JP7346885B2 - Shape generation device and shape generation method - Google Patents

Description

The present disclosure relates to a shape generation device and a shape generation method.

In the mechanical design of objects such as aircraft wings and turbine blades, it is necessary to pursue a shape with high performance while satisfying the required specifications for the object. For example, in prior document 1, principal components are extracted by principal component analysis of a plurality of shapes of the object, including shapes that have performance equal to or higher than a preset target performance value, and based on the principal components, It is described that a new shape of an object is generated and its performance is evaluated.

Japanese Patent Application Publication No. 2013-003971

However, in the technique of Patent Document 1, a new shape is not associated with a performance value. In order to understand the performance value of such a new shape, it was necessary to derive the performance value anew from the new shape by CFD (Computational Fluid Dynamics) calculation or the like. In this case, even if the goal is to meet the performance value, the performance value can only be determined from the result of generating a new shape, so it is necessary to generate multiple shapes by hand and check the performance value each time to get closer to the goal. There was a need. Therefore, even if the technique of Patent Document 1 is applied, although the number of calculations is reduced compared to the case where it is not applied, it is not sufficient to shorten the design time.

Therefore, an object of the present disclosure is to provide a shape generation device and a shape generation method that can shorten design time.

In order to solve the above problems, a shape generation device according to an aspect of the present disclosure includes an encoding module that generates feature amounts by encoding learning shape data and learning performance values corresponding to the learning shape data. , and a decoding module that decodes feature quantities and learning performance values to generate shape data corresponding to the learning shape data, and includes a target performance value and a decoding module for a generative model that has been subjected to machine learning multiple times . The learning shape data includes point sequence data corresponding to the outer shape of the learning shape based on the learning shape data, and the learning shape data includes point sequence data corresponding to the outer shape of the learning shape based on the learning shape data. has a relatively high density in areas where the learning shape has a high curvature, and has a relatively low density in areas where the learning shape has a low curvature.

In order to solve the above problems, a shape generation device according to an aspect of the present disclosure includes an encoding module that generates feature amounts by encoding learning shape data and learning performance values corresponding to the learning shape data. , and a decoding module that decodes feature quantities and learning performance values to generate shape data corresponding to the learning shape data, and includes a target performance value and a decoding module for a generative model that has been subjected to machine learning multiple times. A decoding module processing section that generates target shape data by inputting arbitrary feature quantities, and a smoothing processing section that smoothes the target shape based on the generated target shape data, and generates a smoothed target shape. The data includes point sequence data corresponding to the outer shape of the target shape, and the point sequence data has a relatively high density in areas where the curvature of the target shape is high, and a relatively low density in areas where the curvature of the target shape is low.

In order to solve the above problems, in a shape generation method according to one aspect of the present disclosure, a computer generates feature amounts by encoding learning shape data and learning performance values corresponding to the learning shape data. It includes an encoding module and a decoding module that decodes feature values and learning performance values to generate shape data corresponding to the learning shape data, and is targeted at the decoding module in a generative model that has undergone multiple machine learnings. A shape generation method that generates target shape data by inputting performance values and arbitrary feature quantities, the learning shape data including point sequence data corresponding to the outer shape of the learning shape based on the learning shape data. , the point sequence data has a relatively high density in areas where the learning shape has a high curvature, and has a relatively low density in areas where the learning shape has a low curvature .

In order to solve the above problems, in a shape generation method according to one aspect of the present disclosure, a computer generates feature amounts by encoding learning shape data and learning performance values corresponding to the learning shape data. It includes an encoding module and a decoding module that decodes feature values and learning performance values to generate shape data corresponding to the learning shape data, and is targeted at the decoding module in a generative model that has undergone multiple machine learnings. A shape generation method that generates target shape data by inputting performance values and arbitrary feature quantities, and smoothes the target shape based on the generated target shape data , the smoothed target shape data is The point sequence data includes point sequence data corresponding to the outer shape of the shape, and the point sequence data has a relatively high density in areas where the curvature of the target shape is high, and a relatively low density in areas where the curvature of the target shape is low.

