Simulator, simulation method, and trained model for simulator

Abstract

This system smoothly couples predictions using neural networks with conventional simulations. [Solution] The simulator 1 stores a trained model 20 corresponding to the space K that is the target of the simulation, and a numerical model N configured to perform numerical simulations of a dynamic part 30 that performs dynamic actions that affect the fluid R present in space K. The trained model 20 is machine-trained to output data from a second time point, which shows the physical quantities of the fluid in the future, when data from a first time point, which shows the physical quantities of the fluid R present in space K, is input. The simulator 1 includes one or more computers 2 that perform processing to couple the data of the boundary G between space K and the dynamic part 30 using the trained model 20 and the numerical model N when simulating the physical quantities of the fluid R present in space K.

Description

Technical Field

[0001] Embodiments of the present invention relate to simulation technology.

Background Art

[0002] In recent years, the application of machine learning has been progressing in various academic fields. In the field of fluids, machine learning has been applied in many aspects such as shape optimization. In particular, numerical simulations using general-purpose fluid analysis software take a long time for prediction and thus cannot be used in real time, and the application of machine learning is expected. Also, when performing, for example, fluid simulations using machine learning called convolutional neural network (CNN) or long short-term memory (LSTM), a technique for predicting the flow behavior in space in real time corresponding to changes in the arrangement of structures in space is known.

[0003] Also, a technique is known in which a large number of sub-neural networks learned in appropriately sized small regions are prepared in advance and combined appropriately like puzzle pieces to evaluate the physical behavior of the system to be solved. This evaluates the physical behavior instantaneously by combining a large number of pre-prepared small regions, rather than by dividing the target calculation system determined in advance. As a result, without performing learning for each evaluation target, a large-scale system can be predicted and evaluated by combining small regions. However, the small regions to be combined form a system in which space is divided into a large number of computational grids according to fluid analysis software.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Fluid simulations using general-purpose fluid analysis software require a very large number of computational grids, for example, when evaluating the ventilation and air conditioning system of a factory or building as a whole. Simultaneously, dynamic equipment such as pumps and fans also need to be simulated, resulting in a significant amount of computation time. Even with the conventional techniques mentioned earlier, learning based on a very large number of computational grids is necessary, requiring considerable training time and computing resources.

[0006] Furthermore, while large spaces to be evaluated can be divided and evaluated separately, coupling them with dynamic equipment such as pumps or fans, and networks simulating piping or ducts, is difficult. These networks aim to simulate the overall flow of factories, buildings, etc., rather than focusing on the minute flow within the piping or ducts. Therefore, while the characteristics of dynamic equipment are simulated, the configuration of piping and ducts is simulated in one dimension, and the behavior of the entire network is simulated. Consequently, such simulations do not have a computational grid like fluid simulations, and coupled evaluation is not possible.

[0007] The embodiments of the present invention have been made in consideration of these circumstances, and aim to smoothly couple predictions using neural networks with conventional simulations. [Means for solving the problem]

[0008] A simulator according to an embodiment of the present invention stores a trained model corresponding to a space to be simulated and a numerical model configured to perform numerical simulation of a dynamic part that performs dynamic actions affecting a fluid present in the space. The trained model is machine-trained to output data from a second time point, indicating the physical quantity of the fluid present in the space, when data from a first time point indicating the physical quantity of the fluid present in the space is input. The simulator includes one or more computers that perform processing to couple data of the boundary between the space and the dynamic part using the trained model and the numerical model when simulating the physical quantity of the fluid present in the space. [Effects of the Invention]

[0009] According to embodiments of the present invention, it is possible to smoothly couple predictions using neural networks with conventional simulations. [Brief explanation of the drawing]

[0010] [Figure 1] A block diagram showing the simulator. [Figure 2] A diagram showing the trained model. [Figure 3] An explanatory diagram showing two spaces in the first embodiment. [Figure 4] An explanatory diagram showing the state in which the two spaces of the first embodiment are combined. [Figure 5] An explanatory diagram showing the space and dynamic parts of the second embodiment. [Figure 6] An explanatory diagram showing the space and dynamic parts of the third embodiment. [Figure 7] An explanatory diagram showing the space and dynamic parts of the fourth embodiment. [Figure 8] An explanatory diagram showing the space and dynamic parts of the fifth embodiment. [Figure 9] An explanatory diagram showing the space and dynamic parts of the sixth embodiment.

[0011] (First Embodiment) The following describes in detail embodiments of the simulator, simulation method, and trained model for the simulator, with reference to the drawings. First, the first embodiment will be described using Figures 1 to 4.

