarity between each of the plurality of learning condition data and the input condition data; and an output means for outputting the predicted physical property data and outputting the similarity as the accuracy of the predicted physical property data. This allows the degree of certainty of a physical property prediction to be known for each prediction using a computer.

FIG. 1 is a diagram illustrating a configuration of a physical property prediction device according to an embodiment of the present disclosure. FIG. 2 is a functional block diagram showing an example of functions implemented in the property prediction device. 4 is a diagram showing an example of a plurality of learning data stored in a learning data storage unit; FIG. FIG. 13 is a diagram showing an example of a plurality of input data. FIG. 2 is a diagram illustrating an example of machine learning performed by a learning unit. FIG. 13 is a diagram illustrating an example of prediction using a machine learning model performed by a prediction unit. FIG. 13 is a diagram illustrating an example of a prediction result for input data. FIG. 11 is a diagram illustrating an example of a method for determining similarity between learning data and input data. FIG. 13 is a diagram showing another example of a prediction result for input data. FIG. 2 is a diagram showing an example of a flow of a prediction process performed by a property prediction device. 13 is a diagram showing an example of attribute data generated for attributes of condition item data. FIG. 13 is a diagram showing an example of attribute data generated for attributes of condition item data. FIG. 13 is a diagram showing an example of attribute data generated for attributes of condition item data. FIG.

Below, an embodiment of the present disclosure will be described with reference to the drawings. In this embodiment, input data indicating the conditions for generating and inputting a substance (conditions related to raw materials, the generation process, and the measurement of physical properties, etc.) is input into a trained machine learning model, which outputs predicted physical property data indicating the properties of the generated substance, and also outputs information regarding the accuracy of the predicted physical property data. Below, an example will be described in which the substance is a mixture (e.g., tire material) generated by multiple raw materials and a generation process, but the substance is not limited to a mixture of multiple raw materials, and may be anything generated based on specified conditions.

[1. Hardware Configuration]
Fig. 1 is a diagram showing a configuration of a physical property prediction apparatus 10 according to an embodiment of the present disclosure. The physical property prediction apparatus 10 according to the present embodiment is a computer such as a personal computer, a general-purpose computer, or a mobile information terminal, and includes a processor 11, a storage unit 12, a communication unit 13, a display unit 14, and an operation unit 15, as shown in Fig. 1. The physical property prediction apparatus 10 may also include an optical disk drive for reading an optical disk, a USB (Universal Serial Bus) port, and the like.

The processor 11 is a program-controlled device such as a CPU (Central Processing Unit) that operates according to a program installed in the property prediction device 10, which is, for example, a computer. The memory unit 12 is a storage element such as a ROM (Read Only Memory) or a RAM (Random Access Memory), or a hard disk drive. The memory unit 12 stores data such as programs executed by the processor 11. The communication unit 13 is a communication interface such as a network board. The display unit 14 is a display device such as a liquid crystal display, and displays various images according to instructions from the processor 11. The operation unit 15 is a user interface such as a keyboard or mouse, and accepts operation input from the user and outputs a signal indicating the content to the processor 11.

2. Functional Blocks
The physical property prediction device 10 outputs predicted physical property data indicating, for example, predicted performance of the properties of a material that is a mixture (for example, hardness of the material, durability and rolling resistance when used as a tire, etc.), and also outputs information regarding the accuracy of the predicted physical property data. The output of the predicted physical property data and information regarding the accuracy by the physical property prediction device 10 will be described below.

FIG. 2 is a functional block diagram showing an example of functions implemented in the physical property prediction device 10. As shown in FIG. 2, the physical property prediction device 10 functionally includes a learning data storage unit 20 (learning condition data storage unit), an input data acquisition unit 30 (input condition data acquisition unit), a machine learning model 40, a similarity determination unit 50, a predicted physical property data output unit 60, and an item name output unit 70. Note that the physical property prediction device 10 does not need to implement all of the functions shown in FIG. 2, and functions other than the functions shown in FIG. 2 may be implemented.

[2-1. Learning data storage unit]
The learning data storage unit 20 stores a plurality of learning condition data. The "learning condition data" is data indicating conditions for generating and measuring a substance, and is teacher data used for learning a machine learning model 40 described later. Hereinafter, the learning condition data is also simply referred to as "learning data." The learning data storage unit 20 may be realized mainly by the storage unit 12, but is not limited thereto, and may be realized by other storage devices such as an external storage device or a NAS (Network Attached Storage) connected to the physical property prediction device 10 by wire or wirelessly.

