Machine-learning method and condition search method

Abstract

The present invention efficiently predicts a physical property value of a resin composition with good accuracy by using an appropriate explanatory variable. In order to solve this problem, a machine-learning method comprises: selecting an explanatory variable to be used for machine learning from training data that includes explanatory variables, which indicate at least one among information about raw materials of a resin composition, a blending amount of the raw materials, a manufacturing condition, and a physical property measurement condition, and that includes objective variables which indicate the physical property of the resin composition corresponding to the explanatory variables; and constructing, from the training data, a learning model by using the selected explanatory variable and an objective variable corresponding to the explanatory variable.

Description

Machine learning method and condition exploration method

[0001] The present disclosure relates to a machine learning method and a condition exploration method.

[0002] In recent years, methods for designing resin compositions and optimizing manufacturing conditions using the learning results of machine learning have been proposed. By using such techniques, for example, it becomes possible to predict physical properties according to the blending amount of raw materials, kneading conditions, etc., shorten the material development period, and suppress the development cost.

[0003] For example, in Patent Document 1, by machine learning the raw material blending ratio of a fiber-reinforced polyphenylene sulfide resin composition and the blending conditions of a fibrous filler, etc., a learning model for designing a fiber-reinforced polyphenylene sulfide resin composition having desired properties is provided. It is described. Further, in Patent Document 2, a method for determining manufacturing conditions using a dataset of the impact resistance of a polyarylene sulfide resin composition, blending components, mixing conditions, manufacturing conditions, etc. has been proposed. Patent Document 3 describes a physical property prediction device for a flame-retardant resin composition that accurately predicts the elongation and tensile strength of a resin composition. Patent Document 4 describes a method of learning the properties of a material composition containing a plurality of polymers and a plurality of blending agents for preset properties and predicting the properties of the material composition to be evaluated.

[0004] Japanese Patent Application Laid-Open No. 2023-051839 International Publication No. 2021 / 172127 Japanese Patent Application Laid-Open No. 2024-035451 Japanese Patent Application Laid-Open No. 2017-188032

[0005] However, when performing machine learning using various variables at the time of generating a learning model, variables having a small influence on the predicted value are also used in calculations, so there is a limit to improving the quality of the learning model, and it is difficult to predict physical property values with good accuracy. There is a problem.

[0006] In other words, when a large number of explanatory variables are set during machine learning, explanatory variables that have little influence on the target variable are also used in the machine learning process. This can result in the generation of a learning model in which the predicted value is constant in certain algorithms, or the learning model may become unstable or its interpretability may decrease due to multicollinearity. As a result, the accuracy of predicting the physical properties of resin compositions using the learning model may not be sufficiently good, hindering the efficiency of resin composition design.

[0007] Furthermore, while predicting the physical properties of resin compositions using a learning model can provide predicted values ​​for resin composition properties resulting from raw material combinations that humans could not anticipate, if the learning model becomes complex due to the influence of interactions between explanatory variables, the time required to search for an appropriate raw material combination may increase.

[0008] The technology disclosed herein has been developed in view of the above, and aims to provide a machine learning method and a condition search method that can efficiently predict the physical properties of a resin composition with good accuracy using appropriate explanatory variables.

[0009] According to one aspect of this disclosure, a machine learning method includes selecting explanatory variables to be used for machine learning from training data which includes explanatory variables that indicate at least one of the following: information on the raw materials of the resin composition, the amount of raw materials used, manufacturing conditions, and physical property measurement conditions, and an objective variable that indicates the physical properties of the resin composition corresponding to the explanatory variables; and constructing a learning model using the selected explanatory variables and the objective variable corresponding to the explanatory variables from the training data.

