Plastic compounding properties prediction system and plastic compounding properties prediction method

Abstract

The challenge lies in achieving the appropriate plastic formulation. [Solution] The system includes a transformation unit 102 that calculates a mapping to a latent space having fewer dimensions than the explanatory variables stored in the material DB 110 by manifold learning based on the plastic material DB 110; a learning unit 103 that generates a learning model by performing regression learning of material properties from the data stored in the material DB 110; a similar record ranking calculation unit 203 that maps the explanatory variables to the latent space using the mapping and calculates the similarity between the result of the mapping and the record data stored in the material DB 110; and an AWE prediction model generation unit 205 that generates an ensemble model as a plastic formulation property prediction model from the result of inputting the values ​​of the required explanatory variables, which are explanatory variables required as the formulation properties of the plastic, into the learning model using weighting based on the similarity.

Description

Technical Field

[0005]

[0001] The present invention relates to the technology of a plastic formulation property prediction system and a plastic formulation property prediction method.

Background Art

[0002] The market for plastic material recycling has been booming. In such a market, there is high expectation for data-driven MI (Materials Informatics) technologies such as property prediction according to an input recipe and recipe prediction for achieving target performance. Along with such growing expectation, it has become an urgent task to establish MI technologies that can accurately predict various recycling sources. In data preparation for MI, since the number of experimental man-hours is limited, it is necessary to incorporate formulation data based on publicly available literature to ensure diversity. However, the data sets obtained from the literature contain many unmeasured values and are sparse.

[0003] Techniques for mapping to a latent space are known to eliminate such sparsity. Also, in model creation for MI, it is effective to use specialized models according to recycling sources to handle diversity. As such a specialized model, a method of appropriately adjusting the weights of the outputs of each machine learning model according to input data by the Adaptive weight ensemble (AWE) technique is known.

[0004] However, a specific method for constructing an MI model that can accurately predict with various inputs from a sparse data set is unknown and remains an issue.

[0005] ]Patent Document 1 discloses a latent space mapping technique for complementing sparseness. Patent Document 1 discloses a system and method for training a neural network having an autoencoder architecture to recover missing data (see abstract). "An autoencoder includes an encoder for encoding its input into a latent space and a decoder for decoding the encoding from the latent space. The method includes creating a first training set containing a multidimensional valid dataset, and training the encoder and decoder using the first training set in a first training stage to reduce the difference between the valid dataset provided to the encoder and the dataset decoded by the decoder. The method further includes creating a second training set containing an invalid dataset, and training the encoder using the second training set in a second training stage to reduce the difference between the encoding of valid data instances and the encoding of their corresponding invalid data instances."

[0006] Patent Document 2 discloses AWE technology. Patent Document 2 describes a prediction device, data processing device, prediction method, data processing method, computer program, and recording medium (see abstract). It states that "Z prediction models are each associated with Z sets of different explanatory variables. Each of the Z sets of explanatory variables is a combination of p different explanatory variables out of N explanatory variables. The N explanatory variables represent molecular structure. Each of the Z prediction models calculates the value of the objective variable representing the properties of the target molecular compound by substituting the values ​​corresponding to the molecular structure of the target molecular compound into the set of explanatory variables associated with the prediction model. The model selection unit 222 selects A prediction models from the Z prediction models according to the values ​​of the set of explanatory variables corresponding to the molecular structure of the target molecular compound. The prediction execution unit 223 uses the A prediction models to predict the properties of the target molecular compound."

[0007] Furthermore, Non-Patent Document 1 discloses a method for searching for similar data within verification data, calculating the prediction error, and adjusting the weights based on the set similarity criteria. [Prior art documents] [Patent Documents]

[0008] [Patent Document 1] Japanese Patent Publication No. 2022-140294 [Patent Document 2] Japanese Patent Publication No. 2024-25242 [Non-patent literature]

[0009] [Non-Patent Document 1] H. Guo et al., Processes, 2024, Vol.12, p.1854 [Overview of the Initiative] [Problems that the invention aims to solve]

[0010] Patent documents 1, 2, and 1 do not contain any explanatory variables that are effective in predicting plastic compounding properties.

[0011] In light of this background, the present invention was made, and its objective is to achieve an appropriate plastic compound. [Means for solving the problem]

[0012] To solve the aforementioned problems, the present invention comprises: a mapping calculation unit that calculates a mapping to a latent space having fewer dimensions than the explanatory variables stored in a material database of plastics by manifold learning based on the material database; a learning unit that generates a learning model by performing learning to regress material properties from data stored in the material database; a similarity calculation unit that maps the explanatory variables to the latent space using the mapping and calculates the similarity between the result of the mapping and the record data stored in the material database; and an ensemble model generation unit that generates an ensemble model as a model for predicting the composition properties of the plastic from the result of inputting the values ​​of required explanatory variables, which are explanatory variables required as composition properties of the plastic, into the learning model using weighting based on the similarity. Other solutions will be described as appropriate in the embodiments. [Effects of the Invention]

[0013] According to the present invention, it is possible to achieve an appropriate plastic compound. [Brief explanation of the drawing]

[0014] [Figure 1] This figure shows an example configuration of the plastic compounding properties prediction system according to this embodiment. [Figure 2] This diagram shows the hardware configuration of a plastic compounding properties prediction system. [Figure 3] This figure shows a schematic example of a materials database. [Figure 4] This is a diagram (part 1) showing the effective characteristics database. [Figure 5] This is a diagram (part 2) showing the effective characteristics database. [Figure 6] This is a diagram (part 3) showing the effective characteristics database. [Figure 7] This is a diagram (part 4) showing the effective characteristics database. [Figure 8] This figure shows an example of a latent space database. [Figure 9] This figure shows an example of a data lake in the conversion section. [Figure 10] It is a diagram showing an example of a data lake for evaluating the performance of a learning model. [Figure 11] It is a flowchart showing the procedure of processing by the model construction unit. [Figure 12] It is a flowchart showing the procedure of processing by the blending property prediction unit. [Figure 13] It is a diagram showing an example of a screen for setting prediction target variables. [Figure 14] It is a diagram showing an example of a screen for displaying a feature vector. [Figure 15] It is a diagram showing an example of a screen for displaying predicted blending properties. [Figure 16] It is a flowchart showing the procedure of processing by the blending recipe generation unit. [Figure 17] It is an example showing an example of a target property / condition input screen. [Figure 18] It is a diagram showing an example of an alert display screen. [Figure 19] It is a diagram (part 1) showing an example of a screen for displaying a predicted blending recipe. [Figure 20] It is a diagram (part 2) showing an example of a screen for displaying a predicted blending recipe.

Embodiments of the Invention

[0015] Next, embodiments for carrying out the present invention (referred to as "embodiments") will be described in detail with appropriate reference to the drawings.

[0016] [System Configuration Diagram] FIG. 1 is a diagram showing a configuration example of a plastic blending property prediction system Z according to the present embodiment.

[0017] The plastic blending property prediction system Z includes a model construction unit 1, a blending property prediction unit 2, and a blending recipe generation unit 3.