According to the present disclosure, it is possible to shorten design time.

FIG. 1 is a block diagram showing a schematic configuration of a shape generation device. FIG. 2 is a flowchart showing a general flow of the shape generation method by the shape generation device. FIG. 3 is an explanatory diagram for explaining the schematic configuration of the machine learning section. FIG. 4 is an explanatory diagram for explaining the schematic configuration of the shape generation section. FIG. 5 is an explanatory diagram for explaining the first method. FIG. 6 shows a verification of machine learning using the first method. FIG. 7 is an explanatory diagram for explaining the second method. FIG. 8 is an explanatory diagram for explaining self-intersection. FIG. 9 shows a verification of machine learning using the second method. FIG. 10 shows a verification of machine learning using the second method. FIG. 11 is an explanatory diagram for explaining the schematic configuration of the machine learning section. FIG. 12 is an explanatory diagram for explaining the schematic configuration of the shape generation section.

An embodiment of the present disclosure will be described below with reference to the accompanying drawings. The dimensions, materials, and other specific numerical values shown in these embodiments are merely examples for facilitating understanding of the present disclosure, and are not intended to be limiting unless otherwise specified. In this specification and drawings, elements having substantially the same functions and configurations are designated by the same reference numerals to omit redundant explanation, and elements not directly related to the present disclosure are omitted from illustration. do.

In this disclosure, in order to generate a new shape with the goal of satisfying a predetermined performance value, machine learning is performed using shape data indicating the shape (learning shape data) and performance values (learning performance value) as input nodes. use As for machine learning, a learning model called a generative model is being considered. Such a generative model is a method of generating an image that does not actually exist by, for example, learning feature amounts from a plurality of image data and specifying a numerical value called a label. In the present disclosure, among generative models such as VAE (Variational Auto-Encoder), CVAE (Conditional Variational Auto-Encoder), GAN (Generative Adversarial Network), and Conditional GAN, for example, CVAE is used, and the "image" and " Instead of "label", "shape" and "performance value" will be used.

VAE (Variational Auto-Encoder) corresponds to the basic configuration of CVAE. Since VAE is an existing technology and various technical documents can be referred to, a detailed explanation thereof will be omitted here. In simple VAE, data is generated based on random variables, but in CVAE of the present disclosure, a performance value (label) is attached to this data to generate a desired type of data. Therefore, CVAE differs from simple VAE in that performance value nodes are added to both input nodes and feature amounts. In CVAE, as will be described later, various shapes can be reproduced by incorporating a variance σ and a scale factor ε in addition to the expected value μ.

FIG. 1 is a block diagram showing a schematic configuration of a shape generation device 100, and FIG. 2 is a flowchart showing a schematic flow of a shape generation method by the shape generation device 100. The shape generation device 100 functions as a machine learning section 110 and a shape generation section 120 using a computer including a central processing unit (CPU), a ROM in which programs and the like are stored, and a RAM as a work area. As shown in FIG. 2, the machine learning unit 110 generates a generative model (for example, An encoding module and a decoding module (to be described later) are subjected to machine learning (machine learning phase S1). The shape generation unit 120 uses a part of the generative model learned in this way (for example, a decoding module) to create a new shape (target (shape generation phase S2).

FIG. 3 is an explanatory diagram for explaining the schematic configuration of the machine learning section 110. Machine learning section 110 includes a performance value deriving section 112, an encoding module learning section 114, and a decoding module learning section 116. Here, solid arrows indicate the flow of numerical values, and dashed arrows indicate the direction of control.

The performance value deriving unit 112 extracts learning shape data (a plurality of parameters that can identify the learning shape) of iteration 1 from the plurality of types of learning shapes 10 representing the target object prepared as samples, and performs CFD. The performance value (learning performance value) of the learning shape 10 is derived by calculation or the like. The performance value is expressed by one or more numerical values, and the presence or absence of units does not matter. Note that here, for convenience of explanation, a numerical value of 1 is used as the performance value. The performance value deriving unit 112 stores the derived performance value for learning in association with the shape data for learning. In this way, learning performance values are associated with all of the plurality of learning shapes. Note that when the performance value of the learning shape 10 is known in advance, the performance value deriving unit 112 uses the known performance value as the learning performance value to generate the learning shape data without deriving the learning performance value. It is stored in association with.