[0010] Furthermore, according to another aspect of the present disclosure, the condition search method includes: selecting explanatory variables to be used for machine learning from learning data which includes explanatory variables that indicate at least one of the following: information on the raw materials of the resin composition, the amount of raw materials used, manufacturing conditions, and physical property measurement conditions, and an objective variable that indicates the physical properties of the resin composition corresponding to the explanatory variables; constructing a learning model using the selected explanatory variables and the objective variable corresponding to the explanatory variables from the learning data; inputting the explanatory variables into the constructed learning model to calculate the objective variable, and searching for a combination of explanatory variables in which the calculated objective variable achieves a predetermined goal.

[0011] Figure 1 is a block diagram showing the configuration of an information processing device according to one embodiment. Figure 2 is a flowchart showing a condition search method according to one embodiment. Figure 3 is a block diagram showing an example of the hardware configuration of an information processing device.

[0012] An embodiment of the present disclosure will be described below with reference to the attached drawings. The embodiment described below is illustrative and should not be interpreted as limiting.

[0013] Figure 1 is a block diagram showing the configuration of an information processing device 100 according to one embodiment. The information processing device 100 shown in Figure 1 includes a learning data storage unit 110, a variable selection unit 120, a machine learning unit 130, a learning model storage unit 140, a condition setting unit 150, a prediction unit 160, and a result output unit 170.

[0014] The learning data storage unit 110 stores learning data for generating a learning model by machine learning. Specifically, the learning data storage unit 110 stores combinations of explanatory variables and target variables as learning data. Here, the explanatory variables include information on the raw materials of the resin composition, the amount of raw materials used, manufacturing conditions, and physical property measurement conditions, while the target variables include physical properties such as mechanical properties, electrical properties, thermal properties, long-term properties, weather resistance, chemical resistance, and sliding properties of the resin composition manufactured under the conditions indicated by each explanatory variable. The learning data may also be data obtained by manufacturing the resin composition under the conditions indicated by each explanatory variable and measuring the physical properties of the resin composition.

[0015] The variable selection unit 120 refers to a portion of the training data stored by the training data storage unit 110 and selects explanatory variables to be used for machine learning from among the explanatory variables of the training data. At this time, the variable selection unit 120 may select different combinations of explanatory variables for each target variable. By having the variable selection unit 120 select the explanatory variables to be used for machine learning, fewer explanatory variables are selected than all the explanatory variables stored as training data in the training data storage unit 110, thereby reducing the number of explanatory variables used in machine learning.

[0016] The variable selection unit 120, for example, if the absolute value of the correlation coefficient between two explanatory variables is large, selects only one explanatory variable to be used as an explanatory variable for machine learning. The variable selection unit 120 may also select explanatory variables to be used for machine learning by using decision tree-based machine learning algorithms, genetic algorithms, linear machine learning algorithms, wrapper methods, etc., either individually or in combination. The variable selection performed by the variable selection unit 120 is also called feature selection, and features may be used as explanatory variables, or features created by combinations of explanatory variables or transformations of explanatory variables may be used.

[0017] The machine learning unit 130 obtains a dataset from the training data stored by the training data storage unit 110, consisting of explanatory variables selected by the variable selection unit 120 and the target variable corresponding to these explanatory variables, and constructs a learning model by performing machine learning. In other words, the machine learning unit 130 generates a learning model that outputs the target variable when explanatory variables are input. Therefore, the machine learning unit 130 generates a learning model that takes different combinations of explanatory variables as input for each target variable.

[0018] The learning model storage unit 140 stores the learning model generated by the machine learning unit 130.

[0019] The condition setting unit 150 sets the conditions for manufacturing an unknown resin composition in order to predict its physical properties. Specifically, the condition setting unit 150 sets various explanatory variables to be used when manufacturing the resin composition, and the target of the objective variable that the physical properties of the manufactured resin composition should satisfy. Then, the condition setting unit 150 changes the set explanatory variables in response to a request from the prediction unit 160.

[0020] The prediction unit 160 uses the learning model stored in the learning model storage unit 140 to predict the physical properties of an unknown resin composition and determines a combination of explanatory variables that satisfies the goal set by the condition setting unit 150. Specifically, the prediction unit 160 determines a combination of explanatory variables that satisfies the goal, for example, by Bayesian optimization.

