Information processing device, information processing method, and information processing program

Abstract

To provide an information processing device capable of spending less time on CAE calculation.SOLUTION: There is provided an information processing device comprising: a latent vector estimation unit which outputs a latent vector, which is a vector representing a feature of a shape, from a first learned model on the basis of an entered shape parameter; a decoder unit which generates a collection of coordinates of the shape corresponding to the latent vector from a second learned model; and a distribution estimation unit which estimates, with the use of the second learned model, the distribution of feature amounts of the shape on the basis of the latent vector and the collection of coordinates.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This disclosure relates to an information processing device, an information processing method, and an information processing program. [Background technology]

[0002] In the field of mechanical engineering technology development, CAE (Computer-Aided Engineering) calculations are performed to simulate performance values ​​such as stress distribution in objects with a certain three-dimensional shape. Non-patent documents 1 to 3 disclose simulation methods that use machine learning to generate the shape of a three-dimensional structure and predict performance values ​​such as stress distribution, stiffness value, and Cd (Constant Drag) value of the three-dimensional structure. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Haoqiang Fan, Hao Su, Leonidas J. Guibas. "A Point Set Generation Network for 3D Object Reconstruction from a Single Image", In Proceedings of the IEEE conference on computor vision and pattern recognition, pp. 605-613. 2017. [Non-Patent Document 2] Wen Xiao, Zhenan Fan, Qiuyan Liu. "Point Cloud Generation via Variational Auto-encoder" [Non-Patent Document 3] Nikita Durasov, Artem Lukoyanov, Jonathan Donier, Pascal Fua. "DEBOSH: Deep Bayesian Shape Optimization", arXiv preprint arXiv:2109.13337 (2021).

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, conventional CAE calculations require a very large amount of calculation time and calculation resources, and it is required to reduce the time spent on CAE calculations.

[0005] The present disclosure has been made in view of the above points, and an object thereof is to provide an information processing apparatus, an information processing method, and an information processing program capable of reducing the time required for CAE calculations as compared with the prior art.

Means for Solving the Problems

[0006] A first aspect of the present disclosure is an information processing apparatus including a latent vector estimation unit that outputs, from a first learned model, a latent vector that is a vector representing the characteristics of a shape based on input shape parameters, a decoder unit that generates, from a second learned model, a set of coordinates of a shape corresponding to the latent vector, and a distribution estimation unit that estimates, by the second learned model, a distribution of feature amounts of the shape based on the latent vector and the set of coordinates.

[0007] According to the first aspect of the present disclosure, it is possible to obtain the distribution of feature amounts with respect to a shape without having to redo CAE calculations from scratch.

[0008] A second aspect of the present disclosure is the information processing apparatus according to the first aspect, further including a performance value estimation unit that estimates, by a second learned model, a performance value of the shape based on the latent vector.

[0009] According to a second aspect of this disclosure, performance values ​​for a given shape can be obtained without having to redo the CAE calculation from scratch.

[0010] A third aspect of this disclosure is an information processing device according to the second aspect, further comprising a presentation unit that presents information relating to the performance value of the shape estimated by the performance value estimation unit.

[0011] According to a third aspect of this disclosure, it becomes possible to allow the user of the information processing device to confirm the performance values ​​for the shape.

[0012] A fourth aspect of this disclosure is an information processing method, wherein a processor causes a first trained model to output a latent vector, which is a vector representing the features of a shape, based on input shape parameters; generates a set of coordinates of the shape corresponding to the latent vector from a second trained model; and performs a process in which the second trained model estimates the distribution of the feature quantities of the shape based on the latent vector and the set of coordinates.

[0013] According to the fourth aspect of this disclosure, the distribution of feature quantities with respect to shape can be determined without having to redo the CAE calculation from scratch.

[0014] A fifth aspect of this disclosure is an information processing program which causes a computer to output a latent vector, which is a vector representing the features of a shape, from a first trained model based on input shape parameters; generate a set of coordinates of the shape corresponding to the latent vector from a second trained model; and execute a process in which the distribution of the feature quantities of the shape is estimated by the second trained model based on the latent vector and the set of coordinates.