Figure 3 is a diagram showing an example of multiple learning data stored in the learning data storage unit 20. In the example shown in Figure 3, data identified by an individual ID (1, 2, N, etc.) (one row of data in a table) corresponds to one piece of learning data.

The learning data includes multiple condition item data as data for multiple items indicating the conditions for generating and measuring a substance. Each condition item data is data for an item related to the generation conditions or measurement conditions of a substance. In the example shown in FIG. 3, the learning data includes, as multiple condition item data, values of multiple raw materials (raw material 1 to raw material R), values of multiple process conditions (process 1 to process condition S), and values of multiple measurement conditions (measurement condition 1 to measurement condition T).

A "raw material" is a raw material contained in a substance that is a mixture, and "raw material data" is data such as values that indicate the degree of a raw material contained in a substance (mass, weight, blending ratio, etc.). For example, learning data identified by ID "1" indicates that raw material 1 (the first raw material) has a value of "44", raw material 2 (the second raw material) has a value of "10", raw material P (the Pth raw material) has a value of "7", raw material P+1 (the P+1th raw material) has a value of "3", raw material Q (the Qth raw material) has a value of "1", raw material Q+1 (the Q+1th raw material) has a value of "2", and raw material R (the Rth raw material) has a value of "3". Additionally, in the learning data identified by ID "2", it is shown that the value of ingredient 1 is "43", the value of ingredient 2 is "10", the value of ingredient P is "6", the value of ingredient P+1 is "3", the value of ingredient Q is "1", the value of ingredient Q+1 is "2", and the value of ingredient R is "0".

"Process conditions" are the conditions for processing (such as machining) when producing a substance from raw materials, and "process condition data" is data that indicates the conditions when processing a substance (for example, the temperature and time when heating or cooling). For example, in learning data identified by an ID of "1", it is shown that the value of process condition 1 (the first process condition) is "148" and the value of process condition S (the Sth process condition) is "189". In addition, in learning data identified by an ID of "2", it is shown that the value of process condition 1 is "147" and the value of process condition S is "178".

"Measurement conditions" are conditions for measuring the physical properties of the generated substance, and "measurement condition data" is data indicating the conditions for measuring the substance (e.g., measurement method and measurement time). For example, in learning data identified by ID "1", it is shown that measurement condition 1 (first measurement condition) has a value of "2" and measurement condition 2 (second measurement condition) has a value of "3". Similarly, in learning data identified by ID "2", it is shown that measurement condition 1 has a value of "2" and measurement condition 2 has a value of "3".

The reason why the condition item data for the learning data includes data on measurement conditions is that the data measured as physical properties (physical property values in Figure 3) differ depending on the measurement conditions for the generated substance.

In this embodiment, one of a plurality of attributes is assigned to each condition item data of the learning data. In the example shown in FIG. 3, five attributes are assigned to the learning data. More specifically, among the plurality of condition item data included in the learning data, "attribute 1" is assigned to raw material 1 to raw material P, "attribute 2" is assigned to raw material 1 and raw material P+1 to raw material Q, "attribute 3" is assigned to raw material Q+1 to raw material R, "attribute 4" is assigned to process condition 1 to process condition S, and "attribute 5" is assigned to measurement condition 1 to measurement condition T. In the example shown in FIG. 3, one of attributes 1 to 3 is assigned according to the type of condition item data of the raw material (raw material 1 to raw material R) (for example, type such as polymer material, filler, or other). In the example shown in FIG. 3, one of attributes 1 to 3 is assigned to the data of the plurality of raw materials, but the attribute assigned to the raw material data may be only attribute 1, or may be other attributes other than attributes 1 to 3.

The learning data also includes physical property data indicating the physical properties (hardness, etc.) of the material generated under the multiple conditions included therein (conditions indicated by multiple condition item data). In the example shown in FIG. 3, the physical property data is represented as a physical property value and is indicated by a numerical value. For example, the learning data identified by ID "1" has a physical property value of "1.2", and the learning data identified by ID "2" has a physical property value of "1.3". Note that the physical property data is not limited to numerical values such as physical property values, and may be anything that can identify the physical property. The physical property data may be indicated by a number sequence, or by symbols or letters (hard, soft, etc.).