[0021] In other words, the prediction unit 160 inputs the explanatory variables set by the condition setting unit 150 into the learning model to calculate the target variable, which is a predicted value of the physical properties, and also calculates an acquisition function that shows the probability of obtaining the target variable within the target range for each explanatory variable. Once all candidate combinations of explanatory variables have been set, the prediction unit 160 calculates the target variable and the acquisition function using the learning model, and notifies the result output unit 170 of the combinations of explanatory variables set based on the acquisition function.

[0022] The result output unit 170 outputs a combination of explanatory variables that is highly likely to yield a predicted value that satisfies the target, as a result of the prediction by the prediction unit 160.

[0023] Next, the machine learning method and condition search method using the information processing device 100 configured as described above will be explained with specific examples, referring to the flowchart shown in Figure 2.

[0024] When a learning model is generated by machine learning, it is determined which target variable the learning model will be generated for, and the determined target variable is set in the variable selection unit 120 and the machine learning unit 130 (step S101). The physical properties of the resin composition stored as learning data are used as the target variable. More specifically, examples of physical properties that can be used as target variables include tensile strength, tensile fracture strain, tensile modulus, flexural strength, flexural modulus, Charpy impact strength, Izod impact strength, temperature of deflection under load, density, coefficient of linear expansion, Poisson's ratio, flammability, Rockwell hardness, Shore hardness, dielectric breakdown strength, tracking strength, volume resistivity, thermal conductivity, dielectric constant, dielectric loss tangent, bar flow, melt flow rate (MFR), melt viscosity, shrinkage rate, creep properties, fatigue properties, weather resistance, and moist heat properties.

[0025] Once the objective variable is set, the variable selection unit 120 selects explanatory variables to be used in machine learning of the learning model corresponding to the objective variable (step S102). That is, a combination of explanatory variables to be used in machine learning is selected from the explanatory variables of the learning data. Explanatory variables include information on the raw materials of the resin composition stored as learning data, the mixing ratio of the raw materials, manufacturing conditions, and physical property measurement conditions. More specifically, raw material information that serves as explanatory variables includes melting point, glass transition temperature, melt flow rate, melt viscosity, crystallization temperature, density, particle size, thickness, aspect ratio, fiber length, and fiber diameter. Manufacturing conditions that serve as explanatory variables include the discharge rate of the extruder, cylinder temperature, screw rotation speed, screw pattern, and the input position of each raw material. Physical property measurement conditions that serve as explanatory variables include parameters in various physical property measurements (for example, in tensile properties, the shape of the test piece, tensile speed, test temperature, etc.).

[0026] The selected explanatory variables may be features, or features created by combining or transforming explanatory variables. In selecting explanatory variables, a portion of the training data stored in the training data storage unit 110 is used, and a necessary and sufficient number of explanatory variables are selected by using, for example, a method based on the correlation function between explanatory variables, a decision tree-based machine learning algorithm, a genetic algorithm, a linear machine learning algorithm, or a wrapper method, either alone or in combination.

[0027] When a method based on the correlation function between explanatory variables is used, the correlation coefficient between two explanatory variables is calculated, and the method is repeated until the absolute value of the correlation coefficient falls below a certain threshold, by eliminating one of the two explanatory variables with the highest correlation coefficient. This prevents the contribution of explanatory variables with similar meanings from being underestimated or the learning model from becoming unstable by not selecting them.

[0028] When decision tree-based machine learning algorithms are used, the data is divided into feature nodes within the decision tree used for classification and prediction. After the initial division is performed by the root node using the entire dataset, the data is divided by each node, and this division is repeated until all data is classified into the same class or the division no longer makes sense. Then, explanatory variables are selected based on their importance calculated from the constructed model.

