JP2023051839A - Learning model generation program, compounding condition setting program, learning model generation method, compounding condition setting method, and fiber-reinforced polyphenylene sulfide resin composition - Google Patents

Abstract

To provide a learning model generation program capable of efficiently designing a fiber-reinforced polyphenylene sulfide resin composition having desired characteristics, a compounding condition setting program, a learning model generation method, a compounding condition setting method, and a fiber-reinforced polyphenylene sulfide resin composition.SOLUTION: A learning model generation program causes a computer to execute a learning model generation step of generating a learning model by executing machine learning using a compounding ratio of raw materials constituting a known fiber-reinforced polyphenylene sulfide resin composition, and data for learning, in which a compounding condition consisting of a value concerning a shape of a fibrous filler is an explanatory variable and a mechanical property value of the fiber-reinforced polyphenylene sulfide resin composition is an objective variable.SELECTED DRAWING: Figure 1

Description

TECHNICAL FIELD The present invention relates to a learning model generation program, a blending condition setting program, a learning model generation method, a blending condition setting method, and a fiber-reinforced polyphenylene sulfide resin composition.

Polyphenylene sulfide (hereinafter abbreviated as "PPS") resin is an engineering plastic excellent in heat resistance, flame retardancy, chemical resistance, electrical insulation, heat and humidity resistance, mechanical strength and dimensional stability. PPS resin can be molded into various molded products, fibers, films, etc. by various molding methods such as injection molding and extrusion molding, so it is practically used in a wide range of fields such as electrical and electronic parts, mechanical parts and automobile parts. ing. Furthermore, it is known that the mechanical strength and dimensional stability can be improved by adding fibrous fillers such as glass fibers and carbon fibers, and inorganic fillers such as calcium carbonate, talc and mica to PPS resins. It is For example, in Patent Documents 1 to 3, glass fibers and various additives are blended with PPS resin for the purpose of expressing properties such as high tensile strength, high impact resistance, high weld strength, high dimensional accuracy, and low dielectric constant. A fiber reinforced PPS resin composition and a molded article are described.

JP 2014-28917 A JP 2019-195909 A Japanese Patent Application Laid-Open No. 2021-31613

However, in the fiber-reinforced PPS resin compositions described in Patent Documents 1 to 3, the compounding ratio and the like are designed by a person in charge of research and development through trial and error, and it has been desired to make the design more efficient.

The present invention has been made in view of the above, and provides a learning model generation program, a compounding condition setting program, and a learning model generation method that enable efficient design of a fiber-reinforced polyphenylene sulfide resin composition having desired properties. , to provide a method for setting compounding conditions and a fiber-reinforced polyphenylene sulfide resin composition.

In order to solve the above-mentioned problems and achieve the object, the learning model generation program according to the present invention provides a computer with a mixing ratio of raw materials constituting a known fiber-reinforced polyphenylene sulfide resin composition and a fibrous filler. A learning model generation step of generating a learning model by executing machine learning using learning data in which the compounding condition consisting of the value related to the shape is an explanatory variable and the mechanical property value of the fiber-reinforced polyphenylene sulfide resin composition is an objective variable. , is executed.

In the learning model generation program according to the present invention, in the above invention, the learning model generation step applies a part of the learning data to a plurality of mutually different statistical models and performs machine learning respectively to generate a plurality of verifications. an optimal statistical model is selected using the plurality of learning models for verification; and the selected statistical model and the learning data are used to generate the learning model.

In the learning model generation program according to the present invention, in the above invention, the learning model generation step selects the statistical model by evaluating accuracy of the learning model for verification.

In the learning model generation program according to the present invention, in the above invention, the learning model generation step calculates an accuracy evaluation index for the verification learning model, and selects the verification learning model having the best accuracy evaluation index. Select the statistical model used for generation.

In the learning model generation program according to the present invention, in the above invention, the accuracy evaluation index is defined using a coefficient of determination calculated by cross-validation that evaluates the generalization performance of the verification learning model for unlearned data. is the value to be

In the learning model generating program according to the present invention, in the above invention, the learning model generating step generates a plurality of learning models by executing machine learning using mutually different mechanical property values as the objective variables.

In the learning model generation program according to the present invention, in the above invention, the value related to the shape of the fibrous filler includes at least one selected from fiber length, fiber diameter and fiber length distribution of the fibrous filler.

In the above invention, the learning model generation program according to the present invention further includes at least one of a spectral intensity, a process condition value at the time of manufacturing, and a physical quantity different from the mechanical property value as the explanatory variable.

The compounding condition setting program according to the present invention stores in a computer a compounding condition consisting of a compounding ratio of raw materials constituting a known polyphenylene sulfide resin composition and a value related to the shape of a fibrous filler as an explanatory variable, and the fiber-reinforced polyphenylene sulfide. An input step of inputting a plurality of compounding conditions, which are candidates for manufacturing conditions to be set, to a learning model generated by machine learning using the mechanical property values of a resin composition as an objective variable, and outputting the learning model. a setting step of setting, as compatible compounding conditions, compounding conditions that provide mechanical physical property values that match predetermined conditions among the plurality of mechanical property values to be obtained.

The mixing condition setting program according to the present invention is the above invention, wherein the plurality of mixing conditions has a range of values that can be taken for each raw material, and the setting step sets at least a part of the plurality of mixing conditions. is used to set the compatible compounding conditions.

The mixing condition setting program according to the present invention is the above invention, wherein the input step inputs a plurality of mixing conditions to a plurality of learning models generated with mutually different mechanical property values as the objective variables, respectively; The step sets, as the compatible compounding condition, a compounding condition in which a combination of mechanical property values output by the plurality of learning models for a common compounding condition conforms to a predetermined condition.