[0015] According to the fifth aspect of this disclosure, the distribution of feature quantities with respect to shape can be determined without having to redo the CAE calculation from scratch. [Effects of the Invention]

[0016] According to this disclosure, it is possible to provide an information processing device that can reduce the time spent on CAE calculations compared to conventional technology. [Brief explanation of the drawing]

[0017] [Figure 1] This figure shows an example configuration of an information processing device according to an embodiment of the disclosed technology. [Figure 2] This figure shows a specific example of the configuration of an information processing device during the learning phase. [Figure 3] This figure shows a specific example of the encoder unit's configuration. [Figure 4] This figure shows a specific example of the decoder unit's configuration. [Figure 5] This figure shows a specific example of the configuration of the distribution estimation unit. [Figure 6] This figure shows a specific example of the configuration of an information processing device during the inference phase. [Modes for carrying out the invention]

[0018] Hereinafter, an example of an embodiment of this disclosure will be described with reference to the drawings. In each drawing, identical or equivalent components and parts are given the same reference numerals. Also, the dimensional ratios in the drawings are exaggerated for illustrative purposes and may differ from the actual ratios.

[0019] Figure 1 shows an example of the configuration of an information processing device according to this embodiment. The information processing device 10 shown in Figure 1 is a device that applies deep learning to quickly calculate feature distributions and performance values ​​for a three-dimensional shape. In the information processing device 10, the CPU 11 functions as an encoder unit 101, a decoder unit 102, a distribution estimation unit 103, a latent vector estimation unit 104, a performance value estimation unit 105, and a presentation unit 106. Features are values ​​such as pressure and fluid force.

[0020] The CPU 11 is a central processing unit that executes various programs and controls each component. Specifically, the CPU 11 reads a program from the ROM 12 or storage 14 and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various calculations according to the program recorded in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores an information processing program that applies deep learning and performs information processing to quickly calculate feature distributions and performance values ​​for three-dimensional shapes.

[0021] ROM12 stores various programs and data. RAM13 temporarily stores programs or data as a working area. Storage14 consists of a storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory, and stores various programs, including the operating system, and various data.

[0022] When executing the above information processing program, the information processing device 10 uses the above hardware resources to implement various functions. The functional configuration implemented by the information processing device 10 will now be described.

[0023] As shown in Figure 1, the information processing device 10 has the following functional configuration: an encoder unit 101, a decoder unit 102, a distribution estimation unit 103, a latent vector estimation unit 104, a performance value estimation unit 105, and a presentation unit 106. Each functional configuration is realized by the CPU 11 reading and executing an information processing program stored in the ROM 12 or storage 14.

[0024] The encoder unit 101 generates a latent vector z that represents the features of the target three-dimensional shape using a trained model. The decoder unit 102 reconstructs the target three-dimensional shape from the latent vector z generated by the encoder unit 101 using the trained model. The encoder unit 101 and the decoder unit 102 form a variational autoencoder, and by connecting a neural network that estimates the feature distribution to the variational autoencoder, the information processing device 10 can learn both the three-dimensional shape and the feature distribution simultaneously.

[0025] The distribution estimation unit 103 estimates the feature distribution of the three-dimensional shape using a trained model. The latent vector estimation unit 104 estimates latent vectors from the shape parameters of the target three-dimensional shape using a trained model. The performance value estimation unit 105 estimates the performance value of the target three-dimensional shape using a trained model. The presentation unit 106 presents the performance value of the three-dimensional shape estimated by the performance value estimation unit 105. The presentation unit 106 presents the performance value of the three-dimensional shape estimated by the performance value estimation unit 105, for example, in the form of a GUI (Graphical User Interface).

[0026] The information processing device 10 is characterized by taking a set of vertex coordinates of a three-dimensional shape as input, learning a model that outputs a set of feature quantities for each vertex, and using the learning results to input the set of vertex coordinates of the three-dimensional shape into the model, thereby outputting performance values ​​of the three-dimensional shape from the model. The specific configuration of the information processing device 10 will be described below.