[0012] In Figure 1, symbol 1 represents the simulator. The simulation method is performed using this simulator 1. Simulator 1 is a system for simulating the state of matter in a given region using machine learning with a neural network. In other words, simulator 1 is a system that uses a trained model 20 (Figure 2) generated by machine learning.

[0013] Simulator 1 performs, for example, fluid dynamics simulations. As shown in Figures 3 and 4, an example of an analysis system for a rectangular sub-region (space K) targeting two-dimensional fluid simulations is provided. The interior of each space K is divided into two-dimensional computational grids for calculation.

[0014] An example of a region to be simulated is a small area K inside a designated tank located in a designated plant. The space K shown in Figures 3 and 4 has a simplified tank shape. Space K may also be the interior of a designated manufacturing apparatus. An example of a substance to be simulated is a fluid R flowing through space K. An example of fluid R is a gas such as air. This fluid R may also be a liquid such as water, or a fluid powder. Changes in physical quantities such as the direction, velocity, pressure, viscosity, and temperature of the fluid R are then predicted.

[0015] The space K to be simulated is constructed within the virtual space of the prediction computer 2, which is a predetermined information processing device. For example, two spaces K are constructed within the virtual space of the prediction computer 2. A corresponding trained model 20 (Figure 2) for each of the two spaces K is generated in advance. In the following description, one space K may be referred to as the first space K1, and the other space K as the second space K2.

[0016] These spaces K are, for example, enclosed by walls H, through which fluid R passes from one to the other. The streamlines L of the fluid R are pre-programmed using machine learning. Furthermore, the two spaces K are connected to each other (Figure 4). For example, the first space K1 has an inflow of fluid R from outside the fluid simulation domain, and the second space K2 has an outflow of fluid R to outside the fluid simulation domain.

[0017] Also, assume that there is a boundary G between spaces K. The boundary G is an entrance / exit through which fluid R flows from one space K to the other. That is, the boundary G is a part (opening) where the fluid R inside the space K is affected by an external force. For example, the fluid R in the first space K1 affects the fluid R in the second space K2. Also, the fluid R in the second space K2 affects the fluid R in the first space K1.

[0018] In FIGS. 3 and 4, for the sake of understanding, data of the fluid R near this boundary G is illustrated as first boundary data D1 and second boundary data D2. For example, the first boundary data D1 is affected by the fluid R in the second space K2. The second boundary data D2 is affected by the fluid R in the first space K1. Here, the learned model 20 in the first space K1 is affected by the fluid R in the second space K2 as the first boundary data D1. The learned model 20 in the second space K2 is affected by the fluid R in the first space K1 as the second boundary data D2.

[0019] That is, the first space K1 and the second space K2 each have two regions for coupling. For example, the first space K1 has two regions: the second boundary data D2 used as its own boundary condition and the first boundary data D1 used as the boundary condition of the second space K2. The second space K2 has two regions: the first boundary data D1 used as its own boundary condition and the second boundary data D2 used as the boundary condition of the first space K1.

[0020] As shown in FIG. 4, when a fluid simulation region is set such that these two regions overlap, the first space K1 and the second space K2 will each calculate the boundary data D1 and D2 of the other in their own regions. By repeating the process of recalculating each region using the calculated boundary data D1 and D2 to update the boundary data D1 and D2, the simulation results of the two regions will be smoothly joined.

[0021] Furthermore, a predetermined object Q may be placed inside space K. Object Q is a solid component fixedly placed in space, and is assumed to be a predetermined obstacle or structure that affects the behavior of fluid R. In other words, an object Q that affects the behavior of fluid R exists in space K. In addition, the conditions for fluid R flowing through boundary G are set as the velocity of fluid R when it enters space K and the pressure of fluid R when it leaves space K.

[0022] The user selects several spaces K from a set of multiple spaces K. These selected spaces K are then combined to construct the overall computational domain. In the example in Figure 4, two rectangular spaces K are combined to construct one overall computational domain. Here, the behavior of the fluid R (streamlines L) in each individual space K has been trained using machine learning, eliminating the need to perform new machine learning on the behavior of the fluid R in the large overall computational domain. The prediction computer 2 can then instantly derive the simulation results for the overall computational domain based on multiple trained models 20 (Figure 2) corresponding to each of the multiple spaces K.

