PREDICTION MODEL LEARNING METHOD, PREDICTION MODEL, AND LEARNING DEVICE

Abstract

To provide a learning method, a prediction model learned by the learning method, and a learning device that can update a parameter of the prediction model which predicts aerodynamic performance of an object considering a shape of the object.SOLUTION: A processor 10 that is a learning device sets a virtual space A including a vehicle V and a virtual cross section in which the virtual space A is cut off, calculates a virtual area obtained by eliminating an area of a cross section of the vehicle V from a total area of the virtual cross section, calculates an error of prediction data of a prediction model 100 relative to teacher data, and updates a parameter of the prediction model 100 to cause an evaluation value derived from a loss function to be close to a minimum value, on the basis of the virtual area and the error of the prediction data relative to the teacher data in the virtual cross section, and the loss function is defined such that the larger the error of the prediction data relative to the teacher data in the virtual cross section is, the larger the evaluation value is, and the larger the virtual area is, the smaller the evaluation value is.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a learning method for a prediction model that predicts the aerodynamic performance of an object, a trained prediction model trained by the learning method, and a learning device. [Background technology]

[0002] The prediction model described in Patent Document 1 predicts the aerodynamic performance of a vehicle based on feature quantities extracted from design data of the vehicle. Patent Document 1 also describes a learning method that calculates an evaluation value that quantitatively indicates the suitability of the prediction model based on the predicted value of the aerodynamic performance predicted by the prediction model and the aerodynamic performance measured by CFD (simulation analysis), and updates the parameters of the prediction model using the evaluation value. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2014-006812 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, the method of learning the predictive model described in Patent Document 1 does not take into account the shape of the vehicle when calculating the evaluation value, which makes it difficult to improve the accuracy of predicting aerodynamic performance using the predictive model.

[0005] The problem that the present invention aims to solve is to provide a learning method for updating parameters of a predictive model that predicts the aerodynamic performance of an object by taking into account the shape of the object, and a trained predictive model and learning device that have been trained using the learning method. [Means for solving the problem]

[0006] The present invention solves the above problem by setting a virtual space including an object and a virtual cross section cut through the virtual space, calculating a virtual area by subtracting the area of ​​the object's cross section from the total area of ​​the virtual cross section, and updating the parameters of a prediction model based on a loss function defined so that the larger the error of predicted data relative to training data, the larger the evaluation value, and the smaller the evaluation value, as the virtual area increases. Effect of the Invention

[0007] According to the present invention, an evaluation value is calculated according to the size of a virtual area reflecting the shape of an object, thereby achieving the effect of updating parameters of a predictive model that predicts the aerodynamic performance of an object, taking into account the shape of the object. [Brief description of the drawings]

[0008] [Figure 1] 1 is a block diagram showing a configuration of a learning device according to a first embodiment. [Diagram 2] 2 is a block diagram showing the configuration of a prediction model learned by the learning device shown in FIG. 1. [Diagram 3] FIG. 3(a) is a diagram showing an example of a virtual space set by the learning device shown in FIG. 1 and a virtual cross section perpendicular to direction X, and FIGS. 3(b) and 3(c) are diagrams showing the virtual areas of the two virtual cross sections shown in FIG. 3(a) by hatching. [Figure 4] FIG. 4(a) is a diagram showing an example of a virtual space set by the learning device shown in FIG. 1 and a virtual cross section perpendicular to direction Y, and FIGS. 4(b) and (c) are diagrams showing the virtual areas of the two virtual cross sections shown in FIG. 4(a) with diagonal lines. [Diagram 5] FIG. 5(a) is a diagram showing an example of a virtual space set by the learning device shown in FIG. 1 and a virtual cross section perpendicular to direction Z, and FIGS. 5(b) and (c) are diagrams showing the virtual areas of the two virtual cross sections shown in FIG. 5(a) with diagonal lines. [Figure 6] FIG. 11 is a block diagram showing the configuration of a learning device according to a second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. First Embodiment A method for learning the prediction model 100 by the processor 10 according to the first embodiment will be described with reference to FIGS. The processor 10 is a learning device that trains a prediction model 100 that predicts the aerodynamic performance of an object. The processor 10 compares prediction data output by the prediction model 100 with pre-stored teacher data, and evaluates the error of the prediction data relative to the teacher data based on a pre-defined loss function. The processor 10 then updates the parameters of the prediction model 100 so that the evaluation value derived from the loss function approaches a minimum value. That is, the learning method executed by the processor updates the parameters of the prediction model that predicts the aerodynamic performance of the object based on the loss function. Furthermore, the aerodynamic performance is an index that indicates the velocity distribution and pressure distribution of the flow field around the object. In the following description, the present invention will be described on the assumption that the object whose aerodynamic performance is predicted by the prediction model 100 is a vehicle.