The encoding module learning unit 114 performs machine learning on the encoding module 114a that generates feature quantities by encoding learning shape data and learning performance values associated with the learning shape data, which are input nodes. . In the encoding module 114a, the feature quantity is a variable sequence probabilistically determined according to a normal distribution with an expected value μ, a variance σ, and a scale factor ε. The variables of such feature quantities are each shown in a range of 0 to 1. Here, the encoding module learning unit 114 uses a neural network based on CVAE to generate the encoding module 114a, for example, by multiplying each input node or by a weighting coefficient. Module 114a can be represented by a complex nonlinear function.

The decoding module learning unit 116 is a decoding module that generates shape data indicating a new shape 20 by decoding the feature amount obtained by the encoding module 114a and the learning performance value associated with the learning shape data. Machine learning 116a. Here, the decoding module learning unit 116 uses a neural network based on CVAE to multiply performance values, feature quantities, or performance values and feature quantities, or multiplies them by a weighting coefficient, so that the decoding module 116a, the decoding module 116a can be expressed by a complex nonlinear function.

The decoding by the decoding module learning section 116 corresponds to the inverse calculation of the encoding by the encoding module learning section 114. Therefore, if the encoding module 114a and the decoding module 116a can be estimated correctly, as long as the same performance value is input, the learning shape data indicating the learning shape 10 input to the encoding module 114a will be transferred to the decoding module. 116a, to form shape data that corresponds to (is similar to) the learning shape data. In other words, the encoding module learning section 114 and the decoding module learning section 116 train the encoding module 114a and the decoding module 116a so that the generated new shape 20 is the same as the input learning shape 10. Machine learning.

FIG. 4 is an explanatory diagram for explaining the schematic configuration of the shape generation section 120. The shape generation section 120 includes a decoding module processing section 122 and a smoothing processing section 124. Note that the smoothing processing unit 124 will be described in detail later. Here, solid arrows indicate the flow of numerical values, and dashed arrows indicate the direction of control.

The decoding module processing unit 122 refers to the learned decoding module 116a and decodes target shape data indicating the new shape 20 from the target performance value and the feature amount. The smoothing processing unit 124 smoothes the target shape based on the generated shape data. The smoothing processing section 124 will be described in detail later. Here, the performance value is a target performance value targeted for design (required in design), and the shape data is target shape data indicating a new shape (target shape) that satisfies the target performance value.

Further, the feature amount has the same format as the feature amount that is the result of the encoding module 114a, and is expressed as a numerical value string arbitrarily determined by the designer. Each numerical value in such arbitrary numerical value string may be any numerical value (random numerical value) as long as it is in the range of 0 to 1. Here, no matter what numerical value is input as the feature quantity, theoretically, target shape data that satisfies the target performance value should be derived. Therefore, the designer can adjust the new shape 20 by changing the feature amount. In adjusting the shape, for example, design, workability, production cost, weight, etc. are taken into consideration.

A specific design example of the machine learning section 110 and the shape generation section 120 will be described below. Here, the cross-sectional shape of a wing is cited as the shape that is the object of design. Note that in the above-mentioned CVAE, learning shape data as an input node can be expressed in various formats. For example, the cross-sectional shape of a wing can be represented as an image.

However, when representing the cross-sectional shape of a wing as an image, the learning shape data that is the input node is the brightness of each pixel in the image, and the new shape data that is generated is also represented by brightness, and the boundary line of the shape is It becomes vague. In this case, it is difficult to obtain a clear outline necessary for mechanical design. Furthermore, in order to capture subtle differences in shape, it becomes necessary to increase the resolution, which may unnecessarily increase the dimensionality of the input node. Therefore, in the present disclosure, the cross-sectional shape of the blade is shown by an outline line, and specifically, the outline line is approximated by a polygonal line and represented by point sequence data (a plurality of points expressed in a row).