[0023] As shown in the block diagram of Figure 1, the simulator 1 includes a prediction computer 2. This prediction computer 2 includes an input unit 3, an output unit 4, a communication unit 5, a processing circuit 6, and a storage unit 7. The prediction computer 2 is a computer in which information processing by software is realized using hardware resources by executing various programs. Furthermore, the simulation method of this embodiment is realized by having the prediction computer 2 execute various programs.

[0024] Each component of Simulator 1 does not necessarily have to be installed on a single computer. For example, one Simulator 1 may be implemented on multiple computers connected to each other via a network. For instance, the function for storing the trained model 20, the function for constructing the overall computing domain, and the function for executing the simulation may each be installed on separate computers.

[0025] Furthermore, the configuration of Simulator 1 may be implemented as a cloud service. In other words, the computers that make up Simulator 1 may be servers on the cloud. For example, not only the configuration that performs memory processing, but the entire configuration that performs the main processing may reside on the cloud, and the user may only configure Simulator 1 and check its input / output via an API (Application Programming Interface) or a web browser.

[0026] Input unit 3 receives predetermined information in response to the user's actions using the prediction computer 2. This input unit 3 includes input devices such as a mouse, keyboard, and touch panel. In other words, predetermined information is input to the prediction computer 2 in response to the operation of these input devices.

[0027] Output unit 4 outputs predetermined information. Output unit 4 includes a device for displaying images, such as a display, which outputs the analysis results. The display may be separate from or integrated with the main body of the prediction computer 2. Additionally or alternatively, a display on another computer connected via a network may output the analysis results.

[0028] In this embodiment, a display is exemplified as the output unit 4, but other configurations are also possible. For example, a head-mounted display or a projector may be the output unit 4. Furthermore, a printer that prints information on paper may also be the output unit 4.

[0029] The communication unit 5 communicates with other information processing devices via a communication line such as the Internet. In this embodiment, the prediction computer 2 and the other computers are connected to each other via the Internet, but other configurations are also possible. For example, the prediction computer 2 and the other computers may be connected to each other via a LAN (Local Area Network), WAN (Wide Area Network), or mobile communication network. Alternatively, each device may be connected to each other via a bus.

[0030] The processing circuit 6 is, for example, a circuit equipped with a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a dedicated or general-purpose processor. This processor realizes various functions by executing various programs stored in the memory unit 7. The processing circuit 6 may also be composed of hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit). Various functions can also be realized by this hardware. Furthermore, the processing circuit 6 can realize various functions by combining software processing by the processor and programs with hardware processing.

[0031] The memory unit 7 stores a predetermined program to be executed by the processing circuit 6. The memory unit 7 also stores various information necessary when executing the simulation method. For example, the memory unit 7 pre-stores multiple trained models 20 (Figure 2) corresponding to multiple spaces K, each having a shape that is at least partially combined with one another. Each trained model 20 corresponds to one space K.

[0032] Predictive computer 2 includes artificial intelligence (AI) that performs machine learning. Furthermore, predictive computer 2 includes a deep learning unit that extracts specific patterns from multiple patterns based on deep learning.

[0033] Analysis using the predictive computer 2 can utilize analytical techniques based on artificial intelligence learning. For example, pre-trained models 20 generated by machine learning using neural networks, pre-trained models 20 generated by other machine learning methods, and mathematical algorithms such as deep learning algorithms can be used. Furthermore, forms of machine learning include forms such as deep learning.

[0034] The simulator 1 of this embodiment may consist of a single computer equipped with a neural network, or it may consist of multiple computers equipped with neural networks.

[0035] Here, a neural network is a model in which artificial neurons (nodes) forming a network change their connection strength through learning, thereby acquiring problem-solving abilities. Furthermore, neural networks acquire problem-solving abilities through deep learning.

[0036] As shown in Figure 2, the trained model 20 includes an input layer 21, an intermediate layer 22, and an output layer 23. Input data is input to the input layer 21. The parameters of the intermediate layer 22 are pre-trained using the training data. The output layer 23 outputs output data that shows the results processed by the intermediate layer 22 in response to the input data input to the input layer 21.

[0037] The input layer 21 receives data at a first time point representing physical quantities of fluid R in space K. The output layer 23 outputs data at a second time point representing physical quantities of fluid R at a future time point than the first time point, in response to the data input to the input layer 21. The hidden layer 22 has its parameters pre-machine-trained using training data that takes at least one of the data at the first time point or data simulating it as input and at least one of the data at the second time point or data simulating it as output. This hidden layer 22 performs calculations so that the physical quantities of fluid R of multiple trained models 20 that make up the entire computational domain are correlated with each other. Note that simulated data includes not only actually acquired data but also data artificially created (simulated) through numerical analysis such as simulations.