[0010] The processor 10 includes a database 11, a virtual space setting unit 12, a virtual area calculation unit 13, an input unit 14, a prediction data acquisition unit 15, a teacher data acquisition unit 16, an error calculation unit 17, and a parameter update unit 18. The processor 10 is a computer equipped with a CPU, a ROM, and a RAM. The database 11 is stored in the memory of the processor 10. The virtual space setting unit 12, the virtual area calculation unit 13, the input unit 14, the prediction data acquisition unit 15, the teacher data acquisition unit 16, the error calculation unit 17, and the parameter update unit 18 are programs for executing the functions of the processor 10. The prediction model 100 may be a part of the program of the processor 10, or may be a program of hardware separate from the processor 10.

[0011] The database 11 stores vehicle shape data to be used for learning the prediction model 100. CAD data, which is design data of the vehicle, is converted into SDF data (SDF: Signed Distance Function) or voxel data (grid unit data) and stored in the database 11. That is, the vehicle shape data stored in the database 11 is SDF data or voxel data. The database 11 also stores teacher data including the vehicle shape data and aerodynamic performance associated with the shape data. The aerodynamic performance of the teacher data is data that has been analyzed in advance by computational fluid analysis based on the shape data of the object. The vehicle design data indicates the structure data of the vehicle.

[0012] The virtual space setting unit 12 sets a virtual space A based on the shape data of the vehicle acquired from the database 11 (see Figs. 3(a), 4(a), and 5(a)). The virtual space A is a space having a width in each of three mutually perpendicular directions X, Y, and Z. The virtual space setting unit 12 also sets a plurality of virtual cross sections Cx1, Cx2, Cy1, Cy2, Cz1, and Cz2 obtained by cutting the virtual space A in directions perpendicular to each of the three directions X, Y, and Z. Details of the method for setting the virtual space A and the virtual cross sections Cx1, Cx2, Cy1, Cy2, Cz1, and Cz2 will be described later.

[0013] Next, the virtual area calculation unit 13 calculates virtual areas Cxa1, Cxa2, Cya1, Cya2, Cza1, and Cza2 by subtracting the area of ​​the vehicle cross section from the total area of ​​the virtual cross sections Cx1, Cx2, Cy1, Cy2, Cz1, and Cz2 (see Figs. 3(b), (c), 4(b), (c), 6(b), (c)). The method of calculating the virtual areas will be described in detail later.

[0014] On the other hand, the input unit 14 inputs the vehicle shape data acquired from the database 11 to the prediction model 100. The prediction model 100 is a prediction model of aerodynamic performance that uses a neural network. The prediction model 100 outputs the aerodynamic performance of the vehicle to the prediction data acquisition unit 15 based on the input shape data.

[0015] As shown in FIG. 2, the prediction model 100 includes a shape data input unit 1, a feature extraction unit 2, a feature matching unit 3, an output result generation unit 4, and an output unit 5. The shape data input unit 1 acquires the shape data of the vehicle as SDF data or voxel data. When the original shape data is CAD data, the shape data of the vehicle is converted into SDF data or voxel data before being input to the shape data input unit 1. The feature extraction unit 2 extracts the feature amount of the shape data acquired from the shape data input unit 1. Specifically, the feature extraction unit 2 repeats three-dimensional convolution and activation function processing on the shape data of the vehicle to extract multiple intermediate layer features. The feature matching unit 3 matches the intermediate layer features extracted by the feature extraction unit 2 to generate a one-dimensional feature vector. The output result generating unit 4 generates the aerodynamic performance of the vehicle, i.e., the three components of the velocity (Ux, Uy, Uz) and the pressure (P) of the flow field around the vehicle, based on the intermediate layer features extracted by the feature extracting unit 2 and the one-dimensional feature vector generated by the feature matching unit 3. The three components of the velocity (Ux, Uy, Uz) correspond to the directions X, Y, and Z shown in Figs. 3(a), 4(a), and 5(a). The output result generating unit 4 also generates output results for each of the multiple unit space regions (three-dimensional regions in a cubic shape of 1 cm x 1 cm x 1 cm) constituting the virtual space A shown in Figs. 3(a), 4(a), and 5(a). Specifically, when the virtual space A is a space in a rectangular parallelepiped shape of 10 m x 4 m x 3 m, the aerodynamic performance (Ux, Uy, Uz, P) is predicted for each of 100 x 40 x 30 = 120,000 unit space regions constituting the virtual space A. The output unit 5 outputs the output results (Ux, Uy, Uz, P) generated by the output result generation unit 4. The unit space region is not limited to a three-dimensional region having a cubic shape of 1 cm×1 cm×1 cm, and may be, for example, a three-dimensional region having a cubic shape of 1 mm×1 mm×1 mm. In addition, when the shape data of the vehicle is voxel data, the unit space region may be a voxel (grid unit).