[0029] When genetic algorithms are used, explanatory variables are selected using methods such as GAPLs (Genetic Algorithm-based Partial Least Squares) or GASVR (Genetic Algorithm-based Support Vector Regression). GAPLs is a method that combines a genetic algorithm with partial least squares regression, while GASVR is a method that combines a genetic algorithm with support vector regression.

[0030] When using a wrapper method, explanatory variables are selected while the performance of the learned model is evaluated. That is, explanatory variables are added or removed one by one, and the combination that best performs the learned model is searched for. Examples of wrapper methods include forward selection and backward elimination.

[0031] These methods allow for the selection of explanatory variables used in machine learning based on the target variable, preventing an excessive increase in the number of explanatory variables and improving the efficiency of machine learning. As a result, a high-quality learning model can be obtained, enabling efficient prediction of the physical properties of resin compositions with good accuracy.

[0032] Once explanatory variables are selected using some of the training data, the machine learning unit 130 retrieves all the explanatory variables and target variable from the training data (step S103). That is, a dataset of combinations of the set target variable and the explanatory variables selected for this target variable is obtained from all the training data stored in the training data storage unit 110. Then, machine learning is performed using the obtained dataset to construct a learning model for the set target variable (step S104). The constructed learning model is stored in the learning model storage unit 140.

[0033] Once the learning model is constructed, the values ​​of the explanatory variables that optimize the physical properties of the unknown resin composition (i.e., the objective variable) are searched for through prediction using the learning model. For example, Bayesian optimization may be used for this optimization. Specifically, first, the target that the objective variable should satisfy and the explanatory variables as initial data are set by the condition setting unit 150 (step S105), and notified to the prediction unit 160. The explanatory variables as initial data may be explanatory variables included in the learning data.

[0034] Then, the prediction unit 160 inputs the explanatory variables notified by the condition setting unit 150 into the learning model, and calculates the target variable, which is the predicted value of the physical properties, and also calculates the acquisition function (step S106). This acquisition function indicates the probability of achieving the target value of the target variable with respect to the explanatory variables, or the maximum value of the target variable, and explanatory variables with a high acquisition function are notified to the result output unit 170 as explanatory variable candidates (step S107).

[0035] When a candidate explanatory variable with a high acquisition function is output from the result output unit 170, new data is measured for this combination of candidate explanatory variables, and the measured value of the target variable in the candidate explanatory variable is obtained (step S108). Then, it is determined whether the obtained measured value of the target variable achieves the target (step S109).

[0036] If the measured value of the target variable meets the target (step S109 Yes), it means that a resin composition that achieves the target physical properties can be designed from the combination of candidate explanatory variables, and the process ends. On the other hand, if the target is not achieved (step S109 No), the obtained experimental results, namely the candidate explanatory variables and the measured value of the target variable, are added to the learning data storage unit 110, and the process is repeated from the construction of a learning model by the machine learning unit 130 (step S104).

[0037] As described above, according to this embodiment, explanatory variables to be used in machine learning are selected using a portion of the training data, and a learning model is constructed by machine learning the selected explanatory variables and the target variable from all the training data. Then, explanatory variables that yield a target variable that satisfies the target are searched for while inputting explanatory variables into the constructed learning model and calculating the target variable. For this reason, not all of the numerous explanatory variables included in the training data are used in machine learning, and the efficiency of machine learning can be improved by using appropriate explanatory variables. As a result, a high-quality learning model is obtained, and the physical properties of resin compositions can be predicted efficiently with good accuracy.

[0038] [Explanatory Variables] The explanatory variables in the above embodiment can take on various values. For example, raw materials used in the resin composition include thermoplastic resins, fillers, flame retardants, stabilizers, lubricants, colorants, plasticizers, and nucleating agents. Examples of thermoplastic resins include crystalline thermoplastic resins and amorphous thermoplastic resins. Examples of crystalline thermoplastic resins include polyacetal resin (POM), polybutylene terephthalate resin (PBT), polyethylene terephthalate resin (PET), polyphenylene sulfide resin (PPS), liquid crystal polymer (LCP), and polyamide resin (PA). Examples of amorphous thermoplastic resins include polycarbonate resin (PC), acrylic resins, styrene resins, and cyclic olefin (co)polymers (COP, COC).

