Property prediction program, learned model generation program, and property prediction system

Abstract

To provide a property prediction program, a learned model generation program, and a property prediction system capable of predicting properties without lowering accuracy even under a condition that a destruction mode becomes complicated.SOLUTION: A property prediction program according to the present invention causes a computer to execute: a shape parameter estimation step of inputting a blending condition of a thermoplastic resin composition to be estimated, and a manufacturing condition and a test condition to a shape parameter prediction model in which a blending condition of raw materials constituting a thermoplastic resin composition, and a manufacturing condition and a test condition are explanatory variables and a shape parameter defining a shape of a measurement datum obtained by a dynamic test is a target variable; and a property prediction step of inputting the shape parameters into the shock absorbing property prediction model having shape parameters as explanatory variables and shock absorbing properties as target variables to acquire the shock absorbing properties.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a characteristic prediction program, a trained model generation program, and a characteristic prediction system. [Background technology]

[0002] Thermoplastic resin compositions are one type of engineering plastic that excel in heat resistance, flame retardancy, chemical resistance, electrical insulation, moist heat resistance, mechanical strength, and dimensional stability. The mechanical properties of these resin compositions vary depending on their formulation and manufacturing conditions, so formulations and manufacturing conditions are selected to match the desired properties. For example, a technique is known for predicting the impact absorption properties of a thermoplastic resin composition by inputting the formulation and manufacturing conditions of the resin composition into a trained model generated by machine learning and obtaining the predicted values ​​output by the trained model (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2024-3697 Summary of the Invention [Problem to be solved by the invention]

[0004] In Patent Document 1, impact absorption properties are predicted directly from compounding conditions and manufacturing conditions. However, in impact tests in which the load energy on the material is very high or in impact tests conducted under high or low temperature conditions, the fracture mode of the thermoplastic resin composition when impacted becomes complex, and the prediction accuracy of the impact absorption properties using the trained model may be low.

[0005] The present invention has been made in consideration of the above, and aims to provide a characteristic prediction program, a trained model generation program, and a characteristic prediction system that can predict characteristics without reducing accuracy even under conditions where the failure mode becomes complex. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems and achieve the object, the property prediction program of the present invention is a property prediction program for predicting the properties of a thermoplastic resin composition, and causes a computer to execute the following steps: a shape parameter prediction model is read from memory; the shape parameter prediction model has as explanatory variables the blending conditions of the raw materials constituting the thermoplastic resin composition, the manufacturing conditions and test conditions of a molded article used in dynamic testing of the thermoplastic resin composition, and as objective variables the shape parameters that define the shape of the measurement data obtained by dynamic testing of the thermoplastic resin composition; the blending conditions of the thermoplastic resin composition to be estimated, and the manufacturing conditions and test conditions of a molded article to be used in dynamic testing of the thermoplastic resin composition are input into the shape parameter prediction model, and the shape parameters of the thermoplastic resin composition to be estimated are acquired; and a property prediction step is read from memory; the shape parameters of the thermoplastic resin composition are input as explanatory variables and the impact absorption properties of the thermoplastic resin composition are input into the impact absorption prediction model, and the impact absorption properties of the thermoplastic resin composition to be estimated are acquired.

[0007] In addition, in the property prediction program according to the present invention, in the above invention, the dynamic test is any one of a drop weight impact test, a high-speed compression test, a high-speed tensile test, a puncture impact test, a Charpy impact test, and an Izod impact test.

[0008] In addition, in the characteristic prediction program according to the present invention, in the above invention, the shape parameter is at least one of the load value and displacement value corresponding to the maximum point, the load value and displacement value corresponding to the minimum point, the number of extreme values, and the coefficient and constant term of a function obtained by polynomial approximation of a specified data interval, in the measurement data and function-processed data obtained by first-order or second-order differentiation of the measurement data.

[0009] In addition, in the characteristic prediction program according to the present invention, in the above invention, the datasets used for training the shape parameter prediction model and the impact absorption prediction model are datasets that have been processed in advance to exclude, as outliers, values ​​that are more than three times the standard deviation from the average value of each shape parameter.

[0010] In addition, in the property prediction program according to the present invention, in the above invention, the shape parameters are selected based on the variable importance calculated using a trained model obtained by machine learning using the shape parameters of a known thermoplastic resin composition as explanatory variables and the impact absorption properties of the thermoplastic resin composition as objective variables, and the variable importance of the explanatory variables of the trained model is calculated.

[0011] In addition, in the property prediction program of the present invention, in the above invention, the shape parameters are selected based on the combination of variables that provides the highest accuracy by repeatedly adding or reducing the number of explanatory variables of the trained model obtained by machine learning using the shape parameters of a known thermoplastic resin composition as explanatory variables and the impact absorption properties of the thermoplastic resin composition as objective variables, using a trained model.

[0012] In addition, the trained model generation program of the present invention causes a computer to execute the following steps: a shape parameter selection step of reading from memory a trained model obtained by machine learning, in which shape parameters that define the shape of measurement data obtained by dynamic testing of a known thermoplastic resin composition are used as explanatory variables and the impact absorption characteristics of the thermoplastic resin composition are used as objective variables, and selecting a shape parameter from multiple shape parameters based on the importance of the explanatory variables in the trained model; a shape parameter prediction model generation step of generating, for each selected shape parameter, a shape parameter prediction model that outputs the shape parameter by machine learning using the blending conditions of the raw materials that make up the known thermoplastic resin composition, the manufacturing conditions and test conditions of the molded product used in dynamic testing of the thermoplastic resin composition as explanatory variables, and each shape parameter selected in the previous step for the thermoplastic resin composition as objective variables; and an impact absorption prediction model generation step of generating, by machine learning using the selected shape parameters as explanatory variables and the impact absorption characteristics of the thermoplastic resin composition as objective variables, an impact absorption prediction model that outputs the impact absorption characteristics of the thermoplastic resin composition.

[0013] Furthermore, the property prediction system according to the present invention is a property prediction system for predicting the properties of a thermoplastic resin composition, and includes: a shape parameter estimation unit that inputs the compounding conditions of the thermoplastic resin composition to be estimated and the manufacturing conditions and test conditions of the molded product to be used in the dynamic testing of the thermoplastic resin composition into a shape parameter prediction model having, as explanatory variables, the compounding conditions of the raw materials constituting the thermoplastic resin composition, and the manufacturing conditions and test conditions of the molded product to be used in the dynamic testing of the thermoplastic resin composition, and the shape parameters that define the shape of the measurement data obtained by the dynamic testing of the thermoplastic resin composition as objective variables, and acquires the shape parameters of the thermoplastic resin composition to be estimated; and a property prediction unit that inputs the shape parameters estimated by the shape parameter estimation unit into an impact absorption prediction model having, as explanatory variables, the shape parameters of the thermoplastic resin composition and the impact absorption properties of the thermoplastic resin composition as objective variables, and acquires the impact absorption properties of the thermoplastic resin composition to be estimated. [Effects of the Invention]

