Prediction system, prediction method, and program

Abstract

To more easily predict a time-series change in fluid behavior in a space even when a shape in the space changes.SOLUTION: A prediction system according to an embodiment includes a prediction unit. The prediction unit outputs second state data indicating a fluid behavior of a space at a second time different from a first time based on a prediction model in which a parameter is machine-learned using teacher data output by numerical analysis for performing calculation of a predetermined relational expression with first state shape data indicating a state at the first time with respect to a shape of the space that changes with time as an input.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] FIELD Embodiments of the present invention relate to a prediction system, a prediction method, and a program. [Background technology]

[0002] Conventional fluid simulations require, for example, the generation of a computational grid and the numerical convergence of the Navier-Stokes equations. Therefore, evaluating the behavior of fluids that change over time generally requires a significant amount of computational time. Therefore, performing a simulation every time the shape of a space changes and grasping the state of the fluid each time is not practical, as it involves cumbersome work and takes a long time to calculate. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-127035 [Patent Document 2] Japanese Patent Application Laid-Open No. 2014-006812 [Patent Document 3] Japanese Patent Publication No. 2022-148420 Summary of the Invention [Problem to be solved by the invention]

[0004] With conventional techniques, it has been difficult to more easily predict time-series changes in fluid behavior within a space when the shape within the space changes. [Means for solving the problem]

[0005] A prediction system according to an embodiment includes a prediction unit that receives as input first-state shape data representing a state at a first time of a space that changes over time, and outputs second-state data representing a fluid behavior in the space at a second time different from the first time, based on a prediction model in which parameters are machine-learned using training data output by numerical analysis that calculates a predetermined relational expression. [Brief explanation of the drawings]

[0006] [Figure 1] 1 is a diagram showing an example of a functional configuration and a processing flow of a prediction system according to an embodiment; [Figure 2] FIG. 10 is a diagram illustrating an example of a prediction model according to the embodiment. [Figure 3] FIG. 10 is a diagram illustrating an example of a neural network of a prediction model according to an embodiment. [Figure 4] FIG. 2 is a diagram illustrating an example of an LSTM according to an embodiment. [Figure 5A] FIG. 2 is a diagram showing Example 1 of the shape of the agitator according to the embodiment. [Figure 5B] FIG. 10 is a diagram showing a second example of the shape of the agitator according to the embodiment. [Figure 6A] FIG. 10 is a contour diagram showing Example 1 of the wall distance according to the embodiment. [Figure 6B] FIG. 10 is a contour diagram showing Example 2 of the wall distance according to the embodiment. [Figure 7] FIG. 10 is a diagram for explaining prediction accuracy of the prediction system according to the embodiment. [Figure 8] FIG. 1 is a diagram illustrating an example of a hardware configuration of a prediction system according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0007] Hereinafter, embodiments of a prediction system, a prediction method, and a program will be described in detail with reference to the accompanying drawings.

[0008] In recent years, various technologies using machine learning have emerged. The prediction time using neural networks is much shorter than the calculation time required for thermal fluid simulation. For example, a flow behavior prediction technology has been provided that can predict the behavior of fluids in space using a prediction model based on machine learning.

[0009] However, conventional flow behavior prediction technologies could only make predictions within the scope of their learning. In other words, learning was required for each analysis condition, and the training data was created using computational fluid dynamics (CFD). This meant that, for example, it was necessary to change the shape, dimensions, and position of structures inside a space, such as the flow field of a rotating body, which required the creation and training of a huge amount of training data. Furthermore, the accuracy of flow field predictions using conventional flow behavior prediction technologies was not as high as that of CFD (computational fluid dynamics).

[0010] In the following embodiment, a prediction system will be described that performs numerical analysis by calculating one or more of the momentum conservation equation, pressure equation, turbulence model equation, and energy conservation equation using shape data of the space to be numerically analyzed that changes over time. In the prediction system 1 of the embodiment described below, a prediction model that has undergone deep learning using velocity distribution (Ux, Uy, Uz) as training data is input with shape data of the space that changes over time, and the velocity distribution (Ux, Uy, Uz) is output. This makes it possible to more easily predict, for example, time-series changes in velocity distribution.

[0011] First, the configuration of the main functional units of the prediction system of the embodiment and an example of the processing flow will be described in detail.