[0016] The relationship between the shape data (F) input to the prediction model 100 and the aerodynamic performance (Ux, Uy, Uz, P) output by the prediction model 100 is expressed by the following formula (1). Note that w is a parameter representing the weight, and b is a parameter representing the bias.

number

[0017] The prediction data acquisition unit 15 shown in FIG. 1 acquires, as prediction data, the aerodynamic performance (Ux, Uy, Uz, P) output by the prediction model 100 based on the vehicle shape data (F) input by the input unit 14. Specifically, the prediction data acquisition unit 15 acquires, as prediction data, the aerodynamic performance (Ux, Uy, Uz, P) for each unit area region constituting each virtual cross section based on the aerodynamic performance (Ux, Uy, Uz, P) output by the prediction model 100 for each unit space region. Here, the unit area region constituting each virtual cross section is a square region Q of 1 cm×1 cm as shown in FIG. 3(b). The unit area region can be defined in advance. The unit area is not limited to this, and may be, for example, a square area of ​​1 mm×1 mm.

[0018] 1 acquires, from the database 11, teacher data including vehicle shape data and aerodynamic performance associated with the shape data. The aerodynamic performance of the teacher data is data analyzed by computational fluid analysis based on the vehicle shape data. The shape data included in the teacher data acquired by the teacher data acquisition unit 16 is the same as the shape data input by the input unit 14 to the prediction model 100. The teacher data acquisition unit 16 also acquires the aerodynamic performance for each unit area region Q constituting each virtual cross section based on the teacher data in the database 11.

[0019] The error calculation unit 17 calculates the difference (error) between the aerodynamic performance of the predicted data and the aerodynamic performance of the teacher data for each of the three components of velocity (Ux, Uy, Uz) and pressure (P). The error calculation unit 17 calculates the error of the predicted data with respect to the teacher data for each virtual cross section set by the virtual space setting unit 12. The error of the predicted data with respect to the teacher data for each virtual cross section is the sum of the errors in each unit area region Q constituting the virtual cross section.

[0020] Next, the parameter update unit 18 updates the parameters (parameter w representing a weight and parameter b representing a bias) of the prediction model 100 based on the virtual area calculated by the virtual area calculation unit 13 and the error calculated by the error calculation unit 17. A specific method of updating the parameters w and b of the prediction model 100 will be described later.

[0021] A method for setting the virtual space A and a method for calculating the virtual area will be described in detail with reference to FIGS. 3(a) to (c), 4(a) to (c), and 5(a) to (c). First, as shown in Figs. 3(a), 4(a), and 5(a), the virtual space setting unit 12 sets a virtual space A that includes 3D data of a vehicle V therein. In the following description, the length direction of the virtual space A shown in Figs. 3(a), 4(a), and 5(a) is defined as direction X, the width direction (depth direction) as direction Y, and the height direction as direction Z. The three directions X, Y, and Z are perpendicular to each other. Furthermore, the length of the virtual space A in direction X is L, the width in direction Y is W, and the height in direction Z is H.

[0022] 3(a) to (c) show examples of two virtual cross sections Cx1 and Cx2 obtained by cutting the virtual space A perpendicularly to the direction X. Note that the virtual cross sections set by the virtual space setting unit 12 are not limited to the virtual cross sections Cx1 and Cx2, but are set at unit length (1 cm) intervals along the direction X. As shown in FIG. 3(a), the virtual cross section Cx1 cuts the front part (bonnet) of the vehicle V, and the virtual cross section Cx2 cuts the central part of the vehicle V. The unit length is not limited to 1 cm and may be, for example, 1 mm.

[0023] For the virtual cross section Cx1 shown in Fig. 3(b), an area Cxa1 is calculated by subtracting the area Cxv1 of the cross section of the vehicle V from the overall area H·W of the virtual cross section Cx1. For the virtual cross section Cx2 shown in Fig. 3(c), an area Cxa2 is calculated by subtracting the area Cxv2 of the cross section of the vehicle V from the overall area H·W of the virtual cross section Cx2. The virtual area Cxa1 is larger than the virtual area Cxa2.