[0038] The data at the first and second time points include, for example, the direction of the fluid R, the velocity of the fluid R, the pressure of the fluid R, the momentum flux of the boundary G, and the mass flux of the boundary G. Other information may also be included in the data at the first and second time points.

[0039] The second point in time, which is in the future compared to the first point in time, could be 1 second later, 1 minute later, 10 minutes later, 1 hour later, or 1 day later. The time difference between the first and second points in time can be arbitrarily set by the user in advance.

[0040] Alternatively, the data from the second time point output from the output layer 23 may be input to the input layer 21 as the data from the first time point. By repeatedly outputting the data from the second time point from the output layer 23 and inputting it to the input layer 21 as the data from the first time point, it is possible to predict the future state of the fluid R, which changes in a time-series order.

[0041] The trained model 20 causes the prediction computer 2 to function by inputting data from the first time point into the input layer 21, performing calculations in the hidden layer 22, and outputting data from the second time point through the output layer 23.

[0042] For example, a neural network is provided with an intermediate layer 22 having multiple layers. Each layer of this intermediate layer 22 is composed of multiple units. Furthermore, by pre-training the multi-layer neural network using training data (supervisory data), it is possible to automatically extract features from the patterns of change in the state of spatial K (streamline L). In addition, the number of intermediate layers, the number of units, the learning rate, the number of training iterations, and the activation function of the multi-layer neural network can be set arbitrarily on the user interface.

[0043] Furthermore, a reward function may be set for each information item to be learned, and deep reinforcement learning, in which the information item with the highest value is extracted based on the reward function, may be used in the neural network.

[0044] In this embodiment, Physics-informed neural networks (PINNs) and Conservative physics-informed neural networks (CPINNs) are used. Convolutional neural networks (CNNs) may also be used.

[0045] Furthermore, autoencoders, LSTMs (Long Short-Term Memory), SDFs (Signed Distance Functions), GANs (Generative Adversarial Networks), and RNNs (Recurrent Neural Networks) may also be applied to the machine learning in this embodiment.

[0046] The numerical simulation using the neural network in this embodiment constructs a network that can compute instantaneously by training the neural network based on data of the physical quantities of the fluid R within the computational system.

[0047] In particular, CPINNs divide the entire computational domain under analysis into multiple subdomains (space K). CPINNs solves the physical quantities of the solutions output by subneural networks (trained models 20) generated for each subdomain, while maintaining the continuity of the flux of those physical quantities across the subdomains. For example, a large number of subneural networks, each trained on the changes in appropriately sized subdomains (space K), are pre-configured. These are then combined appropriately, like puzzle pieces, to evaluate the physical behavior of the system to be solved (the entire computational domain).

[0048] This embodiment is based on CPINNs, but differs from conventional CPINNs in that, instead of dividing a predetermined target computation system (the entire computation domain), it instantaneously evaluates physical behavior using a combination of numerous pre-prepared sub-domains (space K). This allows for instantaneous evaluation even when there are large-scale changes in the arrangement or shape of objects Q within an arbitrary computation system.

[0049] The trained model 20 (Figure 2) is pre-trained to output data from the second time point when data from the first time point is input. Furthermore, the trained model 20 is pre-trained to output data from the second time point based on the behavior (streamlines L) of the fluid R flowing around object Q (Figure 4) when data from the first time point is input. In this way, the trained model 20 can output data from the second time point that takes into account the influence that object Q in space K has on the fluid R. The prediction computer 2 then pre-stores the trained model 20 in the memory unit 7 (Figure 1).

[0050] In the first embodiment described above, a large number of subneural networks trained in a space K (small region) of appropriate size are prepared in advance, and the physical behavior of the system to be solved is evaluated by combining them appropriately like puzzle pieces.

[0051] The flow velocity evaluated at the boundary G (periphery) of the first space K1 is used as the boundary condition for the adjacent second space K2, thereby creating a continuous connection of the flow in space K.

[0052] Furthermore, flow simulations often use pressure boundaries and flow boundaries as boundary conditions. A flow boundary is defined by the flow rate at boundary G, while a pressure boundary is defined by pressure, not flow rate.