[0014] According to the present invention, it is possible to predict characteristics without reducing accuracy even under conditions where the failure mode becomes complex. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a diagram showing a schematic configuration of a characteristic prediction system according to one embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram showing the configuration of a learning device included in a characteristic prediction system according to one embodiment of the present invention. [Figure 3] FIG. 3 is a flowchart showing the flow of a trained model generation process performed by a learning device according to one embodiment of the present invention. [Figure 4] FIG. 4 is a flowchart showing the flow of the parameter selection process in the process of generating a trained model. [Figure 5] FIG. 5 is a flowchart showing the flow of the shape parameter prediction model generation process in the trained model generation process. [Figure 6] FIG. 6 is a block diagram showing the configuration of a characteristic prediction device included in a characteristic prediction system according to one embodiment of the present invention. [Figure 7] FIG. 7 is a flowchart showing the flow of the characteristic prediction process performed by the characteristic prediction device according to one embodiment of the present invention. [Figure 8] FIG. 8 is a diagram (part 1) for explaining the energy absorbing member used in the drop weight test. [Figure 9] FIG. 9 is a diagram (part 2) for explaining the energy absorbing member used in the drop weight test. [Figure 10] FIG. 10 is a diagram (part 3) for explaining the energy absorbing member used in the drop weight test. [Figure 11] FIG. 11 is a diagram for explaining the drop weight test. [Figure 12] FIG. 12 is a diagram for explaining calculation of the deformation amount of the energy absorbing member. [Figure 13] FIG. 13 is a graph showing a curve showing the load-displacement relationship obtained by a drop weight test. [Figure 14] FIG. 14 is a diagram in which the load-displacement curve shown in FIG. 13 is converted into a curve showing the relationship between the amount of absorbed energy and the displacement. [Figure 15] FIG. 15 is a diagram in which the curve showing the relationship between the amount of absorbed energy and the displacement shown in FIG. 14 is converted into a curve showing the relationship between the amount of absorbed energy and the amount of deformed resin. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, an embodiment of a property prediction system according to the present invention will be described in detail with reference to the drawings, although the present invention is not limited to this embodiment.

[0017] (Embodiment) 1 is a diagram showing a schematic configuration of a property prediction system according to one embodiment of the present invention. The property prediction system 1 includes a learning device 2 that creates training data and generates a trained model trained using the created training data, a property prediction device 3 that predicts the properties of a resin composition using the trained model generated by the learning device 2, a display device 4 that displays information including the prediction results of the property prediction device 3, and an input device 5.

[0018] The properties predicted by the property prediction device 3 include impact absorption properties, such as load-displacement curve data from any one of dynamic tests such as a drop weight impact test, a high-speed compression test, a high-speed tensile test, a puncture impact test, a Charpy impact test, and an Izod impact test, or functional processing data obtained by processing the same into functions.

[0019] The thermoplastic resin composition is not particularly limited, but examples thereof include styrene-based resins, fluororesins, polyoxymethylene, polyamide, polyester, polyamideimide, vinyl chloride, olefin-based resins, polyacrylate, polyphenylene ether, polycarbonate, polyethersulfone, polyetherimide, polyetherketone, polyetheretherketone, polyarylene sulfide, cellulose derivatives, liquid crystal resins, and modified resins thereof, etc. Two or more of these may be contained.

[0020] Examples of styrene-based resins include PS (polystyrene), HIPS (high impact polystyrene), AS (acrylonitrile / styrene copolymer), AES (acrylonitrile / ethylene-propylene-non-conjugated diene rubber / styrene copolymer), ABS (acrylonitrile / butadiene / styrene copolymer), and MBS (methyl methacrylate / butadiene / styrene copolymer). Here, " / " indicates a copolymer, and the same applies below. Two or more of these may be contained. Among these, ABS is particularly preferred.

[0021] Specific examples of polyamides include polycaproamide (nylon 6), polyhexamethylene adipamide (nylon 66), polypentamethylene adipamide (nylon 56), polytetramethylene adipamide (nylon 46), polyhexamethylene sebacamide (nylon 610), polyhexamethylene dodecamide (nylon 612), polyundecane amide (nylon 11), polydodecanamide (nylon 12), polycaproamide / polyhexamethylene adipamide copolymer (nylon 6 / 66), polycaproamide / polyhexamethylene terephthalamide copolymer (nylon 6 / 6T), polyhexamethylene adipamide / polyhexamethylene isophthalamide copolymer (nylon 6 / 6T), and polyhexamethylene adipamide / polyhexamethylene isophthalamide copolymer (nylon 6 / 6T). Examples of such copolymers include nylon 66 / 6I, polyhexamethylene terephthalamide / polyhexamethylene isophthalamide copolymer (nylon 6T / 6I), polyhexamethylene terephthalamide / polydodecanamide copolymer (nylon 6T / 12), polyhexamethylene adipamide / polyhexamethylene terephthalamide / polyhexamethylene isophthalamide copolymer (nylon 66 / 6T / 6I), polyxylylene adipamide (nylon XD6), polyhexamethylene terephthalamide / poly-2-methylpentamethylene terephthalamide copolymer (nylon 6T / M5T), polynonamethylene terephthalamide (nylon 9T), and copolymers thereof. Two or more of these may be used.

[0022] Examples of olefin resins include polypropylene, polyethylene, ethylene / propylene copolymer, ethylene / 1-butene copolymer, ethylene / propylene / non-conjugated diene copolymer, ethylene / ethyl acrylate copolymer, ethylene / glycidyl methacrylate copolymer, ethylene / vinyl acetate / glycidyl methacrylate copolymer, propylene-g-maleic anhydride copolymer, ethylene / propylene-g-maleic anhydride copolymer, methacrylic acid / methyl methacrylate / glutaric anhydride copolymer, etc. Two or more of these may be contained.

[0023] Preferred polyesters are polymers or copolymers whose main structural units are residues of dicarboxylic acids or their ester-forming derivatives and diols or their ester-forming derivatives. Among these, aromatic polyesters such as polyethylene terephthalate, polypropylene terephthalate, polybutylene terephthalate, polycyclohexanedimethylene terephthalate, polyethylene naphthalate, polypropylene naphthalate, polybutylene naphthalate, polyethylene isophthalate / terephthalate, polypropylene isophthalate / terephthalate, polybutylene isophthalate / terephthalate, polyethylene terephthalate / naphthalate, polypropylene terephthalate / naphthalate, and polybutylene terephthalate / naphthalate are particularly preferred, with polybutylene terephthalate being the most preferred. Two or more of these may be contained. In these polyesters, the ratio of terephthalic acid residues to all dicarboxylic acid residues is preferably 30 mol% or more, more preferably 40 mol% or more.