[0018] (Model Construction Unit 1) The model construction unit 1 calculates a mapping that maps explanatory variables to a latent space, and performs learning using the data in the material DB 110.

[0019] Model building unit 1 includes a material DB 110, an effective properties DB 120, an effective properties acquisition unit 101, a transformation unit 102, a latent space DB 130, and a data lake 140 for the transformation unit 102. Model building unit 1 also includes a learning unit 103, a feature vector calculation unit 104, and a data lake 150 for evaluating the performance of the learned model.

[0020] The material database DB110, which is a database of plastic materials, stores compounding experiment data 111 and catalog data 112 (see Figure 3), which are data collected from literature, catalogs, or experiments.

[0021] The effective properties DB120 stores information about the dependent variable and the corresponding explanatory variables used when generating an ensemble model to predict the values ​​of explanatory variables related to plastic formulations. A property is the dependent variable and the value of the explanatory variable corresponding to the dependent variable.

[0022] The effective property acquisition unit 101 acquires the effective properties, which are the objective and explanatory variables necessary for mapping from the material DB 110 and effective property DB 120 to the latent space.

[0023] The transformation unit 102, which is the mapping calculation unit, calculates a mapping to a latent space having fewer dimensions than the explanatory variables stored in the material DB 110, based on the material DB 110, using manifold learning. The transformation unit 102 stores the calculated mapping in its data lake 140. Furthermore, the transformation unit 102 stores the result of substituting the record data 113 of the material DB 110 into the mapping in the latent space DB 130. The record data 113 of the material DB 110 will be described later, but it is data collected for a particular material.

[0024] The learning unit 103 generates a learning model by performing learning that regresses material properties from the data stored in the material DB 110. The learning unit 103 reflects the contribution rate calculated during learning in the effective properties DB 120.

[0025] The feature vector calculation unit 104 calculates a feature vector by substituting the verification data used during training by the learning unit 103 into the mapping stored in the data lake 140 of the transformation unit 102.

[0026] The results of the processing performed by the learning unit 103 and the feature vector calculation unit 104 are then stored in the data lake 150 for evaluating the performance of the learning model.

[0027] (Formulation characteristic prediction unit 2) The blending characteristic prediction unit 2 generates a blending characteristic prediction model, which is an ensemble model, using the input explanatory variables to be predicted and feature vectors. The mark for output device 405 shown in the blending characteristic prediction unit 2 and the blending recipe generation unit 3 indicates that the processing unit to which this mark is attached will output (display) from output device 405. The blending characteristic prediction model, which is an ensemble model, is output by the AWE prediction model generation unit 205, which is an ensemble model generation unit, and the output from the AWE prediction model generation unit 205 is the final output of the blending characteristic prediction unit 2.

[0028] The blending characteristic prediction unit 2 includes a predictable explanatory variable input unit 201, a feature vector acquisition unit 202, a similar record ranking calculation unit 203, a partial output calculation unit 204, and an AWE prediction model generation unit 205. The blending characteristic prediction unit 2 also includes a prediction accuracy / effectiveness calculation unit 206 and a predicted blending characteristic output unit 207.

[0029] In the predictable explanatory variable input unit 201, the values ​​of the explanatory variables required for the plastic compounding characteristics (required explanatory variables) (required explanatory variable data as appropriate) are input. Then, the feature vector acquisition unit 202 acquires feature vectors from the data lake 140 of the transformation unit 102.

[0030] The similarity calculation unit, the similarity record ranking calculation unit 203, then obtains a mapping from the data lake 150 used for evaluating the performance of the learning model. The similarity record ranking calculation unit 203 then uses the obtained mapping to map the explanatory variables input in the prediction target explanatory variable input unit 201 to the latent space. Subsequently, the similarity record ranking calculation unit 203 calculates the similarity between the result of this mapping and the record data 113 (see Figure 3) stored in the material DB 110. The similarity record ranking calculation unit 203 also ranks the record data 113 based on the similarity.

[0031] The partial output calculation unit 204, which is an ensemble model generation unit, substitutes the required explanatory variable data input to the predictable explanatory variable input unit 201 into the learning model generated by the learning unit 103. Furthermore, the partial output calculation unit 204 calculates partial outputs weighted by the error rate of the learning model. The error rate is defined as (100 - accuracy rate).

[0032] The AWE prediction model generation unit 205, which is the ensemble model generation unit, generates an ensemble model from the results of inputting the values ​​of the required explanatory variables (required explanatory variable data), which are explanatory variables required as plastic compounding characteristics, into the learning model using weighting based on similarity. The generated ensemble model becomes the plastic compounding characteristics prediction model. In this embodiment, the ensemble model becomes the AWE prediction model.

[0033] (Formulation recipe generation unit 3) The formulation recipe generation unit 3, which is the recipe generation unit, generates a plastic formulation recipe using the ensemble model generated by the formulation characteristic prediction unit 2.

[0034] The blending recipe generation unit 3 includes a target characteristic / condition input unit 301, a constraint condition setting unit 302, a first appropriate value search unit 303, a predictable characteristic determination unit 304, a second appropriate value search unit 305, and a Pareto solution candidate selection unit 306. Furthermore, the blending recipe generation unit 3 includes an error / prediction accuracy calculation unit 307, a Pareto solution determination unit 308, and a predicted blending recipe output unit 309.

[0035] In the target characteristic / condition input unit 301, the user inputs the target characteristics of the compounded plastic and the conditions to be imposed on the explanatory variables.

[0036] Then, the constraint condition setting unit 302 sets the constraint conditions for the explanatory variables based on the conditions entered in the target characteristic / condition input unit 301.

[0037] Furthermore, the feature vector acquisition unit 202 to the AWE prediction model generation unit 205 of the blending characteristic prediction unit 2 generate an ensemble model (AWE prediction model in this embodiment) based on the target characteristics input in the target characteristic / condition input unit 301 (dashed arrow in Figure 1).

[0038] The first appropriate value search unit 303, which is the update processing unit, uses the record data 113 of the highest predetermined rank in the ranking as the initial value and updates the explanatory variables a predetermined number of times (so that the distance becomes smaller than a predetermined threshold described later) so that the constraint conditions for the explanatory variables in the blending recipe are satisfied and the distance between the target characteristic and the ensemble model value becomes smaller. The ensemble model value is the value of the explanatory variables input into the ensemble model. Generally, the predetermined number of times is input by the user. The predetermined threshold is also input by the user.

[0039] The update processing unit, the predictable characteristic determination unit 304, determines whether, among a predetermined number of updates, there are any pairs of explanatory variables in which the distance between the target characteristic and the ensemble model value does not reach a predetermined threshold, such as when a local minimum is reached.

[0040] The second appropriate value search unit 305, which is the update processing unit, continues updating after removing the target characteristic corresponding to the set of explanatory variables if, during a predetermined number of updates, there is a set of explanatory variables whose distance from the target characteristic to the ensemble model value does not reach a predetermined threshold.

