Method for determining gate position of injection molded article, device for determining gate position of injection molded article, and program for functioning as method for determining gate position of injection molded article

Abstract

According to the present invention, in order to collect analysis data used for training a neural network regression model, resin flow analysis is performed repeatedly, the neural network regression model is trained using the analysis data that has been collected, and at least one of a vector value of a warpage amount of an injection molded article, a volume shrinkage difference of the injection molded article, and an injection pressure, is output as output information, by inputting nodal coordinates of a minute element of an injection molded article as inference data, nodal coordinates of a gate position, and a set value of resin flow analysis, into a trained analysis surrogate model.

Description

Gate position determination method for injection molded products, gate position determination device for injection molded products, and program for functioning as a gate position determination method for injection molded products

[0001] The present disclosure relates to a gate position determination method for an injection-molded product, a gate position determination device for an injection-molded product, and a program for causing the method to function as a gate position determination method for an injection-molded product.

[0002] Injection molding, a method for forming plastic parts (hereinafter referred to as "molded products") by injecting molten resin material into a mold, is widely used. In injection molding, gates, which are filling holes for injecting resin, are installed. The location and number of these gates significantly affect the filling pressure when filling the mold with resin. A lower filling pressure is generally desirable, and the size of the molding machine is selected based on the predicted filling pressure. Reducing the filling pressure value allows for a smaller molding machine and reduces the cost of the molded product. Furthermore, reducing the filling pressure value and uniforming the pressure distribution reduces residual strain and warpage of the molded product. Furthermore, when glass fibers are used as an additive in the resin material, the orientation of the glass fibers can cause significant warpage. Therefore, it is important to select gate locations that minimize the influence of glass fiber orientation and minimize warpage.

[0003] Traditionally, gate locations have often been determined by the experience and intuition of skilled engineers who design injection molding dies. In recent years, it has become common to determine gate locations by predicting the flow of resin using resin flow analysis, a type of numerical analysis. However, resin flow analysis studies often rely on the experience and intuition of the engineer, making it difficult to determine the optimal gate location.

[0004] Furthermore, there is a prior art method for determining gate positions that combines resin flow analysis and an optimization technique. The optimization technique used is, for example, a general optimization technique such as quadratic programming, which uses a local gradient to search for the minimum and maximum points that provide the optimal solution, or a genetic algorithm, which searches for the optimal solution by repeating the process of reproduction, crossover, and mutation, mimicking the evolutionary process of living organisms (see, for example, Patent Documents 1 and 2).

[0005] JP 2003-80575 A JP 2010-5893 A

[0006] Determining the appropriate gate location to meet the required quality of a molded product (e.g., warpage, sink, filling pressure, weld position, etc.) is generally performed by a skilled designer with ample knowledge and experience, using resin flow analysis. When using resin flow analysis, the calculation time for the analysis is greatly affected by the shape and size of the molded product. For example, for a large molded product such as an air conditioner housing, the calculation time for resin flow analysis can take anywhere from several hours to over a dozen hours. This means that the process of repeated trial and error in resin flow analysis is extremely time-consuming. Furthermore, if the designer's technical level is insufficient, the required quality may not be achieved.

[0007] Furthermore, even when determining gate positions using optimization techniques such as those described in the prior art, it is necessary to repeat the analysis at least several dozen times before obtaining an optimal solution, which poses the problem of taking a long time to determine the gate positions.

[0008] The present disclosure discloses technology for solving the above-mentioned problems, and aims to provide a gate position determination method for injection-molded products, an apparatus for determining gate positions for injection-molded products, and a program for causing the method to function as a gate position determination method for injection-molded products, which can easily obtain appropriate gate positions that meet the quality required of injection-molded products in a short period of time, regardless of the technical level of the designer.

[0009] The gate position determination device for injection-molded products disclosed herein comprises an analytical model generation unit that divides CAD data of the injection-molded product into minute elements and generates an analytical model for performing resin flow analysis; an analysis unit that performs resin flow analysis on the analytical model generated by the analytical model generation unit in accordance with predetermined analytical parameters; a proxy model generation unit that learns, in accordance with a neural network regression model, the relationship between the nodal coordinates of the minute elements of the injection-molded product analyzed by the resin flow analysis, the nodal coordinates of the gate position, and the setting values ​​of the resin flow analysis and the predicted value of the required quality; and a calculation unit that inputs the nodal coordinates of the minute elements of the injection-molded product, the nodal coordinates of the gate position, and the setting values ​​of the resin flow analysis as inference data into the learned analytical proxy model, and outputs at least one of a vector value of the amount of warpage of the injection-molded product, a volumetric shrinkage difference of the injection-molded product, or an injection pressure as output information. The gate position determination method for an injection-molded product disclosed herein comprises: a resin flow analysis process that divides CAD data of the injection-molded product into minute elements, generates an analytical model for performing resin flow analysis, and performs resin flow analysis on the analytical model in accordance with preset analytical parameters; an analytical surrogate model learning process that learns, in accordance with a neural network regression model, the relationship between the nodal coordinates of the minute elements of the injection-molded product analyzed in the resin flow analysis process, the nodal coordinates of the gate position, and the setting values ​​of the resin flow analysis and the predicted value of the required quality; and an analytical surrogate model inference process that inputs the nodal coordinates of the minute elements of the injection-molded product, the nodal coordinates of the gate position, and the setting values ​​of the resin flow analysis as inference data into the learned analytical surrogate model, and outputs at least one of a vector value of the amount of warpage of the injection-molded product, a volumetric shrinkage difference of the injection-molded product, or an injection pressure as output information.Furthermore, the program for functioning as the gate position determination method for injection-molded products disclosed herein causes a computer to execute the steps of repeatedly performing resin flow analysis to collect analysis data to be used for training a neural network regression model, training the neural network regression model using the collected analysis data, and inferring the quality required for the injection-molded product using the trained neural network regression model based on the nodal coordinates of the calculation model and the nodal coordinates of the gate position.

[0010] According to the disclosed gate position determination method for injection-molded products, gate position determination device for injection-molded products, and program for functioning as a gate position determination method for injection-molded products, it is possible to easily obtain appropriate gate positions that meet the quality requirements for injection-molded products in a short period of time, regardless of the technical level of the designer.

[0011] 1 is a block diagram showing the configuration of a gate position determination device for an injection-molded product according to embodiment 1. FIG. 2 is a block diagram showing an example of the hardware configuration of the gate position determination device for an injection-molded product according to embodiment 1. FIG. 3 is a diagram showing an example of an analytical model of an injection-molded product. FIG. 4 is a diagram showing an example of a warpage displacement plot after analysis of an injection-molded product. FIG. 5 is a diagram showing an example of a warpage displacement plot after analysis of an injection-molded product. FIG. 6 is a diagram showing an example of a contact point extracted from a molded product shape divided into minute elements. FIG. 7 is a diagram showing an example of a neural network. FIG. 8 is an image diagram showing an example of supervised learning for realizing the gate position determination method according to embodiment 1. FIG. 9 is a flowchart related to a learning process of a neural network. FIG. 10 is a flowchart showing a series of processes for a trained neural network to obtain an output. FIG. 11 is a flowchart of a method for determining a gate position. FIG. 12 is a flowchart showing a series of learning steps of an analytical surrogate model. FIG. 13 is a flowchart showing a series of inference steps of an analytical surrogate model. FIG. 14 is a schematic diagram showing the structure of an analytical surrogate model. FIG. 15 is a diagram showing an example of an analytical model of an injection-molded product. FIG. 16 is a diagram showing an example of a predicted displacement plot using a surrogate model of an injection-molded product. FIG. 17 is a diagram showing an example of a predicted displacement plot using a surrogate model of an injection-molded product. 1 is a schematic diagram of nodes obtained by dividing a molded product into FEM meshes, which are expressed as an adjacency matrix, and 2 is a schematic diagram showing an analytical surrogate model (neural network regression model) that is employed.