[0024] 4(a) to 4(c) show examples of two virtual cross sections Cy1 and Cy2 obtained by cutting the virtual space A perpendicularly to the direction Y. Note that the virtual cross sections set by the virtual space setting unit 12 are not limited to the virtual cross sections Cy1 and Cy2, but virtual cross sections are set at unit length (1 cm) intervals along the direction Y. As shown in FIG. 4(a), the virtual cross section Cy1 cuts the vehicle V along the central axis, and the virtual cross section Cy2 cuts a portion where the vehicle V is not present without cutting the vehicle V.

[0025] For the virtual cross section Cy1 shown in Fig. 4(b), the area Cya1 is calculated by subtracting the area Cyv1 of the cross section of the vehicle V from the total area H·L of the virtual cross section Cy1. For the virtual cross section Cy2 shown in Fig. 4(c), the total area Cya2 (=H·L) of the virtual cross section Cy2 is calculated as the virtual area. The virtual area Cya1 is smaller than the virtual area Cya2.

[0026] 5(a) to 5(c) show examples of two virtual cross sections Cz1 and Cz2 obtained by cutting the virtual space A perpendicularly to the direction Z. Note that the virtual cross sections set by the virtual space setting unit 12 are not limited to the virtual cross sections Cz1 and Cz2, but are set at unit lengths (1 cm) along the direction Z. As shown in FIG. 5(a), the virtual cross section Cz1 cuts an upper portion of the vehicle V (a portion corresponding to the height at which the window is provided) approximately horizontally, and the virtual cross section Cz2 cuts a lower portion of the vehicle V (a portion corresponding to the height at which the bumper is provided) approximately horizontally.

[0027] For the virtual cross section Cz1 shown in Fig. 5(b), an area Cza1 is calculated by subtracting the area Czv1 of the cross section of the vehicle V from the total area W·L of the virtual cross section Cz1. Also, for the virtual cross section Cz2 shown in Fig. 5(c), an area Cza2 is calculated by subtracting the area Czv2 of the cross section of the vehicle V from the total area W·L of the virtual cross section Cz2. The virtual area Cza1 is larger than the virtual area Cza2.

[0028] Next, a method in which the parameter update unit 18 updates the parameters w and b of the prediction model 100 based on the virtual area and the error of the predicted data relative to the teacher data in the virtual cross section will be described in detail.

[0029] First, as described above, the relationship between the shape data (F) input to the prediction model 100 and the aerodynamic performance (Ux, Uy, Uz, P) output by the prediction model 100 is expressed by the following formula (2).

number

[0030] The parameter update unit 18 updates the parameters w and b of the prediction model 100 so as to bring an evaluation value (Loss) derived from a predefined loss function closer to a minimum value. Here, in the present invention, the loss function is defined by the following formula (3).

number

[0031] In the above-mentioned loss function, the error of the predicted data with respect to the teacher data in one virtual cross section is expressed by the following formula (4) for the virtual cross sections Cx1 and Cx2 shown in Figs. 3(a) to (c). That is, the error of the predicted data with respect to the teacher data in the virtual cross section is the sum of the errors of the predicted data with respect to the teacher data calculated for each unit area region Q of the virtual cross section. In the example of the virtual cross section Cx1 shown in Fig. 3(b), the number of unit area regions Q of 1 cm x 1 cm squares is W x H. That is, the error calculation unit 17 calculates the error of the predicted data with respect to the teacher data in the virtual cross section by adding up all the errors of the predicted data with respect to the teacher data calculated for each of the W x H unit area regions Q.

number

[0032] Similarly, for example, in the virtual cross sections Cy1 and Cy2 shown in FIGS. 4(a) to 4(c), the error of the predicted data with respect to the teacher data in each virtual cross section is expressed by the following formula (5).

number

[0033] Similarly, for example, in the virtual cross sections Cz1 and Cz2 shown in FIGS. 5(a) to 5(c), the error of the predicted data with respect to the teacher data in each virtual cross section is expressed by the following formula (6).

number

[0034] The loss function also includes a first loss evaluation element obtained by dividing the error of predicted data relative to teacher data in the virtual cross section by the virtual area. In the virtual cross sections Cx1 and Cx2 shown in Fig. 3(a) to (c), the first loss evaluation element is expressed by the following formula (7).