In this way, when representing the outline of the cross-sectional shape of a wing using point sequence data, there are two methods: the first method in which one of the X and Y axes is restricted, and the other in which both axes are restricted, depending on the axis restriction of the point sequence data. For example, a second method in which the method is not performed. Both methods will be considered below.

FIG. 5 is an explanatory diagram for explaining the first method. In the first method, one of the X-axis and Y-axis (here, the X-axis) of the point sequence data is restricted. That is, in the first method, as shown in FIG. 5 by x0, x1, x2, x3, x4, x5,... The axis coordinates are free variables. In addition, a lift coefficient is cited as a learning performance value. Here, the lift coefficient is a dimensionless numerical value that represents the difference in lift caused by the shape and angle of an object when the object moves in the air or is placed in an air flow.

Therefore, among the input nodes of the encoding module 114a, the input nodes (learning shape data) regarding the cross-sectional shape (shape) of the wing are (x0, y0), (x1, y1), (x2 , y2), (x3, y3), (x4, y4), (x5, y5), (x1, y6), (x2, y7), (x3, y8), (x4, y9), (x5, y10 ),... Here, two Y-axis coordinates are assigned to the same X-axis coordinate.

Furthermore, the performance value deriving unit 112 derives a lift coefficient from the point sequence data. The derived lift coefficient is also added to the input node. The encoding module learning unit 114 performs machine learning on the encoding module 114a that encodes all of the learning shape data indicating the cross-sectional shape of the wing and multiple combinations (for example, 189 samples) of lift coefficients to generate feature quantities. do. Furthermore, the decoding module learning unit 116 performs machine learning on the decoding module 116a, which decodes all combinations of feature values and lift coefficients obtained by the encoding module 114a to generate shape data indicating the cross-sectional shape of the wing.

At this time, the encoding module learning unit 114 and the decoding module learning unit 116 operate the encoding module 114a and the decoding module learning unit 116 so that the cross-sectional shape of the generated new wing is the same as the input cross-sectional shape of the learning wing. Machine learning is performed on the decryption module 116a.

As the machine learning of the encoding module 114a and the decoding module 116a progresses in this way, the decoding module processing unit 122 refers to the learned decoding module 116a and calculates the lift coefficient as the target performance value and the arbitrarily determined feature quantity. Based on this, target shape data that newly indicates the cross-sectional shape of the wing is decoded. In this way, a new wing cross-sectional shape is obtained that satisfies the required lift coefficient.

FIG. 6 shows a verification of machine learning using the first method. Here, it is assumed that the target lift coefficients are 1.0 and 1.2, and the designer arbitrarily adjusts the feature amounts to obtain one wing cross-sectional shape for each. For example, (a) in FIG. 6 shows the cross-sectional shape of a wing obtained with a lift coefficient of 1.0, and (b) in FIG. 6 shows a cross-sectional shape of a wing obtained with a lift coefficient of 1.2. Here, it can be seen that the cross-sectional shape of the wing changes depending on the lift coefficient.

In addition, for verification purposes, the lift coefficients of the newly obtained cross-sectional shapes of each wing were further derived for the purpose of confirmatory target shape data. Here, in order to derive the lift coefficient, Xfoil (two-dimensional subsonic viscous flow code), which is a calculation program for laminar airfoils, is used.

In the verification, the lift coefficient was found to be 0.9860 for the cross-sectional shape of the wing shown in (a) in FIG. Therefore, it can be understood that the cross-sectional shape of the new wing is formed to satisfy the lift coefficient of 1.0, which was input as the target. In addition, in the cross-sectional shape of the wing shown in (b) in FIG. 6, the lift coefficient is 1.2237, and the new wing cross-sectional shape is formed to satisfy the lift coefficient of 1.2 input as the target. It can be understood.