[0012] FIG. 1 is a diagram illustrating an example of a functional configuration and processing flow of a prediction system 1 according to an embodiment. The prediction system 1 according to the embodiment is a thermal fluid simulation system that uses a fluid behavior prediction model. The prediction system 1 according to the embodiment includes a numerical analysis unit 2, a prediction unit 3, and a machine learning unit 4.

[0013] [Numerical Analysis Department] The numerical analysis unit 2 includes a grid creation unit 5, a boundary condition input unit 6, a momentum conservation equation calculation unit 7, a pressure equation calculation unit 8, a turbulence model equation calculation unit 9, a convergence determination unit 10, and a spatial shape data calculation unit 11.

[0014] The grid creation unit 5 receives analysis conditions such as the shape / dimensions of the space to be numerically analyzed, fluid temperature, shape change rate (for example, rotation number or rotation speed), and fluid components via an input device such as a keyboard. In this embodiment, the shape / dimensions of the space to be numerically analyzed are the shape / dimensions of the blades of the mixer, the shape / dimensions of the container used in the mixer, and the shape / dimensions of the rotating shaft of the mixer. The grid creation unit 5 creates shape grid data by dividing the space inside the mixer into a grid according to the analysis conditions.

[0015] The boundary condition input unit 6 accepts input of boundary conditions, including the initial flow velocity in the space divided into a grid, the fluid temperature, and the shape change rate of the structure in the space (in this embodiment, the impeller rotation speed of the mixer), via an input device such as a keyboard. The initial flow velocity may be any value; for example, the simulation may be started from a flow velocity of 0.

[0016] The momentum conservation calculation unit 7 receives the boundary conditions received by the boundary condition input unit 6 and the flow velocity distribution (Ux, Uy, Uz) predicted by the prediction unit 3 (described later). The momentum conservation calculation unit 7 numerically solves the fluid behavior in space using the Navier-Stokes equations for each grid.

[0017] The pressure equation calculation unit 8 calculates the pressure equation derived from the mass conservation equation.

[0018] The turbulence model equation calculation unit 9 calculates the turbulence model equation and outputs the analysis results.

[0019] The convergence determination unit 10 determines the convergence of the analysis results (calculation results) by the turbulence model equation calculation unit 9. The convergence determination is performed by solving the thermal fluid motion equation for the input and output of each grid, and determining whether the solution approaches the input and output between adjacent grids (for example, whether the difference between the solution of the thermal fluid motion equation and the input and output between adjacent grids is equal to or less than a threshold).

[0020] If the analysis results have not converged, the convergence determination unit 10 returns the flow behavior in the space (flow velocity distribution (Ux, Uy, Uz)) to the momentum conservation equation calculation unit 7. The processes of the momentum conservation equation calculation unit 7, the pressure equation calculation unit 8, and the turbulence model equation calculation unit 9 are repeatedly executed until the analysis results converge.

[0021] When the analysis results converge, the convergence determination unit 10 outputs the flow velocity distribution (Ux, Uy, Uz). For example, the flow velocity distribution (Ux, Uy, Uz) is output as output information together with spatial shape data (Wall Distance) at the simulated time t and position (x, y, z). The output information is displayed three-dimensionally on a display device, for example.

[0022] Furthermore, when the analysis results have converged, the convergence determination unit 10 stores the analysis results (converged time t, lattice positions x, y, z, wall distances indicating the shapes of the lattice positions x, y, z, and velocity distributions Ux, Uy, Uz) together with the simulated analysis conditions (input values) in the storage unit 18 of the machine learning unit 4. The analysis results are accumulated in the storage unit 18 of the machine learning unit 4 as time-series data.

[0023] The spatial shape data calculation unit 11 calculates, for each position, parameters indicating the spatial shape at time t (position of the blades of the agitator) for which the time-series change is being simulated, based on the shape grid data. Specifically, the spatial shape data calculation unit 11 calculates the shortest distance (wall distance) from the container wall or the surface of the agitator, which is a structure inside the container, to each position (x, y, z) in the space of the fluid area.

[0024] The spatial shape data calculation unit 11 inputs the parameters indicating the calculated spatial shape to the prediction unit 3 and the machine learning unit 4 as first state shape data (t, x, y, z, Wall Distance) indicating the spatial shape that changes over time.

[0025] [Prediction section] The prediction unit 3 includes a spatial shape data acquisition unit 12, a feature extraction unit 13, a time-series feature extraction unit 14, a generation unit 15, and an output unit 16. The feature extraction unit 13, the time-series feature extraction unit 14, and the generation unit 15 are realized by a prediction model 17 that is updated by the machine learning unit 4 described below.