number

[0035] Here, the following formula (8), which is the denominator of the above formula (7), represents the virtual areas Cxa1 and Cxa2 of the virtual cross sections Cx1 and Cx2. Specifically, for each unit area region Q of the virtual cross sections Cx1 and Cx2, the virtual area calculation unit 13 sets the region where the cross section of the vehicle V exists as "0" and the region where the cross section of the vehicle V does not exist as "1", and calculates the sum of "0" and "1" for W×H unit area regions Q. In this way, the virtual area calculation unit 13 can calculate the virtual areas Cxa1 and Cxa2 for each of the virtual cross sections Cx1 and Cx2.

number

[0036] Therefore, for example, since the virtual area Cxa1 in FIG. 3(b) is larger than the virtual area Cxa2 in FIG. 3(c), the first loss evaluation element (Equation (7)) in each virtual cross section is larger for the virtual cross section Cx2 than for the virtual cross section Cx1.

[0037] The virtual area calculation unit 13 calculates the virtual area Cya1 of Figure 4(b), the virtual area Cya2 of Figure 4(c), the virtual area Cza1 of Figure 5(b), and the virtual area Cza2 of Figure 5(c) in the same manner as the virtual areas Cxa1 and Cxa2 of the virtual cross sections Cx1 and Cx2.

[0038] In the virtual cross sections Cy1 and Cy2 shown in Figures 4(a) to 4(c), the first loss evaluation element is expressed by the following formula (9): Therefore, for example, since the virtual area Cya1 in Figure 4(b) is smaller than the virtual area Cya2 in Figure 4(c), the first loss evaluation element in each virtual cross section is smaller in the virtual cross section Cy2 than in the virtual cross section Cy1.

number

[0039] In addition, in the virtual cross sections Cz1 and Cz2 shown in Figures 5(a) to (c), the first loss evaluation element is expressed by the following formula (10): Therefore, for example, since the virtual area Cza1 in Figure 5(b) is larger than the virtual area Cza2 in Figure 5(c), the first loss evaluation element in each virtual cross section is larger in the virtual cross section Cz2 than in the virtual cross section Cz1.

number

[0040] Furthermore, the loss function includes second loss evaluation elements expressed by the following formulas (11), (12), and (13). That is, the second loss evaluation elements are expressed by formulas that calculate the average values ​​of the first loss evaluation elements of a plurality of virtual cross sections set at unit length (1 cm) intervals along each of the directions X, Y, and Z. The following formulas (11), (12), and (13) of the second loss evaluation elements correspond to the directions X, Y, and Z, respectively.

number

number

number

[0041] Furthermore, the loss function used by the processor 10 to update the prediction model 100 includes a third loss evaluation factor obtained by adding the above-mentioned second loss evaluation factors (equations (11) to (13)) corresponding to each of the three directions X, Y, and Z. In this embodiment, the equation of the loss function for calculating the evaluation value (Loss) is the same as equation (14) expressing the third loss evaluation factor.

number

[0042] As described above, the loss function is defined so that the evaluation value (Loss) increases as the error of the predicted data with respect to the teacher data in the virtual cross section increases, and the evaluation value decreases as the virtual area increases. The purpose of the learning method executed by the processor 10 is to bring the evaluation value (Loss) derived from the loss function shown in the above formula (3) closer to the minimum value.

[0043] In order to bring the evaluation value derived from the loss function closer to the minimum value, the parameter update unit 18 of the processor 10 calculates the gradient of the loss function with respect to the prediction data, and updates the parameters w and b of the prediction model based on the calculated gradient.

[0044] Specifically, the parameters w and b are updated by the following equations (15) and (16), respectively, to generate new parameters w' and b', where η is a preset learning rate.

number

number

[0045] Here, when updating parameters w and b for predicting the velocity component (Ux) in the direction X, the above equations (15) and (16) can be expanded to the following equations (17) and (18), respectively, using the chain rule.

number

number

[0046] δLoss / δUx is the gradient of the loss function for the component (Ux) of the velocity in the direction X of the predicted data. In addition, δUx / δw and δUx / δb, which are the gradients of the predicted data for the parameters w and b, can be calculated by the backpropagation method based on Equation (19).

number

[0047] That is, the processor 10 calculates the gradient (ΔLoss / ΔUx) of the loss function for the prediction data (Ux), and updates the parameters w and b of the prediction model 100 based on this gradient. In addition, the parameters w and b for predicting the velocity components (Ux, Uz) in the Y and Z directions and the pressure (P) can also be updated in a similar manner.