Although not shown here, even when the target lift coefficient is set to 0.8, which is less than 1.0, and 1.4, which is more than 1.2, the lift of the newly generated cross-sectional shape of the wing is The coefficients are 0.7900 and 1.3929, respectively, and it can be seen that the cross-sectional shape of the new wing is formed to satisfy the lift coefficients of 0.8 and 1.4 input as targets. In this way, it becomes possible to shorten the design time.

However, according to the first method in which one of the X-axis and Y-axis is restricted, the value of one axis is fixed, so the distance between the polygonal lines forming the cross-sectional shape of the wing is fixedly increased. Although there are parts of the cross-sectional shape of the wing that can be approximated by a straight line, there are also parts that have a high degree of curvature in order to obtain lift. If the first method is applied to such a region with high curvature, the target lift coefficient may not be obtained, or the cross-sectional shape of the generated new wing may not be a realistic shape. Therefore, the second method will be considered.

FIG. 7 is an explanatory diagram for explaining the second method. In the second method, the point sequence data is not limited in either the X axis or the Y axis. That is, in the second method, as shown in FIG. 7, the density of point sequence data is set relatively high in a region 130 with high curvature in a predetermined learning shape, and the density of point sequence data is set relatively high in a region 132 with low curvature. Keep it relatively low.

Therefore, among the input nodes of the encoding module 114a, the input nodes (learning shape data) regarding the cross-sectional shape (shape) of the wing are (x0, y0), (x1, y1), (x2 , y2), (x3, y3), (x4, y4), (x5, y5), (x6, y6), (x7, y7), (x8, y8), (x9, y9), (x10, y10 ), (x11, y11), (x12, y12), (x13, y13), (x14, y14), and so on. Here, the X-axis and Y-axis coordinates are assigned independently.

Furthermore, the performance value deriving unit 112 derives a lift coefficient (performance value) from the point sequence data. The derived lift coefficient is also added to the input node. The encoding module learning unit 114 performs machine learning on the encoding module 114a that generates feature quantities by encoding all of the multiple combinations (for example, 4560 samples) of learning shape data indicating the cross-sectional shape of the wing and lift coefficients. . Furthermore, the decoding module learning unit 116 performs machine learning on the decoding module 116a, which decodes all combinations of feature values and lift coefficients obtained by the encoding module 114a to generate shape data indicating the cross-sectional shape of the wing.

At this time, the encoding module learning unit 114 and the decoding module learning unit 116 operate the encoding module 114a and the decoding module learning unit 116 so that the cross-sectional shape of the generated new wing is the same as the input cross-sectional shape of the learning wing. Machine learning is performed on the decryption module 116a.

As the machine learning of the encoding module 114a and the decoding module 116a progresses in this way, the decoding module processing unit 122 refers to the learned decoding module 116a and calculates the lift coefficient as the target performance value and the arbitrarily determined feature quantity. Based on this, target shape data that newly indicates the cross-sectional shape of the wing is decoded. In this way, a new wing cross-sectional shape is obtained that satisfies the required lift coefficient. Furthermore, in the second method, a region with high curvature can be represented by a high-density polygonal line that is close to a curve.

However, with the second method, as mentioned above, while it is possible to obtain areas with high curvature in the cross-sectional shape of the wing with high accuracy (high resolution), the point sequence data is freely discretized into multiple axes, so it is possible to obtain self-intersecting becomes more likely to occur. Here, self-intersection means that the polygonal lines are formed irregularly and intersect in a loop.

FIG. 8 is an explanatory diagram for explaining self-intersection. FIG. 8 shows an enlarged view of the cross-sectional shape of the wing newly generated using the second method, particularly the trailing edge portion 134. In the second method, the point sequence data has degrees of freedom regarding two axes, the X-axis and the Y-axis. Therefore, the newly generated point sequence data of the cross-sectional shape of the blade is not restricted to other points and appears at any position. In this case, it will not be a problem in areas that are close to straight lines, but in areas with high curvature and high density of point sequence data, such as the trailing edge 134 in FIG. Self-intersection occurs without forming a broken line. Such self-intersections, even if extremely small, are unacceptable because they make CFD calculations impossible.