[0041] The Pareto solution processing unit, the Pareto solution candidate selection unit 306, calculates (selects) multiple Pareto solution candidates based on the update history.

[0042] The error / prediction accuracy calculation unit 307 calculates the error and prediction accuracy of the Pareto solution candidate.

[0043] The Pareto solution processing unit, Pareto solution determination unit 308, sorts the Pareto solutions based on the error and prediction accuracy of each candidate Pareto solution. Furthermore, the Pareto solution determination unit 308 outputs the sorting result as a plastic formulation recipe to the output device 405.

[0044] [Hardware Configuration Diagram] Figure 2 shows the hardware configuration of the plastic compounding properties prediction system Z.

[0045] The plastic compounding properties prediction system Z comprises a memory 401, a computing unit 402, a storage device 403, an input device 404, an output device 405, and a communication device 406.

[0046] The memory 401 is composed of volatile storage devices such as RAM (Random Access Memory). The arithmetic unit 402 is composed of a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), and the storage device 403 is composed of an HDD (Hard Disk Drive) and an SSD (Solid State Drive), etc.

[0047] The input device 404 consists of a keyboard, mouse, etc. The output device 405, which is the output section, consists of a sprayer, etc.

[0048] The material DB110, effective properties DB120, latent space DB130, data lake 140 for the conversion unit 102, and data lake 150 for learning model performance evaluation shown in Figure 1 are constructed on the storage device 403.

[0049] Furthermore, the program stored in the storage device 403 is loaded into the memory 401, and the loaded program is executed by the arithmetic unit 402. This brings into practice the effective characteristic acquisition unit 101 to the feature vector calculation unit 104, the prediction target explanatory variable input unit 201 to the predicted blending characteristic output unit 208, and the target characteristic / condition input unit 301 to the predicted blending recipe output unit 909 shown in Figure 1.

[0050] In Figure 1, the input unit 201 for predicting explanatory variables and the input unit 301 for target characteristics and conditions receive information input via the input device 404.

[0051] (Material DB110) Figure 3 shows a schematic example of material DB110.

[0052] The materials database 110 includes formulation experiment data 111, which is created by collecting information from literature such as papers or by conducting experiments oneself, and catalog data 112, which is collected from manufacturers' catalogs.

[0053] In the example shown in Figure 3, the records (rows) represent data for a particular material. The data in each record is appropriately referred to as record data 113. In the example shown in Figure 3, the columns represent the material's performance. In the example shown in Figure 3, the material's performance includes base material performance, formulation, fillers, additives, and compounding performance, but other performance characteristics may also be used.

[0054] Furthermore, in the example shown in Figure 3, for example, the base material performance is composed of multiple columns, indicating that the base material performance is further composed of multiple properties (base material melting point, base material tensile stress, etc.). The same applies to the blending, filler, additive, and compound performance. In the case of dry blending, several additional columns showing the base material performance are added.

[0055] Furthermore, even a single cell in the example shown in Figure 3 may consist of multiple items.

[0056] Furthermore, if the same material is described in different documents or catalog data 112, it is possible that the same material may be stored in different record data 113 in the material DB 110.

[0057] In the example shown in Figure 3, the shaded areas (reference numeral 114) indicate that detailed data is stored, while the white areas (reference numeral 115) indicate that the data is completely missing. The dotted areas (reference numeral 116) indicate that some data is missing. Partial data omission means that, in Figure 3, when a single cell contains multiple items, data for a particular item is not stored.

[0058] Thus, the material database DB110 is sparse data with missing data points.

[0059] (Effective characteristics DB120) Figures 4 to 7 show the effective characteristics DB120.

[0060] In Figures 4 to 7, the range indicated by symbol 121 represents the characteristic used as the dependent variable. The range indicated by symbol 122 represents the range of the independent variable.

[0061] In particular, the inventors found that the explanatory variables shown in white within the range indicated by symbol 122 are effective as properties to be mapped to the latent space.

[0062] In other words, as shown in Figure 4, when the objective variable of the learning model and the ensemble model is the crystallinity of the plastic after compounding, at least one of the following is included as an explanatory variable: base material melting point, base material tensile stress, base material heat of fusion, base material thermal decomposition temperature, base material tensile elongation at fracture, base material tensile modulus, and base material MFR.

[0063] Furthermore, as shown in Figure 4, when the dependent variable is the tensile modulus of elasticity after compounding the plastic, at least one of the following is included as an explanatory variable: filler density, matrix MFR, matrix tensile yield stress, and matrix tensile stress.

[0064] As shown in Figure 4, when the dependent variable is the tensile stress after compounding the plastic, at least one of the following is included as an explanatory variable: the tensile modulus of the base material, the base material MFR, the base material elongation at break, the base material melting point, and the base material density.

[0065] Furthermore, as shown in Figure 4, when the predictive model objective variable is the tensile yield stress after compounding the plastic, at least one of the following is included as an explanatory variable: base material elongation at break, base material tensile modulus, filler density, and base material density.

[0066] Furthermore, as shown in Figure 5, when the dependent variable is the tensile elongation at fracture after compounding the plastic, at least one of the following is included as an explanatory variable: the tensile yield stress of the base material, the tensile stress of the base material, the MFR of the base material, the filler density, the tensile modulus of the base material, and the crystallinity of the base material.

[0067] As shown in Figure 5, when the dependent variable is the flexural modulus of the plastic after compounding, at least one of the following is included as an explanatory variable: the tensile modulus of the base material, the filler density, the base material density, and the tensile stress of the base material.

[0068] Furthermore, as shown in Figure 5, when the objective variable is the bending stress after compounding, at least one of the following is included as an explanatory variable: base material Rockwell hardness, presence or absence of filler surface treatment, base material density, base material flexural modulus, and base material MFR.

[0069] As shown in Figure 5, when the dependent variable is the Izod impact value after the plastic has been compounded, at least one of the following is included as an explanatory variable: the tensile modulus of the base material, the base material MFR, the tensile elongation at fracture of the base material, the tensile yield stress of the base material, and the melting point of the base material.

[0070] Furthermore, as shown in Figure 6, when the dependent variable is MFR after plastic compounding, at least one of the following is included as an explanatory variable: base material tensile elongation at fracture, base material tensile yield stress, filler size, filler density, base material tensile stress, and base material density.

[0071] As shown in Figure 6, when the dependent variable is the melting point of the plastic after compounding, at least one of the following is included as an explanatory variable: base material tensile elongation at break, base material MFR, base material tensile modulus, filler size, base material glass transition temperature, base material crystallization temperature, filler weight ratio, base material weight ratio, and base material volume ratio.

[0072] Furthermore, as shown in Figure 6, when the dependent variable is the heat of fusion after compounding the plastic, at least one of the following is included as an explanatory variable: the melting point of the base material, the tensile elongation at fracture of the base material, the MFR of the base material, the tensile modulus of the base material, the crystallinity of the base material, and the tensile stress of the base material.