In the learning model generation method according to the present invention, in the above invention, the computer uses the blending ratio of raw materials constituting a known fiber-reinforced polyphenylene sulfide resin composition and the value related to the shape of the fibrous filler as explanatory variables, the fiber reinforcement A learning model is generated by acquiring from the storage unit learning data having the mechanical property values of the polyphenylene sulfide resin composition as objective variables and executing machine learning.

The method for setting compounding conditions according to the present invention is characterized in that, in the above invention, the computer sets the compounding conditions consisting of values relating to the compounding ratio of the raw materials constituting the known fiber-reinforced polyphenylene sulfide resin composition and the shape of the fibrous filler as explanatory variables. , for a learning model generated by machine learning with the mechanical property value of the fiber-reinforced polyphenylene sulfide resin composition as an objective variable, read out from the storage unit a plurality of compounding conditions that are candidates for the manufacturing conditions to be set, and respectively Among a plurality of mechanical property values input and output by the learning model, a compounding condition that gives a mechanical property value that matches a predetermined condition is set as a compatible compounding condition.

The fiber-reinforced polyphenylene sulfide resin composition according to the present invention comprises raw materials having the compatible compounding conditions according to the present invention.

According to the present invention, a fiber-reinforced polyphenylene sulfide resin composition having desired properties can be efficiently designed.

FIG. 1 is a diagram showing a schematic configuration of a blending condition setting system according to Embodiment 1 of the present invention. FIG. 2 is a block diagram showing the configuration of a learning device included in the compounding condition setting system according to Embodiment 1 of the present invention. FIG. 3 is a block diagram showing the configuration of a mixing condition setting device included in the mixing condition setting system according to Embodiment 1 of the present invention. FIG. 4 is a diagram for explaining the flow of the mixing condition setting process performed by the mixing condition setting system according to the first embodiment of the present invention. FIG. 5 is a flowchart showing an overview of learning processing performed by the learning device according to the first embodiment of the present invention. FIG. 6 is a flow chart showing an outline of a mixing condition setting process performed by the mixing condition setting device according to Embodiment 1 of the present invention. FIG. 7 is a flowchart showing an overview of learning processing performed by the learning device according to the second embodiment of the present invention. FIG. 8 is a flow chart showing an overview of a mixing condition setting process performed by a mixing condition setting device according to Embodiment 3 of the present invention.

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of a blending condition setting system according to the present invention will be described below in detail based on the drawings. It should be noted that the present invention is not limited by this embodiment.

(Embodiment 1)
FIG. 1 is a diagram showing a schematic configuration of a blending condition setting system according to Embodiment 1 of the present invention. The blending condition setting system 1 creates learning data, uses a learning device 2 that creates a learning model learned using the created learning data, and uses the learning model created by the learning device 2 to create a user's desired A mixing condition setting device 3 for setting mixing conditions that provide mechanical property values to be used, a display device 4 for displaying information including the setting results of the mixing condition setting device 3, and an input device 5 are provided.

The compounding conditions set by the compounding condition setting device 3 are values relating to the compounding ratio of raw materials constituting the fiber-reinforced polyphenylene sulfide (PPS) resin composition and the shape of the fibrous filler.
In addition, mechanical properties include tensile strength, tensile elongation at break, tensile elastic modulus, bending strength, bending breaking strain, bending elastic modulus, weld strength, Charpy impact strength, linear expansion coefficient, melt flow rate, spiral flow length, etc. mentioned.

The fiber-reinforced PPS resin composition is composed of a PPS resin, a fibrous filler, a resin other than the PPS resin, and additives other than the fibrous filler. The PPS resin is a polymer having a repeating unit represented by the following formula (a), and a polymer containing 50 mol% or more, preferably 90 mol% or more of a polymer containing a repeating unit represented by the formula (a). is preferred. About less than 30 mol % of the repeating units of the PPS resin may be composed of repeating units having any of the structures represented by the following formulas (b) to (h).

The weight average molecular weight of the PPS resin in the present embodiment is not particularly limited, but from the viewpoint of obtaining excellent mechanical properties, the weight average molecular weight is preferably 20,000 to 150,000, more preferably 25,000 to 130,000, and 30,000. ~90000 is more preferred. Here, when the weight average molecular weight is less than 20,000, the mechanical properties of the PPS resin itself deteriorate. On the other hand, when the weight-average molecular weight exceeds 150,000, the melt viscosity is significantly increased, which tends to be unfavorable in molding.

In addition, the weight average molecular weight in the present invention is a value obtained by converting a measured value measured by gel permeation chromatography (GPC) into polystyrene. For this measurement, for example, a GPC device manufactured by Senshu Scientific Co., Ltd. can be used.

Examples of fibrous fillers constituting the fiber-reinforced PPS resin composition include glass fiber, milled glass fiber, flat glass fiber, modified cross-section glass fiber, glass cut fiber, flat glass fiber, stainless steel fiber, aluminum fiber, brass fiber, and lock. 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 fiber, silicon carbide fiber, asbestos fiber, gypsum fiber, ceramic fiber, zirconia fiber, silica fiber, titanium oxide fiber, silicon carbide fiber, etc., and these fibers can be used in combination of two or more. Among them, glass fiber and carbon fiber are preferred, and glass fiber is particularly preferred. In addition, pretreatment of these fibrous fillers with a coupling agent such as an isocyanate-based compound, an organic silane-based compound, an organic titanate-based compound, an organic borane-based compound, an epoxy compound, or the like results in better mechanical strength. In particular, it is preferable to treat with a sizing agent containing an epoxy group.