[0024] The polyester may also contain one or more residues selected from hydroxycarboxylic acids or their ester-forming derivatives, and lactones. Examples of hydroxycarboxylic acids include glycolic acid, lactic acid, hydroxypropionic acid, hydroxybutyric acid, hydroxyvaleric acid, hydroxycaproic acid, o-hydroxybenzoic acid, m-hydroxybenzoic acid, p-hydroxybenzoic acid, and 6-hydroxy-2-naphthoic acid. Examples of lactones include caprolactone, valerolactone, propiolactone, undecalactone, and 1,5-oxepan-2-one. Examples of polymers or copolymers containing these residues as structural units include aliphatic polyester resins such as polyglycolic acid, polylactic acid, polyglycolic acid / lactic acid, and polyhydroxybutyric acid / β-hydroxybutyric acid / β-hydroxyvaleric acid.

[0025] Polycarbonates can be obtained by the phosgene method, in which phosgene is blown into a bifunctional phenolic compound in the presence of a caustic alkali and a solvent, or by the transesterification method, in which a bifunctional phenolic compound is transesterified with diethyl carbonate in the presence of a catalyst. Examples of polycarbonates include aromatic homopolycarbonates and aromatic copolycarbonates.

[0026] Examples of bifunctional phenolic compounds include 2,2'-bis(4-hydroxyphenyl)propane, 2,2'-bis(4-hydroxy-3,5-dimethylphenyl)propane, bis(4-hydroxyphenyl)methane, 1,1'-bis(4-hydroxyphenyl)ethane, 2,2'-bis(4-hydroxyphenyl)butane, 2,2'-bis(4-hydroxy-3,5-diphenyl)butane, 2,2'-bis(4-hydroxy-3,5-dipropylphenyl)propane, 1,1'-bis(4-hydroxyphenyl)cyclohexane, 1-phenyl-1,1'-bis(4-hydroxyphenyl)ethane, etc. Two or more of these may be used.

[0027] Examples of polyarylene sulfides include polyphenylene sulfide (PPS), polyphenylene sulfide sulfone, polyphenylene sulfide ketone, random copolymers thereof, block copolymers thereof, etc. Two or more of these may be used.

[0028] Examples of cellulose derivatives include cellulose acetate, cellulose acetate butyrate, ethyl cellulose, etc. Two or more of these may be contained.

[0029] The thermoplastic resin composition may contain reinforcing fibers, non-fibrous inorganic fillers, and the like.

[0030] Examples of reinforcing fibers include glass fibers, milled glass fibers, flat glass fibers, modified cross-section glass fibers, cut glass fibers, flat glass fibers, stainless steel fibers, aluminum fibers, brass fibers, rock wool, PAN (Polyacrylonitrile)-based and pitch-based carbon fibers, carbon nanotubes, carbon nanofibers, calcium carbonate whiskers, wollastonite whiskers, potassium titanate whiskers, barium titanate whiskers, aluminum borate whiskers, silicon nitride whiskers, aramid fibers, alumina fibers, silicon carbide fibers, asbestos fibers, gypsum fibers, ceramic fibers, zirconia fibers, silica fibers, titanium oxide fibers, and silicon carbide fibers, and two or more of these can be used in combination. Among these, glass fibers and carbon fibers are preferred.

[0031] Non-fibrous inorganic fillers include talc, wollastonite, zeolite, sericite, mica, kaolin, clay, pyrophyllite, bentonite, asbestos, silicates such as alumina silicate and hydrotalcite, silicon oxide, glass powder, magnesium oxide, aluminum oxide (alumina), silica (crushed and spherical), quartz, glass beads, glass flakes, crushed and irregularly shaped glass, glass microballoons, molybdenum disulfide, aluminum oxide (crushed), translucent alumina (fibrous, plate-like, flake-like, granular, irregularly shaped, crushed), acid Examples of suitable inorganic fillers include titanium dioxide (crushed), oxides such as zinc oxide (fibrous, plate-like, flaky, granular, irregularly shaped, and crushed), carbonates such as calcium carbonate, magnesium carbonate, and zinc carbonate, sulfates such as calcium sulfate and barium sulfate, hydroxides such as calcium hydroxide, magnesium hydroxide, and aluminum hydroxide, silicon carbide, carbon black and silica, graphite, aluminum nitride, translucent aluminum nitride (fibrous, plate-like, flaky, granular, irregularly shaped, and crushed), calcium polyphosphate, graphite, metal powder, metal flakes, metal ribbons, and metal oxides. Specific examples of metals (metal powders, metal flakes, and metal ribbons) include silver, nickel, copper, zinc, aluminum, stainless steel, iron, brass, chromium, and tin. Other inorganic fillers include carbon powder, graphite, carbon flakes, flaky carbon, fullerenes, and graphene. These may be hollow, and two or more of these inorganic fillers can be used in combination. Of these, calcium carbonate, carbon black, and graphite are preferred.

[0032] Other additives include, for example, silane compounds (epoxy group-containing alkoxysilane compounds such as γ-glycidoxypropyltrimethoxysilane, γ-glycidoxypropyltriethoxysilane, and β-(3,4-epoxycyclohexyl)ethyltrimethoxysilane; mercapto group-containing alkoxysilane compounds such as γ-mercaptopropyltrimethoxysilane and γ-mercaptopropyltriethoxysilane; γ-ureidopropyltriethoxysilane, γ-ureidopropyltrimethoxysilane, and γ-(2-ureidoethyl)aminopropyltrimethoxysilane; ureido group-containing alkoxysilane compounds such as γ-isocyanate propyltriethoxysilane, γ-isocyanate propyltrimethoxysilane, γ-isocyanate propylmethyldimethoxysilane, γ-isocyanate propylmethyldiethoxysilane, γ-isocyanate propylethyldimethoxysilane, γ-isocyanate propylethyldiethoxysilane, γ-isocyanate propyltrichlorosilane, and other alkoxysilane compounds containing an isocyanate group; γ-(2-aminoethyl)aminopropylmethyldimethoxysilane silane, γ-(2-aminoethyl)aminopropyltrimethoxysilane, γ-aminopropyltrimethoxysilane, γ-aminopropyltriethoxysilane, and other amino group-containing alkoxysilane compounds, and hydroxyl group-containing alkoxysilane compounds such as γ-hydroxypropyltrimethoxysilane and γ-hydroxypropyltriethoxysilane), antioxidants and heat stabilizers (hindered phenols, hydroquinones, phosphorus-based, phosphite-based, amine-based, sulfur-based, and their substituted derivatives, etc.), weathering agents (resorcinol-based, salicylate-based, benzophenone-based, and other substituted phenols), benzotriazole-based, benzophenone-based, hindered amine-based, etc.), release agents and lubricants (montanic acid and its metal salts, its esters, its half esters, stearyl alcohol, stearamide, stearates, bisurea, polyethylene wax, etc.), pigments (cadmium sulfide, phthalocyanine, coloring carbon black, etc.), dyes (nigrosine, etc.), crystal nucleating agents (talc, silica, kaolin, clay, etc.), plasticizers (octyl p-oxybenzoate, N-butylbenzenesulfonamide, etc.), antistatic agents (alkyl sulfate-type anionic antistatic agents,Examples of typical additives include quaternary ammonium salt-type cationic antistatic agents, nonionic antistatic agents such as polyoxyethylene sorbitan monostearate, and betaine-type amphoteric antistatic agents; flame retardants (for example, red phosphorus, phosphate esters, melamine cyanurate, hydroxides such as magnesium hydroxide and aluminum hydroxide, ammonium polyphosphate, brominated polystyrene, brominated polyphenylene ether, brominated polycarbonate, brominated epoxy resins, and combinations of these brominated flame retardants with antimony trioxide); heat stabilizers; lubricants such as calcium stearate, aluminum stearate, and lithium stearate; strength improvers such as bisphenol epoxy resins such as bisphenol A, novolac phenolic epoxy resins, and cresol novolac epoxy resins; ultraviolet inhibitors; colorants; flame retardants; and foaming agents.