[0073] As shown in Figure 6, when the dependent variable is the crystallization temperature after compounding the plastic, at least one of the following is included as an explanatory variable: matrix crystallization heat, phase solvent density, matrix fusion heat, filler size, and filler density.

[0074] Furthermore, as shown in Figure 7, when the dependent variable is the heat of crystallization after compounding the plastic, at least one of the following is included as an explanatory variable: the melting point of the base material, the filler density, the tensile stress of the base material, the crystallinity of the base material, the filler size, and the tensile elongation at fracture of the base material.

[0075] As shown in Figure 7, when the objective variable is the glass transition temperature after compounding the plastic, at least one of the following is included as an explanatory variable: tensile stress of the base material, tensile modulus of the base material, MFR of the base material, density of the base material, and tensile elongation at fracture of the base material.

[0076] Furthermore, as shown in Figure 7, when the dependent variable is the thermal decomposition temperature after compounding the plastic, at least one of the following must be included as an explanatory variable: tensile stress of the base material, tensile elongation at fracture of the base material, tensile modulus of elasticity of the base material, MFR of the base material, filler size, base material density, base material weight ratio, tensile yield stress of the base material, and base material volume ratio.

[0077] In Figures 4 to 7, the numbers assigned to each explanatory variable represent their ranking based on their contribution rate. A smaller number indicates a higher contribution rate. In this way, explanatory variables in the effective characteristics DB120 should be ranked by their contribution rate to the dependent variable. The contribution rate is calculated using SHAP values, etc.

[0078] (Latent space DB130) Figure 8 shows an example of the latent space DB130.

[0079] The latent space DB130 is a dictionary-type data structure where "data" is the key (code 131), and each key corresponds to a corresponding value (code 132). The keys store the same data as the keys (code 131) in Figure 10. The values ​​corresponding to the keys also store feature vectors. "Data" corresponds to record data 113.

[0080] (Data lake 140 of conversion unit 102) Figure 9 shows an example of the data lake 140 of the conversion unit 102.

[0081] The data lake 140 of the transformation unit 102 is a dictionary-type data set where the target variable ("crystallinity", "tensile modulus", etc.) is used as a key (code 141), and each key corresponds to a corresponding value (code 142). M mappings to latent spaces ("L1, L2, ..., LN") are stored corresponding to each of the target variables that serve as keys.

[0082] (Data lake 150 for evaluating the performance of learning models) Figure 10 shows an example of a data lake 150 used for evaluating the performance of a learning model.

[0083] In the example shown in Figure 10, the data lake 150 for evaluating the performance of the learning model has a dictionary-type data structure in which "data" consists of a key (code 151) and a value corresponding to the key (code 152). The "data" which is the key (code 151) corresponds to the verification data described later among the record data 113 in the material DB 110 in Figure 3.

[0084] Furthermore, the data lake 150 used for evaluating the performance of the learning model does not necessarily have to be structured as dictionary-type data.

[0085] In the data lake 150 used for evaluating the performance of the learning model, the feature vector (code 153) and the external model (code 154), which is the prediction model, are stored as values ​​(code 152) corresponding to the key (code 151). Furthermore, the error rate (code 155), prediction accuracy (code 156), and effectiveness (code 157) are stored as values ​​(code 152). The feature vector is the same as the feature vector (code 132) in Figure 8.

[0086] The external model (indicated by 154) is a predictive model used to regress the material properties, which are the target variable, from the material DB110 data. The external model (indicated by 154) will be described later, but it is a predictive model that is constructed using validation data. The material DB110 data is divided into training data and validation data for training.

[0087] Furthermore, the effectiveness (symbol 157) is the degree to which the explanatory variables contributed to the prediction results in the prediction model (external model (symbol 154)), i.e., the contribution rate. Specifically, the effectiveness (symbol 157) is calculated using SHAP values, etc.

[0088] [Method for predicting plastic compounding properties] Next, with reference to Figures 11 to 20, the details of the plastic compounding properties prediction method performed by the plastic compounding properties prediction system Z will be explained. Refer to Figures 1 to 10 as appropriate.

[0089] <Processing by Model Construction Unit 1> Figure 11 is a flowchart showing the processing procedure performed by the model building unit 1.

[0090] First, a materials database (DB110) is created (S101). The materials database (DB110) can be created by manual input or by a generation AI. As shown in Figure 3, the materials database (DB110) stores compounding experiment data (111) from papers and experiments, and record data (113) on material properties collected from catalogs. The properties used include the properties of the material before compounding, compounding recipes created in the past, and the properties of the material after compounding.

[0091] The learning unit 103 performs learning using the data stored in the material DB 110 (S111). Decision trees such as gradient boosting decision trees are preferably used for learning. However, methods used for regression, such as neural networks or support vector machines, may also be used, not limited to decision trees. In step S111, learning is performed to infer materials from the data in the material DB 110.

[0092] In step S111, the learning unit 103 divides each of the record data 113 (Figure 3) stored in the material DB 110 into training data and validation data. The learning unit 103 then performs training and cross-validation using the training data to generate a predictive model. The predictive model generated using the training data is called the internal model.

[0093] The learning unit 103 then calculates the error rate, prediction accuracy, effectiveness, etc., by applying the validation data to the prediction model generated using the training data. Prediction accuracy is the probability (accuracy rate) that the result predicted by the prediction model is correct. In this case, correct means that the predicted value falls within a certain range for the value of the target variable. The accuracy rate is, for example, 30%. Effectiveness is the degree to which the explanatory variables contributed to the prediction (contribution rate).

[0094] Furthermore, a predictive model is generated by fixing the hyperparameters of the internal model and then performing further training using validation data. The predictive model generated using validation data is called the external model.

[0095] The learning unit 103 performs the above process N times, changing the validation data each time. In other words, the learning unit 103 performs cross-validation N times. The method of change can be random, but it is preferable to change it by annotations such as filler type or additive type in the data record. Then, Na (≦N) is the number of times that a particular record data 113 from the material DB 110 is selected as the validation data for cross-validation during the N cross-validations. Na will be a different value for each record data 113 in the material DB 110. Incidentally, the validation data consists of multiple record data 113.

[0096] Step S111 is a "learning step".

[0097] Next, the learning unit 103 calculates the error rate, prediction accuracy, and effectiveness for each verification data in the material DB 110 (S112).

[0098] Furthermore, the order of the explanatory variables shown in Figures 4 to 6 will be changed depending on the effectiveness calculated in step S112.

[0099] Meanwhile, the effective properties acquisition unit 101 acquires data on effective properties from the effective properties DB 120 and the material DB 110 (S121). In Figures 4 to 6, the explanatory variables that are not hatched are the effective properties.

[0100] Subsequently, the conversion unit 102 calculates M (M is a natural number) mappings "L" for dimensionality reduction using appropriate explanatory variables for each target variable (S122). In step S122, the conversion unit 102 calculates the mapping "L" by using autoencoders or manifold learning such as UMAP. The mapping "L" is a mapping for mapping the explanatory variable data corresponding to the effective characteristic DB120 into the latent space. Also, "M" is a value determined by the calculation method of the mapping "L". At the same time, the conversion unit 102 stores the calculated mapping "L" in the latent space DB130. Note that if "P" is the number of explanatory variables, then "M < P". Step S122 is a mapping calculation step.