[0048] As described above, in the learning method according to the present embodiment, the loss function is defined such that the evaluation value (Loss) is larger as the error of the predicted data with respect to the teacher data in the virtual cross section is larger, and the evaluation value (Loss) is smaller as the virtual area is larger. That is, the larger the cross-sectional area of ​​the vehicle V, the greater the weighting of the effect of the error of the predicted data with respect to the teacher data on the evaluation value (Loss) of the loss function. Therefore, the evaluation value (Loss) regarding the aerodynamic performance around the vehicle V can be reflected intensively in the update of the prediction model 100 according to the shape of the vehicle V. That is, the processor 10 calculates the evaluation value (Loss) according to the size of the virtual area reflecting the shape of the vehicle V, and therefore the parameters w, b of the prediction model predicting the aerodynamic performance of the vehicle V can be updated taking the shape of the vehicle V into consideration. As a result, the processor 10 can increase the efficiency of updating the parameters w, b of the prediction model 100 and improve the prediction accuracy of the prediction model 100.

[0049] More specifically, the processor 10 calculates the gradient (δLoss / δUx) of the loss function for the predicted data, and updates the parameters w and b of the prediction model 100 based on this gradient so as to bring the evaluation value (Loss) derived from the loss function closer to a minimum value. This allows the prediction model 100 to be efficiently trained so as to reduce the error of the predicted data of the prediction model 100 for the teacher data.

[0050] Moreover, the error of the prediction data with respect to the teacher data in the virtual cross section is the sum of the errors of the prediction data with respect to the teacher data calculated for each unit area region Q of the virtual cross section. This allows the processor 10 to comprehensively evaluate the errors of the prediction data with respect to the flow field around the vehicle and reflect them in the calculation of the evaluation value (Loss) of the loss function. Furthermore, by using the virtual cross section set by the virtual space setting unit 12 in this way, the error calculation unit 17 can calculate the error of the prediction data with respect to the teacher data in the area around the vehicle V in a manner that is easy to use for learning the prediction model 100.

[0051] In addition, the aerodynamic performance of the training data is data analyzed by computational fluid analysis based on shape data of the vehicle V. This makes it possible to use highly accurate training data for learning the prediction model 100.

[0052] The loss function also includes a first loss evaluation element obtained by dividing the error of the predicted data with respect to the teacher data in the virtual cross section by the virtual area. This allows the loss function to be defined so that the evaluation value (Loss) increases as the error of the predicted data with respect to the teacher data in the virtual cross section increases, and the evaluation value (Loss) decreases as the virtual area increases.

[0053] Furthermore, the loss function includes a second loss evaluation element for calculating an average value of the first loss evaluation elements of the multiple virtual cross sections along each direction, whereby an average evaluation value of the error of the predicted data with respect to the teacher data for each direction can be reflected in the evaluation value (Loss) of the loss function.

[0054] In addition, the loss function includes a third loss evaluation element obtained by adding the second loss evaluation elements corresponding to each of the three directions X, Y, and Z. This allows the evaluation value (Loss) of the loss function to reflect a three-dimensional evaluation value of the error of the predicted data against the training data.

[0055] In addition, since the aerodynamic performance is an index that indicates the speed distribution and pressure distribution of the flow field around the vehicle V, it is possible to comprehensively predict the influence of air resistance when the vehicle V actually travels. Note that the aerodynamic performance may be an index that indicates only the speed distribution of the flow field around the vehicle V, or may be an index that indicates only the pressure distribution.

[0056] The virtual space A is a space that includes the vehicle V and has a width in each of three mutually perpendicular directions X, Y, and Z. This makes it possible to calculate an evaluation value (Loss) of a loss function that three-dimensionally reflects the shape of the vehicle V in each of the directions X, Y, and Z.

[0057] The shape data of the vehicle V is SDF data or voxel data. Therefore, the shape data input by the processor 10 to the prediction model 100 is in a format that the prediction model 100 can easily process.

[0058] The processor 10 converts the structural data, which is the design data of the vehicle V, into shape data of the vehicle V. This makes it possible to use the prediction data of the aerodynamic performance predicted by the prediction model 100 based on the shape data in the design of the vehicle V. It should be noted that the shape data input to the prediction model 100 is not limited to shape data of a vehicle, but may be shape data of other objects.

[0059] The trained prediction model 100 trained by the processor 10 outputs an output result (aerodynamic performance) for each of a plurality of unit space regions constituting the virtual space A based on the shape data of the vehicle V. This allows the prediction model 100 to predict in detail the aerodynamic performance of the flow field around the vehicle. In addition, the error calculation unit 17 of the processor 10 can calculate the error of the prediction data relative to the teacher data for each unit area region Q of the virtual cross section of the virtual space A. Therefore, when the processor 10 further updates the parameters of the prediction model 100, the efficiency of learning of the prediction model 100 can be improved.

