KR102079027B1 - Method for topology optimization using deep learning - Google Patents

Abstract

The present invention relates to an improved method of phase optimization design using deep learning technology. The present invention comprises the steps of inputting a node variable consisting of a plurality of channels, an element variable consisting of a plurality of channels and at least one scalar variable, performing a neural network for the node variable and the element variable, and the result of the neural network And calculating the latent variable by combining the scalar variable and the first variable, and calculating the first phase design result by decoding the latent variable, wherein the calculating of the latent variable comprises the Neural Network result value and the scalar variable. Calculating a mean value vector and a standard deviation vector, and calculating the latent variable by combining the mean value vector and the standard deviation vector. to provide.

Description

Topology Optimization Using Deep Learning Technology {METHOD FOR TOPOLOGY OPTIMIZATION USING DEEP LEARNING}

The present invention relates to an improved method of phase optimization design using deep learning technology.

Topology optimization technology is a design methodology that can design a structure with infinite degrees of freedom. However, the following limitations exist in the current phase optimization design methodology. First, a lot of time is usually required to find the detailed structure through the optimal design process. Second, in order to proceed with optimal design, in most cases, calculation of mathematical sensitivity is required. Therefore, the variables that can be used in the design are limited. Third, post-processing is required for the design result in order to obtain a smooth shape, and often the processing of the result is impossible.

The present invention proposes an improved phase optimization design technique than the conventional phase optimization design technique by applying various deep learning techniques recently developed.

 Kingma and Welling, “Auto-Encoding Variational Bayes”, ICLR, 2014

It is an object of the present invention to provide a method that can shorten the design time of a phase optimized design using a deep learning technique.

In addition, an object of the present invention is to provide a phase optimization design method using the variable learning freely using the variable.

In addition, an object of the present invention is to provide a phase optimization design method capable of processing the result.

According to an aspect of the present invention to achieve the above or another object, the present invention comprises the steps of inputting a node variable consisting of a plurality of channels, an element variable consisting of a plurality of channels and at least one scalar variable, the node variable and the Performing a neural network on an element variable, calculating a latent variable by combining the neural network result value and the scalar variable, and decoding the latent variable to calculate a first phase design result; The calculating of the latent variable includes combining the Neural Network result and the scalar variable, calculating an average vector and a standard deviation vector, and combining the average vector and the standard deviation vector to calculate the latent variable. Phase design method using deep learning technology characterized in that The.

In an embodiment, the performing of the neural network on the node variable and the element variable may include performing a neural network on the node variable, performing a neural network on the element variable, and performing the node variable. The method may include performing a neural network by combining a neural network result for and a neural network result for the element variable.

In one embodiment, the node variable defines at least one of a force and a boundary condition, the element variable defines a region with a substance and a region without a substance, and the scalar variable relates to information related to a substance and a filter. Information and penalty constants can be defined.

In an embodiment, the step of calculating the first phase design result by decoding the latent variable may be performed through a De-Convolution Neural Network.

According to an embodiment, the present invention may include calculating a second phase design result from the first phase design result and calculating a similarity between the second phase design result and the actual phase design data. The phase design result can be recalculated using the similarity.

In one embodiment, the calculating of the second phase design result and the calculating of the similarity between the second phase design result and the actual phase design data are repeatedly performed, and the calculated similarity is determined by the second phase. It may be used in at least one of calculating a design result and calculating the similarity.

According to the present invention, since the optimal structure can be found instantaneously through the repeatedly learned artificial neural network, the design time can be shorter than the conventional optimal design technique.

On the other hand, some differences (errors) may occur with the structure obtained through the phase optimization design performed according to the present invention. If you need to find a more accurate structure, if you start the optimal design of the shape inferred by the neural network with the initial shape and efficiently tune the topological optimization algorithm through reinforcement learning, you can get the optimal shape significantly faster than the existing methodology. have.