[0026] The spatial shape data acquisition unit 12 receives the shape data in the first state described above from the spatial shape data calculation unit 11 of the numerical analysis unit 2. The shape data in the first state includes lattice positions x, y, and z at analysis time t, and wall distances that indicate the shape at the lattice positions x, y, and z.

[0027] The feature extraction unit 13 calculates feature quantities of the first-state shape data (Wall Distance distribution) acquired together with the analysis time t from the spatial shape data acquisition unit 12. Specifically, in this embodiment, the first-state shape data (Wall Distance distribution) is subjected to three-dimensional convolution processing using a CNN (Convolutional Neural Network) encoder and activation function processing using Max Pooling, repeatedly, to extract features of multiple intermediate layers.

[0028] The time-series feature extraction unit 14 extracts features of time-series changes in shape, i.e., long-term dependencies, while effectively retaining short-term information and long-term memory using LSTM (Long-Short-Term Memory).The time-series feature extraction unit 14 predicts a shape that is in the first state or in the future beyond the first state.

[0029] The generation unit 15 generates a flow velocity distribution (Ux, Uy, Uz) as data of a first state or a second state indicating the behavior of the fluid in the future relative to the first state. Specifically, the generation unit 15 generates features of the flow velocity distribution (Ux, Uy, Uz) from the features of the shape time series change in the intermediate layer extracted by the above-mentioned CNN and LSTM. Next, the generation unit 15 decodes (converts) the features of the flow velocity distribution (Ux, Uy, Uz) using CNN to obtain a flow velocity distribution (Ux, Uy, Uz) that changes over time relative to the shape that changes over time.

[0030] The output unit 16 returns the flow velocity distribution (Ux, Uy, Uz) generated by the generation unit 15 to the numerical analysis unit 2. The flow velocity distribution (Ux, Uy, Uz) returned to the numerical analysis unit 2 is input to the momentum conservation equation calculation unit 7 as the flow velocity distribution (Ux, Uy, Uz) at time t when the analysis is being performed by the numerical analysis unit 2.

[0031] [Machine Learning Department] The machine learning unit 4 updates the parameters of the prediction model 17 through machine learning using training data based on the results of the numerical analysis by the numerical analysis unit 2. Specifically, the machine learning unit 4 includes a memory unit 18, a division unit 19, a spatial shape data acquisition unit 20, a flow velocity data acquisition unit 21, a hyperparameter setting unit 22, a feature extraction unit 23, a time-series feature extraction unit 24, a generation unit 25, an output unit 26, a comparison unit 27, an accuracy determination unit 28, and an update unit 29.

[0032] The storage unit 18 stores the analysis results input from the numerical analysis unit 2 as training data. The training data includes the position in space at time t when the numerical analysis was performed, shape data for the position in space at time t, and the fluid velocity at the position in space at time t. In this embodiment, the storage unit 18 accumulates the analysis results input from the numerical analysis unit 2 (time t when the numerical analysis was performed, lattice positions x, y, z, wall distance indicating the shape of the lattice positions x, y, z, and velocity distributions Ux, Uy, Uz) for each shape of space and analysis condition.

[0033] The dividing unit 19 divides the analysis results (time t, lattice positions x, y, z, wall distance indicating the shape of the lattice positions x, y, z, and velocity distribution Ux, Uy, Uz) stored in the storage unit 18 into training data and verification data for each spatial shape and analysis condition. The training data is data used to train the prediction model 17. The verification data is data used to verify the prediction model 17.

[0034] Specifically, the division unit 19 processes the analysis results so that they can be processed within an allowable learning time depending on the computer performance (e.g., GPU (Graphics Processing Unit) memory) that operates the prediction system 1, and divides the processed analysis results into learning data and verification data.

[0035] The spatial shape data acquisition unit 20 acquires spatial shape (wall distance in this embodiment) data from the learning data.

[0036] The flow velocity data acquisition unit 21 acquires the flow velocities Ux, Uy, and Uz from the learning data.

[0037] The hyperparameter setting unit 22 performs machine learning on the shape data and flow velocity data to set parameters for the process of constructing the prediction model 17. The parameters for the process of constructing the prediction model 17 are parameters for the feature extraction process in the feature extraction unit 23, parameters for the time-series feature extraction process in the time-series feature extraction unit 24, and parameters for the process of generating an output result in the generation unit 25.