As shown in FIG. 3, the condition item data included in each of the multiple learning data are different from each other. Therefore, the physical property values of each learning data are basically different from each other, but due to rounding processing such as rounding off the decimals of the physical property values, the physical property values may match between the multiple learning data.

[2-2. Input data acquisition unit]
The input data acquisition unit 30 acquires input condition data to be input to the machine learning model 40 described later. The "input condition data" is data indicating the conditions for generating and measuring a substance, and is data created by a user to predict the physical properties (hardness, etc.) of the substance generated and measured under the conditions. Hereinafter, the input condition data is also simply referred to as "input data." The input data acquisition unit 30 may be realized mainly by the processor 11 of the physical property prediction device 10, or may be realized by a processor of another information processing device connected to the physical property prediction device 10 via a network.

Figure 4 shows an example of multiple input data created by a user. In the example shown in Figure 4, as in Figure 3, data identified by an individual ID (1, 2, M, etc.) (one row of data in a table) corresponds to one piece of input data.

Like the learning data, the input data also includes multiple condition item data, each of which is data for multiple items indicating the conditions for generating and measuring a substance. In the example shown in FIG. 4, like FIG. 3, the input data includes data on multiple raw materials (raw material 1 to raw material R), data on multiple process conditions (process 1 to process condition S), and data on multiple measurement conditions (measurement condition 1 to measurement condition T), as multiple condition item data. The content of each condition item data is the same as in the learning data, so a description is omitted. The user can create the input data, for example, by inputting values for the condition item data using the operation unit 15 of the property prediction device 10.

In addition, like the condition item data included in the learning data, each condition item data in the input data is assigned one of a number of attributes. The contents of the attributes assigned to each condition item data are the same as those in the learning data, so a description is omitted.

As described above, the input data is intended to predict the physical properties (hardness, etc.) of the material generated and measured using the condition item data contained therein. Therefore, unlike the learning data, the input data does not include physical property data such as physical property values. By creating input data including multiple generation and/or conditions and inputting it to the machine learning model 40 described below, the user can learn information about the physical properties predicted by the machine learning model 40 (more specifically, the prediction unit 43).

[2-3. Machine learning model]
The machine learning model 40 is a model that is trained by a plurality of pieces of training data stored in the training data storage unit 20, and is a model that outputs predicted physical property data (described later) when given input data is input. As shown in FIG. 3, the machine learning model 40 includes a parameter storage unit 41, a training unit 42, and a prediction unit 43. The machine learning model 40 is a model that is trained by machine learning using a given neural network such as a deep neural network (DNN) or a convolutional neural network (CNN). Alternatively, for example, the machine learning model 40 may be a model that is trained by machine learning using a given statistical method, linear regression, or the like.

[2-3-1. Parameter storage unit]
The parameter storage unit 41 stores parameters of the machine learning model 40. More specifically, the parameter storage unit 41 stores parameters of the machine learning model 40 that outputs predicted physical property data (described later) when given input data is input to the machine learning model 40. Note that the parameter storage unit 41 may be realized mainly by the storage unit 12 of the physical property prediction device 10, or may be realized by another storage device or the like.

When the machine learning model 40 is realized by a neural network, the parameter storage unit 41 may store, as parameters of the machine learning model 40, the multiple nodes constituting the neural network, the weighting of each node, the number of layers, the number of nodes used in each layer, and the like. In addition, for example, the parameter storage unit 41 may store, as parameters of the machine learning model 40, a formula for calculating a physical property value from multiple condition item data and the coefficients of the formula.

[2-3-2. Learning section]
The learning unit 42 updates the parameters of the machine learning model 40 stored in the parameter storage unit 41 by performing machine learning using a plurality of pieces of learning data stored in the learning data storage unit 20. That is, the parameters of the machine learning model 40 stored in the parameter storage unit 41 are learned (updated) by the plurality of pieces of learning data. The learning unit 42 may be realized mainly by the processor 11 of the physical property prediction device 10, or may be realized by a processor of another information processing device connected to the physical property prediction device 10 via a network.

Figure 5 is a diagram showing an example of machine learning performed by the learning unit 42. In this embodiment, as shown in Figure 5, the machine learning model 40 is composed of multiple machine learning models that are trained for multiple attributes (for example, attributes 1 to 5 shown in Figure 3). In other words, the parameter storage unit 41 stores parameters of the machine learning model for each of the multiple attributes.