[0027] (Learning Phase) First, we will explain the specific configuration of the information processing device 10 in the learning phase. Figure 2 is a diagram showing a specific example of the configuration of the information processing device 10 in the learning phase. In the following explanation, in mathematical formulas, characters with a "~" attached to a symbol (for example, X) may be represented as ~X, etc.

[0028] The encoder unit 101 is a set of vertex coordinates {x i} i=1,2,3,···,NAn encoder NNq that takes [input] and outputs a latent vector z that represents the shape features φ is p(z|x). Note that NN stands for Neural Network.

[0029] The decoder part 102 is a decoder NNp that takes the latent vector z as input and outputs a set of vertex coordinates of the three-dimensional shape {~x i} i=1,2,3,···,N is p(x|z). θ

[0030] The distribution estimation part 103 is a neural network f(x,z) that takes a set of vertex coordinates of an arbitrary three-dimensional shape {x i} i=1,2,3,···,N and a pair of the latent vector z of the three-dimensional shape as input, and outputs feature quantities {~y i} i for each vertex x i=1,2,3,···,N is p(y|x,z). η

[0031] The latent vector estimation part 104 is a regression model g(v) that takes the shape parameter v of the three-dimensional shape, which is used when generating a shape with CAD software or the like, as input and outputs the latent vector z of the three-dimensional shape.

[0032] The performance value estimation part 105 is a regression model h(z) that takes the latent vector z of the three-dimensional shape as input and outputs the performance value e of the three-dimensional shape.

[0033] Among these parts, the encoder part 101, the decoder part 102, and the distribution estimation part 103 are composed of one neural network, and the latent vector estimation part 104 and the performance value estimation part 105 are independent regression models, respectively.

[0034] FIG. 3 is a diagram showing a specific configuration example of the encoder part 101. The encoder part 101 applies a multi-layer perceptron MLP1 to each element of a set of vertex coordinates of an arbitrary three-dimensional shape {x i} i=1,2,3,···,N to obtain a feature vector x' i =MLP1(x i ​​) is calculated. The set of feature vectors for each point is {x' i} i=1,2,3,···,N The configuration of the multilayer perceptron MLP1 is as follows: 3 input layers, 64 output layers, and 64,128,128,256 hidden layers.

[0035] Next, the encoder unit 101 processes the set of feature vectors into {x' i} i=1,2,3,···,N Find the average value of the elements and let it be the vector z'.

[0036]

number

[0037] Next, the encoder unit 101 applies two multilayer perceptrons, MLP2 and MLP3, to the vector z', and their outputs are vectors ρ and μ. The multilayer perceptrons MLP2 and MLP3 have 64 input layers, 64 output layers, and 128 hidden layers. By combining these vectors σ and μ with a vector ε whose elements are random values ​​obtained according to a standard normal distribution N(0,1), we obtain the latent vector z. z = σ * ε + μ This is how it is calculated. Note that the operator * represents element-wise multiplication.

[0038] Each multilayer perceptron is given as follows, where M is the number of hidden layers:

[0039]

number

[0040] In the above formula, ξ i This is the vector of the i-th layer, FC i LayerNorm is a fully connected layer that takes the i-th layer as input and outputs to the (i+1)th layer, LayerNorm is a layer normalization layer, and GELU is a Gaussian Error Linear Unit.

[0041] Figure 4 shows a specific example of the decoder unit 102. The decoder unit 102 applies a multilayer perceptron MLP4 to the latent vector z to obtain a vector ~X with 3N elements. The multilayer perceptron MLP4 has 64 input layers, 3N output layers, and 256 hidden layers. The multilayer perceptron MLP4 is given by the formula shown above.

[0042] Next, the decoder unit 102 rearranges the vector ~X into an N×3 matrix (reshape), and assigns the vertex coordinates of the shape ~x to each row. i The set of vertex coordinates of a three-dimensional shape {~x i} i=1,2,3,···,N We seek.