[0039] Fillers include inorganic and organic fillers added to resin compositions to improve their mechanical properties and reduce warping. Inorganic fillers include fibrous fillers, plate fillers, and powder / granular fillers. Examples of fibrous fillers include inorganic fibers such as glass fibers, asbestos fibers, carbon fibers, silica fibers, alumina fibers, silica-alumina fibers, aluminum silicate fibers, zirconia fibers, potassium titanate fibers, silicon carbide fibers, and whiskers (whiskers made of silicon carbide, alumina, silicon nitride, etc.). Examples of plate fillers include talc, mica, glass flakes, and graphite. Examples of powder / granular fillers include glass beads, glass powder, milled fibers (e.g., milled glass fibers), and wollastonite. Examples of organic fillers include organic fibers such as aliphatic or aromatic polyamides, aromatic polyesters, fluororesins, acrylic resins such as polyacrylonitrile, and fibers made of rayon.

[0040] Examples of flame retardants include halogen-based flame retardants, phosphorus-based flame retardants, antimony-based flame retardants, nitrogen-based flame retardants, and metal hydrate flame retardants. Fluororesins, used as dripping inhibitors, are also included in the category of flame retardants.

[0041] Examples of halogenated flame retardants include halogenated acrylic resins (halogenated polybenzyl (meth)acrylate resins, brominated polybenzyl (meth)acrylates such as poly(pentabromobenzyl (meth)acrylate), and halogenated benzyl (meth)acrylates such as poly(pentachlorobenzyl (meth)acrylate), either alone or copolymers thereof), halogenated styrene resins (halogenated styrene resins obtained by halogenating styrene resins (such as chlorine treatment, bromine treatment, or chlorinated bromine treatment) (such as brominated polystyrene, chlorinated polyethylene, etc.), either alone or copolymers thereof of halogenated styrene monomers, halogenated polycarbonate resins (halogenated polycarbonates such as brominated polycarbonate and chlorinated polycarbonate, etc.), halogenated epoxy compounds (halogenated epoxy resins such as brominated epoxy resin and chlorinated epoxy resin, halogenated phenoxy resins such as brominated phenoxy resin, etc.), halogenated polyaryl ether compounds (e.g., octa Bis(aryl halogens) ethers such as decabromodiphenyl ether, octa-decachlorodiphenyl ether, etc. (e.g., bis(phenyl halogens) ethers, etc.), halogen-containing polyphenylene oxide resins such as brominated polyphenylene ether, halogenated aromatic imide compounds (e.g., alkylene bisbrominated phthalimide (e.g., C2-6 alkylene bisbrominated phthalimide such as ethylene bisbrominated phthalimide, etc.) and other brominated aromatic imide compounds (e.g., bisimide compounds, etc.), halogenated bisaryl compounds (e.g., bis(C6-10 aryl halogens) such as brominated diphenyl, bis(C6-10 aryl halogens) C1-4 alkanes such as brominated diphenylmethane, halogenated bisphenols such as brominated bisphenol A or their derivatives (e.g., brominated polyesters obtained by polymerizing ethylene oxide adducts of halogenated bisphenols), halogenated alicyclic hydrocarbons (crosslinked cyclic saturated or unsaturated halogenated alicyclic hydrocarbons, e.g., dodecachloropentacyclooctadeca-7,Halogenated polycycloalkadienes such as 1,5-diene, halogenated tri(aryloxy)triazine compounds (such as brominated tri(azoleoxy)triazine compounds such as brominated triphenoxytriazine), etc. are included. Halogen-based flame retardants can be used alone or in combination of two or more.