[0012] Embodiment 1. This embodiment relates to a technique for determining a gate position for injecting resin during injection molding. First, the configuration for performing the gate position determination method will be described with reference to FIG. 1. FIG. 1 is a block diagram showing the configuration of a gate position determination device, which is a device required to implement the gate position determination method. In FIG. 1, the gate position determination device 100 includes a communication unit 110, a control processing unit 120, resin flow analysis software 130, an analytical proxy model control unit 140, a display input / output unit 150, and a storage unit 160. The gate position determination device 100 may be a single device, or may be configured as multiple devices or a system connected via a network such as a WAN (WIDE AREA NETWORK) or LAN (LOCAL AREA NETWORK). The gate position determination device 100 may be implemented as a system using distributed computing or cloud computing, or may even be implemented as multiple computer devices.

[0013] The communication unit 110 includes, for example, a communication interface such as a network interface card (NIC) or a direct memory access (DMA) controller. The communication unit 110 can communicate with a cloud system or other computer devices via a network such as a wide area network (WAN) or a local area network (LAN). The control processing unit 120 includes, for example, a processor such as a central processing unit (CPU) or a graphics processing unit (GPU) that executes programs stored in the storage unit 160. The components of the control processing unit 120 may be implemented by hardware such as a field programmable gate array (FPGA), or may be configured by both software and hardware.

[0014] FIG. 2 is a block diagram showing an example of the hardware configuration of the gate position determination device 100. The hardware of the gate position determination device 100 includes a processor 600 and a storage device 601. The processor 600 corresponds to the control processing unit 120, and the storage device 601 corresponds to the storage unit 160. Although not shown, the storage device 601 includes a volatile storage device such as a random access memory and a non-volatile auxiliary storage device such as a flash memory. Alternatively, a hard disk auxiliary storage device may be used instead of the flash memory. The processor 600 executes a program input from the storage device 601. In this case, the program is input to the processor 600 from the auxiliary storage device via the volatile storage device. The processor 600 may output data such as calculation results to the volatile storage device of the storage device 601, or may store the data in the auxiliary storage device via the volatile storage device. The processor 600 then reads and executes the program stored in the storage device 601.

[0015] The resin flow analysis software 130 includes a CAD (Computer-Aided Design) model reading unit 131, an analysis model generation unit 132, an analysis parameter setting unit 133, and an analysis unit 134. The analysis model generation unit 132 reads CAD data of an injection-molded product X created using three-dimensional CAD software (e.g., "NX" by SIEMENS Corporation) through the CAD model reading unit 131, divides the data into minute elements called meshes as shown in FIG. 3, and generates an analysis model for performing a resin flow analysis, which is a numerical analysis. FIG. 3 is a diagram showing an example of an analysis model of the injection-molded product X. In FIG. 3, the three-dimensional shape of the injection-molded product X along the X, Y, and Z axes is shown, with the numerical values ​​in mm. G is also shown as an example of a gate position.

[0016] The analysis model can be created using general-purpose resin flow analysis software, such as "MOLDFLOW" from AUTODESK Corporation, "3D TIMON" from Toray Engineering D Solutions, Inc., or "MOLDEX 3D" from CoreTech System Co., Ltd. The analysis parameter setting unit 133 performs settings for resin flow analysis (e.g., physical properties of the resin material (density, specific heat, thermal conductivity, viscosity coefficient, etc.), resin temperature, injection speed, holding pressure, etc.). Specifically, if the material used is PBT (POLY BUTYLENE TEREPHTALATE) resin, the material injection temperature is set to 250°C, the mold temperature is set to 80°C, the material injection time is set to 1.5 seconds, the initial gate number is set to two, and the coordinates of the gate positions are input.

[0017] The analysis unit 134 performs a resin flow analysis on the analysis model generated by the analysis model generation unit 132 in accordance with the analysis parameters set by the analysis parameter setting unit 133. After the analysis is completed, analysis results (e.g., warpage displacement amounts at each nodal coordinate) are obtained at each nodal coordinate, as shown in FIG. 4 . FIG. 4 shows an example of a warpage displacement plot after analysis of injection-molded product X. In FIG. 4 , the area indicated by G is the gate position, and the area indicated by H is the location of maximum warpage displacement. If the gate position, one of the input parameters, is different, analysis data different from that shown in FIG. 4 can be obtained at each nodal coordinate, as shown in FIG. 5 . In this way, different analysis result information for each input is stored in the storage unit 160. The resin flow analysis can use the finite element method, finite volume method, or boundary element method to analyze the flow, heat, and shrinkage behavior of the resin in a mold during the filling, pressure-holding, and cooling processes.

[0018] The analytical proxy model control unit 140 includes a learning data acquisition unit 141, a proxy model generation unit 142, a proxy model reading unit 143, an inference input data acquisition unit 144, a calculation unit 145, and a proxy model prediction result output unit 146. The analytical proxy model control unit 140 has a function of controlling an analytical proxy model that utilizes a known learning algorithm that simulates the behavior of the resin flow analysis software 130. The proxy model generation unit 142 acquires the analysis results of the resin flow analysis software 130 stored in the memory unit 160 as learning data via the learning data acquisition unit 141. Thereafter, the analytical proxy model is generated using the learning data.

[0019] The input data for the analytical surrogate model are analytical parameters based on the three-dimensional coordinates of the X-, Y-, and Z-axes of nodes extracted from the molded product shape divided into infinitesimal elements as shown in Figure 3, as shown in Figure 6. These include the three-dimensional coordinates of the X-, Y-, and Z-axes of nodes on the molded product shape that can be set as gate positions from the analytical parameter setting unit 133, as well as analytical parameters such as the physical properties of the resin material, the resin temperature, the injection speed, and the holding pressure. Figure 6 shows an example of a contact point extracted from the molded product shape divided into infinitesimal elements, and an example of a gate position G is shown. The three-dimensional coordinates of the X-, Y-, and Z-axes of nodes on the molded product shape that can be set as gate positions can be freely selected from all the nodes shown in Figure 6. However, if, for example, two gate positions are located separately on the front and back sides of the molded product shape (mold), the mold will not function as an injection molding die. Therefore, it is preferable to select them from a single surface so that the mold structure is valid.

[0020] The output data of the analytical surrogate model is a predicted value of the quality required for the molded product analyzed by the resin flow analysis software 130. Specifically, the output data may be the amount of warpage, which indicates deformation of the molded product, or the amount of sink marks, which indicates minute indentations on the design surface. One or more types of output data may be used. The surrogate model reading unit 143 accesses the surrogate model structure information 161 and the surrogate model parameter information 162 stored in the memory unit 160 to read an analytical surrogate model of an algorithm that has previously learned the behavior of the resin flow analysis software 130. The designer also inputs the three-dimensional coordinates of the nodes of the molded product shape divided into minute elements from the analytical model generation unit 132 to the inference input data acquisition unit 144, and also inputs the three-dimensional coordinates of the nodes of the molded product shape that can be set at gate positions, as well as analysis parameters such as the physical properties of the resin material, resin temperature, injection speed, and holding pressure from the analysis parameter setting unit 133. This allows the analytical surrogate model to perform calculations that simulate analysis.