[0033] Next, the configuration of the learning device 2 will be described with reference to Fig. 2. The learning device 2 is electrically connected to the characteristic prediction device 3. The learning device 2 generates and outputs a trained model by learning using training data.

[0034] 2 is a block diagram showing the configuration of a learning device provided in the characteristic prediction system according to this embodiment. The learning device 2 includes a parameter selection unit 21, a learning unit 22, a control unit 23, and a storage unit 24.

[0035] The parameter selection unit 21 acquires a dataset and selects shape parameters to be used for learning. Here, the dataset is a set of blending conditions for the raw materials constituting the thermoplastic resin composition, manufacturing conditions and test conditions for a molded product used in a dynamic test of the thermoplastic resin composition, and shape parameters that define the shape of measurement data obtained by the dynamic test of the thermoplastic resin composition. In this embodiment, the dataset refers to a set of data that has been pre-processed for the selection process described below or a set of data before the selection process is executed. On the other hand, the learning data is a set of data selected by the parameter selection unit 21 and is a set of data to be read into the learning unit 22. Here, the dataset may be pre-processed to remove outliers in advance in order to improve the prediction accuracy of the trained model. Here, an outlier is a value of a shape parameter that is more than three standard deviations away from the mean value of the dataset for each shape parameter.

[0036] The shape parameter defines the shape of measurement data obtained by dynamic testing of a thermoplastic resin composition. The shape data is at least one of the measurement data and the load and displacement values ​​corresponding to the maximum points, the load and displacement values ​​corresponding to the minimum points, the number of extreme values, and the coefficients and constant terms of the function obtained by polynomial approximation of a specified data interval in the function-processed data obtained by first-order or second-order differentiation of the measurement data. Specifically, the shape parameter is at least one of the following: The maximum load value and / or the corresponding displacement value, the minimum load value and / or the corresponding displacement value, the number of extreme values, and the coefficients and / or constant terms of the function obtained by polynomial approximation of the specified data interval of the measurement data The load value and / or displacement value of the measurement data corresponding to the maximum point of the first derivative value, the load value and / or displacement value of the measurement data corresponding to the minimum point of the first derivative value, the number of extreme values, and the coefficients and / or constant terms of the function obtained by polynomial approximation of a specified data interval, of the function-processed data obtained by first differentiating the measurement data The load value and / or displacement value of the measurement data corresponding to the maximum point of the quadratic differential value, the load value and / or displacement value of the measurement data corresponding to the minimum point of the quadratic differential value, the number of extreme values, and the coefficients and / or constant terms of the function obtained by polynomial approximation of a specified data interval, of the function-processed data obtained by quadratically differentiating the measurement data.

[0037] The parameter selection unit 21 uses a trained model obtained by machine learning using shape parameters of a known thermoplastic resin composition as explanatory variables and the impact absorption properties of the thermoplastic resin composition as objective variables to calculate the importance (variable importance) of the explanatory variables of the trained model and selects shape parameters based on this variable importance. Variable importance indicates the degree of influence that shape parameters have on the objective variable with respect to the properties of the target to be predicted. For example, when using a decision tree or a machine learning algorithm configured using a decision tree, impurity can be used. Variable importance can also be measured using a partial dependency plot (PDP), which marginalizes the model output to quantify the relationship between features and the objective variable, a shapley additive explanation (SHAP), which indicates the contribution of explanatory variables to the objective variable, or permutation feature importance. Note that variable importance is not limited to these, and values ​​calculated by known methods can also be used. The trained model used here may also be an impact absorption prediction model, which will be described later.

[0038] As a feature selection method, sequential feature selection or a similar feature selection algorithm can be applied. Sequential feature selection refers to a method of sequentially adding or deleting features from a set of candidate features, evaluating performance, and determining an optimal feature set. Specifically, using a trained model obtained by machine learning using the shape parameters of a known thermoplastic resin composition as explanatory variables and the impact absorption properties of the thermoplastic resin composition as a target variable, the number of explanatory variables of the trained model is repeatedly added or deleted one by one, and the shape parameters are selected based on the combination of variables that provides the highest prediction accuracy. In this embodiment, the selection of a specific feature selection method is adjusted appropriately depending on the data to be applied and the purpose.

[0039] The learning unit 22 generates a trained model using the data set selected by the parameter selection unit 21, specifically, learning data in which shape parameters are used as explanatory variables and characteristics corresponding to those shape parameters are used as objective variables.

[0040] A known learning method can be adopted for the learning performed by the learning unit 22. Examples of statistical models adopted for learning include a linear regression model, a general additive model, a random forest, rule-fit regression, a gradient boosting tree, an extra tree, a support vector regression, a Gaussian process regression, a k-nearest neighbor regression, a kernel ridge regression, and a neural network. Examples of linear regression models include a Lasso regression model, a Ridge regression model, and an Elastic Net regression model.

[0041] For example, when the learning unit 22 generates a trained model through learning using regularization, the learning unit 22 provides multiple candidate values ​​for hyperparameters of the trained model, performs learning for each of the provided candidate values ​​of the hyperparameters, generates one trained model for one target variable, and stores the trained model in the storage unit 24. The learning unit 22 then calculates prediction errors by cross-validation or holdout validation using the training data for the models obtained by learning using each candidate value, and selects the model that provides the smallest prediction error. The hyperparameters referred to here are parameters that are set in advance by the learning unit 22 for learning, and include, for example, regularization coefficients. In addition, in the case of a trained model using a neural network, the hyperparameters also include, for example, the number of layers in the neural network.

[0042] Returning to FIG. 2, the control unit 23 controls the overall operation of the learning device 2.

[0043] The memory unit 24 stores various programs for operating the learning device 2 and data including various parameters necessary for the operation of the learning device 2. The various programs include a learning data generation program that generates learning data for generating a trained model, and a trained model generation program that generates a trained model by learning using the training data. The various parameters include hyperparameters and parameters acquired by the learning unit 22 through learning. The memory unit 24 also stores data for constituting the learning data (for example, shape parameters indicating the shape of the resin composition and characteristics).