The fiber-reinforced PPS resin composition in the present embodiment may be used by blending a resin other than the PPS resin. Resins that can be blended are not particularly limited, but specific examples include polyamides such as nylon 6, nylon 66, nylon 610, nylon 11, nylon 12, aromatic nylon, polyethylene terephthalate, polybutylene terephthalate, and polycyclohexyl. Polyester resins such as dimethylene terephthalate and polynaphthalene terephthalate, polyethylene, polypropylene, polytetrafluoroethylene, olefin copolymers having functional groups such as carboxyl groups, carboxylic acid ester groups, acid anhydride groups and epoxy groups, polyolefin elastomers , polyetherester elastomer, polyetheramide elastomer, polyamideimide, polyacetal, polyimide, polyetherimide, polyethersulfone, modified polyphenylene ether resin, polysulfone resin, polyarylsulfone resin, polyketone resin, polyarylate resin, liquid crystal polymer, Polyether ketone resins, polythioether ketone resins, polyether ether ketone resins, polyamideimide resins, polyethylene tetrafluoride resins, epoxy group-containing polyolefin copolymers, and the like can be mentioned.

The fiber-reinforced PPS resin composition of the present embodiment may be used by blending additives other than fibrous fillers. Examples of additives other than fibrous fillers include non-fibrous inorganic fillers.

Examples of non-fibrous inorganic fillers include talc, wollastonite, zeolite, sericite, mica, kaolin, clay, pyrophyllite, bentonite, asbestos, alumina silicate, silicates such as hydrotalcite, silicon oxide, and glass powder. , magnesium oxide, aluminum oxide (alumina), silica (crushed/spherical), quartz, glass beads, glass flakes, crushed/irregular shaped glass, glass microballoons, molybdenum disulfide, aluminum oxide (crushed), transparent Optical alumina (fibrous, plate-like, scale-like, granular, irregular shape, crushed product), titanium oxide (crushed), zinc oxide (fibrous, plate-like, scale-like, granular, irregular shape, crushed product), etc. oxides, calcium carbonate, carbonates such as magnesium carbonate and zinc carbonate, sulfates such as calcium sulfate and barium sulfate, calcium hydroxide, hydroxides such as magnesium hydroxide and aluminum hydroxide, silicon carbide, carbon Black and silica, graphite, aluminum nitride, translucent aluminum nitride (fibrous, plate-like, scale-like, granular, irregular shape, crushed product), calcium polyphosphate, graphite, metal powder, metal flake, metal ribbon, metal oxide etc. Specific examples of metal species (metal powder, metal flakes, metal ribbons) include silver, nickel, copper, zinc, aluminum, stainless steel, iron, brass, chromium, and tin. Other inorganic fillers include carbon powder, graphite, carbon flakes, scale-like carbon, fullerene, graphene, etc. These may be hollow, and two or more of these inorganic fillers may be used in combination. is also possible. Among them, calcium carbonate, carbon black and graphite are preferred.

Other additives include epoxy compounds such as silane compounds (γ-glycidoxypropyltrimethoxysilane, γ-glycidoxypropyltriethoxysilane, β-(3,4-epoxycyclohexyl)ethyltrimethoxysilane, etc.). Group-containing alkoxysilane compounds, mercapto group-containing alkoxysilane compounds such as γ-mercaptopropyltrimethoxysilane, γ-mercaptopropyltriethoxysilane, γ-ureidopropyltriethoxysilane, γ-ureidopropyltrimethoxysilane, γ-( ureido group-containing alkoxysilane compounds such as 2-ureidoethyl)aminopropyltrimethoxysilane; isocyanate group-containing alkoxysilane compounds such as ethoxysilane, γ-isocyanatopropylethyldimethoxysilane, γ-isocyanatopropylethyldiethoxysilane, γ-isocyanatopropyltrichlorosilane, γ-(2-aminoethyl)aminopropylmethyldimethoxysilane, amino group-containing alkoxysilane compounds such as γ-(2-aminoethyl)aminopropyltrimethoxysilane, γ-aminopropyltrimethoxysilane, γ-aminopropyltriethoxysilane; -hydroxyl group-containing alkoxysilane compounds such as hydroxypropyltriethoxysilane), antioxidants and heat stabilizers (hindered phenol, hydroquinone, phosphorus, phosphite, amine, sulfur and substituted products thereof) etc.), weathering agents (resorcinol-based, salicylate-based, 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, ALERT, bisurea and polyethylene wax, etc.), pigments (cadmium sulfide, phthalocyanine, carbon black for coloring, etc.), dyes (nigrosine, etc.), crystal nucleating agents (talc, silica, kaolin, clay, etc.), plasticizers (p-oxy octyl benzoate, N-butylbenzenesulfonamide, etc.), antistatic agents (alkyl sulfate type anionic antistatic agents, quaternary ammonium salt type cationic antistatic agents, nonionic agents such as polyoxyethylene sorbitan monostearate) Antistatic agents, betaine-based amphoteric antistatic agents, etc.), flame retardants (e.g., red phosphorus, phosphoric esters, melamine cyanurate, magnesium hydroxide, hydroxides such as aluminum hydroxide, ammonium polyphosphate, brominated polystyrene, bromine polyphenylene ether, brominated polycarbonate, brominated epoxy resin, or a combination of these brominated flame retardants and antimony trioxide, etc.), heat stabilizers, lubricants such as calcium stearate, aluminum stearate, lithium stearate, bisphenol Ordinary additives such as strength improvers such as bisphenol epoxy resins such as type A, novolac phenol epoxy resins, and cresol novolac epoxy resins, ultraviolet inhibitors, colorants, flame retardants, and foaming agents can be used.

Next, the configuration of the learning device 2 will be described with reference to FIG. The learning device 2 is electrically connected to the compounding condition setting device 3 . The learning device 2 selectively extracts learning data, generates and outputs a learning model through learning using the extracted learning data.

FIG. 2 is a block diagram showing the configuration of a learning device included in the compounding condition setting system according to Embodiment 1 of the present invention. The learning device 2 has an extraction unit 21 , a learning unit 22 , a control unit 23 and a storage unit 24 .