As shown in FIG. 5, the learning unit 42 learns the machine learning model belonging to attribute 1 using condition item data (the data of ingredients 1 to P shown in FIG. 3) to which attribute 1 is assigned among the multiple condition item data included in the learning data. That is, among the multiple parameters stored in the parameter storage unit 41, the parameters of the machine learning model to which attribute 1 is assigned are updated by the condition item data classified as attribute 1 in the learning data. Similarly, the machine learning models belonging to attributes 2 to 5 are also trained by the condition item data to which the corresponding attributes are assigned in the learning data. That is, the parameters of the machine learning models to which attributes 2 to 5 are assigned are also updated by the condition item data classified as the corresponding attributes in the learning data.

[2-3-3. Prediction section]
When given input data (see FIG. 4 ) is input to the machine learning model 40, the prediction unit 43 outputs predicted physical property data for the input data. The "predicted physical property data" is data indicating a prediction of the properties of a substance generated and measured under the conditions indicated by the input data (more specifically, a plurality of condition item data included in the input data). In this embodiment, the "predicted physical property value" shown in FIG. 7 corresponds to the predicted physical property data.

Figure 6 is a diagram showing an example of prediction using a machine learning model performed by the prediction unit 43. As described above, in this embodiment, the machine learning model 40 is composed of multiple machine learning models trained for multiple attributes (e.g., attributes 1 to 5). For this reason, the prediction unit 43 classifies multiple condition item data included in the input data acquired by the input data acquisition unit 30 into multiple attributes (attributes 1 to 5). Then, the prediction unit 43 inputs the condition item data of the input data classified by attribute into the machine learning model of the corresponding attribute to calculate predicted physical property data for the input data.

FIG. 7 is a diagram showing an example of a prediction result for input data. As shown in FIG. 7, the prediction unit 43 calculates predicted physical property values, which are numerical values indicating the properties of a substance, as predicted physical property data for each of a plurality of input data. In the example shown in FIG. 7, it is shown that the predicted physical property value of the input data identified by an ID of "1" is "1.2", and the predicted physical property value of the input data identified by an ID of "2" is "1.2". Note that the predicted physical property data is not limited to numerical values such as predicted physical property values, and may be anything that can identify the predicted physical property. For example, the predicted physical property data may be represented by a number sequence, or may be represented by symbols, letters, etc.

[2-4. Similarity determination unit, predicted property data output unit]
The similarity determination unit 50 determines the similarity (such as similarity or degree of similarity) between the learning data and the input data. The predicted physical property data output unit 60 outputs the predicted physical property data predicted by the prediction unit 43 of the machine learning model 40, and outputs information based on the similarity determined by the similarity determination unit 50 as the accuracy of the predicted physical property data. The predicted physical property data output unit 60 outputs the predicted physical property value and the accuracy of the predicted physical property value to a display device such as the display unit 14, as shown in FIG. 7, for example. The similarity determination unit 50 and the predicted physical property data output unit 60 may be realized mainly by the processor 11 of the physical property prediction device 10, or may be realized by a processor of another information processing device connected to the physical property prediction device 10 via a network.

More specifically, the similarity determination unit 50 determines the similarity between each of the multiple learning data and the input data. For example, the similarity determination unit 50 determines the similarity between each of the multiple learning data identified by IDs "1", "2", ... "N" in FIG. 3 and the input data identified by ID "1" in FIG. 4. The predicted physical property data output unit 60 outputs the predicted physical property data and outputs information based on one of the multiple similarities determined by the similarity determination unit 50 as the accuracy of the predicted physical property data. The predicted physical property data output unit 60 outputs predicted physical property data predicted by the prediction unit 43 of the machine learning model 40 for the input data identified by ID "1" in FIG. 4. In addition, the similarity determination unit 50 outputs information based on one of the multiple similarities determined by the similarity determination unit 50 regarding the similarity between the input data and each of the multiple learning data identified by IDs "1", "2", ... "N" in FIG. 3.

In this embodiment, the similarity determination unit 50 determines whether the condition item data included in the learning data is similar to the condition item data included in the input data and corresponds to the condition item data of the learning data. Then, based on the similarity of the condition item data, the similarity between each of the multiple learning data and the input data is determined.