[0044] The storage unit 24 is configured using a ROM (Read Only Memory) in which various programs etc. are pre-installed, a RAM (Random Access Memory) for storing calculation parameters and data for each process, an HDD (Hard Disk Drive), an SSD (Solid State Drive), etc.

[0045] The various programs can also be widely distributed by recording them on computer-readable recording media such as HDDs, flash memories, CD-ROMs, DVD-ROMs, Blu-ray (registered trademark), etc. The communication network referred to here is configured using, for example, an existing public line network, a LAN (Local Area Network), a WAN (Wide Area Network), etc., and may be wired or wireless.

[0046] The learning device 2 having the above functional configuration is a computer configured using one or more pieces of hardware such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field Programmable Gate Array).

[0047] Here, the process of generating a trained model performed by the learning unit 22 will be described with reference to Fig. 3 to Fig. 5. Fig. 3 is a flowchart showing the flow of the process of generating a trained model performed by a learning device according to one embodiment of the present invention. In this embodiment, an example of generating an impact absorption prediction model that predicts impact absorption characteristics from shape parameters will be described.

[0048] First, when the learning device 2 receives an input instruction to generate a trained model, the parameter selection unit 21 acquires shape parameters from the memory unit 24 and executes a process of selecting shape parameters to be learned from the shape parameters (step S11: shape parameter selection step).

[0049] 4 is a flowchart showing the flow of the parameter selection process in the generation process of a trained model. In the selection process of shape parameters, the parameter selection unit 21 acquires a data set by reading it from the storage unit 24 or the like (step S111).

[0050] The parameter selection unit 21 calculates the variable importance of the shape parameters for the acquired data set (step S112).

[0051] Thereafter, the parameter selection unit 21 selects shape parameters with high variable importance (step S113). The parameter selection unit 21 selects, for example, a preset number of shape parameters in descending order of importance. In this way, the parameter selection unit 21 selects shape parameters in accordance with the set conditions.

[0052] Returning to FIG. 3, the learning device 2 generates a shape parameter prediction model for the selected shape parameters (step S12: shape parameter prediction model generation step). FIG. 5 is a flowchart showing the flow of the shape parameter prediction model generation process in the trained model generation process. In the shape parameter prediction model generation process, the learning unit 22 generates a shape parameter prediction model for each selected shape parameter. In the following description, it is assumed that each selected shape parameter is assigned a different number N (N=1, 2, . . . , n-1, n). In the following description, the maximum value of N is defined as n MAX Let's say.

[0053] The learning unit 22 sets the number N of the shape parameter to N=1 (step S121).

[0054] The learning unit 22 performs learning for the Nth shape parameter using the blending conditions of the raw materials constituting the thermoplastic resin composition, the manufacturing conditions and test conditions of the molded product used in the dynamic test of the thermoplastic resin composition as explanatory variables, and the shape parameter that defines the shape of the measurement data obtained by the dynamic test as objective variables, to generate a shape parameter prediction model numbered N (step S122). Any of the above methods is adopted for learning by the learning unit 22.

[0055] After that, the learning unit 22 increases N by 1 (step S123).

[0056] Then, the learning unit 22 determines whether N>n MAX The learning unit 22 determines whether N≦n MAX If it is determined that N>n is satisfied (step S124: No), the process returns to step S122, and a shape parameter prediction model is generated for the increased N-th shape parameter. MAX If it is determined that this is the case (step S124: Yes), the process of generating the shape parameter prediction model ends.

[0057] By the above-described process of generating a shape parameter prediction model, a shape parameter prediction model is generated for each of the shape parameters selected by the parameter selection unit 21.

[0058] 3, the learning device 2 generates an impact absorption prediction model (step S13: impact absorption prediction model generation step). The learning unit 22 generates the impact absorption prediction model by learning using the shape parameters of the thermoplastic resin composition as explanatory variables and the impact absorption properties (e.g., energy absorption (EA) efficiency) of the molded product formed by the shape parameters as objective variables.

[0059] As described above, a shape parameter prediction model and an impact absorption performance prediction model are generated in the learning device 2. The generated trained models are stored in the storage unit 24.

[0060] Here, if a shape parameter prediction model has already been generated, the processes of steps S11 and S12 can be omitted. That is, the shape parameter prediction model generation process of steps S11 and S12 and the shock absorption prediction model generation process of step S13 can be executed independently of each other.

[0061] Next, the characteristic prediction device 3 will be described with reference to Fig. 6. Fig. 6 is a block diagram showing the configuration of the characteristic prediction device included in the characteristic prediction system according to this embodiment. The characteristic prediction device 3 is electrically connected to the learning device 2 and the display device 4. The characteristic prediction device 3 has a shape parameter estimation unit 31, a characteristic prediction unit 32, a control unit 33, and a storage unit 34.

[0062] In this embodiment, the property prediction device 3 predicts the impact absorption property of a thermoplastic resin composition using a shape parameter prediction model and an impact absorption property prediction model.

[0063] The shape parameter estimation unit 31 estimates shape parameters corresponding to the blending conditions, manufacturing conditions, and test conditions using a shape parameter prediction model acquired from the learning device 2 or read from the storage unit 34. The shape parameter estimation unit 31 inputs the blending conditions, manufacturing conditions, and test conditions into the shape parameter prediction model, and acquires the shape parameters output by the shape parameter prediction model.

[0064] The characteristic prediction unit 32 obtains the impact absorption characteristics predicted from the shape parameters using the impact absorption prediction model obtained from the learning device 2 or read out from the storage unit 34. The characteristic prediction unit 32 inputs the shape parameters estimated by the shape parameter estimation unit 31 into the impact absorption prediction model, and obtains the impact absorption characteristics output by the impact absorption prediction model.

[0065] The control unit 33 comprehensively controls the operation of the property prediction device 3. The control unit 33 has a display control unit 331 that displays the processing results (prediction results) of the property prediction unit 32 on the display device 4. The display control unit 331 may also display, in addition to the prediction results, shape parameters, information on the thermoplastic resin composition corresponding to these shape parameters, and the like on the display device 4.

[0066] The storage unit 34 stores various programs for operating the characteristic prediction device 3 and data including various parameters necessary for the operation of the characteristic prediction device 3. The various programs include a characteristic prediction program executed using a trained model. The storage unit 34 is configured using a ROM in which the various programs are pre-installed, and a RAM, HDD, SSD, etc. that store calculation parameters and data for each process.

[0067] The various programs can be recorded on computer-readable recording media such as HDDs, flash memories, CD-ROMs, DVD-ROMs, and Blu-ray (registered trademark) and distributed widely. The characteristic prediction device 3 can also acquire the various programs via a communications network. The communications network referred to here is configured using, for example, an existing public line network, LAN, WAN, etc., and may be wired or wireless.

[0068] The characteristic prediction device 3 having the above-described functional configuration is a computer configured using one or more pieces of hardware such as a CPU, a GPU, an ASIC, an FPGA, and the like.