The extraction unit 21 extracts from the data stored in the storage unit 24 a plurality of raw materials constituting the fiber-reinforced PPS resin composition, the mixing ratio of the raw materials, and the shape of the fibrous filler among the raw materials. Extract data in which compounding conditions and mechanical property values of the fiber-reinforced PPS resin composition produced under the compounding conditions are associated with each other. That is, the extracted data are data of known fiber-reinforced PPS resin compositions whose mechanical property values are known. Note that the extraction unit 21 may extract data according to conditions input via the input device 5 .
The data set extracted by the extraction unit 21 is used as learning data with the compounding condition as the explanatory variable and the mechanical property value as the objective variable.

Here, the compounding ratio of the known fiber-reinforced PPS resin composition refers to the data accumulated in the past, which expresses the composition ratio of the raw materials in the fiber-reinforced PPS resin composition in terms of weight ratio. Including resin ratio and additive ratio accordingly.

Further, the value related to the shape of the fibrous filler refers to the value of the fiber length, fiber diameter, fiber length distribution, etc. of the fibrous filler. contains at least one The fiber length may be treated as number average fiber length, weight average fiber length, or the like. As the fiber length distribution, a value of dispersion and deviation of the fiber lengths, or, when the fiber length of the fibrous filler follows some statistical distribution, a parameter representing the distribution may be used. Furthermore, it may be handled as a frequency distribution belonging to each section divided for each fiber length.

The learning unit 22 performs learning using the learning data to generate a learning model. A known learning method can be adopted for the learning performed by the learning unit 22 . Statistical models employed for learning include, for example, simple linear regression model, Ridge regression, Lasso regression, Elastic Net regression, general additive model, random forest regression, rule-fit regression, gradient boosting tree, extra tree, and support vector regression. , Gaussian process regression, k-nearest neighbor regression, kernel ridge regression, neural network, and the like.

For example, when the learning unit 22 generates a trained model by learning using regularization, the learning unit 22 provides a plurality of hyperparameter candidate values of the learning model, and for each given hyperparameter candidate value: , and one learning model is generated for one objective variable (machine physical property value) and stored in the storage unit 24 . After that, the learning unit 22 calculates the prediction error by cross-validation or holdout verification for the model obtained by learning with each candidate value, using the learning data, and selects the learning model that gives the minimum prediction error. do. Note that the hyperparameters here are parameters set in advance for the learning unit 22 to perform learning, and include, for example, regularization coefficients. Hyperparameters for a learning model using a neural network also include the number of layers of the neural network.

The control unit 23 centrally controls the operation of the learning device 2 .

The storage unit 24 stores various programs for operating the learning device 2 and data including various parameters required for the operation of the learning device 2 . Various programs include a learning data generation program for generating learning data for generating a learning model, and a learning model generation program for generating a learning model by learning using the learning data. The various parameters include hyperparameters, parameters acquired through learning by the learning unit 22, and the like. The storage unit 24 also stores data (for example, the above-described compounding conditions and mechanical property values) for forming learning data.

The storage unit 24 includes a ROM (Read Only Memory) in which various programs are installed in advance, a RAM (Random Access Memory) for storing calculation parameters and data for each process, a HDD (Hard Disk Drive), and an SSD (Solid State Drive). ), etc.

Various programs can be recorded on computer-readable recording media such as HDD, flash memory, CD-ROM, DVD-ROM, and Blu-ray (registered trademark) and widely distributed. The communication network here is configured using, for example, an existing public line network, LAN (Local Area Network), WAN (Wide Area Network), etc., and may be wired or wireless.

The learning device 2 having the above functional configuration uses one or a plurality 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). It is a computer composed of

Next, the mixing condition setting device 3 will be described with reference to FIGS. 3 and 4. FIG. FIG. 3 is a block diagram showing the configuration of a mixing condition setting device included in the mixing condition setting system according to Embodiment 1 of the present invention. The compounding condition setting device 3 is electrically connected to the learning device 2 and the display device 4 . The compounding condition setting device 3 has a setting section 31 , a control section 32 and a storage section 33 .

The blending condition setting device 3 uses the blending conditions of the fiber-reinforced PPS resin composition as setting candidates and the learning model acquired from the learning device 2 to set blending conditions having mechanical property values (predicted values) suitable for the conditions. set.

FIG. 4 is a diagram for explaining the flow of the mixing condition setting process performed by the mixing condition setting system according to the first embodiment of the present invention. The setting unit 31 acquires the learning model 100 from the learning device 2 . Using this learning model 100, the setting unit 31 outputs predicted values OP of the mechanical property values from the blending conditions IP of the raw materials constituting the fiber-reinforced PPS resin composition. Each compounding condition IP (combination conditions 1 to N) is a compounding condition candidate for the fiber-reinforced PPS resin composition. The compounding conditions 1 to N here differ from each other in terms of compounding conditions and/or information (values) regarding the shape of the fibrous filler. The predicted value OP is mechanical property values (predicted values) 1 to N generated from compounding conditions 1 to N. The setting unit 31 extracts a predicted value suitable for the input mechanical property value, and sets the extracted compounding condition candidate as a compounding condition. The compounding condition set at this time may be one of the most suitable compounding conditions among a plurality of compounding conditions whose predicted values satisfy the conditions, or may be a plurality of compounding conditions whose predicted values satisfy the conditions. good too.

Extraction of the optimum predicted value means that the target mechanical property value (predicted value) becomes the most preferable value. In the case of the coefficient of expansion, it is synonymous with minimization. If there is a preferable range, the mechanical property value (predicted value) should fall within that range. The optimal blending condition is the blending condition that gives the most favorable predicted value for the objective variable at this time.