[0038] The parameters of the feature extraction unit 23 are, for example, the number and size of convolution kernels of the CNN encoder.

[0039] Furthermore, for example, the parameters of the feature extraction unit 23 are the selection of a pooling method for extracting features of fluid behavior (velocity distribution Ux, Uy, Uz) and shape data (wall distance distribution), and the setting parameters of the pooling method. Specifically, the setting parameters are parameters indicating the setting of the pooling size for Max pooling, the setting of the stride (the size of the step for sliding the pooling range), and the setting of whether or not to perform padding (a process of filling the periphery of input data with 0).

[0040] The parameters of the time-series feature extraction unit 24 are parameters that set, for example, the number of hidden layer (intermediate layer (LSTM)) units in LSTM (Long-Short-Term Memory), the learning rate, and the batch size (learning size).

[0041] Furthermore, for example, the encoding parameters of the feature extraction unit 23 are used as the parameters of the generation unit 25. That is, the generation unit 25 performs decoding according to the encoding parameters of the feature extraction unit 23.

[0042] The feature extraction unit 23 extracts features of the fluid behavior (velocity distribution Ux, Uy, Uz) and shape data (wall distance distribution) based on the parameters set by the hyperparameter setting unit 22.

[0043] Based on the parameters set by the hyperparameter setting unit 22, the time series feature extraction unit 24 extracts the features of time series changes, i.e., long-term dependencies, while effectively retaining short-term information and long-term memory in the LSTM.

[0044] The generation unit 25 generates decoded data by performing decoding according to the encoding parameters of the feature extraction unit 23. The decoded data includes time t for each analysis condition, lattice positions x, y, z, wall distance indicating the shape of the lattice positions x, y, z, and velocity distributions Ux, Uy, Uz.

[0045] The output unit 26 outputs the data decoded by the generation unit 25 .

[0046] The comparison unit 27 calculates the difference between the flow velocity distributions Ux, Uy, Uz output as predicted values ​​from the output unit 26 and the time average values ​​of the time-series changes in the flow velocity distributions included in the verification data.

[0047] The accuracy determination unit 28 determines whether or not the prediction accuracy of the flow velocity distributions Ux, Uy, and Uz output as predicted values ​​from the output unit 26 is sufficient. Specifically, the accuracy determination unit 28 determines that the prediction accuracy is sufficient when the difference calculated by the comparison unit 27 is within, for example, 5% of the time average value of the time-series changes in the flow velocity distribution included in the verification data.

[0048] If the prediction accuracy is insufficient, the accuracy determination unit 28 requests the hyperparameter setting unit 22 to adjust the parameters. If the prediction accuracy is sufficient, the accuracy determination unit 28 notifies the update unit 29 that the prediction accuracy of the prediction model 30 is sufficient.

[0049] Upon receiving the notification from the accuracy determining unit 28, the updating unit 29 updates the prediction model 17 of the predicting unit 3 with the parameters of the prediction model 30.

[0050] 2 is a diagram illustrating an example of a prediction model 17 according to an embodiment. As shown in FIG. 2, the prediction model 17 is created by combining a CNN encoder, a CNN decoder, and an LSTM. The prediction model 17 is updated by parameters of a prediction model 30 learned by the machine learning unit 4.

[0051] 3 is a diagram showing an example of a neural network of the prediction model 17 of the embodiment. The function of the feature extraction unit 13 described above is realized by the first half of the nodes (encoders) included in the intermediate layer. The function of the time series feature extraction unit 14 described above is realized by an LSTM. The function of the generation unit 15 described above is realized by the second half of the nodes (decoders) included in the intermediate layer and the nodes in the output layer.

[0052] FIG. 4 is a diagram illustrating an example of an LSTM according to an embodiment. In the first unit, input X t-1 For output h t-1 The second unit outputs the input X t For output h t The third unit outputs the input X t+1 For output h t+1 The output is a chain of units, which transmits information from long-term and short-term memories. The configuration shown in Figure 4 is a standard example of an LSTM configuration, so a detailed explanation will be omitted.

[0053] 5A is a diagram showing Example 1 of the shape of an agitator according to an embodiment. FIG. 5A shows an example of an agitator having a Rushton blade shape. When an agitator having a Rushton blade shape is used, the grid creation unit 5 receives input of analysis conditions in which the Rushton blade shape is specified as the shape of the agitator blade.