[0069] The display device 4 is a display made of liquid crystal or organic EL (Electro Luminescence) or the like, and is electrically connected to the characteristic prediction device 3. The display device 4 acquires and displays display data output from the characteristic prediction device 3 under the control of the display control unit 331. The display device 4 may also have an audio output function such as a speaker.

[0070] The input device 5 receives input of various information including information on the compounding conditions, manufacturing conditions, and test conditions of the thermoplastic resin composition, and outputs the received information to the learning device 2 and the property prediction device 3. The input device 5 is configured using a user interface such as a keyboard, mouse, microphone, and touch panel.

[0071] 7 is a flowchart showing the flow of the characteristic prediction process performed by the characteristic prediction device according to one embodiment of the present invention. Upon receiving an input of a characteristic prediction instruction, the characteristic prediction device 3 executes the characteristic estimation process.

[0072] First, shape parameter estimation unit 31 acquires a shape parameter prediction model by reading it from storage unit 34 or outputting it from learning device 2, and estimates shape parameters (step S21: shape parameter estimation step). Shape parameter estimation unit 31 inputs the blending conditions of the thermoplastic resin composition to be estimated, and the manufacturing conditions and test conditions of a molded product to be used in dynamic testing of the thermoplastic resin composition, into the shape parameter prediction model, and acquires the shape parameters of the thermoplastic resin composition to be estimated (step S21: shape parameter estimation step). Shape parameter estimation unit 31 estimates shape parameters by acquiring the shape parameters output by the shape parameter prediction model.

[0073] Thereafter, the characteristic prediction unit 32 acquires the impact absorption prediction model by reading it from the storage unit 34 or outputting it from the learning device 2, and predicts the impact absorption characteristics of the estimated shape parameters (step S22: characteristic prediction step). The characteristic prediction unit 32 predicts the impact absorption characteristics by acquiring the impact absorption characteristics output by the impact absorption prediction model. In this way, the property prediction device 3 predicts the impact absorption property under the conditions based on the blending conditions, manufacturing conditions, and test conditions of the thermoplastic resin composition to be estimated.

[0074] In the embodiment described above, shape parameters corresponding to the blending conditions, manufacturing conditions, and test conditions are estimated using a shape parameter prediction model that uses the blending conditions of the raw materials constituting the thermoplastic resin composition, the manufacturing conditions and test conditions of the molded product used in dynamic testing of the thermoplastic resin composition as explanatory variables, and the shape parameters that define the shape of the measurement data obtained by dynamic testing of the thermoplastic resin composition as response variables.The estimated shape parameters are then input into an impact absorption prediction model that uses the shape parameters of the thermoplastic resin composition as explanatory variables and the impact absorption properties of the thermoplastic resin composition as response variables, and the properties of the shape parameters are predicted.According to this embodiment, even if the test conditions, in addition to the blending conditions and manufacturing conditions, are conditions that result in complex fracture modes, shape parameters appropriate for these conditions are estimated, and the estimated shape parameters are used to predict properties (here, impact absorption properties), so that the properties can be predicted without a decrease in accuracy.

[0075] (Other embodiments) In the above-described embodiment, when generating a trained model, the learning unit 22 may generate a trained model for verification using multiple different statistical models, extract the statistical model with the best prediction accuracy of the predicted value in the trained model for verification, and generate a trained model using the extracted statistical model.

[0076] Specifically, the learning unit 22 uses, for example, as an index for evaluating the accuracy of the trained model for validation (hereinafter referred to as the accuracy evaluation index), the average value of five coefficients of determination calculated using the trained model for validation generated using five training subsets in five-fold cross-validation and the untrained validation subset in each trained model for validation.

[0077] Here, cross-validation is a method for evaluating the predictive performance of a trained model for untrained data, in other words, its generalization performance. In the case of 5-fold cross-validation, the training dataset is divided into five subsets. Four of these subsets are used as training data to generate a trained model for validation, and one is used as an untrained validation subset to verify the accuracy of the generated trained model for validation. This process is repeated five times, with different subsets of the five subsets used as the validation subset, to evaluate the generalization performance for untrained data.

[0078] The learning unit 22 uses, as an accuracy evaluation index for evaluating the accuracy of the trained model for validation, for example, the coefficient of determination R i 2 Average value of R bar is used. R bar =Σ i R i 2 / 5···(1) where the coefficient of determination R i 2 is defined by the following equation (2). i means summation with respect to i. R i 2 =1-Σ j (y j -f i (x j )) 2 / Σ j (y j -y bar ) 2 ···(2) In the above equation (2), y j is the actual value of the characteristic in the validation subset (j = 1 to n, where n is the number of actual values ​​in the validation subset), f i (x j ) is y j The untrained mixture ratio x j The characteristic value (predicted value) output when input to the trained model i is y bar is the measured value y j is the average value of Σ j means the sum over j.

[0079] The process of generating a trained model in this case follows the flow below. First, the learning unit 22 generates a trained model for validation for a subset of the acquired data set using multiple statistical models. The learning unit 22 generates a trained model for validation using multiple different statistical models selected from the multiple statistical models described above, for example. The statistical models selected at this time may be all statistical models that can be used to construct a trained model, multiple preset statistical models, or a statistical model specified by the user.

[0080] The learning unit 22 calculates the coefficient of determination for the generated trained model for validation, and calculates the average value of the coefficients of determination (here, the coefficients of determination R of the five trained models for validation) as an accuracy evaluation index. i 2 Average value of R bar ) is calculated.

[0081] The learning unit 22 selects the statistical model used to generate the trained model for validation with the best accuracy evaluation index. Specifically, the learning unit 22 selects the statistical model used to generate the trained model for validation with the best accuracy evaluation index. bar Calculate the average value R bar Select a statistical model that provides a trained model for validation that satisfies a predetermined condition. For example, as a predetermined condition, the average value R bar is preferably 0.5 or more, more preferably 0.7 or more, and even more preferably 0.9 or more. This threshold value is set in advance and stored in the storage unit 24. The learning unit 22 selects a statistical model that provides a trained model for verification that satisfies the threshold condition. Note that the average value R that satisfies the threshold condition bar If there are multiple coefficients of determination, the statistical model with the largest value can be selected. This allows a trained model based on a statistical model suitable for prediction to be selected. Note that instead of the average value of the coefficients of determination, the coefficients of determination themselves, the mode of the coefficients of determination, etc. may be used. In addition, the learning unit 22 may extract statistical models whose accuracy evaluation indexes satisfy predetermined conditions and display the extraction results on the display device 4, thereby allowing the user to select a statistical model that meets the desired conditions.

[0082] The learning unit 22 then generates a trained model by performing machine learning on the dataset using the selected statistical model. In this manner, the learning unit 22 generates a shape parameter prediction model and / or an impact absorption prediction model. [Example]

[0083] The present invention will be described below with reference to examples, but the present invention is not limited to these examples.