1 is a conceptual diagram illustrating a Variational Autoencoder (VAE).
2 is a conceptual diagram for calculating a first phase design result according to the present invention.
3 is a conceptual diagram illustrating an input variable including a plurality of channels.
4 is a conceptual diagram illustrating a process of calculating a latent variable from input variables according to the present invention.
5 is a conceptual diagram for calculating a first phase design result from a latent variable according to the present invention.
6 is a result of calculating the first phase design result by the method according to the present invention.
7 is a conceptual diagram illustrating a Generative Adversarial Network (GAN).
8 and 9 are conceptual views illustrating an embodiment using a GAN structure.
10 is a result of calculating the second phase design result by the method according to the present invention.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings, and the same or similar components will be given the same reference numerals regardless of the reference numerals, and redundant description thereof will be omitted. In the following description of the embodiments disclosed herein, when it is determined that the detailed description of the related known technology may obscure the gist of the embodiments disclosed herein, the detailed description thereof will be omitted. In addition, the accompanying drawings are intended to facilitate understanding of the embodiments disclosed herein, but are not limited to the technical spirit disclosed herein by the accompanying drawings, all changes included in the spirit and scope of the present invention. It should be understood to include equivalents and substitutes.

Hereinafter, a phase design method according to the present invention will be described. According to the present invention, three types of phase design results can be obtained. For convenience of explanation, the above three kinds of results are referred to as first and second phase design results.

The first phase design result may be calculated by introducing predetermined input values into the first algorithm, and the second phase design result may be calculated by introducing the first phase design result into the second algorithm.

Hereinafter, a method of calculating each of the first and second phase design results will be described.

First, the step of calculating the first phase design result will be described. Prior to this, Variational Autoencoder (VAE) will be described.

As shown in FIG. 1, the general VAE structure has the same input and output images, and a network having a probability distribution having a latent vector defined while minimizing an error between the input and the output is learned. If the unsupervised learning is carried out by inserting a number of phase-optimal design results into the input / output stage of the VAE, it is possible to accurately define the structure generated according to all the optimal design conditions with a small number of latent vectors. This technique can be applied as a dimension reduction technique of engineering, which can effectively perform the design.

In the present invention, the structure described in FIG. 1 is modified, and as shown in FIG.

In this case, the loss function used for network learning is defined as minimizing the error between the network generated structure and the structure used for learning. The training data is generated by a phase-optimal design code, and randomly determines information such as force conditions, boundary conditions, and material ratios to generate data for learning to generate various data. If this VAE network is well designed and successfully learned, it is possible to instantaneously find the optimal structure without the topology optimization process.

The information at the input consists of node variables, element variables and scalar variables. Here, the node variable and element variable may be composed of a plurality of channels. On the other hand, the node variable defines at least one of force and boundary conditions, and the element variable defines an area with material and an area without material. In addition, the scalar variable defines information related to a substance, information related to a filter, and information related to a penalty constant. However, the present invention is not limited thereto, and the node variable, the element variable, and the scalar variable may each define other information as necessary.

For example, in order to make any optimal design condition into a form that can be learned, conditions such as force, boundary condition, and the like can be discretized. In detail, in the node illustrated in FIG. 3, a magnitude value of a force is input to a node to which a force is added, and 0 is input to an area where no force is added. On the other hand, the outermost node inputs a value of -1 as the boundary condition.

Then, the variables are analyzed through Neural Network. Here, the neural network may be any one of a convolutional neural network (CNN), a fully connected layer (FCL), and a recurrent neural network. In the present specification, a CNN is described as an example of an neural network, but is not limited thereto.

Meanwhile, a plurality of CNNs for processing each of the node variable and the element variable may be provided. If necessary, the CNNs for each of the node variables and the element variables are combined not to be combined at an output stage but to be able to share information with each other in a CNN multilayer section.

For example, when the grid used for the topology optimization design is (N, M), the force and boundary conditions are added to the node, thus forming a (N + 1, M + 1) sized network to represent node information. do. In addition, a network of (N, M) size is configured to define grid information such as the presence or absence of a substance. Finally, we construct a scalar network for information such as the volume fraction, filtering radius, and penalty method constants used in the variables of the phase optimization design.

On the other hand, as shown in Figure 4, the step of performing a CNN for each of the node variable and the element variable may be performed. Thereafter, CNN may be performed by combining the CNN result value for the node variable and the CNN result value for the element variable. The latent variable may be calculated by combining the scalar variable and the CNN result calculated in the above-described manner.

The calculating of the latent variable may include combining the CNN result, calculating an average vector and a standard deviation vector, and combining the average vector and the standard deviation vector to calculate the latent variable.