[0043] Figure 5 shows a specific example of the configuration of the distribution estimation unit 103. The distribution estimation unit 103 is a set of vertex coordinates of an arbitrary three-dimensional shape {x i} i=1,2,3,···,N For each element, the latent vector z obtained by the decoder unit 102 is combined to form the set {(x i ,z)} i=1,2,3,···,N Next, the distribution estimation unit 103 creates the set {(x i ,z)} i=1,2,3,···,N A multilayer perceptron MLP5 is applied to each element, and its output is ~y i Let the coordinate x i The features are defined as follows. The configuration of the multilayer perceptron MLP5 is such that, if the dimension of the features to be estimated is d, there are 67 input layers, d output layers, and 512,256,256 hidden layers.

[0044] In the learning phase, the information processing device 10 determines the parameters of the functions of each of these parts using pre-prepared training data. That is, the set of vertex coordinates {x i}, a set of features obtained by CAE calculations etc. {y i The parameters of the functions of each part are determined using a set of shape parameters v and performance values ​​e obtained from CAE calculations, etc., as training data.

[0045] In the learning phase, first, deep learning is performed to minimize the loss function Loss, and the function q of the encoder unit 101 is improved. φ , the function p of the decoder unit 102 θ , the function f of the distribution estimation unit 103 η To decide.

[0046]

number

[0047] In the above formula, D KL is the Kullback-Leibler divergence, p is a standard normal distribution, and w1 and w2 are non-negative real numbers.

[0048] Next, the function q of the encoder unit 101 is obtained from the shape parameter v. φ The function g(v) for obtaining the latent vector z from is determined using a regression model. Also, the function q of the encoder unit 101 φ The function h(e) for calculating the performance value e from the latent vector z obtained is determined using a regression model.

[0049] The information processing device 10 can infer performance values ​​e from shape parameters v by determining the functions of each part in the learning phase as described above.

[0050] (Inference Phase) Next, we will describe the specific configuration of the information processing device 10 in the inference phase. Figure 6 is a diagram showing a specific example of the configuration of the information processing device 10 in the inference phase. In the inference phase, the information processing device 10 uses the decoder unit 102, distribution estimation unit 103, latent vector estimation unit 104, and performance value estimation unit 105 obtained in the learning phase to obtain the set of coordinates of the vertices of the three-dimensional shape {~x} from the shape parameter v. i} i=1,2,3,···,N , a set of features corresponding to each vertex {~y i} i=1,2,3,···,N And calculate the performance value e.

[0051] The information processing device 10 obtains shape parameters v from an external source and inputs them to the latent vector estimation unit 104 to obtain the latent vector z. Subsequently, the information processing device 10 inputs the obtained latent vector z to the decoder unit 102 and obtains the set of coordinates of the vertices of the three-dimensional shape {~x i} i=1,2,3,···,N We seek.

[0052] Next, the information processing device 10 calculates the set of coordinates {~x i} i=1,2,3,···,N By combining each element with the latent vector z and inputting it into the distribution estimation unit 103, the set of features corresponding to each vertex {~y i} i=1,2,3,···,N The latent vector z is then calculated. The information processing device 10 also inputs the calculated latent vector z to the performance value estimation unit 105 to calculate the performance value e.

[0053] By operating in this manner, the information processing device 10 can predict how the properties of an object will change when the shape parameter v of that object is changed.

[0054] In this embodiment, the same information processing device 10 was used in both the learning phase and the inference phase, but this disclosure is not limited to this example. Different devices may be used in the learning phase and the inference phase. That is, the model learned in the learning phase may be used in a different device to infer the properties of an object. Furthermore, the functions of the information processing device 10 may be implemented by one or more computers. Furthermore, at least some of the functions of the information processing device 10 may be implemented by a virtual machine. Furthermore, at least some of the functions of the information processing device 10 may be implemented using cloud computing.

[0055] In methods that use CAE calculations to determine the distribution of features and performance values, it was necessary to redo the CAE calculation from scratch even for shapes within a defined range of shape parameters. The information processing device 10 according to this embodiment can generate similar shapes or shapes intermediate between two shapes by calculating latent vectors with the encoder unit 101 and generating shapes based on those latent vectors with the decoder unit 102. Furthermore, the information processing device 10 according to this embodiment can determine the distribution of features and performance values ​​for the newly generated shapes using the distribution estimation unit 103 and the performance value estimation unit 105 without having to redo the CAE calculation from scratch. In addition, by associating shape parameters with latent vectors, the information processing device 10 according to this embodiment does not need to generate shapes corresponding to the shape parameters, thus enabling the rapid determination of performance values, etc.