In the example shown in FIG. 3 and FIG. 4, both the learning data and the input data include values of multiple items such as ingredient 1 and ingredient 2 as multiple condition items. The similarity determination unit 50 determines whether the learning data value and the input data value are similar for each of the multiple items. If the difference between the learning data value and the input data value for a certain item (such as ingredient 1) is within a predetermined numerical range (in other words, if the absolute value of the difference is equal to or less than a threshold value), the similarity determination unit 50 determines that the items are similar, and if the difference is not within the predetermined numerical range (in other words, if the absolute value of the difference exceeds a threshold value), the similarity determination unit 50 determines that the items are not similar. In addition, the similarity determination unit 50 may determine the similarity of each of the multiple items in three stages, match, similar, and mismatch, or may calculate a score according to the similarity as a multiple similarity determination in three or more stages.

The predicted physical property data output unit 60 outputs information based on the similarity determined for one of the multiple learning data most similar to the input data as the accuracy of the predicted physical property data predicted for the input data. The predicted physical property data output unit 60 outputs, for example, information based on the similarity with the learning data that has the most items determined to be similar among the multiple learning data. As another example, when the similarity determination unit 50 calculates a score according to the similarity for multiple items, the predicted physical property data output unit 60 may output information based on the similarity with the learning data with the highest total value of the calculated scores.

When the multiple condition item data included in the input data identified by ID "1" in FIG. 4 is compared with the multiple condition item data included in the learning data identified by ID "1" in FIG. 3, it matches raw material 1, raw material 2, raw material P, raw material P+1, raw material Q, raw material Q+1, raw material R, process condition 1, process condition S, and measurement condition 1 (items other than measurement condition T), and is the most matching (similar) condition item data compared to the other learning data. Therefore, the predicted physical property data output unit 60 outputs information based on the similarity with the learning data identified by ID "1" as the learning amount data most similar to the input data identified by ID "1".

The predicted physical property data output unit 60 may output information based on the number of condition item data determined to be similar by the similarity determination unit 50 as the accuracy of the predicted physical property data. In this case, the predicted physical property data output unit 60 outputs, for example, a "circle" when the number of condition item data determined to be similar is equal to or greater than a predetermined number, and outputs an "x" when the number of condition item data determined to be similar does not meet the predetermined number. Without being limited to this, the predicted physical property data output unit 60 may output the number of condition item data determined to be similar, or may output the total score calculated for each of the multiple condition item data.

Figure 8 is a diagram showing an example of a method for determining the similarity between learning data and input data. As described above, in both the learning data and the input data, one of multiple attributes (attribute 1 to attribute 5) is assigned to the condition item data. In this case, the similarity determination unit 50 may determine, for each of the multiple attributes (attribute 1 to attribute 5), the similarity between the condition item data included in the learning data and the corresponding condition item data included in the input data (condition item data whose item names correspond (match) to the condition item data of the learning data), and determine the similarity between the learning data and the input data for each attribute based on the similarity.

The predicted physical property data output unit 60 may output information based on the number of condition item data determined to be similar by the similarity determination unit 50 for each of multiple attributes (attribute 1 to attribute 5) and the total score. In the example shown in FIG. 7, for input data identified by ID "1", the predicted physical property data output unit 60 outputs information such as "o, o, o, o, x" as the accuracy of the predicted physical property value. This indicates that the number of condition item data determined to be similar for each of attributes 1 to 4 (or the total score) is equal to or greater than the threshold value specified for each corresponding attribute, and the number of condition item data determined to be similar for attribute 5 (or the total score) is less than the threshold value specified for attribute 5. In other words, it indicates that attributes 1 to 4 are all similar, but attribute 5 is not similar. For input data identified by ID "2", the predicted physical property data output unit 60 outputs information such as "o, o, x, x, x" as the accuracy of the predicted physical property value. This indicates that the number of condition item data items (or the total score value) determined to be similar for each of attributes 1 and 2 is equal to or greater than the threshold value for each corresponding attribute, and the number of condition item data items (or the total score value) determined to be similar for each of attributes 3 to 5 is less than the threshold value for each corresponding attribute.

Figure 9 is a diagram showing another example of a prediction result for input data. As shown in Figure 9, the predicted physical property data output unit 60 may output the number of attributes (attribute 1 to attribute 5) for which the number of condition item data determined to be similar is equal to or greater than a threshold value as the accuracy of the predicted physical property value. For example, in input data identified by ID "1", the number of condition item data (or the total score) is equal to or greater than the threshold value for four attributes, attribute 1 to attribute 4, so "4" is output as the accuracy of the input data. In input data identified by ID "2", the number of condition item data is equal to or greater than the threshold value for two attributes, attribute 1 to attribute 2, so "2" is output as the accuracy of the input data.

