JP2022148420A - Learning method of prediction model, prediction model and learning device - Google Patents

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

The present invention relates to a prediction model learning method for predicting the aerodynamic performance of an object, a learned prediction model trained by the learning method, and a learning device.

The prediction model described in Patent Literature 1 predicts the aerodynamic performance of a vehicle based on feature values extracted from vehicle design data. In addition, in Patent Document 1, based on the predicted value of the aerodynamic performance predicted by the prediction model and the aerodynamic performance measured by CFD (simulation analysis), an evaluation that quantitatively indicates that the prediction model is appropriate A learning method is described that calculates a value and uses the evaluation value to update the parameters of a prediction model.

JP 2014-006812 A

However, in the prediction model learning method described in Patent Document 1, since the shape of the vehicle is not taken into account in calculating the evaluation value, there is a problem that it is difficult to increase the accuracy of aerodynamic performance prediction using the prediction model.

The problem to be solved by the present invention is a learning method for updating parameters of a prediction model that predicts the aerodynamic performance of an object in consideration of the shape of the object, a learned prediction model trained by the learning method, and a learning device. is to provide

The present invention sets a virtual space containing an object and a virtual cross section obtained by cutting the virtual space, calculates the virtual area by subtracting the area of the cross section of the object from the total area of the virtual cross section, and calculates the error of the prediction data with respect to the teacher data. The above problem is solved by updating the parameters of the prediction model based on a loss function defined such that the larger the value of , the larger the evaluation value, and the larger the virtual area, the smaller the evaluation value.

According to the present invention, since the evaluation value is calculated according to the size of the virtual area that reflects the shape of the object, the parameters of the prediction model that predicts the aerodynamic performance of the object are updated in consideration of the shape of the object. It has the effect of being able to

1 is a block diagram showing the configuration of a learning device according to a first embodiment; FIG. 2 is a block diagram showing the configuration of a prediction model learned by the learning device shown in FIG. 1; FIG. 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 the direction X, and FIGS. It is the figure which showed each virtual area of two virtual cross sections illustrated to (a) with the diagonal line. 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 the direction Y, and FIGS. It is the figure which showed each virtual area of two virtual cross sections illustrated to (a) with the diagonal line. 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 the direction Z, and FIGS. It is the figure which showed each virtual area of two virtual cross sections illustrated to (a) with the diagonal line. FIG. 11 is a block diagram showing the configuration of a learning device according to a second embodiment; FIG.

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below with reference to the drawings.
<<1st Embodiment>>
A method of learning the predictive model 100 by the processor 10 according to the first embodiment will be described with reference to FIGS. 1 to 5. FIG.
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 the prediction data output by the prediction model 100 with pre-stored teacher data, and evaluates the error of the prediction data with respect to the teacher data based on a predefined loss function. Processor 10 then updates the parameters of prediction model 100 so that the evaluation value derived from the loss function approaches the 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. Also, aerodynamic performance is an index that indicates the velocity distribution and pressure distribution of a flow field around an 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.

The processor 10 has a database 11 , a virtual space setting section 12 , a virtual area calculation section 13 , an input section 14 , a prediction data acquisition section 15 , a teacher data acquisition section 16 , an error calculation section 17 and a parameter update section 18 . Note that the processor 10 is a computer including a CPU, ROM, and RAM. A 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 used for executing the functions of the processor 10. It's a program. Note that the prediction model 100 may be a part of the program of the processor 10 or may be a hardware program separate from the processor 10 .

The database 11 stores vehicle shape data for use in learning the prediction model 100 . CAD data, which is vehicle design data, 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 shape data of the vehicle and aerodynamic performance associated with the shape data. The aerodynamic performance of the training data is data analyzed in advance by computational fluid dynamics based on the shape data of the object.
It should be noted that the design data of the vehicle indicates structural data of the vehicle.

The virtual space setting unit 12 sets the 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 widths in each of three directions X, Y, Z perpendicular to each other. 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. The details of the method of setting the virtual space A and the virtual cross sections Cx1, Cx2, Cy1, Cy2, Cz1 and Cz2 will be described later.

Next, the virtual area calculator 13 calculates virtual areas Cxa1, Cxa2, Cya1, Cya2, Cza1, Cza2 by subtracting the area of the cross section of the vehicle from the total area of the virtual cross sections Cx1, Cx2, Cy1, Cy2, Cz1, Cz2. (see FIGS. 3(b), (c), 4(b), (c), 6(b), and (c)). The details of the method for calculating the virtual area will be described later.