As a specific search method for obtaining the optimum compounding conditions, we comprehensively calculate candidate compounding conditions from the range of possible compounding conditions for each raw material, and then obtain the target mechanical property values from the learning model and each compounding ratio candidate. , and from the prediction result, the predicted value of each mechanical property value selects a compounding condition that satisfies the target range. A random search method is included to obtain formulation conditions that maximize/minimize the predicted value. In addition, when the statistical model adopted to generate the learning model can simultaneously calculate the error and distribution of the predicted values, the confidence interval expressed as the mean of the predicted values ± the predicted error, which takes into account the prediction error of the predicted values, is calculated. A Bayesian optimization method can be used as a search method that repeats "utilization" for selecting the vicinity of the extreme value of accumulated data as a target and "search" for selecting a space with few data points.

The control unit 32 controls the operation of the mixing condition setting device 3 in an integrated manner. The control unit 32 has a display control unit 321 that causes the display device 4 to display the calculation result (setting result) of the setting unit 31 . The display control unit 321 may cause the display device 4 to display information such as the raw material, the raw material, and predicted mechanical property values, in addition to the setting result.

The storage unit 33 stores various programs for operating the mixing condition setting device 3 and data including various parameters required for the operation of the mixing condition setting device 3 . Various programs also include a compounding condition setting program that is executed using a learning model. In addition, the storage unit 33 stores candidates for compounding conditions of raw materials constituting the fiber-reinforced PPS resin composition (hereinafter also simply referred to as compounding condition candidates). This compounding condition candidate may be updated in synchronization with the learning device 2 . The storage unit 33 is configured using a ROM in which various programs and the like are installed in advance, and a RAM, HDD, SSD, and the like for storing calculation parameters, data, and the like for each process.

Various programs can be recorded on computer-readable recording media such as HDD, flash memory, CD-ROM, DVD-ROM, and Blu-ray (registered trademark) and widely distributed. Also, the mixing condition setting device 3 can acquire various programs via a communication network. The communication network here is configured using, for example, an existing public line network, LAN, WAN, etc., and may be wired or wireless.

The compounding condition setting device 3 having the above functional configuration is a computer configured using one or a plurality of hardware such as a CPU, GPU, ASIC, and FPGA.

The display device 4 is a display made of liquid crystal, organic EL (Electro Luminescence), or the like, and is electrically connected to the compounding condition setting device 3 . The display device 4 acquires and displays display data output from the compounding condition setting device 3 under the control of the display control section 321 . Note that the display device 4 may have an audio output function such as a speaker.

The input device 5 receives input of various information including information such as set mechanical property values related to the process of setting the blending conditions, and outputs the received information to the learning device 2 and the blending condition setting device 3 . The input device 5 is configured using a user interface such as a keyboard, mouse, microphone, and touch panel.

Next, the flow of learning processing performed by the learning device 2 will be described with reference to FIG. FIG. 5 is a flowchart showing an overview of learning processing performed by the learning device according to the first embodiment of the present invention. First, the learning device 2 refers to the storage unit 24 and extracts a data set to be used for learning (step S11). Here, the extracting unit 21 extracts a data set consisting of a set of compounding conditions including values relating to the compounding ratio and the shape of the fibrous filler, and mechanical property values associated with the compounding conditions. This data set corresponds to learning data for generating a learning model.

In general, the less biased the distribution of values in one explanatory variable in the dataset, the better. For example, if the blending ratio of each raw material is used as an explanatory variable, the blending ratio of raw materials with low appearance frequency will be 0 in many learning data. become difficult to obtain. If there are many types of resins, fillers, and additives used in the learning data, and there is a high percentage of the blending ratio of 0 in one data, each raw material should be categorized and treated as the blending ratio for each category. is also valid. For example, for multiple types of PPS resins, by focusing on the molecular weight, the size of viscosity, the difference in the post-treatment process during polymerization, etc., and integrating them into several groups, the explanatory variables in each data become 0. The ratio decreases (lower sparsity), and the accuracy of the obtained learning model can be expected to improve.

The learning unit 22 generates a learning model by performing machine learning using the learning data generated in step S11 (step S12).

After that, in the compounding condition setting device 3, compounding conditions are set, which include the optimum compounding ratio of the fiber-reinforced PPS resin composition suitable for the mechanical property values to be set and the values relating to the shape of the fibrous filler. FIG. 6 is a flow chart showing an outline of a mixing condition setting process performed by the mixing condition setting device according to Embodiment 1 of the present invention. The setting unit 31 acquires the learning model from the learning device 2, refers to the storage unit 33, and inputs the compounding condition candidate to the learning model (step S21). When each compounding condition candidate is input to the learning model, a mechanical property value (predicted value) predicted for each is output (see FIG. 4).

The setting unit 31 selects the optimum predicted value from the predicted values of each compounding condition candidate (step S22). In the present embodiment, the setting unit 31 selects the optimum predicted value from the set mechanical property values. It should be noted that the description will be made assuming that one optimum predicted value is selected from among the set conditions (here, desired mechanical property values). All the predicted values that satisfy the condition may be selected, or the top few predicted values that satisfy the set condition may be selected.

After that, the setting unit 31 sets the compounding condition candidate corresponding to the selected predicted value to the optimum compounding condition (step S23).

When the display control unit 321 acquires the setting result from the setting unit 31, the display control unit 321 outputs the setting result to the display device 4 and performs display control to display the setting result on the display device 4. FIG. The display device 4 displays the blending conditions of the raw materials of the optimum fiber-reinforced PPS resin composition that are expected to exhibit the mechanical property values to be set.