[0060] Second Embodiment A processor 20 according to the second embodiment is shown in Fig. 6. The processor 20 is obtained by adding a divergence prediction value calculation unit 21 to the configuration of the processor 10 shown in Fig. 1. Like the processor 10, the processor 20 is a learning device that updates parameters w, b of a prediction model 100 that predicts the aerodynamic performance of an object based on a loss function. In the following description, the same reference numerals as those in FIG. 1 indicate the same or similar configurations, and therefore detailed description thereof will be omitted.

[0061] The divergence prediction value calculation unit 21 calculates the divergence prediction value divU of the velocity (Ux, Uy, Uz) of the flow field around the vehicle V by the following equation (20) based on the prediction data acquired by the prediction data acquisition unit 15. That is, the divergence prediction value calculation unit 21 calculates partial derivatives of the three components for each of the three directions in the space of the predicted velocity field acquired from the prediction data acquisition unit 15, and calculates the sum of these values.

number

[0062] Furthermore, the divergence prediction value calculation unit 21 calculates the difference of the divergence prediction value divU with respect to the speed divergence under the uncompressed condition. Here, since the speed divergence under the uncompressed condition is "0", the difference of the divergence prediction value divU with respect to the speed divergence under the uncompressed condition is the absolute value of the divergence prediction value |divU|-0. Then, the divergence prediction value calculation unit 21 outputs the difference of the divergence prediction value with respect to the speed divergence under the uncompressed condition, i.e., the absolute value of the divergence prediction value, to the parameter update unit 18.

[0063] In this embodiment, the processor 20 defines a loss function shown in the following formula (21) in order to update the parameters w and b of the prediction model 100. The loss function shown in formula (21) is defined so that the evaluation value (Loss) increases as the error of the predicted data relative to the teacher data in the virtual cross section increases, the evaluation value (Loss) decreases as the virtual area increases, and the evaluation value (Loss) increases as the difference |divU|-0 between the divergence predicted value for the velocity divergence under non-compressible conditions increases.

number

[0064] The parameter update unit 18 updates the parameters w, b of the prediction model 100 based on the error of the predicted data relative to the teacher data in the virtual cross section, the virtual area, and the difference (|divU|-0) between the divergence predicted value for the velocity divergence under non-compression conditions so that the evaluation value (Loss) derived from the above loss function approaches a minimum value.

[0065] As described above, in the learning method according to the present embodiment, the loss function is defined by taking into consideration the error of the predicted data with respect to the teacher data, the virtual area, and also the difference of the divergence predicted value with respect to the divergence of the speed under the non-compression condition. Here, the condition of the flow field of the actual vehicle V is approximated to the non-compression condition. Therefore, when the speed of the aerodynamic performance of the predicted data diverges, there is a high possibility that an error occurs between the aerodynamic performance of the predicted data and the aerodynamic performance of the actual vehicle V. Therefore, the loss function is defined so that the evaluation value (Loss) is also large when the difference of the divergence predicted value with respect to the divergence of the speed under the non-compression condition is large. This allows the parameter update unit 18 of the processor 20 to update the parameters w, b of the prediction model 100 with higher accuracy.

[0066] In the first and second embodiments, the processors 10 and 20 set multiple virtual cross sections along each of the three directions X, Y, and Z, but this is not limited to this, and the processors 10 and 20 may set virtual cross sections along only one or two of the directions. [Explanation of symbols]

[0067] 10...Processor (learning device) 11. Database 12...Virtual space setting section 13...Virtual area calculation section 14...Input section 15…Prediction data acquisition section 16…Teacher data acquisition section 17...Error calculation section 18…Parameter update section 100…Prediction model A…Virtual space V…Vehicle Cx1, Cx2, Cy1, Cy2, Cz1, Cz2...Virtual cross sections Cxa1, Cxa2, Cya1, Cya2, Cza1, Cza2...Virtual area

Claims

1. A learning method in which a processor updates parameters of a predictive model that predicts an aerodynamic performance of an object based on a loss function, the method comprising: The processor, setting a virtual space including the object and at least one virtual cross section cutting the virtual space; Calculating a virtual area by subtracting the area of ​​the cross section of the object from the total area of ​​the virtual cross section; inputting shape data of the object into the predictive model; acquiring, as prediction data, the aerodynamic performance of the object predicted by the prediction model based on the shape data of the object; acquiring, from a database, teacher data including shape data of the object and the aerodynamic performance associated with the shape data; Calculating an error of the predicted data with respect to the teacher data; updating parameters of the prediction model based on an error of the prediction data with respect to the teacher data in the virtual area and the virtual cross section so as to bring the evaluation value derived from the loss function closer to a minimum value; The loss function is defined such that the evaluation value increases as the error of the predicted data relative to the teacher data in the virtual cross section increases, and the evaluation value decreases as the virtual area increases.