On the other hand, the input unit 14 inputs vehicle shape data acquired from the database 11 to the prediction model 100 . The prediction model 100 is an aerodynamic performance prediction model using 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.

Here, as shown in FIG. 2, the prediction model 100 has 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 vehicle shape data 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 and then input to the shape data input unit 1 . The feature extraction unit 2 extracts feature amounts of 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 a plurality of intermediate layer features. The feature matching unit 3 matches the intermediate layer features extracted by the feature extraction unit 2 and generates a one-dimensional feature vector. Based on the intermediate layer features extracted by the feature extraction unit 2 and the one-dimensional feature vectors generated by the feature matching unit 3, the output result generation unit 4 calculates the aerodynamic performance of the vehicle, that is, the flow field around the vehicle. Generate three components of velocity (Ux, Uy, Uz) and pressure (P). The three components of velocity (Ux, Uy, Uz) correspond to directions X, Y, and Z shown in FIGS. In addition, the output result generation unit 4 generates a plurality of unit space regions (three-dimensional cubic regions of 1 cm × 1 cm × 1 cm) that constitute the virtual space A shown in Figs. ), produces an output result. Specifically, when the virtual space A is a rectangular parallelepiped space of 10 m × 4 m × 3 m, the aerodynamic performance (Ux, Uy, Uz, P) are predicted. The output unit 5 outputs the output result (Ux, Uy, Uz, P) generated by the output result generation unit 4 .
The unit space area is not limited to a cubic three-dimensional area of 1 cm×1 cm×1 cm, and may be, for example, a cubic three-dimensional area of 1 mm×1 mm×1 mm. Further, when the shape data of the vehicle is voxel data, the unit space area may be voxels (grid units).

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

The prediction data acquisition unit 15 shown in FIG. 1 uses 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 as prediction data. get. Specifically, the prediction data acquisition unit 15, based on the aerodynamic performance (Ux, Uy, Uz, P) output for each unit space region by the prediction model 100, Aerodynamic performance (Ux, Uy, Uz, P) is obtained as prediction data. Here, the unit area region forming each virtual cross section is a square region Q of 1 cm×1 cm as shown in FIG. 3(b). The unit area area can be predefined.
In addition, the unit area area is not limited to this, and may be, for example, a square area of 1 mm×1 mm.

Also, the teacher data acquisition unit 16 shown in FIG. 1 acquires from the database 11 teacher data including the shape data of the vehicle and the aerodynamic performance associated with the shape data. The aerodynamic performance of the training data is data analyzed by computational fluid dynamics based on the shape data of the vehicle. The shape data included in the teacher data acquired by the teacher data acquisition unit 16 is the same as the shape data input to the prediction model 100 by the input unit 14 . The teacher data acquisition unit 16 also acquires the aerodynamic performance for each unit area area Q that constitutes each virtual cross section, based on the teacher data in the database 11 .

The error calculator 17 calculates the difference (error) between the aerodynamic performance of the teacher data and the aerodynamic performance of the prediction data for each of the three components of velocity (Ux, Uy, Uz) and the pressure (P). Note that the error calculator 17 calculates the error of the prediction data with respect to the teacher data for each virtual cross section set by the virtual space setting unit 12 . Also, the error of the prediction data with respect to the teacher data in each virtual cross section is the sum of the errors in each of the unit area regions Q forming the virtual cross section.

Next, the parameter updating unit 18 updates the parameters of the prediction model 100 (parameter w representing weight, parameter w representing bias, b) update. A specific method of updating the parameters w and b of the prediction model 100 will be described later.

Based on FIGS. 3(a) to (c), 4(a) to (c), and 5(a) to (c), the method of setting the virtual space A and the method of calculating the virtual area will be described in detail. .
First, as shown in FIGS. 3(a), 4(a), and 5(a), the virtual space setting unit 12 sets a virtual space A including 3D data of the vehicle V therein. In the following description, the length direction of the virtual space A shown in FIGS. do. The three directions X, Y, Z are directions perpendicular to each other. The virtual space A has a length L in the X direction, a width W in the Y direction, and a height H in the Z direction.

3A to 3C illustrate two virtual cross sections Cx1 and Cx2 obtained by cutting the virtual space A perpendicularly to the direction X. FIG. Note that virtual cross sections set by the virtual space setting unit 12 are not limited to the virtual cross sections Cx1 and Cx2, and virtual cross sections are set along the direction X at intervals of unit length (1 cm). As shown in FIG. 3A, the virtual cross section Cx1 cuts the front portion (hood) of the vehicle V, and the virtual cross section Cx2 cuts the central portion of the vehicle V. As shown in FIG.
Note that the unit length is not limited to 1 cm, and may be 1 mm, for example.