[0101] Then, the conversion unit 102 creates a data lake 140 of the conversion unit 102 (S123). At this time, the conversion unit 102 stores M mappings "L" for each target variable in the data lake 140 of the conversion unit 102. As shown in FIG. 9, M mappings "L" for each target variable are stored in the data lake 140 of the conversion unit 102.

[0102] Next, the feature vector calculation unit 104 calculates a feature vector for each verification data based on the "M (M is a natural number)" mappings stored in the data lake 140 of the conversion unit 102 (S131). The feature vector is a vector representing the position in the latent space of the verification data mapped into the latent space by substituting the verification data into the mappings stored in the data lake 140 of the conversion unit 102. The feature vectors are generated in the number of latent spaces (M) × the number of verification data (Na).

[0103] Then, the feature vector calculation unit 104 creates a data lake 150 for evaluating the performance of the learning model, in which for each verification data, there are a prediction model and Na (Na ≦ N: Na is a natural number) accuracies of the prediction model and M feature vectors associated therewith (S132). The accuracy of the prediction model is, for example, an error rate, a prediction accuracy, an effectiveness, etc.

[0104] <Processing by the blending characteristic prediction unit 2> Figure 12 is a flowchart showing the processing procedure by the formulation characteristic prediction unit 2.

[0105] First, at least one target variable is set by the predictive explanatory variable input unit 201, and the required explanatory variable data "Ti" is set (S201). The values ​​of the explanatory variables of the effective properties shown in Figures 4 to 7 are input as the required explanatory variable data. For example, crystallinity, tensile modulus, tensile stress, etc., shown in Figure 4 are input (see Figure 4).

[0106] Next, the feature vector acquisition unit 202 accesses the data lake 140 of the transformation unit 102.

[0107] The feature vector acquisition unit 202 then calculates a feature vector by substituting the required explanatory variable data into the mapping "L" stored in the data lake 140 of the transformation unit 102 (S202). The feature vector acquired in step S202 is represented as "LM(Ti)=(L1(Ti),L2(Ti),···,LM(Ti))". The feature vector acquired in step S202 has M components.

[0108] Next, the similar record ranking calculation unit 203 calculates a group of similar records "{Di,k}" (S203). In step S203, the similar record ranking calculation unit 203 obtains M feature vectors "LM(Ti)" obtained in step S202 and M feature vectors LM(D) from the validation data "D". The feature vectors LM(D) from the validation data "D" were calculated in step S131 in Figure 11. Incidentally, the validation data consists of the record data 113 shown in Figure 3.

[0109] Then, in step S203, the similarity record ranking calculation unit 203 calculates the Euclidean distance (similarity) between the feature vector "LM(Ti)" obtained in step S202 and the M feature vectors LM(D) from the verification data "D".

[0110] The Euclidean distance between the feature vector "LM(Ti)" obtained in step S202 and the M feature vectors LM(D) from the validation data "D" is given by equation (21) below. Incidentally, the Euclidean distance is defined in the latent space.

[0111]

number

[0112] The similar record ranking calculation unit 20 then extracts the nth verification data in ascending order of Euclidean distance to calculate the similar record group "{Di,k}". In other words, the similar record group "{Di,k}" is the verification data (records in Figure 3) that are similar to the requirement explanatory variable data entered in step S201. From the above description, the similar record group "{Di,k}" can be summarized by the following equation (22). In the similar record group "{Di,k}", the verification data is stored in order of similarity to the requirement explanatory variable data entered in step S201.

[0113]

number

[0114] In equation (1), "n" is a number specified by the user. The similar record group "{Di,k}" is a ranking of similar records in the validation data (record data 113) for the required explanatory variable data.

[0115] In equation (22), "i" corresponds to "i" in the required explanatory variable data "Ti" entered in step S201. Also, "j" corresponds to "j" in the mapping "Lj". And "k" is the rank in the similar record group "{Di,k}".

[0116] Step S203 is the "similarity calculation step".

[0117] By calculating such a group of similar records, the subsequent processing in the AWE prediction model generation unit 205 can generate an AWE prediction model using validation data with high similarity.

[0118] Then, the partial output calculation unit 204 calculates the partial output "Oi,k" (S204). In step S204, the internal model error rate "eij" (percentage) for similar records "Di,k" is extracted up to the mth smallest, along with the external model value (external model value) "Fi". As a result, the partial output calculation unit 204 calculates an ensemble partial output "Oi,k" for one similar record "Di,k".

[0119] The ensembled partial output "Oi,k" is represented, for example, by equation (23) below.

[0120]

number

[0121] In equation (23), "i" corresponds to the "i" in the requirement explanatory variable data "Ti" entered in step S201. "j" identifies the external model "Fj". And "k" is the rank in the similar record group "{Di,k}".

[0122] The partial output "Oi,k" is the sum of the external model "Fj(Ti)" weighted by the error rates "ei,j". "Fj(Ti)" is obtained by substituting the required explanatory variable data "Ti" into the external model. Furthermore, it is weighted so that the value increases as the error rate decreases. The denominator of equation (23) is a term used for normalization. Any mathematical formula that can express this characteristic (the value increases as the error rate decreases) is acceptable, not just equation (23).

[0123] Next, the AWE prediction model generation unit 205 calculates the ensemble prediction blending characteristic "Oi" (ensemble model) (S205). The prediction blending characteristic is calculated by weighting the partial output "Oi,k" with the Euclidean distance "L(Ti,Di,k)" of the similar record group "{Di,k}", and then summing them up. The Euclidean distance "L(Ti,Di,k)" is the distance between the required explanatory variable data "Ti" input in step S201 and the elements "Di,k" of the similar record group (i.e., the validation data). The weighting is such that the smaller the Euclidean distance "L(Ti,Di,k)", the larger the weighting. The prediction blending characteristic is calculated by the following equation (24). The generation of the AWE prediction model is completed by the processing in step S205.

[0124]

number

[0125] Note that the denominator of equation (24) is a term used for normalization. The terms "i" and "k" in equation (24) are the same as in equation (23). As before, any formula other than equation (24) that can express this characteristic (the smaller the Euclidean distance "L(Ti,Di,k)", the larger the value) may be used.

[0126] Steps S204 and S205 are "ensemble model generation steps".

[0127] Subsequently, the prediction accuracy and effectiveness calculation unit 206 calculates the prediction accuracy and effectiveness of the prediction made by the AWE prediction model (S206). However, step S206 is optional.

[0128] The predicted blending characteristics output unit 207 then outputs the predicted blending characteristics calculated in step S205 to the output device 405 (S207). The predicted blending characteristics indicate what characteristics can be expected when the required explanatory variable data is applied to the learning results by the learning unit 103.

[0129] (Screen 210 for setting the variable to be predicted) Figure 13 shows an example of the predictor variable setting screen 210 that is displayed on the display device when inputting the target variable and required explanatory variable data in step S201 of Figure 12.