[0042] As the phosphorus-based flame retardant, as long as it is a compound having a phosphorus atom, it is not particularly limited, and examples thereof include organic phosphorus-based flame retardants and inorganic phosphorus-based flame retardants. Examples of the organic phosphorus-based flame retardant include phosphate esters (such as aromatic phosphate esters such as triphenyl phosphate), phosphate ester amides, phosphonitrile compounds ((poly)phenoxyphosphazene, etc.), organic phosphonic acid compounds (such as phosphonic acid esters such as diphenyl methanephosphonate and diethyl phenylphosphonate), organic phosphinic acid compounds (such as methyl phosphinate, etc.), phosphine oxides (such as triphenylphosphine oxide, tricresylphosphine oxide, etc.).

[0043] Examples of the inorganic phosphorus-based flame retardant include non-condensed or condensed (sub)phosphate salts (such as metal salts such as calcium) such as red phosphorus, orthophosphoric acid, phosphorous acid, hypophosphorous acid, polyphosphoric acid (such as metaphosphoric acid, pyrophosphoric acid, tripolyphosphoric acid, tetraphosphoric acid, etc.), polyphosphorous acid (such as metaphosphorous acid, pyrophosphorous acid, etc.).

[0044] Examples of the metal hydroxide flame retardant include magnesium hydroxide (Mg(OH)₂), aluminum hydroxide (Al(OH)₃), hydrotalcite, calcium aluminate hydrate, calcium hydroxide, barium hydroxide, hard clay, etc.

[0045] Examples of the antimony-based flame retardant include antimony trioxide, antimony pentoxide, sodium antimonate, etc.

[0046] Examples of the nitrogen-based flame retardant include salts of triazine-based compounds and cyanuric acid or isocyanuric acid, double salts of nitrogen compounds containing amino groups and polyphosphoric acid, etc.

[0047] Stabilizers include organic substances added for the purpose of improving the heat resistance, weather resistance, and durability of the resin composition, such as antioxidants, secondary antioxidants, hydrolysis inhibitors, decomposition gas absorbers, and the like.

[0048] Lubricants include fatty acid metal salts, low molecular weight polyolefins, alkylene bis fatty acid amides, partially saponified fatty acid esters, lower alcohol fatty acid esters, polyhydric alcohol fatty acid esters, and the like.

[0049] Colorants include dyes and pigments containing inorganic or organic substances. Examples of inorganic pigments include carbon black, calcium carbonate, magnesium carbonate, titanium oxide, zinc white, zinc sulfide, red lead, yellow lead, cadmium sulfide, ultramarine, and the like. Examples of organic dyes and pigments include azo-based, phthalocyanine-based, quinacridone-based, dioxazine-based, perylene-based, perinone-based, anthraquinone-based, indigo-based, and the like, and one or more of these may be combined.

[0050] Plasticizers include ester-based plasticizers (such as aromatic polycarboxylic acid esters, alicyclic polycarboxylic acid esters, polyesters, sucrose fatty acid esters, etc.), phosphate ester-based plasticizers (such as tributyl phosphate, tri-2-ethylhexyl phosphate, triphenyl phosphate, tricresyl phosphate, etc.), acrylic polymers, and the like.

[0051] Examples of nucleation agents include inorganic nucleation agents and organic nucleation agents. Examples of inorganic nucleating agents include talc, mica, synthetic mica, glass flakes, non-swelling mica, fullerene, carbon nanotubes, carbon black, graphite, metal foil, ceramic beads, clay, sericite, zeolite, bentonite, aluminum hydroxide, dolomite, kaolin, silica, fine silicic acid powder, feldspar powder, potassium titanate, shirasu balloons, calcium carbonate, magnesium carbonate, barium sulfate, calcium oxide, aluminum oxide, titanium oxide, magnesium oxide, aluminum silicate, silicon oxide, magnesium hydroxide, gypsum, novaculite, dawsonite, white clay, glass fibers, carbon fibers, graphite fibers, metal fibers, potassium titanate whiskers, aluminum borate whiskers, magnesium-based whiskers, silicon-based whiskers, wollastonite, sepiolite, slag fibers, zonolite, elestadite, gypsum fibers, silica fibers, silica-alumina fibers, zirconia fibers, boron nitride fibers, silicon nitride fibers, and boron fibers. Examples of organonuclearing agents include amide compounds, fatty acid ester compounds, metal salts of amide compounds, hydrazide compounds, metal salts of organic carboxylic acids, metal salts of organic sulfonic acids, metal salts of phenylphosphonic acids, metal salts of phosphate esters, melamine compounds, uracils, and ureas.