In the virtual cross section Cx1 shown in FIG. 3B, an area Cxa1 obtained by subtracting the area Cxv1 of the cross section of the vehicle V from the total area H·W of the virtual cross section Cx1 is calculated as the virtual area. In the virtual cross section Cx2 shown in FIG. 3(c), an area Cxa2 obtained by subtracting the area Cxv2 of the cross section of the vehicle V from the total area H·W of the virtual cross section Cx2 is calculated as the virtual area. The virtual area Cxa1 is larger than the virtual area Cxa2.

4A to 4C illustrate two virtual cross sections Cy1 and Cy2 obtained by cutting the virtual space A perpendicularly to the direction Y. FIG. It should be noted that virtual cross sections set by the virtual space setting unit 12 are not limited to the virtual cross sections Cy1 and Cy2, and virtual cross sections are set along the direction Y at intervals of unit length (1 cm). As shown in FIG. 4A, the virtual cross section Cy1 cuts the vehicle V along the central axis, and the virtual cross section Cy2 cuts a portion without the vehicle V without cutting the vehicle V. there is

In the virtual cross section Cy1 shown in FIG. 4B, an area Cya1 obtained 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 is calculated as the virtual area. Further, in 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.

5A to 5C illustrate two virtual cross sections Cz1 and Cz2 obtained by cutting the virtual space A perpendicularly to the direction Z. FIG. Note that virtual cross sections set by the virtual space setting unit 12 are not limited to the virtual cross sections Cz1 and Cz2, and virtual cross sections are set along the direction Z at intervals of a unit length (1 cm). As shown in FIG. 5(a), the virtual cross section Cz1 cuts the upper part of the vehicle V (the part corresponding to the height where the window is provided) substantially horizontally, and the virtual cross section Cz2 cuts the vehicle V. (the portion corresponding to the height where the bumper is provided) is cut substantially horizontally.

In the virtual cross section Cz1 shown in FIG. 5B, an area Cza1 obtained 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 is calculated as the virtual area. In the virtual cross section Cz2 shown in FIG. 5(c), an area Cza2 obtained 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 is calculated as the virtual area. Virtual area Cza1 is larger than virtual area Cza2.

Next, a method by 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 prediction data with respect to the teacher data in the virtual cross section will be described in detail.

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 equation (2). expressed.

The parameter updating unit 18 updates the parameters w and b of the prediction model 100 so that the evaluation value (Loss) derived from the predefined loss function approaches the minimum value. Here, in the present invention, the loss function is defined by Equation (3) below.

In the loss function described above, the error of the prediction data with respect to the teacher data in one virtual cross section, for example, in the virtual cross sections Cx1 and Cx2 shown in FIGS. . 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. 3B, the number of square unit area regions Q of 1 cm×1 cm is W×H. That is, the error calculator 17 calculates the error of the predicted data with respect to the teacher data in the virtual cross section by adding all the errors of the predicted data with respect to the teacher data calculated for each W×H unit area areas Q.

Similarly, in the virtual cross sections Cy1 and Cy2 shown in FIGS. 4A to 4C, the error of the predicted data with respect to the teacher data in each virtual cross section is expressed by the following equation (5).

Similarly, for virtual cross sections Cz1 and Cz2 shown in FIGS. 5A to 5C, for example, the error of prediction data with respect to teacher data in each virtual cross section is expressed by the following equation (6).

Also, the loss function 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. In the virtual cross sections Cx1 and Cx2 shown in FIGS. 3(a) to 3(c), the first loss evaluation element is represented by Equation (7) below.

Equation (8) below, which is the denominator of Equation (7) above, represents virtual areas Cxa1 and Cxa2 of 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 determines “0” for the region where the cross section of the vehicle V exists, and “1” for the region where the cross section of the vehicle V does not exist. ”, and the sum of “0” and “1” for W×H unit area regions Q is calculated. Thereby, the virtual area calculator 13 can calculate the virtual areas Cxa1 and Cxa2 for each of the virtual cross sections Cx1 and Cx2.