[0084] <Raw materials> (A): Polyamide 6 resin ("Amilan" (registered trademark) manufactured by Toray Industries, Inc.) was used. In this polyamide 6 resin, the viscosity η r is 2.70, melting point is 225°C, and amide group concentration is 6.2 × 10 -2 mmol / g. (B): A layered inorganic compound ("Somasif" (registered trademark) MAE manufactured by Katakura Co-op Agri Co., Ltd.) was used. The cation exchange capacity was 1.3 meq / g. (C): Glass fiber (T-249 manufactured by Nippon Electric Glass Co., Ltd.) was used. (D): Silane coupling agent (Silane coupling agent KBE-9007N manufactured by Shin-Etsu Chemical Co., Ltd.), molecular weight is 247.4 g / mol. (E): An elastomer (Tafmer (registered trademark) M MH7020 manufactured by Mitsui Chemicals, Inc.) was used.

[0085] <Getting the dataset> (1) Compounding conditions of resin composition Each raw material was weighed and pre-blended to satisfy the following compounding conditions relative to a total of 100 parts by weight of the raw materials: 40 to 55 parts by weight of polyamide 6 resin (A), 0 to 4 parts by weight of layered inorganic compound (B), 0 to 45 parts by weight of glass fiber (C), 0 to 0.5 parts by weight of silane coupling agent (D), and 0 to 10 parts by weight of elastomer (E). The extruded gut was pelletized using a twin-screw extruder (TEX30α manufactured by Japan Steel Works) with a cylinder temperature of 250°C and a screw rotation speed of 150 rpm to obtain a resin composition. Using the above manufacturing method, a number of resin compositions satisfying the above blending conditions were prepared.

[0086] (2) Manufacturing conditions for molded products Each of the obtained resin compositions was molded in an injection molding machine under conditions of a cylinder set temperature of 250°C and a mold set temperature of 80°C to produce an energy absorbing member (EA member) 100 having the shape shown in Figures 8 to 10.

[0087] 8 to 10 are diagrams illustrating an energy absorbing member used in a drop weight test. The EA member 100 has a bottom plate portion 101 (length 80 mm × width 80 mm), an energy absorbing portion 102 that stands upright from the bottom plate portion 101 and has a hollow cone shape with a spherical outer surface at its top, a fixing hole 103 provided in the bottom plate portion 103 for fixing the EA member 100 to another member, and a rib 104 provided within the hollow portion of the energy absorbing portion 102. The EA member 100 is designed to receive an impact load, such as a collision load, from the top side of the energy absorbing portion 102. The height H1 of the energy absorbing portion 102 is 117 mm, and the width (length D1 × width D2) of the bottom plate portion 103 is 80 mm × 80 mm.

[0088] (3) Drop weight impact test 11 is a diagram for explaining the drop weight test. The EA member 100 obtained by the method in (2) above was fixed or placed on the base 201 of a drop weight impact tester, and a striker 202 (tip plate), which is a weight of 250 kg, was dropped from a predetermined height to perform the drop weight impact test, and the load and displacement (load-displacement curve) were measured. The EA member 100 used in the test was previously subjected to a water absorption treatment (60°C x 95% RH for 24 hours). The temperature conditions were -30°C, 23°C, and 80°C. For the -30°C measurement, the EA member 100 was first adjusted to -40°C, attached to the testing machine, and then the test was carried out when the target temperature was reached. The drop heights were 0.5 m and 0.75 m.

[0089] (4) Energy absorption (EA) efficiency After obtaining the load (F)-displacement (ΔH) curve from the drop weight test, the load F-ΔH curve was integrated and converted into an energy absorption (EA)-ΔH curve. The volume was then calculated from the cross-sectional area of ​​the EA component corresponding to ΔH, and the volume was multiplied by the specific gravity to calculate the amount of deformed resin (ΔW) corresponding to the displacement. This gave an EA-ΔW curve, and the slope of this curve was calculated using the least squares method to determine the EA efficiency (J / g).

[0090] Fig. 12 is a diagram for explaining how to calculate the deformation amount of an energy absorbing member. Fig. 12 shows an EA member deformed by a drop weight test. In Fig. 12, the top of the energy absorbing part 102 is displaced by a displacement H2 due to the drop weight test. A displaced cross-sectional area S1 is calculated from the cross-sectional area of ​​the EA member 100 before deformation and the displacement H2, and the deformed volume and the deformed amount of resin are calculated based on this cross-sectional area S1.

[0091] Fig. 13 is a diagram showing a curve showing the load-displacement relationship obtained by a drop weight test. Fig. 14 is a diagram obtained by converting the load-displacement curve shown in Fig. 14 into a curve showing the relationship between energy absorption amount and displacement. Fig. 15 is a diagram obtained by converting the curve showing the relationship between energy absorption amount and displacement shown in Fig. 14 into a curve showing the relationship between energy absorption amount and resin deformation amount. Once the load (F)-displacement (ΔH) curve L1, which shows the relationship between the load F and the displacement ΔH shown in FIG. 13, is obtained, the curve L1 is integrated and converted into the energy absorption amount (EA), thereby obtaining the energy absorption amount (EA)-ΔH curve L2 shown in FIG. 14.

[0092] (5) Shape parameters After obtaining a load (F)-displacement (ΔH) curve by the drop weight test, the following shape parameters (a) to (g) were calculated. (a) The linear coefficient of the regression equation obtained by linearly regressing the data in the displacement (ΔH) range of 1 mm to 3 mm using the least squares method. (b) Load value at the maximum load point on the load (F)-displacement (ΔH) curve (c) Displacement value at the maximum load point on the load (F)-displacement (ΔH) curve (d) The linear coefficient of the regression equation obtained by linearly regressing the data in the displacement (ΔH) range of 10 mm to 30 mm using the least squares method. (e) The constant term of the regression equation obtained by linearly regressing the data in the displacement (ΔH) range of 10 mm to 30 mm using the least squares method. (f) Quadratic coefficient of the regression equation obtained by approximating the load (F)-displacement (ΔH) curve to a cubic function using the least squares method (g) The linear coefficient of the regression equation obtained by approximating the load (F)-displacement (ΔH) curve to a cubic function using the least squares method.

[0093] (Example) Datasets (Examples 1 and 2) were prepared by preprocessing to exclude values ​​of the shape parameters (a) to (g) that are more than three times the standard deviation from the mean value of each dataset as outliers, and data sets (Examples 3 and 4) that included these outliers but were not preprocessed were also prepared. Furthermore, using these data sets (Examples 1 and 3), shape parameters were selected based on the variable importance obtained by SHAP of a model obtained by machine learning using the shape parameters (a) to (g) as explanatory variables and EA efficiency as the objective variable. In this example, the criterion for selecting shape parameters was the condition that the absolute value of the SHAP value, which represents the importance of each variable, be 0.2 or more. Furthermore, using the same data sets (Examples 2 and 4) that had been preprocessed / not been preprocessed, a model was obtained by machine learning using the shape parameters (a) to (g) as explanatory variables and EA efficiency as the objective variable, and then the number of explanatory variables was reduced by one variable at a time through sequential feature selection to construct a model, and the coefficient of determination R was used as an accuracy evaluation index. 2 Using the shape parameters selected by each method, we then created a shape parameter prediction model with the blending amounts of each raw material in the resin composition, the temperature conditions in the drop weight test, and the drop weight height as explanatory variables, and each shape parameter as the response variable, as well as an impact absorption prediction model with each shape parameter as the explanatory variable and EA efficiency as the response variable (see Table 1). Using the two prediction models obtained, we predicted the EA efficiency of the training data and the validation data that had not been trained by the trained model. The accuracy of the EA efficiency prediction was evaluated using the coefficient of determination R 2 was calculated for the training data and the validation data, respectively.