Therefore, consider smoothing the target shape based on point sequence data in which such self-intersection occurs. For example, assume that a neural network error is added to points on an ideal curve to generate dispersed point sequence data such as the trailing edge portion 134 in FIG. 8. In order to eliminate the error and create a regression curve, the smoothing processing unit 124 generates an approximate curve using, for example, Gaussian process regression in the present disclosure.

Gaussian process regression is based on the assumption that a curve on two axes has a single value on the other axis for every value on one axis. Therefore, in the outline of the cross-sectional shape of an endless blade, for one axis value, there are at least two values in other axes, so Gaussian process regression can be applied as is. I can't. Therefore, as shown in FIG. 8, among the generated point sequence data of the cross-sectional shape of the blade, the point xmin where the X-axis coordinate is the minimum and the point xmax where the X-axis coordinate is the maximum are the points of the cross-sectional shape of the blade. Divide column data into upper and lower halves. The smoothing processing unit 124 independently generates a point sequence model of the cross-sectional shape of the wing by regressing the Y-axis coordinate from the X-axis coordinate for each of the upper half point sequence data and the lower half point sequence data. do.

By doing so, it becomes possible to suppress self-intersection of the point sequence in a region with high curvature and smooth it. For example, the point sequence data of the trailing edge portion 134 in FIG. 8 becomes a point sequence model 136 by Gaussian process regression. Referring to the point sequence model 136, it can be seen that the approximate curve is smooth.

Here, an example has been described in which the cross-sectional shape of the blade is divided into an upper half and a lower half. However, the present invention is not limited to this case, and the Gaussian process regression may be performed for each curve by further subdividing the shape. In this case, by regarding the tangents of the subdivided curves as one axis, it is possible to specify one value on the other axes for every value on one axis.

Note that if smoothing of the point sequence data is not necessary, the smoothing processing unit 124 does not perform any processing. Therefore, the shape data generated by the decoding module 116a becomes the final shape data.

FIG. 9 shows a verification of machine learning using the second method. Here, as in FIG. 6, it is assumed that the target lift coefficients are 1.0 and 1.2, and the designer arbitrarily adjusts the feature amounts to obtain one wing cross-sectional shape for each. For example, (a) in FIG. 9 shows the cross-sectional shape of a wing obtained with a lift coefficient of 1.0, and (b) in FIG. 9 shows a cross-sectional shape of a wing obtained with a lift coefficient of 1.2. Here, it can be seen that the cross-sectional shape of the wing changes depending on the lift coefficient.

In addition, for verification purposes, the lift coefficients of the newly obtained cross-sectional shapes of each wing were further derived for the purpose of confirmatory target shape data. In the verification, the lift coefficient was found to be 1.0152 for the cross-sectional shape of the wing shown in (a) in FIG. Therefore, it can be understood that the cross-sectional shape of the new wing is formed to satisfy the lift coefficient of 1.0, which was input as the target. In addition, in the cross-sectional shape of the wing in (b) in FIG. 9, the lift coefficient is 1.2028, and it can be seen that the new wing cross-sectional shape is formed to satisfy the lift coefficient of 1.2 that was input as the target. It can be understood. In this way, it becomes possible to shorten the design time.

Here, the second method described above was further applied to a turbine blade for verification. As can be understood with reference to FIG. 10, which will be described later, the width of the turbine blade in the X-axis direction (chord length) differs depending on the target performance value, so the interval between the X-axis coordinates cannot be appropriately determined. First, it is difficult to apply the first method.

FIG. 10 shows a verification of machine learning using the second method. Here, two parameters, the outflow angle and the loss coefficient, are listed as performance values. Here, the outflow angle indicates the angle at which the fluid flows out at the turbine outlet. The loss coefficient is a dimensionless value expressing the frictional loss of fluid energy.