[0021] The analytical simulation parameters may be input manually by the designer through a GUI (Graphical User Interface) of the display input / output unit 150, or may be input automatically using an API (Application Programming Interface) provided in the resin flow analysis software 130. When analytical simulation parameters are input to the loaded analytical proxy model, the calculation unit 145 outputs the results of simulating the resin flow analysis through the proxy model prediction result output unit 146. After the prediction is completed, the prediction results are displayed to the designer through the display input / output unit 150 and are also stored in the storage unit 160.

[0022] The display input / output unit 150 includes a display device such as a liquid crystal display, which can be used by a designer using the gate position determination device 100 to perform settings and operations via a GUI. The storage unit 160 includes, for example, a hard disc drive (HDD), a solid state drive (SSD), a read only memory (ROM), and a random access memory (RAM). The storage unit 160 stores various programs such as firmware and application programs, as well as proxy model structure information 161, proxy model parameter information 162, and molded product information 163.

[0023] The proxy model structure information 161 stores a program describing a learning algorithm, and the proxy model parameter information 162 stores parameters constituting the proxy model. Furthermore, the molded product information 163 stores information such as 3D CAD data of the molded product, the resin material used, and standard molding conditions corresponding to the type of molded product. While the analyst can freely set these standard molding conditions corresponding to the type of molded product, it is desirable to set them within a range that does not cause errors in the analysis, taking into account the actual molding conditions of similar products or the resin temperature recommended by the resin material manufacturer. For example, if the filling time is set to an extremely long value, such as 10 seconds, in injection molding, the molten resin may solidify before filling the mold, resulting in a short shot (poor filling) and an error in the analysis. By implementing the above configuration, the gate position determination method according to this embodiment can be implemented. The settings in the analysis parameter setting unit 133 of the resin flow analysis software 130 affect a series of results, such as the flow, heat, and shrinkage behavior of the resin, analyzed by the analysis unit 134. Furthermore, the "settings in the analysis parameter setting unit 133" and the "series of results, such as resin flow, heat, and shrinkage behavior, analyzed by the analysis unit 134" affect the predicted quality required for the molded product. The surrogate model can capture the characteristics of how these influences affect the molded product. For example, when performing analysis using the resin flow analysis software 130, resin material data containing glass fiber data is used. A large amount of analysis data is collected that predicts the amount of warpage, which indicates the deformation of the molded product, by changing the product shape and gate position. This data is then used to build a surrogate model that can predict the amount of warpage taking into account the effects of changes in fiber orientation due to the product shape and gate position. The required number of training data items is approximately equal to the number of nodes in the molded product (1,024 in the example below). While collecting training data and training the surrogate model takes time, they can be prepared separately from the molded product production design, and can also be automated using an API, etc. Furthermore, quality prediction (inference) using the surrogate model during the production design process can be performed in a short time.The advantage of being able to take into account the content considered in the resin flow analysis described above cannot be obtained through simple machine learning that narrows down input and output data (for example, learning the effect of resin pressure distribution on displacement distribution using a multiple regression model).

[0024] Next, we will explain the neural network. The learning algorithm used in the analytical surrogate model control unit 140 can be a known algorithm used in supervised learning. As an example, we will explain the application of a neural network. Below, we will explain the learning phase and the utilization phase. First, we will explain the method for generating a surrogate model and an overview of the neural network, the supervised learning method, and the learning process for generating a surrogate model in the learning phase. The surrogate model generation unit 142 uses so-called supervised learning, for example, according to a neural network regression model, to learn the relationship between the nodal coordinates of the minute elements of the molded product analyzed by the resin flow analysis software 130 as output, the nodal coordinates of the gate positions, and the predicted value of the required quality relative to the resin flow analysis settings. Here, supervised learning refers to a technique in which pairs of input and output data are provided to a learning device, which learns the characteristics of the learning data and infers the results from the input.

[0025] A neural network consists of an input layer consisting of multiple neurons, an intermediate layer (hidden layer) consisting of multiple neurons, and an output layer consisting of multiple neurons. The intermediate layer may be one layer or two or more layers. For example, in a three-layer neural network as shown in Figure 7, when multiple inputs are input to the input layer (X1 to X3), the values ​​are multiplied by weights W1 (W11 to W16) and input to the intermediate layer (Y1 to Y2), and the result is further multiplied by weights W2 (W21 to W26) and output from the output layer (Z1 to Z3). Neural network weights can be interpreted in the same way as standard weights in the field of injection molding. That is, weights are numerical values ​​expressed as real numbers and are coefficients that determine the level of importance each neuron in the neural network assigns to its input. The output result varies depending on the values ​​of weights W1 and W2.

[0026] FIG. 8 is a conceptual diagram illustrating an example of supervised learning for implementing a gate position determination method. In this embodiment, as shown in FIG. 8, the neural network receives, as input 1, the nodal coordinates of the analysis model, the nodal coordinates of the gate position, and the set values ​​of the resin flow analysis. Input 2 (correct answer) is the resin flow analysis results (e.g., the warpage displacement of the input node, the volumetric shrinkage rate of the input node, the filling time of the input node, and the filling pressure of the input node). Based on the learning data generated based on the combination of inputs 1 and 2, the neural network learns the output resin flow analysis results (e.g., the warpage displacement of the input node, the volumetric shrinkage rate of the input node, the filling time of the input node, and the filling pressure of the input node) through so-called supervised learning. The resin flow analysis results here are, for example, analysis values ​​(warpage displacement, volumetric shrinkage rate, filling time, filling pressure, etc.) at the input nodal coordinates. The warpage displacement represents how the molded product deforms. The volumetric shrinkage rate represents the degree of local shrinkage of the molded product. Additionally, the filling time indicates how the resin is filled into the mold. Here, filling time refers to the filling time of each node (the time it takes for the resin to reach the node). If the filling time of each node is used as the output of the neural network, it becomes possible to determine how many seconds it will take for the resin to reach each input node without running an analysis. Furthermore, the filling pressure indicates the pressure value immediately after the resin has filled the mold.

[0027] That is, the neural network learns by inputting input 1 to the input layer and adjusting weights W1 and W2 so that the result output from the output layer approaches input 2 (correct answer). The surrogate model generation unit 142 generates a trained model by performing the above-mentioned training. The storage unit 160 stores the trained model output from the surrogate model generation unit 142.

[0028] Next, the learning process by which the surrogate model generation unit 142 generates a surrogate model will be described with reference to FIG. 9 . FIG. 9 is a flowchart illustrating the neural network learning process. In step S1, the learning data acquisition unit 141 acquires input 1 and input 2 (correct answer). Note that while input 1 and input 2 (correct answer) are acquired simultaneously, it is sufficient that input 1 and input 2 (correct answer) are associated and input, and the data for input 1 and input 2 (correct answer) may be acquired at different times. Next, in step S2, the surrogate model generation unit 142 learns the output by so-called supervised learning in accordance with learning data created based on the combination of input 1 and input 2 (correct answer) acquired by the learning data acquisition unit 141, and generates a trained model. Next, in step S3, the storage unit 160 stores the trained model generated by the surrogate model generation unit 142.

[0029] Next, regarding the utilization phase, a method for obtaining inference results using a trained surrogate model and a method for using this to determine a gate position will be described. First, a case in which a trained surrogate model is utilized will be described. The inference input data acquisition unit 144 acquires input 1. The calculation unit 145 infers an output obtained using the trained model. That is, by inputting input 1 acquired by the inference input data acquisition unit 144 to this trained surrogate model, an output inferred from input 1 can be output. Note that, although the first embodiment has been described using a surrogate model trained with analysis data of molded product X, it is also possible to infer an output based on a surrogate model trained with analysis data of a molded product other than molded product X (a molded product having a different shape and resin material from molded product X).