[0054] Fig. 5B is a diagram showing Example 2 of the shape of the agitator of the embodiment. Fig. 5A shows an example of an agitator having a Disper blade shape. When an agitator having a Disper blade shape is used, the grid creation unit 5 accepts input of analysis conditions in which Disper blades are specified as the shape of the agitator blades.

[0055] Fig. 6A is a contour diagram showing Example 1 of the wall distance of the embodiment, Fig. 6A shows an example of the wall distance in the case of a mixer having a Rushton blade shape.

[0056] Fig. 6B is a contour diagram showing Example 2 of the wall distance of the embodiment. Fig. 6B shows an example of the wall distance in the case of a mixer having a Disper blade shape.

[0057] 6A and 6B, the distribution of the wall distance at each position in space differs depending on the shape of the agitator blade. Furthermore, the wall distance at each position in space changes depending on the agitator blade (rotation of the agitator blade), which changes with time t.

[0058] Fig. 7 is a diagram for explaining the prediction accuracy of the prediction system 1 of the embodiment. The example of Fig. 7 shows the prediction accuracy under the analysis conditions where a Disper blade is specified. The CFD in the upper part of Fig. 7 shows the verification data (analysis results using thermal fluid analysis software). The lower part of Fig. 7 shows the analysis results (analysis results using a prediction model based on machine learning) by the prediction system 1 of the embodiment.

[0059] As shown in Figure 7, the prediction accuracy of the prediction system 1 of the embodiment is comparable to that of CFD. In CFD, performing a simulation every time the state (shape) in space changes and understanding the state of the fluid each time involves cumbersome work. On the other hand, the prediction system 1 of the embodiment inputs shape data of the space that changes over time and outputs the velocity distribution (Ux, Uy, Uz) using a prediction model that has been deep-learned using velocity distribution (Ux, Uy, Uz) as training data, making predictions easier than with CFD.

[0060] As described above, in the prediction system 1 of the embodiment, the prediction unit 3 receives as input first-state shape data indicating the state at a first time regarding the shape of a space that changes over time, and outputs second-state data indicating the fluid behavior in the space at a second time different from the first time based on a prediction model in which parameters are machine-learned using training data output by numerical analysis that calculates a predetermined relational expression. Then, the numerical analysis unit 2 outputs third-state data indicating the fluid behavior in the space at a third time different from both the first time and the second time by numerical analysis that calculates a predetermined relational expression regarding the fluid behavior based on the first-state shape data and the second-state data.

[0061] As a result, according to the prediction system 1 of the embodiment, it is possible to more easily predict time-series changes in fluid behavior within a space, even when the shape within the space changes. The second and third state data may include at least one of the fluid velocity for each spatial position and the fluid direction for each spatial position. The second and third state data may further include at least one of the fluid temperature for each spatial position, the fluid pressure for each spatial position, and the fluid shear rate for each spatial position.

[0062] Note that when the prediction unit 3 uses a prediction model 17 outside the scope of deep learning and the flow velocity distribution (Ux, Uy, Uz) of the second state is not output, the prediction process by the prediction unit 3 may be omitted. In this case, the flow behavior in space (flow velocity distribution (Ux, Uy, Uz)) is output by performing calculations using only the numerical analysis unit 2 until the data of the third state converges.

[0063] Furthermore, even when the update unit 29 determines that the prediction accuracy of the prediction model 30 is insufficient and is unable to update the prediction model 17, the prediction process by the prediction unit 3 may be omitted. In this case, the flow behavior in the space (flow velocity distribution (Ux, Uy, Uz)) is output by performing calculations using only the numerical analysis unit 2 until the data in the third state converges.

[0064] That is, the prediction unit 3 stops predicting the data in the second state if the machine learning unit 4 does not learn parameters of the prediction model 30 whose prediction accuracy is greater than the threshold. Then, the numerical analysis unit 2 outputs the data in the third state by CFD in which the data in the second state is not used as input.

[0065] The numerical analysis unit 2 may also be implemented using thermo-fluid analysis software. For example, the prediction unit 3 may predict the fluid behavior in the space in the converged state, and input the predicted convergence value of the flow velocity distribution (Ux, Uy, Uz) into the thermo-fluid analysis software. This makes it possible to simulate physical quantities (e.g., temperature, pressure, shear rate, etc.) other than the flow velocity distribution (Ux, Uy, Uz) while significantly shortening the convergence time.