The latent variable may then be decoded via a decoder network, as shown in FIG. 5, to derive the first phase design result. Here, the step of decoding the latent variable and calculating the phase design result may be performed through a De-Convolution Neural Network.

The decoder generates two-dimensional or three-dimensional information according to the information defined in the latent variable. The first phase design result calculated by the decoder is shown in FIG. 6.

The first phase design result described above is used to calculate the second phase design result through the method described below.

Prior to describing the method of calculating the second phase design result, a description will be given of the Generative Adversarial Network (GAN) technology.

As shown in FIG. 7, the GAN is a technology in which an image generating network and a network evaluating the generated image evolve competitively to create an increasingly realistic picture. This technology has been applied to the fields such as converting an original picture into a new style picture or creating new content.

Hereinafter, a method of obtaining a fine second phase design result from the first phase design result by applying the GAN technique will be described. The present invention utilizes producer networks and discriminators.

Specifically, as shown in FIG. 7, the generator network receives the first phase design result as an input and calculates a second phase design result based on the input. In addition, the discriminator receives the actual design data and the second phase design result in pairs, and serves to determine whether the second phase design result or the actual design data is received from the input. The loss function of the generator network is defined to be smaller as it produces outputs that are indistinguishable from the discriminator, and the loss function of the discriminator is smaller as it better distinguishes the second-phase design results from the actual design data. Is defined.

The learning process of the entire network is similar to the general GAN, where the distinguisher learns the distinction of the product / product, and the producer learns to produce a fine grid product that is difficult for the discriminator to distinguish. Thus, as learning continues, the two networks compete with each other to improve, so that the second phase design results generated by the generator network are very similar to the topological design results in a fine grid. In an embodiment, the above-described GAN structure may be utilized as shown in FIGS. 8 and 9. An example of the result generated by learning using the above structure is shown in FIG. 10.

To further develop these results, GAN technology could be used to control the characteristics of structures. For example, in the case of face photos, it is possible to control the angle, facial expression, and pupil shape of the face by adjusting the input variables of the network generated by the GAN. The result of the topology optimization is that there are design parameters such as sensitivity filtering radius as well as load and boundary conditions. The characteristics of these information can be extracted to infer the change of the structure without performing new topology optimization. Can be.

Claims (7)

Inputting, by an artificial neural network, a node variable consisting of a plurality of channels, an element variable consisting of a plurality of channels, and at least one scalar variable;
Performing, by the artificial neural network, a neural network for the node variable and the element variable;
Calculating, by the artificial neural network, a latent variable by combining the neural network result and the scalar variable; And
The artificial neural network decoding the latent variable to produce a first phase design result;
Computing the latent variable by the artificial neural network,
Calculating, by the artificial neural network, the mean value vector and the standard deviation vector by combining the neural network result and the scalar variable; And
The artificial neural network includes calculating the latent variable by combining the mean value vector and the standard deviation vector,
The neural network performing the neural network for the node variable and the element variable,
The neural network performing the neural network on the node variable;
The neural network performing a neural network on the element variable; And
The artificial neural network performing a neural network by combining a neural network result value for the node variable and a neural network result value for the element variable;
The node variable defines at least one of force and boundary condition,
The elemental variable defines a region with and without a substance,
The scalar variable is a phase design method using deep learning technology, characterized in that for defining information related to the material, filter related information, penalty information. delete delete The method of claim 1,
The artificial neural network decoding the latent variable to produce a first phase design result,
Topology design method using deep learning technology, which is carried out through De-Convolution Neural Network. The method of claim 1,
Calculating, by the artificial neural network, a second phase design result from the first phase design result; And
Calculating, by the artificial neural network, a similarity degree between the second phase design result and actual phase design data;
And the second phase design result is recalculated using the similarity. The method of claim 5,
The artificial neural network calculating the second phase design result and the artificial neural network calculating the similarity between the second phase design result and the actual phase design data are repeatedly performed;
The calculated similarity is
And wherein the artificial neural network is used in at least one of calculating the second phase design result and the artificial neural network calculating the similarity. The method of claim 1,
The neural network is a topology design method using deep learning technology, characterized in that the convolutional neural network.

Images