The likelihood of a physical property prediction based on multiple condition item data included in the input data varies for each input data (i.e., for each prediction). In this regard, the similarity determination unit 50 and the predicted physical property data output unit 60 output the similarity between the learning data and the input data as the accuracy of the predicted physical property value, so that the user can know the likelihood of the predicted physical property value for each prediction.

[2-5. Item name output section]
The item name output unit 70 outputs condition item data included in the learning data and corresponding condition item data included in the input data (condition item data whose item names correspond (match) with the condition item data of the learning data). If it is determined that the condition item data are not similar to each other, the item name output unit 70 outputs the item name of the condition item data. It may also be realized by a processor of another information processing device connected to 10 via a network.

As shown in FIG. 9, the item name output unit 70 outputs the item name of the condition item data that is determined to be not similar as a cause of inhibiting the similarity between the input data and the learning data. For example, the absolute value of the difference between the value of the condition item data with the item name "measurement condition T" between the input data identified by the ID "1" in FIG. 3 and the learning data identified by the ID "1" that is most similar to the input data in FIG. 4 is "3", which is larger than a predetermined threshold value (for example, "1"). Therefore, the item name output unit 70 outputs the item name "measurement condition T" for the input data identified by the ID "1". Also, the difference between the value of the item names "raw material Q", "process condition 1", and "measurement condition T" between the input data identified by the ID "2" and the learning data identified by the ID "1" that is most similar to the input data is not within a predetermined numerical range. Therefore, the item name output unit 70 outputs the item names "raw material Q," "process condition 1," and "measurement condition T" for the input data identified by the ID "2."

In this way, the item name output unit 70 outputs the item names of condition item data that are determined to be dissimilar as the cause of the inhibition of the similarity between the input data and the learning data, so the user can know which items of the input data are dissimilar to the learning data, which is known data. The user can use this information to decide whether to prototype a substance or to create (redesign) the input data.

3. Flowchart
Fig. 10 is a diagram showing an example of the flow of the prediction process performed by the property prediction apparatus 10. First, the input data acquisition unit 30 acquires input data including a plurality of condition item data (step S101). The number of input data acquired in step S101 may be a plurality of data as shown in Fig. 4, or may be one.

Next, the prediction unit 43 calculates predicted physical property values of the input data based on the machine learning model for each attribute (more specifically, parameters of the machine learning model learned for each attribute 1 to 5 based on the learning data stored in the learning data storage unit 20) (step S102). At this time, the prediction unit 43 may classify the multiple condition item data included in the input data acquired in step S101 into multiple attributes (attributes 1 to 5) and input the condition item data classified for each attribute to the machine learning model of the corresponding attribute. This causes the predicted physical property values for the input data to be calculated. If multiple input data are acquired in step S101, the predicted physical property values are calculated for each of the multiple input data in step S102.

Next, the similarity determination unit 50 determines the similarity for each attribute between the input data acquired in step S101 and the learning data stored in the learning data storage unit 20 (step S103). The similarity determination unit 50 determines the similarity between the input data and each of the multiple learning data stored in the learning data storage unit 20. More specifically, it determines the similarity between the condition item data included in the learning data and the condition item data included in the corresponding input data (condition item data whose item names correspond (match) to the condition item data of the learning data), and determines the similarity between the input data and the learning data based on the similarity.

The similarity determination unit 50, for example, determines the similarity of the condition item data contained in the input data and the learning data for each of a number of attributes (attributes 1 to 5), and determines the similarity (such as similarity or degree of similarity) for each attribute between the input data and the learning data based on the similarity.

Next, the predicted physical property data output unit 60 outputs information based on the predicted physical property value calculated in step S102 and the similarity determined in step S103 (step S104), and ends the process. The predicted physical property data output unit 60 outputs information based on the similarity determined for the learning data that is most similar to the input data among the multiple learning data stored in the learning data storage unit 20 as the accuracy of the predicted physical property value. The predicted physical property data output unit 60 outputs, for example, information indicating the similarity determined for each of the multiple attributes (attributes 1 to 5). The predicted physical property data output unit 60 may output, for example, characters including "o" and "x" shown in FIG. 7, or may output the number of attributes for which the number of condition item data determined to be similar is equal to or greater than a threshold value.