[0066] Furthermore, in the prediction system 1 of the embodiment, the numerical analysis unit 2 performs the thermo-fluid simulation until the solution converges, even if there is at least some learning data (for example, even if there are unlearned shapes or boundary conditions), thereby achieving analysis accuracy equivalent to that of conventional CFD (thermal fluid simulation).

[0067] Furthermore, in the prediction system 1 of the embodiment, the higher the prediction accuracy of the prediction model 17, the higher the analysis accuracy that can be obtained. However, in the prediction system 1 of the embodiment, there are cases in which the amount of learning is still insufficient during the learning process and the prediction accuracy of the prediction model is somewhat low, but the numerical analysis unit 2 continues to perform the thermal fluid simulation until the solution converges. As a result, analysis accuracy equivalent to that of conventional CFD (thermal fluid simulation) can be obtained.

[0068] In the prediction system 1 of the embodiment, the prediction accuracy of the prediction model 17 improves with an increase in the variation in the shape of the learning range, an increase in the variation in boundary conditions, and the accumulation of learning data. As the prediction accuracy of the prediction model 17 improves, the numerical analysis unit 2 can speed up CFD (thermal fluid simulation) and significantly reduce the analysis time.

[0069] In this embodiment, wall distance is used as data indicating shape change. When general-purpose thermal fluid software is used as the numerical analysis unit 2, it is generally equipped with a function to calculate wall distance. Therefore, when the numerical analysis unit 2 is realized by general-purpose thermal fluid software, it can output data (wall distance) indicating time-series shape change.

[0070] (Modification of the embodiment) Modifications of the above-described embodiment will be described. For example, data other than wall distance may be used as shape data of the space. For example, SDF data (SDF: Signed Distance Function) inside the space may be used as shape data of the space. Furthermore, for example, SDF data converted into voxel data (grid unit data) or the time change of either voxel data may be used as shape data of the space.

[0071] Since SDF data or voxel data more faithfully represents the shape of the interior of a space than Wall Distance, the prediction accuracy of the prediction model 17 is improved and the convergence calculation time of the numerical analysis unit 2 can be further shortened.

[0072] If the thermal fluid software does not have the functionality to calculate SDF data or voxel data, it is necessary to create a program to calculate SDF data or voxel data as a subroutine of the thermal fluid analysis software.

[0073] Finally, an example of the hardware configuration of the prediction system 1 according to the embodiment will be described.

[0074] [Example of hardware configuration] 8 is a diagram illustrating an example of a hardware configuration of the prediction system 1 according to the embodiment. The prediction system 1 according to the embodiment includes a processor 201, a main storage device 202, an auxiliary storage device 203, a display device 204, an input device 205, and a communication device 206. The processor 201, the main storage device 202, the auxiliary storage device 203, the display device 204, the input device 205, and the communication device 206 are connected via a bus 210.

[0075] Note that some of the above components may not be included in the prediction system 1. For example, if the prediction system 1 can use the input function and display function of an external device, the prediction system 1 may not be provided with the display device 204 and the input device 205.

[0076] The processor 201 executes a program read from the auxiliary storage device 203 to the main storage device 202. The main storage device 202 is a memory such as a ROM and a RAM. The auxiliary storage device 203 is a HDD, a memory card, or the like.

[0077] The display device 204 is, for example, a liquid crystal display. The input device 205 is an interface for operating the prediction system 1. The display device 204 and the input device 205 may be realized by a touch panel or the like having a display function and an input function. The communication device 206 is an interface for communicating with other devices.

[0078] For example, the program executed by the prediction system 1 is provided as a computer program product in the form of an installable or executable file recorded on a computer-readable storage medium such as a memory card, hard disk, CD-RW, CD-ROM, CD-R, DVD-RAM, or DVD-R.

[0079] Furthermore, for example, the program executed by the prediction system 1 may be stored on a computer connected to a network such as the Internet, and may be provided by being downloaded via the network.

[0080] Furthermore, for example, the program executed by the prediction system 1 may be provided via a network such as the Internet without being downloaded. Specifically, the processing by the prediction system 1 may be performed by a so-called ASP (Application Service Provider) type service, which realizes the processing function by issuing an execution instruction and obtaining the result from a server computer without transferring the program.