Here, the combinations of the outflow angle and loss coefficient are -67.2 degrees and 0.0394, and -62.3 degrees and 0.0394, and the designer can arbitrarily adjust the feature values. Assume that the cross-sectional shape of one turbine blade is obtained. For example, (a) in FIG. 10 shows the cross-sectional shape of a turbine blade obtained with an outflow angle of -67.2 degrees and a loss coefficient of 0.0394, and (b) in FIG. 10 shows an outflow angle of -62.3 degrees. , shows the cross-sectional shape of the turbine blade obtained with loss coefficient=0.0394. Here, it can be seen that the angle of the trailing edge 134 of the wing changes depending on the outflow angle.

In addition, here, for verification purposes, the outflow angle and loss coefficient of each newly obtained cross-sectional shape of the turbine blade were further derived for the target shape data indicating the cross-sectional shape of the turbine blade. In the verification, for the cross-sectional shape of the blade shown in (a) in FIG. 10, the outflow angle was -66.8 degrees and the loss coefficient was 0.0406. Therefore, it can be seen that the cross-sectional shape of the new turbine blade is formed to satisfy the input target outflow angle = -67.2 degrees and loss coefficient = 0.0394. In addition, in the cross-sectional shape of the turbine blade in (b) in Fig. 10, the outflow angle = -62.0 degrees and the loss coefficient = 0.0389, and the new blade cross-sectional shape is the outflow angle = -62 which was input as the target. It can be seen that it is formed to satisfy the following conditions: .3 degrees and a loss coefficient of 0.0394. In this way, it becomes possible to shorten the design time.

With the shape generation device 100 and shape generation method described above, a new shape (shape data) can be generated with the goal of satisfying a predetermined performance value. Therefore, even if a predetermined performance value is required, a shape that satisfies the target performance value can be easily and quickly identified, making it possible to shorten the design time.

Although one embodiment has been described above with reference to the accompanying drawings, it goes without saying that the present disclosure is not limited to the above embodiment. It is clear that those skilled in the art can come up with various modifications or modifications within the scope of the claims, and it is understood that these naturally fall within the technical scope of the present disclosure. be done.

For example, in the embodiments described above, fluid components that utilize fluids such as air or water, such as aircraft wings and turbine blades, have been described as objects of mechanical design. However, the generative model is not limited to the object and can generate various shapes.

Furthermore, in the above-described embodiments, the cross-sectional shape, that is, the two-dimensional shape, was used as the shape. However, the present invention is not limited to this case, and can be applied to three-dimensional shapes as long as the shape can be quantified as some parameter. In this case, the above point sequence data corresponds to the surface of a three-dimensional shape instead of the outline of a two-dimensional shape.

In addition, in the above-described embodiment, an example of using CVAE among VAE, CVAE, GAN, and Conditional GAN is given as the generative model, but the generative model is not limited to CVAE, and for example, Conditional GAN may be used. I can do it. When using Conditional GAN instead of CVAE, the machine learning section 110 and shape generation section 120 are operated as follows.

FIG. 11 is an explanatory diagram for explaining the general configuration of the machine learning section 110. The machine learning section 110 includes a performance value derivation section 112, a generator learning section 214, a selection section 216, and a discriminator learning section 218. Here, solid arrows indicate the flow of numerical values, and dashed arrows indicate the direction of control.

As described using FIG. 3, the performance value deriving unit 112 extracts learning shape data of iteration 1 from multiple types of learning shapes 10 representing the target object prepared as samples, and performs CFD calculation. etc., the performance value (learning performance value) of the learning shape 10 is derived.

The generator learning unit 214 uses a neural network to machine learn a generator 214a that generates new shape data using a random vector and a learning performance value as input nodes. Such generator 214a plays a role similar to decoding module 116a in FIG. 3.

The selection unit 216 selects either the combination of the learning shape data and the learning performance value derived by the performance value deriving unit 112, or the combination of the shape data and the learning performance value generated by the generator 214a. Select randomly.

The classifier learning unit 218 identifies (determines) whether the combination of shape data and performance values selected by the selection unit 216 is an actual one (genuine) or one generated by the generator 214a (fake). Machine learning is performed on a discriminator 218a.