A person in charge of research and development and a manufacturer produce a fiber-reinforced PPS resin composition having a desired mechanical property value, or a fiber-reinforced PPS resin composition having a desired mechanical property value by manufacturing a fiber-reinforced PPS resin composition consisting of set compounding conditions. A fiber-reinforced PPS resin composition having mechanical properties can be obtained. The method for producing the fiber-reinforced PPS resin composition is not particularly limited, and it can be produced using a known method. For example, each raw material is supplied to a known melt mixer such as a single-screw or twin-screw extruder, a Banbury mixer, a kneader and a mixing roll, and kneaded at a temperature of 280 to 380°C. be able to. There are no particular restrictions on the order in which the raw materials are mixed. After blending all the raw materials, they are melt-kneaded by the above method. or a method of mixing the rest of the raw materials using a side feeder during melt-kneading with a single-screw or twin-screw extruder after blending some of the raw materials. As for the raw material for the additive in small amounts, it is also possible to knead the other raw material by the above-described method or the like, pelletize it, and then add it before molding and subject it to molding.
The fiber-reinforced PPS resin composition of the present invention thus obtained can be subjected to various molding such as injection molding, extrusion molding, blow molding and transfer molding.

In Embodiment 1 described above, the mechanical property values predicted for each candidate compounding condition of the fiber-reinforced PPS resin composition are generated using a learning model, and the compounding condition candidate with the optimum predicted value is determined. Set the compounding conditions. According to Embodiment 1, a person in charge of research and development can efficiently design a fiber-reinforced polyphenylene sulfide resin composition having desired properties without trial and error.

In the first embodiment, the explanatory variables are the blending ratio and the blending conditions consisting of values related to the shape of the fibrous filler. , may further include at least one physical quantity different from the mechanical property value, and may further include a measurement condition value. In that case, these values are additionally set as compounding conditions.

Spectral spectral intensity indicates the intensity of the spectrum at each wavelength obtained by spectroscopic measurement of the fiber reinforced resin composition, for example, each wavelength region such as infrared light, near infrared light, visible light, ultraviolet light, and X-ray spectroscopic measurement, which is measured using a light source of

The process condition values refer to control parameters for producing a resin composition, and include, for example, production conditions during melt-kneading using a twin-screw extruder and production conditions during injection molding. For example, when using a twin-screw extruder, screw rotation speed, feed amount, barrel temperature, resin pressure, resin temperature during discharge, and the like can be mentioned.

If the effect of process conditions on the target physical properties is assumed, the accuracy of the learning model obtained by using the process condition values as explanatory variables in addition to the blending ratio and the shape of the fibrous filler is expected to improve the accuracy of the learning model. can. For example, when the screw speed and cylinder temperature of a twin-screw extruder change the reactivity of additives and affect mechanical properties, using these as explanatory variables can be expected to improve the accuracy of the learning model. Further, when the target mechanical property values are the physical property values of an injection-molded product and are expected to be affected by molding conditions, the manufacturing conditions during injection molding may be included in the process condition values.

In addition, the physical quantity different from the mechanical property value includes the physical property value other than the target mechanical property value, the physical property value and property inherent to each raw material blended in the fiber-reinforced PPS resin composition, and the value that can be calculated from the blending conditions, and Includes any physical quantity such as molecular weight or boiling point. Here, the physical quantity peculiar to each raw material can be calculated from the amount of specific functional groups and the number of partial structures possessed by each raw material, or from the chemical structure.

Moreover, the measurement condition value refers to a value relating to the preconditions for obtaining physical properties. For example, the measurement condition values include any of measurement temperature, measurement humidity, probe movement speed in strength measurement, presence or absence of notch in impact test, and the like.

It is preferable to appropriately select the explanatory variables used for building the learning model according to the target mechanical property value. For example, when predicting weld strength as a mechanical property value, it is preferable to add the molecular weight or melt viscosity of the fiber-reinforced PPS resin as an explanatory variable because the molecular weight and viscosity of the resin affect the formation of the weld portion during molding. For example, in the case of molecular weight, in addition to entering specific molecular weight values as explanatory variables, predictive accuracy can be improved by classifying molecular weight values into “large”, “medium”, and “small” as explanatory variables. obtain.

In addition, when a contribution derived from the chemical structure of each component such as resins and additives contained in the fiber-reinforced PPS resin composition is assumed, not only the blending ratio of each component but also the chemical structural formula of each component By using as explanatory variables the amount of specific functional groups calculated from the original, the number of partial structures and their related values, or the molecular weight that can be calculated from the chemical structure of each component, or physical values including physical property values such as boiling points. , further improvement in the accuracy of the obtained learning model can be expected.

Further, in Embodiment 1, the range of possible blending ratios for each raw material may be set for the blending ratio of the fiber-reinforced PPS resin composition. For example, when the sum of the compounding ratios represented by the weight ratio of the resin and the fibrous filler in the fiber-reinforced PPS resin composition is 100, the total value X ₁ of the compounding ratio of each raw material, various fibrous inorganic fillers, In addition to the total value X ₂ of the compounding ratio of the material and the total value X ₃ of the compounding ratio of various non-fibrous inorganic fillers, X ₄ to X indicating the presence or absence of various additives such as antioxidants and heat stabilizers. Use ₆ to set the following range. For example, _X1 is set in the range of 20 to 60, _X2 is set in the range of 30 to 50, and _X3 is set in the range of 0 to 25. X ₄ to X ₆ indicating the presence or absence of blending of various additives indicate the presence or absence of blending with "1, 0". For example, one compounding ratio is (X ₁ , X ₂ , X ₃ , X ₄ , X ₅ , X ₆ )=(40, 40, 20, 1, 0, 1). When calculating the predicted value, the chemical formula calculated from the process conditions, other physical property values, and the resin composition so as to match the explanatory variables of the learning model constructed based on the blending ratio and the shape of the fibrous filler. Physical quantities, physical property values, etc. derived from the structure may be added as explanatory variables.