Therefore, for example, the virtual area Cxa1 in FIG. 3B is larger than the virtual area Cxa2 in FIG. The virtual cross section Cx2 is larger than

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

In the virtual cross sections Cy1 and Cy2 shown in FIGS. 4(a) to 4(c), the first loss evaluation element is represented by Equation (9) below. Therefore, for example, the virtual area Cya1 in FIG. 4B is smaller than the virtual area Cya2 in FIG. becomes smaller.

Also, in the virtual cross sections Cz1 and Cz2 shown in FIGS. 5(a) to 5(c), the first loss evaluation element is represented by the following formula (10). Therefore, for example, the virtual area Cza1 in FIG. 5B is larger than the virtual area Cza2 in FIG. becomes larger.

Furthermore, the loss function includes a second loss evaluation factor represented by Equations (11), (12), and (13) below. That is, the second loss evaluation element calculates the average value of the first loss evaluation elements of a plurality of virtual cross sections set every unit length (1 cm) along each of the directions X, Y, and Z. is represented by the formula Equations (11), (12), and (13) below for the second loss evaluation element correspond to directions X, Y, and Z, respectively.

Furthermore, the loss function used by the processor 10 to update the prediction model 100 is obtained by adding the above-described second loss evaluation elements (equations (11) to (13)) corresponding to each of the three directions X, Y, and Z Includes a third loss assessment factor. In this embodiment, the equation of the loss function for obtaining the evaluation value (Loss) is the same as the equation (14) representing the third loss evaluation element.

As described above, the loss function is defined such that the evaluation value (Loss) increases as the error of the prediction 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 Equation (3) closer to the minimum value.

In order to bring the evaluation value derived from the loss function closer to the minimum value, the parameter updating 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. update.

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

Here, when updating the parameters w and b for predicting the velocity component (Ux) in the direction X, the above equations (15) and (16) are converted into the following equations ( 17) and (18).

δLoss/δUx is the gradient of the loss function with respect to the velocity component (Ux) in the direction X in the prediction data. Also, δUx/δw and δUx/δb, which are the slopes of the prediction data with respect to the parameters w and b, can be calculated by the error backpropagation method based on Equation (19).

That is, the processor 10 calculates the gradient (δLoss/δUx) of the loss function with respect to the prediction data (Ux), and updates the parameters w and b of the prediction model 100 based on this gradient.
Also, the parameters w, b for predicting the velocity components (Ux, Uz) in the directions Y, Z and the pressure (P) can be similarly updated.

As described above, in the learning method according to the present embodiment, the loss function has a larger evaluation value (Loss) as the error of the prediction data with respect to the teacher data in the virtual cross section is larger, and the evaluation value (Loss) is larger as the virtual area is larger. defined to be small. That is, the greater the cross-sectional area of the vehicle V, the greater the weighting of the influence of the error of the prediction data with respect to the teacher data on the evaluation value (Loss) of the loss function. Therefore, depending on the shape of the vehicle V, it is possible to intensively reflect the evaluation value (Loss) regarding the aerodynamic performance around the vehicle V in updating the prediction model 100 . That is, since the processor 10 calculates the evaluation value (Loss) according to the size of the virtual area reflecting the shape of the vehicle V, the prediction model predicts the aerodynamic performance of the vehicle V in consideration of the shape of the vehicle V. can update the parameters w, b of . Thereby, the processor 10 can improve the efficiency of updating the parameters w and b of the prediction model 100 and improve the prediction accuracy of the prediction model 100 .

More specifically, the processor 10 calculates the gradient (δLoss/δUx) of the loss function for the prediction data, and based on this gradient, the evaluation value (Loss) derived from the loss function is brought closer to the minimum value. , update the parameters w and b of the prediction model 100 . As a result, the prediction model 100 can be learned efficiently so as to reduce the error of the prediction data of the prediction model 100 with respect to the teacher data.

Also, 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. Thereby, the processor 10 can comprehensively evaluate the error of the prediction data for the flow field around the vehicle and reflect it in the calculation of the evaluation value (Loss) of the loss function. In addition, by using the virtual cross section set by the virtual space setting unit 12 in this way, the error calculation unit 17 uses the error of the prediction data with respect to the teacher data in the area around the vehicle V for learning of the prediction model 100. It can be calculated in an easy-to-use manner.

Further, the aerodynamic performance of the teacher data is data analyzed by computational fluid dynamics based on the shape data of the vehicle V. FIG. As a result, highly accurate teacher data can be used for learning the predictive model 100 .

Also, the loss function 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. As a result, the loss function can be defined such that the larger the error of the prediction data with respect to the teacher data in the virtual cross section, the larger the evaluation value (Loss), and the larger the virtual area, the smaller the evaluation value (Loss). .