[0081] Furthermore, for example, the program of the prediction system 1 may be provided by being pre-installed in a ROM or the like.

[0082] The program executed by the prediction system 1 has a modular configuration that includes functions that can be realized by the program among the above-mentioned functional configurations. As for each function, as actual hardware, the processor 201 reads the program from a storage medium and executes it, and the above-mentioned functional blocks are loaded onto the main memory device 202. In other words, the above-mentioned functional blocks are generated on the main memory device 202.

[0083] Note that some or all of the above-described functions may be realized by hardware such as an integrated circuit (IC) rather than by software.

[0084] Furthermore, each function may be realized using a plurality of processors 201, in which case each processor 201 may realize one of the functions, or may realize two or more of the functions.

[0085] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0086] 1. Prediction System 2 Numerical Analysis Department 3. Prediction Department 4. Machine Learning Department 5 Grid creation section 6 Boundary condition input section 7 Momentum conservation equation calculator 8 Pressure equation calculation section 9 Turbulence model calculation section 10 Convergence judgment section 11. Spatial shape data calculation unit 12 Spatial shape data acquisition unit 13 Feature Extraction Unit 14 Time series feature extraction unit 15 Generation part 16 Output section 17 Predictive Models 18 Memory section 19 Division 20 Spatial shape data acquisition unit 21 Flow velocity data acquisition section 22 Hyperparameter setting section 23 Feature Extraction Unit 24 Time series feature extraction unit 25 Generation part 26 Output section 27 Comparison section 28 Accuracy judgment section 29 Update section 201 processor 202 Main storage 203 Auxiliary storage device 204 Display device 205 Input Device 206 Communication Equipment 210 Bus

Claims

1. a prediction unit that receives as input first-state shape data indicating a state at a first time regarding a shape of a space that changes over time, and outputs second-state data indicating a fluid behavior in the space at a second time different from the first time, based on a prediction model in which parameters are machine-learned using teacher data output by numerical analysis that calculates a predetermined relational expression; Prediction system.

2. a numerical analysis unit that performs a numerical analysis to calculate a predetermined relational expression regarding fluid behavior based on the shape data in the first state and the data in the second state, and outputs third state data that indicates fluid behavior in the space at a third time that is different from the first time and the second time, The prediction system of claim 1 .

3. a machine learning unit that updates parameters of the prediction model by machine learning using training data based on the results of the numerical analysis; The prediction system of claim 1 further comprising:

4. the prediction unit stops predicting the data in the second state; The numerical analysis unit outputs the data of the third state by CFD (Computational Fluid Dynamics) in which the data of the second state is not used as an input. The prediction system of claim 2 .

5. The teacher data includes a position in the space at time t when the numerical analysis is performed, shape data of the position in the space at time t, and a fluid velocity at the position in the space at time t. The prediction system of claim 1 .

6. The shape data of the position in the space is the shortest distance from a wall surface of the space or a surface of a structure in the space to the position in the space. The prediction system of claim 1 .

7. The shape data of the position in the space is represented using SDF (Signed Distance Function) data or voxel data. The prediction system of claim 1 .

8. the structure in the space is a mixer having a mixing blade, The shape data of the position in the space changes depending on the time of day. The prediction system of claim 1 .

9. the predetermined relational expression includes one or more of a momentum conservation equation, a pressure equation, a turbulence model equation, and an energy conservation equation; The prediction system of claim 1 .

10. the data in the second state and the data in the third state include at least one of a velocity of the fluid for each position in the space and a direction of the fluid for each position in the space; The prediction system of claim 2 .

11. the second state and the third state data further include at least one of a temperature of the fluid for each spatial location, a pressure of the fluid for each spatial location, and a shear rate of the fluid for each spatial location; The prediction system of claim 10.

12. a prediction system receiving input of first-state shape data indicating a state at a first time regarding the shape of a space that changes over time, and outputting second-state data indicating fluid behavior in the space at a second time different from the first time, based on a prediction model in which parameters have been machine-learned using training data output by numerical analysis that calculates a predetermined relational expression; A forecasting method including:

13. Computer, a prediction unit that receives as input first-state shape data indicating a state at a first time regarding the shape of a space that changes over time, and outputs second-state data indicating a fluid behavior in the space at a second time different from the first time, based on a prediction model in which parameters have been machine-learned using teacher data output by numerical analysis that calculates a predetermined relational expression; A program to function as a

Images