[0030] Next, the processing by the analytical proxy model control unit 140 to obtain an output will be described with reference to FIG. 10 . FIG. 10 is a flowchart showing a series of processing steps by which a trained neural network obtains an output. In step S11, the inference input data acquisition unit 144 acquires input 1. In step S12, the calculation unit 145 inputs input 1 to the trained model read from the storage unit 160 and obtains an output. In step S13, the calculation unit 145 outputs the output obtained by the trained model to the display input / output unit 150. Note that, in the first embodiment, a case has been described in which supervised learning is applied to the learning algorithm used by the proxy model generation unit 142, but this is not limited to this. As for the learning algorithm, in addition to supervised learning, reinforcement learning, unsupervised learning, semi-supervised learning, or the like can also be applied.

[0031] The surrogate model generation unit 142 may also learn the output according to training data created for multiple molded products. The surrogate model generation unit 142 may acquire training data for multiple molded products molded with the same resin material, or may learn the output using training data collected from multiple molded products molded with different resin materials. Furthermore, a trained surrogate model that has learned the output for a certain molded product may be applied to another molded product, and the output may be retrained and updated for the other molded product. The learning algorithm used by the surrogate model generation unit 142 may also be deep learning, which learns to extract feature quantities themselves. Furthermore, machine learning may be performed according to other known methods, such as genetic programming, functional logic programming, and support vector machines.

[0032] Next, a specific method for determining a gate position will be described. A series of processes for determining a gate position in this embodiment will be described with reference to FIG. 11. FIG. 11 is a flowchart showing an example of a method for determining a gate position. Details of S100 in FIG. 11 are shown in FIG. 12 as a series of flowcharts of the learning process of the analytical surrogate model. Details of S200 in FIG. 12 are shown in FIG. 13 as an example of a series of flowcharts of the inference process of the analytical surrogate model. Below, a specific description will be given of a gate position determination method for minimizing warpage of a molded product with two gate positions (generally called a two-point gate) when the required quality when determining a gate position is warpage.

[0033] To determine the gate position, the analytical surrogate model learning process (S100) is initiated. First, the gate position determination device 100 is started, and then the general-purpose resin flow analysis software 130 is started. Then, in S101, the necessary settings for resin flow analysis are made. Three-dimensional CAD data for learning data is read from the molded product information 163 via the CAD model reading unit 131 of the resin flow analysis software 130. Next, the read CAD data is converted into an analytical model such as that shown in FIG. 3 by the analytical model generation unit 132. The analytical model is a divided model, in which multiple infinitesimal elements are joined at points called nodes. The mesh division method is performed using an algorithm adopted in the general-purpose resin flow analysis software 130, but other methods such as the Delaunay triangulation method, the advancing front method, or even the voxel method may also be used.

[0034] After the analysis model is created, the analysis parameter setting unit 133 sets the resin flow analysis parameters (e.g., the physical properties of the resin material, the resin temperature, the injection speed, the holding pressure, etc.). In the first embodiment, since the goal is to minimize warpage, the resin flow analysis is performed using one pattern of standard molding conditions corresponding to the type of molded product stored in the molded product information 163. These standard molding conditions can be freely set by the analyst, but it is desirable to set them within a range that does not cause errors in the analysis, taking into account the actual molding conditions of similar products and the resin temperature recommended by the resin material manufacturer. For example, if an extremely long injection molding time is set, such as a filling time of 10 seconds, the molten resin will solidify before filling the mold, resulting in a short shot (poor filling) and an error in the analysis.

[0035] 6 extracted from the molded product divided into meshes, the IDs of the nodes that can be selected as gates are selected and stored in the molded product information 163. The three-dimensional coordinates of the X-axis, Y-axis, and Z-axis of the nodes of the molded product shape that can be set as gate positions may be freely selected from all the nodes shown in Fig. 6, but since an injection molding die will not be valid if, for example, two gate positions are set separately on the front and back surfaces of the molded product shape (mold), it is desirable to select from a single surface so that the die structure will be valid.

[0036] Next, in S102, the learning analysis data collection program stored in the storage unit 160 is started, and the number of analysis repetitions and the algorithm for the repetitive analysis are set as the initial settings of the program. In the first embodiment, the repetition setting is 100 times, and the algorithm for the repetitive analysis is set randomly. Note that the repetition setting may be set arbitrarily, and the algorithm for the repetitive analysis may be a known optimization algorithm such as Bayesian optimization, particle swarm optimization, or a genetic algorithm.

[0037] Next, in S103, analysis data is collected. The learning analysis data collection program first creates a combination table of gate IDs to be set as gate positions in the analysis model from the IDs of nodes selectable as gates stored in the molded product information 163. In this embodiment, for example, the gate ID combination table is created by randomly selecting two gate IDs from the IDs of nodes selectable as gates, and repeating this process 10,000 times to create a combination table of 10,000 gate IDs. Alternatively, two gate IDs may be selected randomly with a constraint that the two gate positions are at least 30 mm apart. Furthermore, the first gate ID (gate position) of the two gates may be fixed, and the second gate ID (gate position) may be selected randomly to create a combination table of gate IDs.

[0038] Next, the learning analysis data collection program acquires gate IDs to be set as gate positions in the analysis model from the gate ID combination table in order to set two gate positions in accordance with the algorithm setting for the repetitive analysis. In embodiment 1, the algorithm for the repetitive analysis is set randomly, so the acquired gate IDs are selected at random. The learning analysis data collection program sets gate positions in the analysis model based on the acquired gate IDs. After that, a resin flow analysis is performed on the analysis model for which the settings have been completed.

[0039] When the resin flow analysis is completed, the warpage displacement at each node, which is the analysis result, can be obtained for each gate position, which is the analysis setting value, as shown in Figures 4 and 5. For example, when the analysis is performed using the gate position shown in Figure 4, the maximum warpage displacement is 0.40 mm, resulting in a warpage displacement that causes the molded product to deform diagonally. Furthermore, when the analysis is performed using the gate position shown in Figure 5, the maximum warpage displacement is 0.20 mm, which is smaller than the warpage displacement in Figure 4. In this embodiment, the warpage displacement of the molded product is obtained from the resin flow analysis settings (physical properties of the resin material, resin temperature, injection speed, holding pressure, etc.) and the coordinates of all the nodes of the molded product and the coordinates of the nodes of the gate positions (three-dimensional coordinates of the X-axis, Y-axis, and Z-axis), and is stored in the molded product information 163.

[0040] Once the analysis is completed for the preset number of repetitions, an end determination is made in S104. There are two end determinations: the first is that the analysis is completed for the preset number of repetitions, and the second is that the analysis of all molded products prepared as learning data is completed. In this embodiment, 100 types of molded products are prepared as learning data, and S101 to S104 are repeated until repeated analysis is completed for all molded products. Furthermore, once data collection is completed, the learning analysis data collection program is terminated. Then, in S105, learning processing is performed. In this embodiment, the deep learning model in which the neural network described previously is superimposed is used as the analytical surrogate model.

[0041] Figure 14 is a schematic diagram showing the analytical surrogate model used. The analytical surrogate model receives three types of input data: the node coordinates of the mesh of the molded product, the node coordinates of the gate position, and the set values ​​for the resin flow analysis. The analytical surrogate model receives one type of output data: the amount of warpage of the molded product. The analytical surrogate model is structured as follows: a multilayer perceptron layer, which is a multilayer neural network; a coupling layer that combines the outputs of each layer; a feature extraction layer; and finally an output layer.