The calculation of the predicted physical property value is performed based on a machine learning model that has been machine-learned for each attribute (step S102). Therefore, by determining the similarity for each attribute between the input data and the learning data (step S103) and outputting the determined similarity for each attribute as a degree of accuracy (step S104), it is possible to increase the likelihood of accuracy that indicates the accuracy of the predicted physical property value.

In step S104, when it is determined that the condition item data included in the learning data is not similar to the corresponding condition item data included in the input data (condition item data whose item names correspond (match) to the condition item data of the learning data), the item name output unit 70 may output the item name of the condition item data. For example, as shown in FIG. 9, the item name output unit 70 may output the item name of the condition item data determined to be not similar as a cause of inhibiting the similarity between the input data and the learning data. In this way, the user can know in which items the input data is not similar to the learning data, which is known data.

[4. Summary]
As described above, in this embodiment, the similarity determination unit 50 determines the similarity between each of the multiple learning data and the input data. The predicted physical property data output unit 60 outputs predicted physical property data indicating the properties of a substance generated under the generation conditions indicated by the input data, and outputs information based on any of the similarities determined by the similarity determination unit 50 as the accuracy of the predicted physical property data. In this way, the similarity between the learning data and the input data is output as the accuracy of the predicted physical property value, so that the user can know the accuracy of the predicted physical property value for each input data (i.e., each prediction). This can be useful for the user to determine which prediction to rely on to prototype an actual substance.

5. Modifications
The present invention is not limited to the above embodiments.

In the embodiment, an example has been described in which the similarity determination unit 50 determines, for each attribute, the similarity between the condition item data included in the learning data and the condition item data included in the input data (condition item data whose item names correspond (match) to the condition item data in the learning data), and determines the similarity based on the similarity. This is not limiting, and the similarity determination unit 50 may determine the similarity between the learning data and the input data using various methods.

The similarity determination unit 50 may, for example, generate attribute data for each of a plurality of attributes for the learning data and the input data based on the condition item data to which each attribute is assigned, determine whether the attribute data for the learning data is similar to the attribute data for the input data, and determine the similarity between the learning data and the input data based on the similarity. Attribute data is provided for each attribute (for example, for each of a plurality of attributes, Attribute 1 to Attribute 5), and is data including at least one element. The element included in the attribute data for each attribute may be different from or may be the same as the content of the condition item data for the corresponding attribute.

Figures 11A to 11C are diagrams showing examples of attribute data generated for each of attributes 1 to 3 (attributes assigned to ingredient data). The example shown in Figure 11A shows attribute data generated for attribute 1, and includes multiple elements. The example shown in Figure 11A includes "element 1," which is the same element as "ingredient 1," "element 2," which is the same element as "ingredient 2," and "element X," which is the same element as "ingredient P." In other words, the multiple elements included in the attribute data for attribute 1 match the multiple condition item data to which attribute 1 is assigned, as shown in Figures 3 and 4.

The example shown in FIG. 11B shows attribute data generated for attribute 2, and includes "element X+1" which is the same element as "ingredient 1", "element X+2" which is the same element as "ingredient P+1", and "element Y" which is the same element as "ingredient Q". Here, "ingredient 1" is included not only in the elements of attribute 2, but also in the elements of attribute 1. In this way, the similarity determination unit 50 may generate two pieces of attribute data that include specific condition item data.

The example shown in FIG. 11C shows attribute data generated for attribute 3, and includes "element Y+1" which is an element combining "ingredient 1" and "ingredient 2", "element Y+2" which is the same element as "ingredient Q+1", and "element Z" which is the same element as "ingredient R". As shown in FIG. 4, in the input data identified by ID "1", the value of ingredient 1 is "44" and the value of ingredient 2 is "10", so the similarity determination unit 50 sets "element Y+1", which is the element combining "ingredient 1" and "ingredient 2" related to the input data, to "54", which is the sum of "44" and "10". Similarly, the similarity determination unit 50 also sets "54" to "element Y+1" related to the learning data.

In this way, the similarity determination unit 50 may generate attribute data related to the learning data and attribute data related to the input data for attributes 1 to 3. Then, it may determine whether the attribute data related to the learning data and the attribute data related to the input data are similar to each other (more specifically, whether elements with the same item name are similar to each other), and determine the similarity between the learning data and the input data based on the similarity. Note that the similarity determination unit 50 may generate attribute data for the other attributes, attribute 4 (an attribute assigned to data on process conditions) and attribute 5 (an attribute assigned to measurement conditions), in the same way as for attributes 1 to 3, and determine whether they are similar to each other, or may determine whether the condition item data to which attributes 4 and 5 are assigned are similar to each other as described in the embodiment.