[0056] In this embodiment, the information processing device 10 performs machine learning by combining the encoder unit 101, the decoder unit 102, and the distribution estimation unit 103 into a single neural network. This network structure is a combination of the distribution estimation unit 103 and a variational autoencoder model for shape, which consists of the encoder unit 101 and the decoder unit 102. In the case of only the variational autoencoder model, only shape information is included in the latent vector. In contrast, in this embodiment, by adding the distribution estimation unit 103, information on the distribution of feature quantities is also included in addition to the shape information. Therefore, the information processing device 10 in this embodiment can improve the estimation accuracy of quantities such as performance values ​​related to the distribution of feature quantities.

[0057] The information processing device 10 according to this embodiment employs layer normalization layers in the normalization layers of each multilayer perceptron in the encoder unit 101, decoder unit 102, and distribution estimation unit 103. While batch normalization layers are commonly used as normalization layers, batch normalization layers normalize multiple data points as a single data point, requiring a computing environment with large memory for training. In contrast, layer normalization layers normalize each data point individually, enabling training using large data such as 3D data even in environments with limited computing resources. Furthermore, by employing layer normalization compared to batch normalization layers, the information processing device 10 according to this embodiment enables more accurate training.

[0058] Furthermore, the information processing that the CPU reads and executes in each of the above embodiments may be executed by various processors other than the CPU. Examples of such processors include PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) whose circuit configuration can be changed after manufacturing, and dedicated electrical circuits that are processors with circuit configurations specifically designed to execute specific processing, such as ASICs (Application Specific Integrated Circuits). In addition, the information processing may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (for example, multiple FPGAs, and a combination of a CPU and an FPGA). More specifically, the hardware structure of these various processors is an electrical circuit that combines circuit elements such as semiconductor elements.

[0059] Furthermore, while the above embodiments describe a configuration in which the information processing program is pre-stored (installed) in ROM or storage, the invention is not limited to this. The program may be provided in a form recorded on a non-transitory recording medium such as a CD-ROM (Compact Disk Read Only Memory), DVD-ROM (Digital Versatile Disk Read Only Memory), or USB (Universal Serial Bus) memory. Alternatively, the program may be provided in a form that is downloaded from an external device via a network. [Explanation of Symbols]

[0060] 10 Information Processing Devices 101 Encoder section 102 Decoder section 103 Distribution estimation part 104 Latent Vector Estimation Unit 105 Performance Value Estimation Unit 106 Presentation section

Claims

1. A latent vector estimation unit outputs a latent vector, which is a vector representing the shape features, from a first trained model based on the input shape parameters. A decoder unit that generates a set of coordinates of the shape corresponding to the aforementioned latent vector from a second trained model, A distribution estimation unit that estimates the distribution of the shape features using the second trained model based on the latent vector and the set of coordinates, An information processing device equipped with the following features.

2. The information processing apparatus according to claim 1, further comprising a performance value estimation unit that estimates the performance value of the shape based on the latent vector using a second trained model.

3. The information processing apparatus according to claim 2, further comprising a presentation unit that presents information regarding the performance value of the shape estimated by the performance value estimation unit.

4. The processor, Based on the input shape parameters, the first trained model outputs a latent vector, which is a vector representing the shape features. A set of coordinates for the shape corresponding to the aforementioned latent vector is generated from the second trained model. Based on the latent vector and the set of coordinates, the distribution of the shape features is estimated by the second trained model. An information processing method that executes a process.

5. On the computer, Based on the input shape parameters, the first trained model outputs a latent vector, which is a vector representing the shape features. A set of coordinates for the shape corresponding to the aforementioned latent vector is generated from the second trained model. Based on the latent vector and the set of coordinates, the distribution of the shape features is estimated by the second trained model. An information processing program that executes a process.

Images