As described above, for the blending ratio of the fiber-reinforced PPS resin composition, by setting the possible blending ratio range for each raw material, an unrealistic blending ratio such as the additive occupying the largest blending ratio is set. restrain from being

(Embodiment 2)
Next, Embodiment 2 will be described. In the second embodiment, the above-described learning model generation processing (see FIG. 5) is different. Note that the processing other than the generation of the learning model is the same as in the first embodiment. In the second embodiment, a verification learning model is generated using a plurality of mutually different statistical models, a statistical model with the best prediction accuracy of a predicted value in the verification learning model is extracted, and the extracted statistical model is used. to generate a learning model.

In the second embodiment, as an index for evaluating the accuracy of the verification learning model (hereinafter referred to as an accuracy evaluation index), the verification learning model generated using five training subsets in the 5-fold cross validation and the average value of the five coefficients of determination calculated using the unlearned validation subset in each learning model for validation.

Here, cross-validation is a technique for evaluating the prediction performance of a learning model for unlearned data, that is, the generalization performance. For 5-fold cross-validation, the training data set is divided into 5 subsets. Four of these subsets are used as learning data to generate a learning model for verification, and one is used as an unlearned verification subset to verify the accuracy of the generated learning model for verification. By repeating this five times with a different subset out of the five subsets as a verification subset, the generalization performance for unlearned data is evaluated.

In the second embodiment, the average value R _bar of the coefficients of determination R _i ² of the verification learning model i (i=1 to 5) is used as an accuracy evaluation index for evaluating the accuracy of the verification learning model.
R _bar =Σ _i R _i ² /5 (1)
Here, the coefficient of determination R _i ² is defined by the following equation (2). Also, Σ _i means the sum regarding 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 measured value of the mechanical property value in the verification subset (j = 1 to n; n is the number of measured values in the verification subset), f _i (x _i ) forms a pair with y _j The mechanical property value (predicted value) output when the unlearned compounding condition x _j is input to the learning model i, y _bar is the average value of the measured values y _i . Also, Σ _j means the sum regarding j.

FIG. 7 is a flowchart showing an overview of learning processing performed by the learning device according to the second embodiment of the present invention.

First, the learning device 2 extracts a data set used for learning by the extraction unit 21 in the same manner as steps S11 and S12 (step S31).

The learning unit 22 uses a plurality of statistical models to generate a learning model for verification for the subset (step S32). The learning unit 22 generates a learning model for verification using, for example, a plurality of mutually different statistical models selected from the plurality of statistical models described above. The statistical model selected at this time may be selected from all statistical models capable of constructing a learning model, may be selected from a plurality of preset statistical models, or may be specified by the user. You may choose a statistical model based on

The learning unit 22 calculates the coefficient of determination for the verification learning model generated in step S32, and uses the average value of the plurality of determination coefficients as an accuracy evaluation index (here, the determination coefficient R _i of the five verification learning models ² average value R _bar ) is calculated (step S33).

The learning unit 22 selects the statistical model used when generating the verification learning model with the best accuracy evaluation index (step S34). Specifically, the learning unit 22 calculates an average value R _bar of the coefficients of determination, and selects a statistical model that provides a learning model for verification whose average value R _bar 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 learning model for verification that satisfies this threshold condition. Note that if there are a plurality of average values R _bar that satisfy the threshold conditions, the statistical model having the maximum value may be selected. As a result, a learning model based on a statistical model suitable for prediction is selected. The coefficient of determination itself, the mode of the coefficient of determination, or the like may be used instead of the average value of the coefficient of determination.
In addition, the learning unit 22 extracts a statistical model whose accuracy evaluation index satisfies a predetermined condition, and displays the extraction result on the display device 4 so that the user can select a statistical model that meets the desired condition. good too.

After that, the learning unit 22 generates a learning model by performing machine learning on the data set extracted in step S31 using the selected statistical model (step S35). The learning unit 22 generates a learning model, for example, in the same manner as in step S12 shown in FIG.

According to the second embodiment, among the learning models generated by a plurality of statistical models, a learning model with high prediction accuracy is selected, and the blending conditions are set by this learning model, so that the desired mechanical property value is obtained. Compounding conditions can be obtained more accurately.

In addition, when dividing the learning data, divide the data near the extrapolation region into a subset for extrapolation data that is preferentially allocated and the remaining subset for learning, and perform 5-fold cross validation in the subset for learning, A coefficient of determination calculated for the extrapolation data subset may be used as the accuracy evaluation index of the learning model. This makes it possible to preferentially select a learning model that is effective in predicting the extrapolation region.

To enable prediction in the extrapolation domain, the measured temperature, which is the measurement condition value, may be included as an explanatory variable. This makes it possible to obtain physical property values under different measurement conditions. For example, by adding the measured temperatures to explanatory variables for accumulated data on weld strength measured at a plurality of temperatures different from each other, it is possible to calculate predicted values even at temperatures outside the measurement range.

(Embodiment 3)
Next, Embodiment 3 will be described. Embodiment 3 differs from the above-described learning model generation processing (see FIG. 5) and blending condition setting processing (see FIG. 6). The configuration of the compounding condition setting system is the same as that of the first embodiment. In the third embodiment, blending conditions suitable for input conditions are set for each mechanical property value generated by a plurality of learning models that output a plurality of different mechanical property values.

The learning unit 22 generates a plurality of learning models having different mechanical property values as objective variables. The learning unit 22 individually executes the learning process shown in FIG. 5 or the learning process shown in FIG. 7 for each objective variable to generate a learning model for each objective variable. At this time, it is preferable that the learning data be common to the learning of each objective variable. In other words, it is preferable to generate a learning model learned by using a common compounding condition associated with each mechanical property value as an explanatory variable.