[0053] Generally, when two simulations of a flow are coupled, a pressure boundary is combined as a boundary condition to connect the two simulations. For example, in the first simulation, the connecting boundary G is treated as a flow rate boundary, and in the second simulation, the connecting boundary G is treated as a pressure boundary. The two simulations are continuously connected by providing the necessary information for the other boundary condition, such as analysis results for flow rate and pressure. However, it is known that this method can cause oscillations in flow rate and pressure at the connection point. Therefore, the simulator 1, simulation method, and trained model for the simulator of the second embodiment are used.

[0054] (Second Embodiment) Next, a second embodiment will be described with reference to Figure 5. Note that components identical to those shown in the previously described embodiment are denoted by the same reference numerals, and redundant descriptions are omitted. Referencing the aforementioned drawings may be used as appropriate.

[0055] The second embodiment illustrates a configuration in which the internal space K of the tank is connected to a dynamic part 30 such as a pump 31. The dynamic part 30 is a device or component that performs dynamic actions that affect the fluid R present in space K. The dynamic part 30 includes a pump 31 and piping 32 for connecting the pump 31 to space K. A boundary G is set at the end of space K. The dynamic part 30 is connected to the end of space K via components such as piping 32.

[0056] The prediction computer 2 (Figure 1) stores the trained model 20 (Figure 2) and the numerical model N in its memory unit 7. The trained model 20 corresponds to the space K that is the target of the simulation. The numerical model N is configured to perform numerical simulations of the dynamic unit 30.

[0057] The trained model 20 is machine-learned to take data from a first time point representing the physical quantities of fluid R in space K as input, and output data from a second time point representing the physical quantities of fluid R at a future time point.

[0058] When the prediction computer 2 (Figure 1) simulates the physical quantities of the fluid R present in space K, it couples the data of the boundary G between space K and the dynamic part 30 with the trained model 20 and the numerical model N.

[0059] The prediction computer 2 (Figure 1) performs one simulation based on the trained model 20 (Figure 2) and the other simulation based on the numerical model N, and then couples these simulations.

[0060] The boundary G data calculated by the trained model 20 includes second boundary data D2, which is fluid R data located in the region adjacent to the boundary G of the dynamic part 30. The boundary G data calculated by the numerical model N includes first boundary data D1, which is fluid R data located in the region adjacent to the boundary G of space K.

[0061] In the prediction computer 2 (Figure 1), the process of coupling boundary G data includes a first transition process 41 and a second transition process 42.

[0062] The first transition process 41 is the process of converting data of two or more dimensions output from the trained model 20 into one-dimensional data and inputting it into the numerical model N. For example, the first transition process 41 is the process of converting the first boundary data D1 into one-dimensional data and inputting it into the numerical model N.

[0063] The second transition process 42 is the process of converting the one-dimensional data output from the numerical model N into data of two dimensions or more and inputting it into the trained model 20. For example, the second transition process 42 is the process of converting the one-dimensional data output from the numerical model N into second boundary data D2 of two dimensions or more and inputting it into the trained model 20.

[0064] As shown in Figure 5, at the edge of space K, the fluid R data near the boundary G is illustrated as first boundary data D1 and second boundary data D2. The first boundary data D1 is used as the boundary condition for space K itself, and the second boundary data D2 is used as the boundary condition with the dynamic part 30.

[0065] Unlike the simulation of space K using the trained model 20, the simulation using the numerical model N is not divided into a two-dimensional computational grid, and a one-dimensional computational grid in the direction of the fluid flow R is often used. An example of this is given here. Therefore, the computational grids of space K and the dynamic part 30 are significantly different, and the data cannot be directly used interchangeably. To address this, the data can be used interchangeably by performing the first transition process 41 and the second transition process 42 described above.

[0066] The first transition process 41 aggregates the first boundary data D1 calculated by the trained model 20 (Figure 2), and converts this 2D data D1 into 1D data.

[0067] In the first transition process 41, if the numerical model N has cross-sectional area information for a one-dimensional grid, the following transformation is performed. Here, each variable is expressed with the outflow of fluid R from space K as positive.

[0068] Each computational grid in space K: A = flow velocity (m / s), B = cross-sectional area of ​​the computational grid (m²), C = fluid density (kg / m³) Data aggregation at the boundary G of space K: D(kg / m2s)=Σ(A·B·C) / E, E(m2)=Σ(B), f(m / s)=Σ(A·B) / E

[0069] D and C, or F and C, are used as data for the boundary G between space K and the dynamic part 30. These represent flow velocity or flow rate for each cross-sectional area, but the cross-sectional area information is stored in the numerical model N.

[0070] In the first transition process 41, if the numerical model N does not have cross-sectional area information for a one-dimensional grid, the following transformation is performed. Here, each variable is expressed with outflow from space K as positive.