[0042] This section explains the case where the molded part mesh has 1,024 nodes. The input size of the molded part mesh nodes is a (1,024 x 3) matrix because the nodes exist in three-dimensional space along the X, Y, and Z axes. The input size of the gate position nodes is also expressed as a (1,024 x 3) matrix. The values ​​of this matrix are calculated based on the node coordinates of the molded part mesh. The values ​​in the two rows corresponding to the gate position are set to 0, and the values ​​in the other rows represent the distance (coordinate difference) between the gate position and each node coordinate of the molded part mesh. The input size of the resin flow analysis setting values ​​can be expressed as an (N x 1) matrix. N is a positive integer, and can be set to the number of resin flow analysis setting values ​​to be considered in the analytical proxy model being constructed. For example, if you want to consider the effect of mold temperature, you would use a (1 x 1) matrix. If you want to consider mold temperature on both the fixed and movable sides of the mold, you would use a (2 x 1) matrix.

[0043] First, the node coordinates of the mesh of the molded product are input to input layer 1 of the analytical surrogate model, and after passing through a multilayer perceptron, the size becomes (1024 x M). M is a positive integer that varies depending on the dimensionality of the output space of the multilayer perceptron. For example, if the dimensionality of the output space of the multilayer perceptron is 32, M is also 32. Similarly, when the node coordinates of the gate position are input to input layer 2, they are passed through a multilayer perceptron and become (1024 x M) in size. Next, in combination layer 1, the two (1024 x M) matrices output from each multilayer perceptron are combined to output a matrix of size (1024 x 2M). This output is input to the next multilayer perceptron, again obtaining an output of size (1024 x M).

[0044] Next, the resin flow analysis setting values ​​are input to the multilayer perceptron via input layer 3, which outputs an (N x M) matrix. To combine the outputs in combining layer 2, the (N x M) matrix is ​​tiling-processed to convert it into a (1024 x M) matrix. In combining layer 2, a (1024 x M) matrix and a (1024 x M) matrix are combined to output a matrix of size (1024 x 2M). When the output of combining layer 2 is input to the multilayer perceptron, a (1024 x M) matrix is ​​output. This (1024 x M) matrix is ​​then pooled and averaged in the feature extraction layer to obtain a (1024 x 1) matrix. Finally, the output layer passes the data through the multilayer perceptron, which outputs a (1024 x 3) matrix.

[0045] In this way, the output layer outputs the nodal coordinates after displacement, to which the amount of warpage has been added, for the nodal coordinates (three-dimensional coordinates of the X, Y, and Z axes) of the mesh of the molded product that were input to input layer 1. As shown in Figure 14, the analytical proxy model used in this gate position determination method can predict the analysis results that will become the output data by inputting three types of input data separately: the nodal coordinates of the mesh of the molded product, the nodal coordinates of the gate position, and the setting values ​​for the resin flow analysis.

[0046] When the output data is warpage displacement, even if the input data is the same mesh node coordinates for the molded part and the same resin flow analysis settings, if different values ​​are entered for the gate position node coordinates, the warpage displacement for each node will vary between -2.00 and +2.00, resulting in different results for the warpage displacement trends of the molded part. However, the range of -2.00 to +2.00 here is only a guideline and is not a fixed value, as it is greatly affected by the size of the molded part, the shrinkage rate of the resin material used, and the resin settings. Furthermore, if the input data is the same mesh node coordinates for the molded part and the same gate position node coordinates, but different resin flow analysis settings are entered, the warpage displacement trends will not change, but the output results will vary by approximately -2.00 to +2.00.

[0047] To perform the learning process, the surrogate model generation unit 142 acquires three types of input data and one type of output data for the analytical surrogate model stored in the molded product information 163 via the learning data acquisition unit 141. After data acquisition is complete, the surrogate model generation unit 142 learns the analytical surrogate model in accordance with the neural network learning phase described above. Note that learning may be performed within the gate position determination device 100, such as by a CPU (Central Processing Unit), GPU (Graphics Processing Unit), or FPGA (Field Programmable Gate Array) in the control processing unit 120, or by utilizing external computing resources via the communication unit 110.

[0048] After the learning is completed, in S106, the learned analytical surrogate model is stored in the surrogate model structure information 161 and the surrogate model parameter information 162. This completes the analytical surrogate model learning process. After the analytical surrogate model learning process is completed, the analytical surrogate model inference process of S200 is initiated to predict the amount of warpage of another molded product Y. First, as in the analytical surrogate model learning process, settings necessary for resin flow analysis are made (S201). Three-dimensional CAD data for inference data is read from the molded product information 163 via the CAD model reading unit 131 of the resin flow analysis software 130. Next, the read CAD data is converted into an analytical model Y as shown in FIG. 15 by the analytical model generation unit 132. FIG. 15 is a diagram showing an example of an analytical model of an injection-molded product Y. FIG. 15 shows an example of a gate position G.

[0049] After the analysis model is created, resin flow analysis settings (e.g., resin material properties, resin temperature, injection speed, holding pressure, etc.) are made in the analysis parameter setting unit 133. In this embodiment, resin flow analysis settings are made using one pattern of standard molding conditions corresponding to the type of molded product stored in the molded product information 163. Furthermore, IDs of nodes selectable as gates are selected from the nodes extracted from the mesh-divided molded product, as shown in FIG. 15, and stored in the molded product information 163. The three-dimensional coordinates of the X-, Y-, and Z-axes of the nodes of the molded product shape that can be set as gate positions may be freely selected from all the nodes shown in FIG. 15. However, if, for example, two gate positions are located separately on the front and back surfaces of the molded product shape (mold), the injection molding mold will not function. Therefore, it is preferable to select them from a single surface so that the mold structure will function.

[0050] After the settings are complete, in S202, inference data is acquired via the inference input data acquisition unit 144. The acquired inference data is the same as the data input to the analytical proxy model: three items: nodal coordinates of the mesh of the molded product, nodal coordinates of the gate position, and resin flow analysis setting values. Once acquisition of the inference data is complete, the process proceeds to S203. In S203, the learned analytical proxy model is first read from the proxy model structure information 161 and the proxy model parameter information 162 via the proxy model reading unit 143. Then, by inputting the inference data acquired in S202 (nodal coordinates of the mesh of the molded product, nodal coordinates of the gate position, and resin flow analysis setting values) into the learned analytical proxy model read by the calculation unit 145, an inference result output (nodal coordinates after displacement to which the amount of warpage has been added to the nodal coordinates (three-dimensional coordinates of the X-axis, Y-axis, and Z-axis) of the mesh of the input molded product) can be obtained (S204). 16 and 17 are diagrams showing examples of predicted displacement plots using a proxy model of an injection-molded product Y. In Fig. 16 and Fig. 17, an example of a gate position G is shown, along with maximum displacement points 169 and 171.

[0051] The warpage amount acquired in S204 is loaded into the resin flow analysis software 130 in S205 via an API (Application Programming Interface) or GUI (Graphical User Interface) provided in the resin flow analysis software. The loaded warpage amount is mapped as vector information (information on the magnitude and direction of the warpage amount) to the molded product divided into infinitesimal elements and displayed on the display input / output unit 150. Furthermore, the display results change depending on the nodal information of the gate position input into the analytical proxy model, allowing the designer to consider various gate positions. While the above description has been given of the case where the vector value of the warpage amount of the injection-molded product is used as output information, other information such as the volumetric shrinkage difference or the injection pressure value may also be used.