The similarity determination unit 50 determines whether or not the multiple condition item data related to each of the multiple learning data stored in the learning data storage unit 20 is similar to the multiple condition item data related to the input data, and therefore the processing time becomes long when the number of multiple condition item data included in each of the learning data and the input data is huge. In this regard, the similarity determination unit 50 can generate attribute data including elements that combine multiple condition item data, as shown in FIG. 11C, for example, so that the number of elements included in the attribute data can be made smaller than the number of multiple condition item data included in the learning data and the input data. Then, by determining whether or not the attribute data related to each of the learning data and the input data is similar, the processing time can be made shorter than if the multiple condition item data included in the learning data and the input data were directly compared.

The predicted physical property data output unit 60 outputs the similarity between the learning data and the input data determined in this manner as the accuracy of the predicted physical property value. In this way, the user can know the accuracy of the predicted physical property value for each input data.

REFERENCE SIGNS LIST 10 Physical property prediction device, 11 Processor, 12 Memory unit, 13 Communication unit, 14 Display unit, 15 Operation unit, 20 Learning data memory unit, 30 Input data acquisition unit, 40 Machine learning model, 41 Parameter memory unit, 42 Learning unit, 43 Prediction unit, 50 Similarity determination unit, 60 Predicted physical property data output unit, 70 Item name output unit.

Claims (8)

A learning condition data storage means for storing a plurality of learning condition data each indicating a generating condition of a substance;
a parameter storage means for storing parameters of one machine learning model that is learned using the plurality of learning condition data and outputs predicted physical property data indicating the properties of a substance generated under the generation conditions indicated by the input condition data when given input condition data is input;
a similarity determination means for determining a similarity between each of the plurality of learning condition data used in learning the one machine learning model and the input condition data;
an output means for outputting the predicted physical property data and outputting information based on any one of the similarities as the accuracy of the predicted physical property data;
A physical property prediction device comprising: 2. The physical property prediction apparatus according to claim 1,
the output means outputs, as the degree of certainty, information based on the similarity determined for one of the plurality of learning condition data that is most similar to the input condition data.
A physical property prediction device characterized by: 3. The physical property prediction apparatus according to claim 1,
the learning condition data and the input condition data each include a plurality of condition item data,
the similarity determination means determines whether the condition item data included in the learning condition data is similar to the corresponding condition item data included in the input condition data, and determines the similarity based on the similarity.
A physical property prediction device characterized by: 4. The physical property prediction apparatus according to claim 3,
and an item name output means for outputting an item name of the condition item data when it is determined that the condition item data included in the learning condition data and the corresponding condition item data included in the input condition data are not similar to each other.
A physical property prediction device characterized by: 5. The physical property prediction apparatus according to claim 3,
Each of the condition item data is assigned one of a plurality of attributes;
the similarity determination means determines, for each attribute, whether the condition item data included in the learning condition data is similar to the corresponding condition item data included in the input condition data, and determines the similarity based on the similarity;
A physical property prediction device characterized by: 3. The physical property prediction apparatus according to claim 1,
the learning condition data and the input condition data each include a plurality of condition item data,
Each of the condition item data is assigned one of a plurality of attributes;
the similarity determination means generates attribute data for each attribute of the learning condition data and the input condition data based on the condition item data to which the attribute is assigned, determines whether the attribute data related to the learning condition data and the attribute data related to the input condition data are similar to each other, and determines the similarity based on the similarity.
A physical property prediction device characterized by: 7. The physical property prediction apparatus according to claim 1,
the learning condition data and the input condition data are data indicating measurement conditions for a substance,
The predicted physical property data is data indicating properties of a substance to be measured under measurement conditions indicated by the input condition data.
A physical property prediction device characterized by: A learning condition data storage means for storing a plurality of learning condition data each indicating a generating condition of a substance;
a parameter storage means for storing parameters of one machine learning model which is learned using the plurality of learning condition data and outputs predicted physical property data indicating the properties of a substance generated under the generation conditions indicated by the input condition data when given input condition data is input;
a similarity determination means for determining a similarity between each of the plurality of learning condition data used in learning the one machine learning model and the input condition data; and an output means for outputting the predicted physical property data and outputting the similarity as a degree of accuracy of the predicted physical property data.
A program that makes a computer function as a

Images