The blending condition setting device 3 uses the learning model generated by the learning device 2 to set the blending conditions. FIG. 8 is a flow chart showing an overview of the mixing condition setting process performed by the mixing condition setting device according to the third embodiment of the present invention.

The setting unit 31 inputs a common compounding condition candidate to each learning model generated by the learning unit 22 to generate each mechanical property value (predicted value) (step S41). At this time, the predicted value of each objective variable is generated as a set associated with the same compounding condition candidate.

After that, the setting unit 31 selects a combination of predicted values in which each mechanical property value conforms to the setting conditions (step S42). The setting unit 31 selects, for example, a combination in which each predicted value is optimal for each mechanical property value.
Here, the selection of the optimum combination of predicted values depends on the combination of mechanical properties required for individual fiber-reinforced PPS resin compositions. , select the predictors that result in a Pareto-optimal combination.

The setting unit 31 sets the compounding condition candidate corresponding to the selected combination of predicted values to the optimum compounding condition (step S43).

According to the third embodiment, even when obtaining compounding conditions that satisfy desired physical properties for a plurality of mechanical property values, appropriate compounding conditions can be obtained from the prediction results of the learning model with each mechanical property value as the objective variable. can be obtained.

(Other embodiments)
Although the embodiments for carrying out the present invention have been described so far, the present invention should not be limited only to the above-described embodiments. For example, the mixing condition setting device may have the function of the learning section. In this case, the compounding condition setting device successively updates the learning model in addition to generating the objective variable to be set.

1 Blending condition setting system 2 Learning device 3 Blending condition setting device 4 Display device 5 Input device 21 Extractor 22 Learning unit 23, 32 Control unit 24, 33 Storage unit 321 Display control unit

Claims (14)

to the computer,
The mixing ratio of the raw materials constituting the known fiber-reinforced polyphenylene sulfide resin composition and the mixing conditions consisting of values related to the shape of the fibrous filler are used as explanatory variables, and the mechanical properties of the fiber-reinforced polyphenylene sulfide resin composition are used as objective variables. A learning model generation step for generating a learning model by executing machine learning using learning data to
A learning model generation program that runs The learning model generation step includes:
generating a plurality of learning models for verification by applying a portion of the learning data to a plurality of statistical models different from each other and executing machine learning respectively;
Selecting an optimal statistical model using the plurality of learning models for verification,
generating the learning model using the selected statistical model and the learning data;
A learning model generation program according to claim 1. The learning model generation step includes:
selecting the statistical model by evaluating the accuracy of the learning model for validation;
3. The learning model generation program according to claim 2. The learning model generation step includes:
calculating an accuracy evaluation index for the learning model for verification;
selecting the statistical model used to generate the learning model for verification with the best accuracy evaluation index;
A learning model generation program according to claim 3. The accuracy evaluation index is a value defined using a coefficient of determination calculated by cross-validation that evaluates the generalization performance of the verification learning model for unlearned data.
A learning model generation program according to claim 4. The learning model generation step includes:
generating a plurality of learning models by performing machine learning with mutually different mechanical property values as the objective variables;
A learning model generation program according to claim 1. The value related to the shape of the fibrous filler includes at least one selected from fiber length, fiber diameter, and fiber length distribution of the fibrous filler,
A learning model generation program according to claim 1. Further including at least one of a physical quantity different from the spectroscopic spectral intensity, a process condition value at the time of manufacturing, and the mechanical property value as the explanatory variable,
A learning model generation program according to any one of claims 1 to 7. to the computer,
The mixing ratio of the raw materials constituting the known fiber-reinforced polyphenylene sulfide resin composition and the mixing conditions consisting of values related to the shape of the fibrous filler are used as explanatory variables, and the mechanical properties of the fiber-reinforced polyphenylene sulfide resin composition are used as objective variables. an input step of inputting a plurality of compounding conditions, which are candidates for manufacturing conditions to be set, into the learning model generated by machine learning;
a setting step of setting, as a compatible compounding condition, a compounding condition that gives a mechanical physical property value that conforms to a predetermined condition among a plurality of mechanical property values output by the learning model;
Mixing condition setting program to execute For the plurality of blending conditions, a range of values that can be taken for each raw material is set,
The setting step includes:
setting the compatible compounding condition using at least a portion of the plurality of compounding conditions;
The mixing condition setting program according to claim 9. The input step includes:
inputting a plurality of compounding conditions to a plurality of learning models generated with mutually different mechanical property values as the objective variables;
The setting step includes:
setting, as the compatible compounding condition, a compounding condition in which a combination of mechanical property values output by the plurality of learning models conforms to a predetermined condition with respect to a common compounding condition;
The mixing condition setting program according to claim 9. the computer
The mixing ratio of the raw materials constituting the known fiber-reinforced polyphenylene sulfide resin composition and the mixing conditions consisting of values related to the shape of the fibrous filler are used as explanatory variables, and the mechanical properties of the fiber-reinforced polyphenylene sulfide resin composition are used as objective variables. Generate a learning model by acquiring learning data from the storage unit and executing machine learning,
Learning model generation method. the computer
The mixing ratio of the raw materials constituting the known fiber-reinforced polyphenylene sulfide resin composition and the mixing conditions consisting of values related to the shape of the fibrous filler are used as explanatory variables, and the mechanical properties of the fiber-reinforced polyphenylene sulfide resin composition are used as objective variables. read from the storage unit and input a plurality of compounding conditions, which are candidates for the manufacturing conditions to be set, into the learning model generated by machine learning,
setting, as a compatible compounding condition, a compounding condition that gives a mechanical physical property value that conforms to a predetermined condition among a plurality of mechanical property values output by the learning model;
How to set compounding conditions. Made of raw materials having the compatible blending conditions according to any one of claims 9 to 11,
A fiber-reinforced polyphenylene sulfide resin composition.

Images