[0052] Raw material information includes the chemical name of each raw material, as well as, for example, for thermoplastic resins, the number-average molecular weight, weight-average molecular weight, molecular weight distribution, melting point, crystallization temperature, glass transition temperature, specific heat, viscosity, end-functional group content, etc., and structural descriptors calculated from the chemical structure of the raw material, or, in the case of polymers, structural descriptors of repeating units. For fillers, the raw material information may also include average values ​​and distributions of fiber length, fiber diameter, particle size, thickness, aspect ratio, etc.

[0053] One of the explanatory variables, the blending amount, indicates the blending amount of each raw material constituting the resin composition. Furthermore, when predicting any resin composition, the blending amount (blending ratio) of each raw material of the desired resin composition can be set in advance. When setting the blending amount of a resin composition, the sum of the mass percentages of each raw material should be set to 100 mass percentages. The raw materials for which the blending amount is set can be set arbitrarily. For example, in the case of a resin composition requiring flame retardancy, the addition or non-addition of a flame retardant can be selected based on the oxygen index of the resin used. For example, a flame retardant may be added to resins with an oxygen index of less than 21, while it may not be necessary to add a flame retardant to resins with an oxygen index of 21 or higher.

[0054] One of the explanatory variables, the manufacturing conditions, refers to the control parameters for manufacturing the resin composition. For example, it includes values ​​for manufacturing conditions during melt mixing using a single-screw or multi-screw extruder, and manufacturing conditions during injection molding. When using an extruder, control parameters include cylinder temperature, screw speed, raw material supply amount, discharge amount, extruder screw pattern, and vent gauge pressure. When using injection molding, control parameters include cylinder temperature, mold temperature, screw speed, injection speed, and holding pressure. By inputting property data before and after mixing by an extruder, and property data before and after injection molding, the influence of the manufacturing conditions on the resin composition can be reflected in the physical properties.

[0055] One of the explanatory variables, the physical property measurement conditions, refers to the measurement condition parameters for various physical properties. For example, for tensile testing, these include test speed, test temperature, and specimen shape; for dielectric breakdown strength, they include environmental conditions such as atmosphere and the solution used, as well as the boosting speed, applied voltage, and specimen shape; and for bar flow, they include the cylinder temperature of the injection molding machine, mold temperature, injection pressure, and specimen thickness. By allowing the selection of data according to the usage environment and required resistance of the resin composition for each measurement condition, it is possible to make highly accurate estimates of market demands for each objective variable.

[0056] [Target Variable] The target variable in the above embodiment can take on various values. The target variable is various physical properties of the resin composition. Examples of various physical properties include mechanical properties, electrical properties, thermal properties, long-term properties, weather resistance, chemical resistance, sliding properties, and other properties of the resin composition or molded articles made from the resin composition. Specifically, mechanical properties include density, tensile strength, tensile fracture strain, tensile modulus, flexural strength, flexural modulus, flexural fracture strain, impact strength, hardness, and Poisson's ratio. Electrical properties include surface resistivity, volume resistivity, arc resistance, tracking resistance, dielectric breakdown strength, dielectric breakdown voltage, relative permittivity, and dielectric loss tangent. Thermal properties include melting point, glass transition temperature, crystallization temperature, ball pressure temperature, thermal conductivity, load deflection temperature, Vicat softening temperature, melt index, melt viscosity, specific heat, linear expansion coefficient, flash temperature, ignition temperature, oxygen index, and flame retardancy. Long-term properties include creep properties, fatigue properties, stress relaxation properties, and high temperature and humidity resistance. Weather resistance includes sunshine weather tests, ozone resistance, radiation resistance, and UV fade tests. Chemical resistance includes resistance to various inorganic and organic chemicals, as well as penetration. Sliding properties include Taber wear, static and dynamic friction coefficients. Moldability includes shrinkage rate, bar flow, mold deposits, and molding cycle time.