[0130] The prediction target variable setting screen 210 includes a model number input section 211 and a custom input section 212.

[0131] The user can select between inputting via the model number input section 211 or inputting via the custom input section 212 by selecting from the radio buttons 213.

[0132] When the user selects the model number input unit 211 and specifies the model numbers of the base material, reinforcing agent, etc., the explanatory variable input unit 201 sets the explanatory variable data.

[0133] When the user selects the custom input section 212, the explanatory variable data is set by the user's manual input.

[0134] The target variable is set using a target variable setting screen, which is not shown in the diagram.

[0135] (Feature vector display screen 220) Figure 14 shows an example of a screen displaying the feature vectors obtained in step S202 of Figure 12.

[0136] As shown in Figure 14, the feature vector display screen 220 displays the feature vectors. The feature vectors are the result of substituting the verification data into the mapping "L" stored in the data lake 140 of the transformation unit 102 in step S202 of Figure 12.

[0137] (Predicted blending characteristics output screen 230) Figure 15 shows an example of a screen displaying the predicted blending characteristics output in step S207 of Figure 12.

[0138] The predicted blending characteristics output screen 230 consists of a predicted value output unit 231 and a radar chart display unit 232.

[0139] The predicted value output unit 231 displays the predicted value of the target variable, calculated based on the predicted blending characteristics ("Oi" according to equation (24)) calculated for each target variable.

[0140] The radar chart display unit 232 shows the predicted value of the target variable as a dashed line on the radar chart. As shown in Figure 15, the required characteristics (dashed line) and the characteristics of the recycled material before addition (solid line) may also be shown for reference.

[0141] In this embodiment, an ensemble model is generated as a plastic compounding characteristic prediction model from the results of inputting the values ​​of the required explanatory variables, which are explanatory variables required for the compounding characteristics of the plastic, into the learning model. This allows the user to confirm compounding characteristics that are close to the required characteristics.

[0142] <Processing by the formula generation unit 3> Figure 16 is a flowchart showing the processing procedure by the formula recipe generation unit 3.

[0143] First, the user sets the dependent variable and its value (target value) z (≧1) using the target characteristic / condition input unit 301, and inputs the conditions to be imposed on the independent variables (S301). The target value is denoted as "Yi".

[0144] Next, the feature vector acquisition unit 202 to the AWE prediction model generation unit 205 of the blending characteristic prediction unit 2 execute steps S202 to S205 in Figure 12, creating an AWE prediction model "Gi" for each target variable (S311). Note that the AWE prediction model "Gi" is the same as the predicted blending characteristic "Oi" calculated in step S205 in Figure 12.

[0145] Meanwhile, the constraint condition setting unit 302 sets the constraint conditions for parameter attributes (NaN value / user-specified fixed value / search value), search range, and summation rule of component ratios as explanatory variables (S321). The settings performed in step S321 are based on the conditions imposed on the explanatory variables entered in step S301.

[0146] Next, the first optimal value search unit 303 inputs the explanatory variable data "x" into the AWE prediction model "Gi" to calculate the AWE prediction model value "Gi(x)" for each target variable (S331). The explanatory variable data "x" is the initial value of the explanatory variable data set by the method described later.

[0147] The first optimal value search unit 303 sequentially updates the input explanatory variable data u times (S332) so that the sum of the absolute value residues of the AWE prediction model value "Gi(x)" into which the explanatory variable data "x" is substituted and the target value "Yi" becomes small. The sum of the absolute value residues of the AWE prediction model "Gi(x)" into which the explanatory variable data "x" is substituted and the target value "Yi" is given by the following equation (31). The update is performed based on an optimization algorithm, such as Bayesian optimization or a genetic algorithm. Note that equation (31) corresponds to the "distance between the target characteristic and the ensemble model value" mentioned above.

[0148]

number

[0149] In step S332, the first optimal value search unit 303 inputs the explanatory variable "v", the top-ranked similar record obtained in step S312, for the first v (≤ u) of the u updates. This means that the explanatory variable "v", the top-ranked similar record, is set as the initial value for updates such as Bayesian optimization.

[0150] Generally, the initial values ​​for Bayesian optimization are randomly selected, but in this embodiment, the explanatory variable "v", the top-ranked record in the similar record ranking, is used as the initial value. In other words, the record data 113 (verification data) with a predetermined top rank in the ranking is set as the initial value. The ranking is calculated by the similar record ranking calculation unit 203. This makes it possible to improve the efficiency of the update (Bayesian optimization).

[0151] The predictable characteristic determination unit 304 then determines, within u updates, whether a solution was found in which the error in the AWE prediction model value (ensemble model value) is less than or equal to a threshold for the target variable set by the user (S332). The error in the AWE prediction model value is the error calculated when the updated explanatory variable data "x" is input into the AWE prediction model. The error here is the error with respect to the validation data (record data 113) that is closest to the explanatory variable data "x". Specifically, the error is given by equation (31). The threshold is set by the user as appropriate depending on the purpose, etc.

[0152] If a solution below the threshold is found (S332 → No), the predictable characteristic determination unit 304 displays an alert to the user indicating that the calculation will ignore the target variable below the threshold (S333). The blending recipe prediction unit then proceeds to step S341.

[0153] If no solution below the threshold is found (S332 → Yes), the second optimal value search unit 305 updates the explanatory variable data for the target characteristic "Yi" based on Bayesian optimization. In this case, if "No" is determined in step S332, the second optimal value search unit 305 deletes the dependent variable (i.e., the target characteristic) below the threshold. If "Yes" is determined in step S332, the second optimal value search unit 305 does not change the explanatory variable data. The second optimal value search unit 305 then updates the explanatory variable data sequentially for w(u+1) times, starting from the (u+1)th time (S341). This allows for efficient updates.

[0154] Next, the Pareto solution candidate selection unit 306 selects a Pareto solution candidate "x*" from the history of all sequential updates in steps S332 and S341 (S342). The Pareto solution candidate "x*" is "x" that satisfies the following equation (32). The Pareto solution candidate "x*" is a set of explanatory variable data.

[0155]

number

[0156] If there is only one dependent variable, the Pareto solution candidate selection unit 306 selects at least one minimum solution as the Pareto solution "x*".

[0157] Then, the error / prediction accuracy calculation unit 307 calculates the error (error) and prediction accuracy of the Pareto solution candidate (S343). The error of the Pareto solution candidate is the error (distance) between the value obtained by inputting the explanatory variable data "x*" corresponding to the Pareto solution candidate into the AWE prediction model and the target value "Yi".

[0158] Next, the Pareto solution determination unit 308 sorts the Pareto solutions in order of priority based on the error or prediction accuracy of each candidate Pareto solution (S344). The sorting is performed using the Borda or Condorcet method based on the error and prediction accuracy.