2. The learning method according to claim 1 , wherein the processor calculates a gradient of the loss function with respect to the prediction data, and updates parameters of the prediction model based on the gradient.

3. 3 . The learning method according to claim 1 , wherein the error of the predicted data with respect to the teacher data in the virtual cross section is a sum of errors of the predicted data with respect to the teacher data calculated for each unit area region of the virtual cross section.

4. 4. The learning method according to claim 1, wherein the aerodynamic performance of the training data is data analyzed by a computational fluid analysis based on the shape data of the object.

5. 5. The learning method according to claim 1, wherein the loss function includes a first loss evaluation element obtained by dividing an error of the predicted data with respect to the teacher data in the virtual cross section by the virtual area.

6. The processor sets, for at least one direction, a plurality of virtual cross sections perpendicular to the direction; The learning method according to claim 5 , wherein the loss function includes a second loss evaluation element for calculating an average value of the first loss evaluation elements of the plurality of virtual cross sections.

7. The processor sets a plurality of virtual cross sections perpendicular to each of three directions perpendicular to each other; The learning method according to claim 6 , wherein the loss function includes a third loss evaluation element obtained by adding the second loss evaluation elements corresponding to each of the three directions.

8. The learning method according to any one of claims 1 to 7, wherein the aerodynamic performance is an index representing at least one of a velocity distribution and a pressure distribution in a flow field around the object.

9. The learning method according to any one of claims 1 to 8, wherein the virtual space is a space that contains the object and has a width in each of three directions perpendicular to each other.

10. The processor, Obtaining the aerodynamic performance including a velocity distribution of a flow field around the object as prediction data; calculating a divergence prediction of a velocity of a flow field around the object based on the prediction data; Calculating the difference between the predicted divergence and the velocity divergence under incompressible conditions; updating parameters of the prediction model based on an error of the prediction data with respect to the teacher data in the virtual cross section, the virtual area, and a difference of the divergence prediction value with respect to the divergence of velocity under a non-compressible condition so that an evaluation value derived from the loss function approaches a minimum value; The learning method according to any one of claims 1 to 9, wherein the loss function is defined such that the evaluation value is larger as an error of the predicted data relative to the teacher data in the virtual cross section is larger, the evaluation value is smaller as the virtual area is larger, and the evaluation value is larger as a difference of the divergence predicted value relative to the divergence of velocity under uncompressed conditions is larger.

11. The learning method according to any one of claims 1 to 10, wherein the shape data of the object is SDF data or voxel data.

12. The learning method according to any one of claims 1 to 11, wherein the processor converts structural data of the object to create the shape data of the object.

13. The learning method according to any one of claims 1 to 12, wherein the object is a vehicle.

14. The learning method according to claim 12 , wherein the object is a vehicle, and the structural data of the object is design data of the vehicle.

15. A trained prediction model trained using the training method according to any one of claims 1 to 14, A predictive model that generates an output result for each of a plurality of unit spatial regions that constitute the virtual space based on the shape data of the object.

16. A learning device having a processor that updates parameters of a prediction model that predicts an aerodynamic performance of an object based on a loss function, The processor, a virtual space setting unit that sets a virtual space including the object and at least one virtual cross section obtained by cutting the virtual space; a virtual area calculation unit that calculates a virtual area by subtracting an area of ​​the cross section of the object from a total area of ​​the virtual cross section; an input unit for inputting shape data of the object into the prediction model; a prediction data acquisition unit that acquires, as prediction data, the aerodynamic performance of the object predicted by the prediction model based on the shape data of the object; a database that stores training data including shape data of the object and the aerodynamic performance associated with the shape data; A teacher data acquisition unit that acquires the teacher data from the database; an error calculation unit that calculates an error of the predicted data with respect to the teacher data; a parameter update unit that updates parameters of the prediction model based on an error of the prediction data with respect to the teacher data in the virtual area and the virtual cross section so as to bring the evaluation value derived from the loss function closer to a minimum value, The loss function is defined such that the evaluation value increases as the error of the predicted data relative to the teacher data in the virtual cross section increases, and the evaluation value decreases as the virtual area increases.

Images