[0057] Furthermore, the machine learning method and conditional search method according to the above embodiment can be executed by an information processing device. Figure 3 is a block diagram showing an example of the hardware configuration of an information processing device 100 that executes the machine learning method and conditional search method. As shown in Figure 3, the information processing device 100 has a processor 101, a main memory 102, an auxiliary memory 103, an I / O (Input / Output) interface 104, and a network interface (hereinafter abbreviated as "NW interface") 105.

[0058] The processor 101 includes, for example, a CPU (Central Processing Unit), an FPGA (Field Programmable Gate Array), or a DSP (Digital Signal Processor), and provides overall control of the information processing device 100, as well as executing various arithmetic operations.

[0059] The main memory 102 includes, for example, RAM (Random Access Memory) or ROM (Read Only Memory) and stores information used for arithmetic processing performed by the processor 101.

[0060] The auxiliary storage device 103 includes, for example, an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores various programs and data.

[0061] The I / O interface 104 is an interface that allows the user to input information or output information to the user. The I / O interface 104 may include, for example, a keyboard, a display, a touch panel, a microphone, or a speaker.

[0062] The NW interface 105 is an interface that connects to a network via wired or wireless connection.

[0063] The information processing device 100 receives input such as the target variable to be learned and the goals that the target variable should satisfy via the I / O interface 104 and the NW interface 105. The processor 101 then uses the main memory 102 to execute a program stored in the auxiliary memory 103, thereby selecting explanatory variables to be used in machine learning, constructing a learning model, and performing predictions of physical properties using the learning model.

[0064] Furthermore, the processing performed by the information processing device 100 described above can also be written as a program that can be executed by a computer. In this case, this program can be stored in a computer-readable and non-transitory recording medium and installed on the computer. Examples of such recording media include portable recording media such as CD-ROMs, DVD discs, and USB memory, as well as semiconductor memory such as flash memory.

Claims

1. A machine learning method comprising: selecting explanatory variables to be used for machine learning from training data that includes explanatory variables indicating at least one of the following: information on the raw materials of a resin composition, the amount of raw materials used, manufacturing conditions, and physical property measurement conditions, and a target variable indicating the physical properties of the resin composition corresponding to the explanatory variables; and constructing a learning model using the selected explanatory variables and the target variable corresponding to the explanatory variables from the training data.

2. The machine learning method according to claim 1, wherein the selection involves selecting fewer explanatory variables than all explanatory variables stored as training data, using correlation coefficients between explanatory variables, decision tree-based machine learning algorithms, genetic algorithms, linear machine learning algorithms, or wrapper methods.

3. A conditional search method comprising: selecting explanatory variables to be used in machine learning from training data that includes explanatory variables indicating at least one of the following: information on the raw materials of a resin composition, the amount of raw materials used, manufacturing conditions, and physical property measurement conditions, and an objective variable indicating the physical properties of the resin composition corresponding to the explanatory variables; constructing a learning model using the selected explanatory variables and the objective variable corresponding to the explanatory variables from the training data; inputting the explanatory variables into the constructed learning model to calculate the objective variable, and searching for a combination of explanatory variables in which the calculated objective variable achieves a predetermined goal.

4. The conditional search method according to claim 3, wherein the search includes: calculating an acquisition function that shows the probability of obtaining the optimal value of the target variable for each explanatory variable when inputting explanatory variables into the learning model to calculate the target variable; obtaining a target variable corresponding to the candidate explanatory variables determined based on the calculated acquisition function; and determining whether the obtained target variable achieves a predetermined goal.

Images