[0052] Furthermore, the gate location determination method according to this embodiment uses an analytical surrogate model, significantly reducing verification time. Specifically, a typical analytical PC (personal computer) can predict gate locations for 10,000 patterns in this embodiment in just 8.85 seconds. This eliminates the need for sequential optimization, which previously required several hours to several days for resin flow analysis. This allows designers to make decisions more quickly. In addition, the gate location determination method using an analytical surrogate model may test all possible gate location combinations (e.g., for a two-point gate, all combinations of nodes on the molded product where the gate can be placed) when fabricating the mold, or may test a smaller number of patterns. Comparing FIG. 16 and FIG. 17 in this embodiment, the maximum warpage displacement in FIG. 16 is 0.35 mm, while the maximum warpage displacement in FIG. 17 is 0.20 mm. This demonstrates that a 0.15 mm improvement in warpage can be achieved without performing as many analyses as 10,000 patterns.

[0053] Once inference using the analytical surrogate model is complete, the designer selects the gate position that best predicts the required quality from the inference results (e.g., warpage amounts at 10,000 gate position patterns), i.e., the gate position that minimizes warpage, and then performs resin flow analysis using the resin flow analysis software 130 in S300. Next, in S400, it is confirmed whether the required quality of the molded product is met. If so, the gate position is determined and the process ends. If not, an end determination is made in S500. The designer may freely determine the end determination, but in this embodiment, the process ends when the number of verifications using the analytical surrogate model reaches 10. If the required quality is not met and the end determination has not been made, the analytical surrogate model learning process S100 may be performed again using the analysis data acquired in S300. Alternatively, the designer may return to the analytical surrogate model inference process S200 and perform inference using the analytical surrogate model again. Alternatively, the analytical surrogate model designer may switch to a method of directly using resin flow analysis for verification.

[0054] In the above example, two gate IDs are randomly selected from the IDs of nodes selectable as gates. This process is repeated 10,000 times to create a table of 10,000 gate ID combinations. Then, the analytical surrogate model is used to perform inference on these gate combinations. Based on the inference results (e.g., the warpage amounts at the 10,000 gate positions), the gate position that best predicts the required quality—i.e., the gate position that minimizes warpage—is selected. However, if, for example, one selects not only the gate position with the smallest warpage among the 10,000 patterns, but also the gate positions with the second and third smallest warpage amounts, and returns to the analytical surrogate model inference step S200 to perform confirmation by resin flow analysis, performing the resin flow analysis confirmation step S300 in ascending order of the inferred warpage results may result in analysis conditions that are nearly identical to those used in the previous step S300, resulting in nearly identical predicted values ​​for the required quality. For example, to avoid similar gate position patterns (combinations) when there are two gates, the algorithm shown in Figure 18 can be used. This algorithm selects gate positions with the smallest predicted value for the amount of warpage displacement, excluding those close to gate patterns (gate combinations) that have already been analyzed in the past.

[0055] In FIG. 18, P1 and P2 are the previously selected gates (the first is the gate with the smallest amount of warpage, and the second pattern (second set) is the previously selected gate; hereinafter, these are referred to as "Picked Gates"). R1 ​​and R2 are one pattern (hereinafter, "Random Gate") from among gate patterns selected from random patterns (e.g., 10,000 patterns). Dth is a distance threshold for determining that the distance between gate patterns is close.

[0056] In FIG. 18 , in step S801, a distance D11 between P1 and R1 is calculated, and a distance D12 between P1 and R2 is calculated. Next, in step S802, it is determined whether the conditions of distance D11 > Dth and distance D12 > Dth are satisfied. If they are satisfied, it is determined that the two objects are not similar (step S803). If it is determined in step S802 that the conditions are not satisfied, it is determined in step S804 that the condition of distance D11 > D12 is satisfied. If they are satisfied, it is determined in step S805 that a distance D21 between P2 and R1 is calculated. Next, in step S806, it is determined whether the condition of distance D21 > Dth is satisfied, and if so, it is determined that the two objects are not similar (step S803). If it is determined in step S806 that the conditions are not satisfied, it is determined that the two objects are similar (step S807). If it is determined in step S804 that the condition is not satisfied, the distance D22 between P2 and R2 is calculated in step S808. Then, in step S809, it is determined whether the condition of distance D22>Dth is satisfied, and if so, it is determined that the two are not similar (step S803). If it is determined in step S809 that the condition is not satisfied, it is determined that the two are similar (step S807).

[0057] In the algorithm of Figure 18, first, of the two Picked Gates, focus is placed on P1, and if both Random Gate points are far away, it is determined that they are not similar at that point. If even one of the two Random Gate points is close, the process proceeds to the next step. The Random Gate farthest from P1 is selected, and if the selected Random Gate is far from P2 (the Picked Gate not initially focused on), one of the two Random Gate points is far from the Picked Gate, so it is determined that they are not similar. If the selected Random Gate is close to P2, both Random Gate points are close to the Picked Gate, so it is determined that they are similar.

[0058] The above algorithm first selects the gate pattern with the smallest warpage from the inference results of the analytical surrogate model (e.g., warpage amounts at 10,000 gate positions). Then, in S300, the resin flow analysis software 130 actually performs a resin flow analysis. Then, in S400, it is confirmed whether the required quality of the molded product is met. If not, a gate position with the smallest predicted warpage displacement is selected, excluding gate patterns close to previously analyzed gate patterns. Then, in S300, the resin flow analysis software 130 actually performs a resin flow analysis. Then, in S400, it is confirmed whether the required quality of the molded product is met. If not, the above gate pattern selection and resin flow analysis can be repeated. This operation allows efficient selection of two or more conditions that will yield a good predicted value for the required quality from a large number of inference results of the analytical surrogate model. Compared to selecting two or more conditions that will yield a good predicted value for the required quality without excluding conditions close to previously analyzed conditions (which would result in selecting almost identical conditions), conditions that will yield a good predicted value for the required quality can be derived even with fewer resin flow analyses.

[0059] The gate position determination method described above is comprised of a resin flow analysis process, an analytical surrogate model learning process, an analytical surrogate model inference process, and a resin flow analysis confirmation process in which conditions that will result in a good predicted value for the required quality, excluding conditions that are close to conditions that have been analyzed in the past, are selected, and the resin flow analysis is performed under the selected conditions to confirm the predicted value for quality obtained by numerical analysis.

[0060] According to this embodiment, by training an analytical surrogate model and utilizing it to infer appropriate gate locations, gate locations that satisfy the required quality of a molded product can be determined easily and quickly, regardless of the designer's technical level. According to the method for determining gate locations for injection-molded products according to this embodiment, the required quality can be predicted without analysis by inputting the above-mentioned input items into a neural network regression model. Therefore, by utilizing this predicted value, designers can consider gate locations for molded products without repeatedly performing analysis. For example, by using a neural network model to understand the relationship between the required quality and gate locations, gate locations can be determined quickly. Furthermore, even if the molded product being considered changes, gate locations can be determined quickly. By using such technology, gate locations that satisfy the required quality of an injection-molded product can be determined easily and quickly, regardless of the designer's technical level.