[0094] (Comparative Example) As a comparative example, a prediction model was created in which the blending amounts of each raw material in the resin composition, the temperature conditions and the height of the drop weight test were the explanatory variables, and the EA efficiency was the objective variable. Using the obtained prediction model, the EA efficiency of the training data and the validation data was predicted (see Table 1). This prediction model predicts the EA efficiency directly from the blending amounts of the raw materials, the temperature conditions and the height of the drop weight test without estimating the shape parameters. In the comparative example, the EA efficiency of the training data and the validation data that had not been trained by the trained model was predicted in the same manner as in the example, and the coefficient of determination R 2 was calculated.

[0095] [Table 1]

[0096] As shown in Table 1, the coefficient of determination R 2 was higher than the comparative example for both the training data and the validation data. This means that the prediction model according to this example can achieve high prediction accuracy not only for the training data but also for the validation data.

[0097] (Other embodiments) Although the embodiments of the present invention have been described above, the present invention should not be limited to the above-described embodiments. For example, the characteristic prediction device may have the function of a learning unit. In this case, the characteristic prediction device generates a target variable to be predicted and also sequentially updates the trained model. [Explanation of symbols]

[0098] 1. Property prediction system 2 Learning device 3. Characteristics prediction device 4 Display device 5 Input Devices 21 Parameter selection section 22 Learning Department 23, 33 Control section 24, 34 Storage section 31 Shape parameter estimation unit 32 Characteristics prediction section 331 Display control unit

Claims

1. A property prediction program for predicting properties of a thermoplastic resin composition, On the computer, a shape parameter estimation step of reading from memory a shape parameter prediction model in which the blending conditions of raw materials constituting a thermoplastic resin composition, the manufacturing conditions and test conditions of a molded article used in a dynamic test of the thermoplastic resin composition are used as explanatory variables, and the shape parameters that define the shape of measurement data obtained by the dynamic test of the thermoplastic resin composition are used as objective variables, and inputting the blending conditions of the thermoplastic resin composition to be estimated, and the manufacturing conditions and test conditions of the molded article to be used in the dynamic test of the thermoplastic resin composition into the shape parameter prediction model, thereby acquiring the shape parameters of the thermoplastic resin composition to be estimated; a characteristic prediction step of reading from a memory an impact absorption prediction model in which the shape parameters of the thermoplastic resin composition are explanatory variables and the impact absorption characteristics of the thermoplastic resin composition are objective variables, and inputting the shape parameters estimated in the shape parameter estimation step into the impact absorption prediction model to obtain the impact absorption characteristics of the thermoplastic resin composition to be estimated; A characteristic prediction program that executes the above.

2. The dynamic test is any one of a drop weight impact test, a high-speed compression test, a high-speed tensile test, a puncture impact test, a Charpy impact test, and an Izod impact test. The characteristic prediction program according to claim 1 .

3. The shape parameters are at least one of the load value and displacement value corresponding to the maximum point, the load value and displacement value corresponding to the minimum point, the number of extreme values, and the coefficient and constant term of a function obtained by polynomial approximation of a specified data interval in the measurement data and the function-processed data obtained by first-order or second-order differentiation of the measurement data. The characteristic prediction program according to claim 1 .

4. The dataset used for training the shape parameter prediction model and the impact absorption prediction model is a dataset that has been subjected to a process of excluding in advance values ​​that are farther away from the average value of each shape parameter than three times the standard deviation as outliers. The characteristic prediction program according to claim 1 .

5. The shape parameters are selected based on the variable importance calculated using a trained model obtained by machine learning using shape parameters of a known thermoplastic resin composition as explanatory variables and the impact absorption characteristics of the thermoplastic resin composition as a target variable, and the variable importance of the explanatory variables of the trained model is calculated. The characteristic prediction program according to claim 1 .

6. The shape parameters are selected based on the combination of variables that provides the highest accuracy by repeatedly adding or reducing the number of explanatory variables of the trained model, using a trained model obtained by machine learning using the shape parameters of a known thermoplastic resin composition as explanatory variables and the impact absorption characteristics of the thermoplastic resin composition as objective variables. The characteristic prediction program according to claim 1 .

7. On the computer, a shape parameter selection step of reading from a memory a trained model obtained by machine learning, in which shape parameters defining the shape of measurement data obtained by dynamic testing of a known thermoplastic resin composition are used as explanatory variables and the impact absorption characteristics of the thermoplastic resin composition are used as target variables, and selecting a shape parameter from a plurality of shape parameters based on the importance of the explanatory variables in the trained model; a shape parameter prediction model generation step of generating, for each selected shape parameter, a shape parameter prediction model that outputs the shape parameters by machine learning using, as explanatory variables, the blending conditions of raw materials constituting a known thermoplastic resin composition, and the manufacturing conditions and test conditions of a molded product used in a dynamic test of the thermoplastic resin composition, and the shape parameters selected in the previous step of the thermoplastic resin composition as objective variables; an impact absorption prediction model generation step of generating an impact absorption prediction model that outputs the impact absorption characteristics of the thermoplastic resin composition by machine learning using the selected shape parameters as explanatory variables and the impact absorption characteristics of the thermoplastic resin composition as a target variable; A trained model generation program that executes the above.

8. A property prediction system for predicting properties of a thermoplastic resin composition, comprising: a shape parameter estimation unit that inputs the blending conditions of a thermoplastic resin composition to be estimated and the manufacturing conditions and test conditions of a molded product to be used in the dynamic test of the thermoplastic resin composition into a shape parameter prediction model in which the blending conditions of raw materials constituting the thermoplastic resin composition and the manufacturing conditions and test conditions of a molded product to be used in the dynamic test of the thermoplastic resin composition are used as explanatory variables, and the shape parameters that define the shape of measurement data obtained by the dynamic test of the thermoplastic resin composition are used as objective variables, and acquires the shape parameters of the thermoplastic resin composition to be estimated; a property prediction unit that inputs the shape parameters estimated by the shape parameter estimation unit into an impact absorption prediction model having the shape parameters of the thermoplastic resin composition as explanatory variables and the impact absorption properties of the thermoplastic resin composition as objective variables, and acquires the impact absorption properties of the thermoplastic resin composition to be estimated; A characteristic prediction system comprising:

Images