[0159] In steps S342 to S344, multiple candidate Pareto solutions are calculated based on the update history, and the Pareto solutions are sorted based on the error and prediction accuracy of each candidate. The result of this sorting is then output to the output device 405 as a plastic formulation recipe. This allows for the output of a Pareto solution with a small error or high prediction accuracy as the predicted formulation recipe.

[0160] Subsequently, the predicted formula recipe output unit 309 outputs at least one predicted formula recipe to the output device according to its priority order (S345). The predicted formula recipe is the Pareto solution sorted in step S344.

[0161] (Target characteristics / condition input screen 310) Figure 17 shows an example of the target characteristic / condition input screen 310 displayed on the output device 405 in step S301 of Figure 16.

[0162] As shown in Figure 17, the target characteristic / condition input screen 310 is composed of a target characteristic input section 311, a model number input section 312, and a condition input section 313.

[0163] The target characteristic input unit 311 accepts input for the objective variable (in the example shown in Figure 17, "compound characteristics (target compound characteristics)") and multiple target values. In the example shown in Figure 17, the values ​​for "crystallinity," "tensile modulus," and "tensile stress" are entered as target values, indicating that these are specified as target values. Incidentally, items for which no value has been entered indicate that they are not specified as target values.

[0164] The part number input section 312 receives information about the part numbers of the materials used in the product for which the recipe applies. In the example shown in Figure 17, the manufacturer and part numbers of the base material and two reinforcing agents are specified.

[0165] Then, when the user clicks the reflect button 314 with the mouse, candidate conditions are displayed in the condition input unit 313. Of the candidate conditions displayed in the condition input unit 313, those that have already been entered in the target characteristic input unit 311 have values ​​pre-filled. Also, by checking the checkbox in the "Search" item, the conditions to be entered in step S301 in Figure 16 are specified (input). Conditions with checked checkboxes may have specific values ​​entered, but as shown in Figure 17, they do not have to have values ​​entered. This is because, as shown in step S311 in Figure 16, the constraint condition setting unit 302 sets constraint conditions for the conditions with checked checkboxes.

[0166] (Alert display screen 320) Figure 18 shows an example of the alert display screen 320 that appears in step S333 of Figure 16.

[0167] The alert screen 320 displays a message indicating that a solution with an error below the threshold has been found among the user-defined target variables, and therefore, the target variables below the threshold will be excluded (ignored) from subsequent calculations.

[0168] (Predicted recipe display screen 330) Figures 19 and 20 show examples of the predicted breeding recipe display screen 330, which displays a predicted breeding recipe. The predicted breeding recipe display screen 330 is output and displayed in step S345 of Figure 16.

[0169] Figure 19 shows the recipe ranking display screen 331, and Figure 20 shows the prediction characteristic radar chart display screen 332.

[0170] In the recipe ranking display screen 331 shown in Figure 19, predicted recipe combinations are displayed in a ranking. The predicted recipe combinations are the Pareto solutions "x*" selected in step S342 of Figure 16. The ranking is the result of sorting performed in step S344 of Figure 16.

[0171] In Figure 19, "filler size" and "phase solvent density" represent the set of explanatory variables in the Pareto solution.

[0172] Furthermore, in the prediction characteristic radar chart display screen 332 shown in Figure 20, the value of the target variable is displayed in radar chart format for each of the predicted blending recipes shown in the recipe ranking display screen 331 in Figure 19.

[0173] In this embodiment, an ensemble model is generated as a plastic compounding characteristic prediction model from the results of inputting the values ​​of the required explanatory variables, which are explanatory variables required for the plastic compounding characteristics, into a learning model. This allows the user to confirm compounding characteristics that are close to the required characteristics, and to realize an appropriate plastic compound. Furthermore, according to this embodiment, it is possible to significantly improve the accuracy of MI prediction for a wide variety of plastic recycling sources.

[0174] Furthermore, according to this embodiment, by using a mapping to a latent space, it is possible to predict the properties of the compounded plastic with high accuracy even when using a sparse material DB110 as input.

[0175] The present invention is not limited to the embodiments described above, and various modifications are included. For example, the embodiments described above are described in detail to illustrate the present invention clearly, and are not necessarily limited to those having all the configurations described.

[0176] Furthermore, in this embodiment, the model building unit 1, the blending characteristic prediction unit 2, and the blending recipe generation unit 3 are assumed to be running on the same computer. However, at least one of the model building unit 1, the blending characteristic prediction unit 2, and the blending recipe generation unit 3 may be running on a different computer. Also, the material DB 110, the effective characteristic DB 120, the latent space DB 130, the data lake 140 for the conversion unit 102, and the data lake 150 for learning model performance evaluation may be stored on a computer different from the model building unit 1.

[0177] Furthermore, each of the above-mentioned configurations, functions, effective characteristic acquisition unit 101 to feature vector calculation unit 104, prediction target explanatory variable input unit 201 to prediction blending characteristic output unit 207, target characteristic / condition input unit 301 to prediction blending recipe compilation unit 309, storage device 403, etc., may be implemented in hardware, either partially or entirely, by designing them as integrated circuits, for example. Also, as shown in Figure 2, each of the above-mentioned configurations, functions, etc., may be implemented in software by having a processor such as a CPU interpret and execute a program that realizes each function. Information such as programs, tables, and files that realize each function can be stored not only on an HD (Hard Disk), but also in memory, a recording device such as an SSD (Solid State Drive), or a recording medium such as an IC (Integrated Circuit) card, an SD (Secure Digital) card, or a DVD (Digital Versatile Disc).

[0178] Furthermore, in each embodiment, only those control lines and information lines deemed necessary for explanation are shown, and not all control lines and information lines are necessarily shown in the actual product. In practice, it can be assumed that almost all components are interconnected. [Explanation of Symbols]

[0179] 1. Model Construction Department 2. Formulation characteristic prediction unit 3. Recipe Generation Unit (Recipe Generation Unit) 101 Effective characteristic acquisition unit 102 Transformation Unit (Mapping Calculation Unit) 103 Learning Department 104 Feature Vector Calculation Unit 110. Materials Database 121 Sign (dependent variable) 122 Sign (explanatory variable) 203 Similar Record Ranking Calculation Unit (Similarity Calculation Unit) 204 Partial Output Calculation Unit (Ensemble Model Generation Unit) 205 AWE Prediction Model Generation Unit (Ensemble Model Generation Unit) 303 First appropriate value search unit (update processing unit) 305 Second optimal value search unit (update processing unit) 306 Pareto Solution Candidate Selection Unit (Pareto Solution Processing Unit) 308 Pareto solution determination unit (Pareto solution processing unit) 309 Predicted Formula Recipe Output Unit 405 Output device (output section) Z Plastic Compounding Properties Prediction System S111 Learning (Learning Steps) S122 Calculation of the mapping (Mapping calculation step) S203 Calculation of Similar Record Groups (Similarity Calculation Step) S204 Calculation of partial output (Ensemble model generation step) S205 Calculation of Predicted Blending Characteristics (Ensemble Model Generation Step)

Claims

1. A mapping calculation unit calculates a mapping to a latent space having fewer dimensions than the explanatory variables stored in the material database, based on a plastic material database, using manifold learning. A learning unit generates a learning model by performing learning to regress material properties from the data stored in the aforementioned material database. A similarity calculation unit that maps the explanatory variables to the latent space using the mapping and calculates the similarity between the result of the mapping and the record data stored in the material database, An ensemble model generation unit generates an ensemble model as a model for predicting the composition characteristics of the plastic, based on the results of inputting the values ​​of the required explanatory variables, which are explanatory variables required as the composition characteristics of the plastic, into the learning model using weighting based on the aforementioned similarity. A plastic compounding properties prediction system equipped with the following features.