In addition, the loss function includes a second loss factor for calculating the average value of the first loss factor for a plurality of virtual cross sections along each direction. As a result, it is possible to reflect the average evaluation value of the error of the prediction data with respect to the teacher data for each direction in the evaluation value (Loss) of the loss function.

Also, the loss function includes a third loss evaluation factor obtained by adding the second loss evaluation factors corresponding to each of the three directions X, Y, Z. As a result, the value obtained by three-dimensionally evaluating the error of the prediction data with respect to the teacher data can be reflected in the evaluation value (Loss) of the loss function.

Further, since the aerodynamic performance is an index indicating the velocity 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 velocity distribution of the flow field around the vehicle V, or may be an index that indicates only the pressure distribution.

The virtual space A is a space that includes the vehicle V and has widths in each of three directions X, Y, and Z that are perpendicular to each other. Thereby, in each of the directions X, Y, and Z, the evaluation value (Loss) of the loss function reflecting the shape of the vehicle V three-dimensionally can be calculated.

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

The processor 10 converts structural data, which is design data of the vehicle V, into shape data of the vehicle V. FIG. As a result, the prediction data of the aerodynamic performance predicted by the prediction model 100 based on the shape data can be used for the design of the vehicle V. FIG.
The shape data input to the prediction model 100 is not limited to vehicle shape data, and may be other object shape data.

A learned prediction model 100 learned by the processor 10 outputs an output result (aerodynamic performance) for each of a plurality of unit space regions forming the virtual space A based on the shape data of the vehicle V. FIG. This allows the predictive model 100 to predict in detail the aerodynamic performance of the flow field around the vehicle. Further, the error calculator 17 of the processor 10 can calculate the error of the predicted data with respect to the teacher data for each unit area region Q of the virtual cross section of the virtual space A. FIG. Therefore, when the processor 10 further updates the parameters of the prediction model 100, the learning efficiency of the prediction model 100 can be improved.

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

The predicted divergence value calculation unit 21 calculates the predicted divergence value divU of the flow field velocity (Ux, Uy, Uz) around the vehicle V based on the predicted data obtained by the predicted data obtaining unit 15 using the following equation (20). Calculated by That is, the predicted divergence value calculator 21 calculates the partial derivative of the ternary quantity for each of the three directions in the space of the predicted velocity field obtained from the predicted data obtaining unit 15, and calculates the total value thereof.

Furthermore, the predicted divergence value calculator 21 calculates the difference of the predicted divergence value divU with respect to the divergence of the velocity under the non-compressed condition. Here, since the velocity divergence under the uncompressed condition is "0", the difference of the divergence predicted value divU with respect to the velocity divergence under the uncompressed condition is the absolute value |divU|-0 of the divergence predicted value. Then, the predicted divergence value calculator 21 outputs the difference of the predicted divergence value with respect to the divergence of the velocity under the uncompressed condition, ie, the absolute value of the predicted divergence value, to the parameter updater 18 .

In this embodiment, the processor 20 defines the loss function shown in Equation (21) below in order to update the parameters w and b of the prediction model 100. In the loss function shown in Equation (21), the larger the error of the prediction data with respect to the teacher data in the virtual cross section, the larger the evaluation value (Loss), and the larger the virtual area, the smaller the evaluation value (Loss). It is defined such that the evaluation value (Loss) increases as the difference |divU|-0 of the predicted divergence value with respect to the velocity divergence increases.

The parameter updating unit 18 updates the error of the prediction data with respect to the teacher data in the virtual cross section, the virtual area, and the velocity under the non-compressive condition so that the evaluation value (Loss) derived from the loss function approaches the minimum value. Parameters w and b of the prediction model 100 are updated based on the difference (|divU|-0) of the divergence prediction value with respect to the divergence.

As described above, in the learning method according to the present embodiment, the loss function is defined by taking into account the error of the prediction data with respect to the teacher data, the virtual area, and also the difference in the predicted divergence with respect to the divergence of the velocity under the uncompressed condition. It is Here, the flow field conditions of the actual vehicle V are approximated to the incompressible conditions. Therefore, when the speed of the aerodynamic performance of the prediction data diverges, there is a high possibility that there is an error between the aerodynamic performance of the prediction data and the actual aerodynamic performance of the vehicle V. The loss function is defined such that the evaluation value (Loss) also increases when the difference between the predicted divergence values for is large. Thereby, the parameter updating unit 18 of the processor 20 can update the parameters w and b of the prediction model 100 with higher accuracy.