In this way, the classifier 218a plays a role in distinguishing between the shape data generated by the generator 214a and the actual learning shape data. At this time, the generator 214a performs machine learning so as to "deceive" the discriminator 218a, and the discriminator 218a performs machine learning so as not to be "deceived" by the generator 214a. As machine learning progresses, the generator 214a will be able to generate shape data that is substantially equivalent to actual learning shape data. Conditional GAN differs from the above-mentioned CVAE in that it does not explicitly extract and restore feature quantities.

FIG. 12 is an explanatory diagram for explaining the schematic configuration of the shape generation section 120. The shape generation unit 120 refers to the learned generator 214a and generates target shape data from the random vector and the target performance value. Here, no matter what value is input as the random vector, theoretically a shape that satisfies the target performance value should be derived. Therefore, the designer adjusts the shape while checking the performance value by changing the random vector.

In this way, even when Conditional GAN is used, a new shape (shape data) can be generated with the goal of satisfying a predetermined performance value. Therefore, even if a predetermined performance value is required, a shape that satisfies the target performance value can be easily and quickly identified, making it possible to shorten the design time.

The present disclosure can be used in a shape generation device and a shape generation method.

100 Shape generation device 110 Machine learning section 112 Performance value derivation section 114 Encoding module learning section 114a Encoding module 116 Decoding module learning section 116a Decoding module 120 Shape generation section 122 Decoding module processing section 124 Smoothing processing section 214 Generator learning section 214a Generator 216 Selection unit 218 Discriminator learning unit 218a Discriminator

Claims (4)

an encoding module that generates a feature amount by encoding learning shape data and a learning performance value corresponding to the learning shape data; and an encoding module that decodes the feature amount and the learning performance value to generate the learning feature value. a decoding module that generates shape data corresponding to the shape data for use, and inputs a target performance value and an arbitrary feature quantity to the decoding module in a generative model that has undergone machine learning multiple times to generate target shape data. Equipped with a shape generator,
The learning shape data includes point sequence data corresponding to the outer shape of the learning shape based on the learning shape data,
The point sequence data has a relatively high density in a region where the learning shape has a high curvature, and a relatively low density in a region where the learning shape has a low curvature. an encoding module that generates a feature amount by encoding learning shape data and a learning performance value corresponding to the learning shape data; and an encoding module that decodes the feature amount and the learning performance value to generate the learning feature value. a decoding module that generates shape data corresponding to the shape data for use, and inputs a target performance value and an arbitrary feature quantity to the decoding module in a generative model that has undergone machine learning multiple times to generate target shape data. a decryption module processing unit;
a smoothing processing unit that smoothes a target shape based on the generated target shape data;
Equipped with
The smoothed target shape data includes point sequence data corresponding to the outer shape of the target shape,
The point sequence data has a relatively high density in an area where the target shape has a high curvature, and a relatively low density in an area where the target shape has a low curvature. The computer is
an encoding module that generates a feature amount by encoding learning shape data and a learning performance value corresponding to the learning shape data; and an encoding module that decodes the feature amount and the learning performance value to generate the learning feature value. a decoding module that generates shape data corresponding to the shape data for use, and inputs a target performance value and an arbitrary feature quantity to the decoding module in a generative model that has undergone machine learning multiple times to generate target shape data. A shape generation method,
The learning shape data includes point sequence data corresponding to the outer shape of the learning shape based on the learning shape data,
The point sequence data has a relatively high density in an area where the learning shape has a high curvature, and a relatively low density in an area where the learning shape has a low curvature. The computer is
an encoding module that generates a feature amount by encoding learning shape data and a learning performance value corresponding to the learning shape data; and an encoding module that decodes the feature amount and the learning performance value to generate the learning feature value. The method includes a decoding module that generates shape data corresponding to the shape data for use, and generates target shape data by inputting target performance values and arbitrary feature quantities to the decoding module in the generative model that has been subjected to machine learning multiple times. ,
A shape generation method for smoothing a target shape based on the generated target shape data ,
The smoothed target shape data includes point sequence data corresponding to the outer shape of the target shape,
The point sequence data has a relatively high density in an area where the target shape has a high curvature, and a relatively low density in an area where the target shape has a low curvature.

Images