2. If the objective variable of the learning model and the ensemble model is the crystallinity of the plastic after compounding, then at least one of the following is included as an explanatory variable: base material melting point, base material tensile stress, base material heat of fusion, base material thermal decomposition temperature, base material tensile elongation at fracture, base material tensile modulus, and base material MFR. If the objective variable is the tensile modulus of elasticity of the plastic after compounding, then at least one of the filler density, base material MFR, base material tensile yield stress, and base material tensile stress is included in the explanatory variables. If the objective variable is the tensile stress after compounding the plastic, then at least one of the following is included as an explanatory variable: the tensile modulus of the base material, the MFR of the base material, the elongation at break of the base material, the melting point of the base material, and the density of the base material. If the predictive model's dependent variable is the tensile yield stress of the plastic after compounding, then at least one of the following is included as an explanatory variable: base material elongation at break, base material tensile modulus, filler density, and base material density. If the objective variable is the tensile elongation at fracture after compounding the plastic, then at least one of the following is included in the explanatory variables: the tensile yield stress of the base material, the tensile stress of the base material, the MFR of the base material, the filler density, the tensile modulus of the base material, and the crystallinity of the base material. If the objective variable is the flexural modulus of the plastic after compounding, then at least one of the following is included in the explanatory variables: the tensile modulus of the base material, the filler density, the base material density, and the tensile stress of the base material. If the objective variable is the bending stress after compounding, then at least one of the following is included in the explanatory variables: base material Rockwell hardness, presence or absence of filler surface treatment, base material density, base material flexural modulus, and base material MFR. If the objective variable is the Izod impact value after compounding the plastic, then at least one of the following is included in the explanatory variables: the tensile modulus of the base material, the base material MFR, the tensile elongation at fracture of the base material, the tensile yield stress of the base material, and the melting point of the base material. If the objective variable is the MFR after compounding the plastic, then at least one of the following is included in the explanatory variables: base material tensile elongation at fracture, base material tensile yield stress, filler size, filler density, base material tensile stress, and base material density. If the objective variable is the melting point of the plastic after compounding, then at least one of the following explanatory variables is included: base material tensile elongation at break, base material MFR, base material tensile modulus, filler size, base material glass transition temperature, base material crystallization temperature, filler weight ratio, base material weight ratio, and base material volume ratio. If the objective variable is the heat of fusion after compounding the plastic, then at least one of the following is included in the explanatory variables: base material melting point, base material tensile elongation at fracture, base material MFR, base material tensile modulus, base material crystallinity, and base material tensile stress. If the objective variable is the crystallization temperature of the plastic after compounding, then at least one of the following is included in the explanatory variables: matrix crystallization heat, phase solvent density, matrix fusion heat, filler size, and filler density. If the objective variable is the heat of crystallization after compounding the plastic, then at least one of the following is included in the explanatory variables: the melting point of the base material, the filler density, the tensile stress of the base material, the crystallinity of the base material, the filler size, and the tensile elongation at fracture of the base material. If the objective variable is the glass transition temperature after compounding the plastic, then at least one of the following is included in the explanatory variables: tensile stress of the base material, tensile modulus of the base material, MFR of the base material, density of the base material, and tensile elongation at fracture of the base material. If the objective variable is the thermal decomposition temperature of the plastic after compounding, then at least one of the following explanatory variables must be included: base material tensile stress, base material tensile elongation at fracture, base material tensile modulus, base material MFR, filler size, base material density, base material weight ratio, base material tensile yield stress, and base material volume ratio. The plastic compounding properties prediction system according to feature 1.

3. The similarity calculation unit, The record data is ranked based on the aforementioned similarity. The plastic compounding properties prediction system according to feature 1.

4. Recipe generation unit that generates the plastic compounding recipe using the ensemble model. The plastic compounding property prediction system according to claim 3, characterized by comprising the above.

5. The recipe generation unit, An update processing unit updates the explanatory variables a predetermined number of times, using the record data of the highest predetermined rank from the aforementioned ranking as the initial value, so as to satisfy the constraint conditions for the explanatory variables in the aforementioned blending recipe and so as to reduce the distance between the target characteristic and the ensemble model value obtained by inputting the values ​​of the explanatory variables into the ensemble model. The plastic compounding properties prediction system according to claim 4, characterized by comprising the above.

6. The update processing unit, If, during the predetermined number of updates performed by the update processing unit, there is a set of explanatory variables in which the error in the ensemble model value does not reach a predetermined threshold, the target characteristic corresponding to that set of explanatory variables is removed, and the updates continue. The plastic compounding properties prediction system according to claim 5.

7. Based on the update history, the Pareto solution processing unit calculates multiple candidate Pareto solutions, sorts the Pareto solutions based on the error and prediction accuracy of each candidate, and outputs the sorting result as a plastic formulation recipe to the output unit. The plastic compounding properties prediction system according to claim 5, characterized by comprising the following:

8. A mapping calculation unit calculates a mapping to a latent space having fewer dimensions than the explanatory variables stored in the material database, based on a plastic material database, using manifold learning. A learning unit generates a learning model by performing learning to regress material properties from the data stored in the aforementioned material database. A similarity calculation unit that maps the explanatory variables to the latent space using the mapping and calculates the similarity between the result of the mapping and the record data stored in the material database, An ensemble model generation unit generates an ensemble model as a model for predicting the composition characteristics of the plastic, based on the results of inputting the values ​​of the required explanatory variables, which are explanatory variables required as the composition characteristics of the plastic, into the learning model using weighting based on the aforementioned similarity. A recipe generation unit that generates a plastic compounding recipe using the aforementioned ensemble model, A plastic compounding properties prediction system equipped with the following features.

9. A plastic compounding properties prediction system that predicts the compounding properties of plastics The learning step involves generating a learning model by performing regression learning on material properties from data stored in a materials database, and A mapping calculation step that calculates a mapping to a latent space having fewer dimensions than the explanatory variables stored in the material database, based on a plastic material database, using manifold learning, A similarity calculation step involves mapping the explanatory variables to the latent space using the mapping, and calculating the similarity between the result of the mapping and the record data stored in the material database. An ensemble model generation step is performed in which, using weighting based on the aforementioned similarity, an ensemble model is generated as a model for predicting the composition characteristics of the plastic, based on the results of inputting the values ​​of the required explanatory variables, which are explanatory variables required as the composition characteristics of the plastic, into the learning model, and A method for predicting plastic compound properties.

Images