[0061] Embodiment 2. In embodiment 2, a computer program for specifically determining a gate position using the gate position determination method described in embodiment 1 above is described. This application provides a program for causing a computer to function as a gate position determination device. This program causes the computer to execute the steps of repeatedly performing resin flow analysis to collect analysis data used to train a neural network regression model, and training the neural network regression model using the collected analysis data. The program also causes the computer to execute the step of inferring the required quality of a molded product using the trained neural network regression model based on the nodal coordinates of the calculation model and the nodal coordinates of the gate position. The purpose of this embodiment is to provide a program for causing a computer to function as a gate position determination device for determining an optimal gate position in a short period of time. Next, a program for causing a computer to function as the gate position determination method for injection-molded products described in paragraph

[0059] above is described. Such a program causes a computer to perform the following steps: repeatedly performing resin flow analysis to collect analysis data to be used in training the neural network regression model; training the neural network regression model using the collected analysis data; inferring the quality required for the injection molded product using the trained neural network regression model based on the nodal coordinates of the calculation model and the nodal coordinates of the gate position; and selecting conditions that will result in a good predicted value for the required quality, excluding conditions that are close to conditions that have been analyzed in the past, and performing resin flow analysis under the selected conditions to confirm the predicted value of the quality obtained by numerical analysis.

[0062] Embodiment 3. This embodiment shows an embodiment in which a convolutional graph neural network (hereinafter abbreviated as GNN) is used in the learning algorithm (structure of the analytical surrogate model) used in the surrogate model generation unit 142 of the gate position determination method described in the above-mentioned embodiment 1, or in the neural network regression model described in embodiments 1 and 2. By using a GNN in the structure of the analytical surrogate model, an effect is obtained in which the accuracy with which the analytical surrogate model predicts the quality required of injection-molded products is improved.

[0063] A GNN is a type of neural network that applies convolutional operations to graph data. While conventional convolutional neural networks (CNNs) are effective for grid-like data such as images, they cannot be directly applied to unstructured data such as graph data. GNNs were developed to perform convolutional operations on such unstructured data. GNNs receive information about the nodes and edges of graph data as input and perform convolutional operations. Specifically, information is propagated across the graph data by combining and updating the features of each node with the features of its neighboring nodes. The convolutional operations typically use the graph data's adjacency matrix. An adjacency matrix is ​​a matrix that represents the relationship between two adjacent nodes in the graph data, each corresponding to a small element. GNNs can extract local features from graph data, thereby improving prediction accuracy.

[0064] Below, we will provide an overview of the adjacency matrix and then explain a specific method for creating the adjacency matrix in this analytical surrogate model. First, we will explain the overview of the adjacency matrix with reference to FIGS. 19 and 20. FIG. 19 is a schematic diagram of nodes obtained by dividing a molded product into a mesh using FEM (finite element method). If the nodes obtained by dividing a molded product into a mesh using FEM are as shown in FIG. 19, nodes 1 and 2, 2 and 3, and 2 and 4 are adjacent. This can be expressed as an adjacency matrix as shown in FIG. 20. FIG. 20 is the adjacency matrix for FIG. 19. Each node number corresponds to a row and column, and matrix elements corresponding to adjacent nodes are set to 1, and other elements are set to 0. However, the matrix elements are not limited to 1 or 0, and may also be characteristic quantities (such as distance) of the edges of the graph data.

[0065] Next, we will explain the specific method for creating an adjacency matrix in this analytical surrogate model. First, an arbitrary number of nodes in the mesh of the molded part (n in this example) are sampled. Next, the distance is derived for all combinations of selecting two nodes from the sampled n nodes. Next, a threshold is set for the distance between nodes, and nodes closer than the threshold are determined to be adjacent. The threshold for the distance between nodes is determined as a percentage of the maximum distance between nodes (the distance between diagonal nodes) for each molded part. For example, if the threshold is set to 30% for a molded part with a diagonal dimension of 100 mm, two nodes within 100 mm x 30% = 30 mm are determined to be adjacent.

[0066] Next, an nxn matrix is ​​created, and the elements of the matrix corresponding to adjacent nodes are set to 1, and the other elements are set to 0. Next, the adjacency matrix is ​​normalized using the following formula (1) using the method described in the reference (T. N. Kipf, SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS, The International Conference on Learning Representations, 2017.).

[0067]

[0068] Here, A ＊ is the normalized adjacency matrix, Dｗ／ｓｅｌｆ is a matrix derived by the following equation (2), A ｗ／ｓｅｌｆ is a matrix derived by the following equation (3), where A is the adjacency matrix before normalization, and IN is the identity matrix.

[0069]

[0070]

[0071] This A ＊ is used as the adjacency matrix of this analytical surrogate model. Figure 21 is a schematic diagram showing the analytical surrogate model (neural network regression model) used. The analytical surrogate model receives three types of input data: the node coordinates of the molded part mesh, the node coordinates of the gate position, and the resin flow analysis settings. The analytical surrogate model receives one type of output data: the amount of warpage of the molded part. The analytical surrogate model is structured as follows: an adjacency matrix generation layer that generates an adjacency matrix using the above method; a connection layer that aggregates the input data; a multilayer perceptron layer with multiple layers of neural networks; an inner layer that calculates the inner product of the adjacency matrix and the multilayer perceptron layer; and a multilayer perceptron layer, leading to the final output layer. Note that inputting data into the "multilayer perceptron layer with multiple layers of neural networks" described here is equivalent to inputting data into the affine transformation matrix (fully connected layer) in a machine learning model. This model was designed with reference to a paper on convolutional graph networks (T. N. Kipf, SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS, The International Conference on Learning Representations, 2017).

[0072] First, we will explain the input data. Here, we will specifically explain the case where the number of nodes in the molded part mesh is n. The input size of the nodes in the molded part mesh is an (n x 3) matrix because the nodes exist in three-dimensional space along the X, Y, and Z axes. The input size of the nodes for the gate position is also expressed as an (n x 3) matrix, and if there are two gates, the total becomes (n x 6). The values ​​in this matrix are calculated based on the node coordinates of the molded part mesh. The values ​​in the two rows corresponding to the gate position are set to 0, and the values ​​in the other rows represent the distance (coordinate difference) between the gate position and each node coordinate of the molded part mesh. The input size of the setting values ​​for the resin flow analysis can be expressed as an (n x N) matrix. N is a positive integer, and you can set as many as the number of setting values ​​for the resin flow analysis you want to consider in the analytical proxy model you are constructing. For example, if you want to consider the effect of mold temperature, you would use an (n x 1) matrix. If you want to consider mold temperature on both the fixed and movable sides of the mold, you would use an (n x 2) matrix. Also, an (n×n) adjacency matrix is ​​created using the above method.

[0073] Next, the processing content of each layer will be specifically explained using an example in which no resin flow analysis settings are input and the number of gates is two. First, the combining layer combines the node coordinates of the molded product mesh with the node coordinates of the gate positions to output a matrix of size (n × 9). This output is input to the next multilayer perceptron to obtain an output of size (n × M1). M is a positive integer that varies depending on the dimensionality of the output space of the multilayer perceptron layer. For example, if the dimensionality of the output space of the multilayer perceptron is 32, M is also 32. Next, an inner stacking layer takes the inner product of the matrix output from the previous multilayer perceptron layer and the adjacency matrix to obtain a matrix of size (n × M1). This output is input to the next multilayer perceptron to obtain an output of size (n × M2). Finally, the output layer passes it through a perceptron layer to output an (n × 3) matrix.

[0074] In this way, the output layer outputs the warpage displacement or the nodal coordinates after warpage displacement is added to the nodal coordinates (3D coordinates of the X, Y, and Z axes) of the molded part mesh entered in Input Layer 1. The analytical proxy model used in this gate position determination method can predict the analysis results (output data) by inputting three types of input data: the nodal coordinates of the molded part mesh, the nodal coordinates of the gate location, and the resin flow analysis settings. When the output data is warpage displacement, even if the input data is the same molded part mesh nodal coordinates and the same resin flow analysis settings, if different values ​​are entered for the nodal coordinates of the gate location, the warpage displacement of each nodal point will vary between -2.00 and +2.00, resulting in different results for the molded part warpage displacement trend. However, this range of -2.00 to +2.00 is merely a guideline and is not a uniform value, as it is greatly affected by the size of the molded part, the shrinkage rate of the resin material used, and the resin settings. In addition, if different settings are entered for the resin flow analysis between the node coordinates of the mesh of the same molded product and the node coordinates of the same gate position, the tendency of the warpage displacement amount will not change, but the warpage displacement amount of each node will differ by approximately -2.00 to +2.00.