Note that in the first embodiment and the second embodiment, the processors 10 and 20 set a plurality of virtual cross sections along each of the three directions X, Y, and Z, but the present invention is not limited to this. Alternatively, virtual cross sections may be set along only one or two directions.

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

Claims (16)

A learning method in which a processor updates parameters of a prediction model that predicts the aerodynamic performance of an object based on a loss function,
The processor
setting a virtual space containing the object and at least one virtual cross section obtained by 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 the shape data of the object into the prediction model;
acquiring the aerodynamic performance of the object predicted by the prediction model based on the shape data of the object as prediction data;
obtaining teacher data including shape data of the object and the aerodynamic performance associated with the shape data from a database;
calculating an error of the prediction data with respect to the training data;
Update the parameters of the prediction model based on the virtual area and the error of the prediction data with respect to the teacher data in the virtual cross section so that the evaluation value derived from the loss function approaches the minimum value;
The learning method, wherein the loss function is defined such that the larger the error of the predicted data with respect to the teacher data in the virtual cross section, the larger the evaluation value, and the larger the virtual area, the smaller the evaluation value. 2. The learning method of claim 1, wherein the processor calculates a slope of the loss function with respect to the prediction data and updates parameters of the prediction model based on the slope. 3. The learning method according to claim 1, wherein the error of said predicted data with respect to said teacher data in said virtual cross section is a sum of errors of said predicted data with respect to said teacher data calculated for each unit area region of said virtual cross section. . The learning method according to any one of claims 1 to 3, wherein said aerodynamic performance of said teacher data is data analyzed by computational fluid analysis based on said shape data of said object. The learning method according to any one of claims 1 to 4, wherein the loss function 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. The processor sets, for at least one direction, a plurality of the virtual cross sections perpendicular to the direction;
6. The learning method according to claim 5, wherein said loss function includes a second loss evaluation factor for calculating an average value of said first loss evaluation factors of said plurality of virtual cross sections. The processor sets, for each of three mutually perpendicular directions, a plurality of the virtual cross sections perpendicular to each of the directions;
7. The learning method according to claim 6, wherein said loss function includes third loss evaluation factors obtained by adding said second loss evaluation factors corresponding to each of said three directions. The learning method according to any one of claims 1 to 7, wherein said aerodynamic performance is an index representing at least one of velocity distribution and pressure distribution of a flow field around said object. The learning method according to any one of claims 1 to 8, wherein the virtual space is a space containing the object and having widths in each of three directions perpendicular to each other. The processor
Acquiring the aerodynamic performance including the velocity distribution of the flow field around the object as prediction data;
calculating a predicted divergence of the velocity of the flow field around the object based on the predicted data;
calculating the difference of the predicted divergence with respect to the divergence of velocity under uncompressed conditions;
The error of the prediction data with respect to the training data in the virtual cross section, the virtual area, and the divergence prediction value for velocity divergence under incompressible conditions so that the evaluation value derived from the loss function approaches the minimum value updating parameters of the prediction model based on the difference;
In the loss function, the larger the error of the prediction data with respect to the teacher data in the virtual cross section, the larger the evaluation value, the larger the virtual area, the smaller the evaluation value. The learning method according to any one of claims 1 to 9, wherein the evaluation value is defined such that the larger the difference between the divergence prediction values, the larger the evaluation value. The learning method according to any one of claims 1 to 10, wherein said shape data of said object is SDF data or voxel data. The learning method according to any one of claims 1 to 11, wherein said processor transforms structural data of said object to create said shape data of said object. The learning method according to any one of claims 1 to 12, wherein the object is a vehicle. 13. 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. A trained prediction model trained using the learning method according to any one of claims 1 to 14,
A prediction model that generates an output result for each of a plurality of unit space regions that constitute the virtual space based on the shape data of the object. A learning device having a processor that updates parameters of a prediction model that predicts the 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 obtained by removing the area of the cross section of the object from the total area of the virtual cross section;
an input unit that inputs shape data of the object to 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 storing 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;
Parameter update for updating the parameters of the prediction model based on the error of the prediction data with respect to the teacher data in the virtual area and the virtual cross section so that the evaluation value derived from the loss function approaches the minimum value. and
The learning device, wherein the loss function is defined such that the larger the error of the prediction data with respect to the teacher data in the virtual cross section, the larger the evaluation value, and the larger the virtual area, the smaller the evaluation value.

Images