[0075] The characteristics of three-dimensional shapes are abstract and difficult to grasp. Therefore, simple machine learning such as multiple regression, or a simple neural network in which point cloud data is converted into one dimension and input to a fully connected layer, or a neural network in which point cloud data is input instead of graph data (such as PointNet (Qi, Charles R., et al. Pointnet: Deep learning on point sets for 3D classification and segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.)) cannot fully grasp the characteristics, and sufficient accuracy in predicting the required quality of molded products cannot be obtained.

[0076] As described above, the neural network regression model of this embodiment takes the nodal coordinates of the microelements of the injection-molded product, the nodal coordinates of the gate positions, and the setting values ​​of the resin flow analysis as input data, creates an adjacency matrix representing the relationship between two adjacent junctions of the graph data corresponding to the microelements from the nodal coordinates of the injection-molded product, and calculates the inner product of the adjacency matrix and the output matrix of the multilayer perceptron layer.By doing so, for each piece of data representing the relationship between the junction of interest and the adjacent junctions, the input data is input into the affine transformation matrix (fully connected layer) in the machine learning model, and the predicted value of the required quality is used as output data.

[0077] Although various exemplary embodiments and examples are described in this disclosure, the various features, aspects, and functions described in one or more embodiments are not limited to the application of a particular embodiment, but may be applied to the embodiments alone or in various combinations. Therefore, countless variations not illustrated are anticipated within the scope of the technology disclosed in this specification. For example, this includes cases where at least one component is modified, added, or omitted, or where at least one component is extracted and combined with components of another embodiment.

[0078] 100 Gate position determination device, 110 Communication unit, 120 Control processing unit, 130 Resin flow analysis software, 131 CAD model reading unit, 132 Analysis model generation unit, 133 Analysis parameter setting unit, 134 Analysis unit, 140 Analysis proxy model control unit, 141 Learning data acquisition unit, 142 Proxy model generation unit, 143 Proxy model reading unit, 144 Inference input data acquisition unit, 145 Calculation unit, 146 Proxy model prediction result output unit, 150 Display input / output unit, 160 Memory unit, 161 Proxy model structure information, 162 Proxy model parameter information, 163 Molded product information.

Claims

1. a representative model generation unit that performs a resin flow analysis, which is a numerical analysis, on an analytical model generated by dividing CAD data of an injection molded product into minute elements, and learns, in accordance with a neural network regression model, a relationship between the nodal coordinates of minute elements in the injection molded product and the nodal coordinates of gate positions, which are obtained, and a predicted value of a required quality; A gate position determination device for an injection-molded product, comprising a calculation unit that outputs at least one of a vector value of the amount of warping of the injection-molded product, a volumetric shrinkage difference of the injection-molded product, or an injection pressure as output information by inputting the nodal coordinates of the microelements in the injection-molded product and the nodal coordinates of the gate position as inference data into a trained analytical proxy model.

2. 2. The gate position determination device for an injection molded product according to claim 1, The neural network regression model is The node coordinates of the minute elements in the injection molded product and the node coordinates of the gate position are input as input data; creating an adjacency matrix representing a relationship between two adjacent nodes of graph data corresponding to the infinitesimal elements from the node coordinates of the infinitesimal elements in the injection molded product; By calculating the inner product of the adjacency matrix and the output matrix of the multilayer perceptron layer, the input data is input to the affine transformation matrix (fully connected layer) in the machine learning model for each data representing the relationship between the node of interest and the adjacent nodes, The output data is the predicted value of the required quality. Gate position determination device for injection molded products.

3. a resin flow analysis process in which CAD data of the injection molded product is divided into minute elements to generate an analysis model for performing a resin flow analysis, which is a numerical analysis, and a resin flow analysis is performed on the analysis model in accordance with predetermined analysis parameters; an analytical surrogate model learning process for learning a relationship between a predicted value of a required quality and the nodal coordinates of a minute element and a nodal coordinate of a gate position in the injection molded product analyzed by the resin flow analysis process according to a neural network regression model; A method for determining a gate position of an injection-molded product, comprising an analytical surrogate model inference process for inputting the nodal coordinates of a microelement in the injection-molded product and the nodal coordinates of the gate position as inference data into a trained analytical surrogate model, and outputting at least one of a vector value of the amount of warpage of the injection-molded product, a volumetric shrinkage difference of the injection-molded product, or an injection pressure as output information.

4. 4. The method for determining a gate position of an injection molded product according to claim 3, comprising the steps of: The neural network regression model is The node coordinates of the minute elements in the injection molded product and the node coordinates of the gate position are input as input data; creating an adjacency matrix representing a relationship between two adjacent nodes of graph data corresponding to the infinitesimal elements from the node coordinates of the infinitesimal elements in the injection molded product; By calculating the inner product of the adjacency matrix and the output matrix of the multilayer perceptron layer, the input data is input to the affine transformation matrix (fully connected layer) in the machine learning model for each data representing the relationship between the node of interest and the adjacent nodes, A method for determining gate locations for injection molded parts, with the predicted value of the required quality as output data.

5. 4. The method for determining a gate position of an injection molded product according to claim 3, comprising the steps of: The resin flow analysis step; the analytical surrogate model learning step; the analytical surrogate model inference step; A method for determining gate positions for injection molded products, comprising a resin flow analysis confirmation process for selecting conditions that will produce a good predicted value for the required quality, excluding conditions that are close to conditions that have been analyzed in the past, and performing a resin flow analysis under the selected conditions to confirm the predicted value of quality obtained by numerical analysis.

6. A program for causing a computer to function as the gate position determining method for an injection molded product according to claim 3, The program includes: A step of repeatedly performing a resin flow analysis to collect analysis data used for learning a neural network regression model; training the neural network regression model using the collected analytical data; and a step of inferring the quality required for the injection molded product using the trained neural network regression model based on the nodal coordinates of the calculation model and the nodal coordinates of the gate position.

7. The program according to claim 6, The neural network regression model is The node coordinates of the minute elements in the injection molded product and the node coordinates of the gate position are input as input data; creating an adjacency matrix representing a relationship between two adjacent nodes of graph data corresponding to the infinitesimal elements from the node coordinates of the infinitesimal elements in the injection molded product; By calculating the inner product of the adjacency matrix and the output matrix of the multilayer perceptron layer, the input data is input to the affine transformation matrix (fully connected layer) in the machine learning model for each data representing the relationship between the node of interest and the adjacent nodes, A program whose output is a prediction of the required quality.

8. A program for causing a computer to function as the gate position determining method for an injection molded product according to claim 5, The program includes: A step of repeatedly performing a resin flow analysis to collect analysis data used for learning a neural network regression model; training the neural network regression model using the collected analytical data; Inferring a quality required for the injection molded product using the trained neural network regression model based on the nodal coordinates of the calculation model and the nodal coordinates of the gate position; A program that executes a step of selecting conditions that will provide a good predicted value for the required quality, excluding conditions that are close to conditions that have been analyzed in the past, and confirming the predicted value of quality obtained by numerical analysis by performing a resin flow analysis under the selected conditions.