[0071] Each computational grid in space K: A = flow velocity (m / s), B = cross-sectional area of ​​the computational grid (m²), C = fluid density (kg / m³) Cross-sectional aggregation of boundary data for small regions: D(kg / s)=Σ(A·B·C), F(m³ / s)=Σ(A·B)

[0072] D or F are used as boundary G data for the numerical model N. These represent the aggregated mass flow rate or volume flow rate of the boundary G data in space K. Since cross-sectional area information is not stored in the numerical model N, the aggregated values ​​are used as boundary conditions.

[0073] The second transfer process 42 transfers data from the numerical model N to the trained model 20 (Figure 2). Here, the boundary G data calculated by the numerical model N is expanded, and the one-dimensional data is transformed into two-dimensional data. The following transformations are performed in the second transfer process 42.

[0074] Numerical model N calculation results: a = flow velocity (m / s), b = cross-sectional area of ​​the computational grid (m²), c = fluid density (kg / m³) Transformation of the calculation results of the numerical model N: f(m / s) = Σ(a·b) Expansion process to trained model 20 (Figure 2): Σ(A·B)·x=f (calculate x that satisfies the above) A' = A·x Here, A' is used as the data for the boundary G of space K.

[0075] In this way, the simulation of space K and the simulation of the dynamic part 30 can be smoothly connected and coupled.

[0076] In the second embodiment described above, flow prediction using a trained neural network is coupled with other simulations, such as simulations of fluid networks like piping systems and duct systems. The flow velocity values ​​used in the neural network's flow prediction are the same as the flow velocity values ​​used as boundary conditions in the other simulations. Conversely, the flow prediction using the neural network is used as the flow velocity values ​​in the analysis results of the other simulations. In this way, the boundary conditions can be matched, and the flow prediction using the neural network and the other simulations can be coupled.

[0077] (Third embodiment) Next, a third embodiment will be described with reference to Figure 6. Note that components identical to those shown in the previously described embodiments are denoted by the same reference numerals, and redundant descriptions are omitted. Referencing the aforementioned drawings may be used as appropriate.

[0078] In the third embodiment, the boundary G is set inside space K. For example, a configuration is conceivable in which a pipe 32 is connected to the vicinity of the center of a tank constituting space K, and this pipe 32 is connected to a pump 31. The connection point between the tank and the pipe 32 is, for example, a circular opening.

[0079] Prediction computer 2 corrects the flow rate of fluid R inflow or outflow into space K against the flow rate of numerical model N.

[0080] Setting of inflow and outflow into space K: α = inflow (m / s), β = outflow (m / s), y = correction ratio (-) α·z·y=A', β·z·(1-y)=A' (Calculate the value of z that satisfies the above conditions) α' = α - α·z·y β' = β - β·z·(1-y) Here, α' and β' are used as correction values ​​for the settings of inflow or outflow into space K.

[0081] In summary, the third embodiment allows for the smooth connection and coupling of the simulation of space K and the simulation of the dynamic part 30.

[0082] (Fourth Embodiment) Next, a fourth embodiment will be described with reference to Figure 7. Note that components identical to those shown in the previously described embodiments are denoted by the same reference numerals, and redundant descriptions are omitted. Referencing the aforementioned drawings may be used as appropriate.

[0083] The numerical model N of the fourth embodiment includes a reaction model 50 which involves at least one of the reactions of consumption or generation of substances constituting the fluid R.

[0084] Reaction model 50 is a model that simulates reactions such as combustion or chemical reactions. Although reaction model 50 does not simulate the flow of fluid R, at least one of the consumption or generation of fluid R often occurs during the reaction.

[0085] The reaction model 50 includes an inflow process 51, a reaction process 52, and an outflow process 53. The inflow process 51 simulates the state in which fluid R flows in from space K. The reaction process 52 simulates the state in which fluid R undergoes a reaction such as combustion or a chemical reaction. The outflow process 53 simulates the state in which fluid R flows out into space K.

[0086] In summary, the fourth embodiment treats the consumption and production in the reaction as the flow rate of the coupled region in the simulation of the piping system, thereby enabling a smooth connection and coupling between the simulation of space K and the simulation of the dynamic part 30.

[0087] (Fifth embodiment) Next, a fifth embodiment will be described with reference to Figure 8. Note that components identical to those shown in the previously described embodiments are denoted by the same reference numerals, and redundant descriptions are omitted. Referencing the aforementioned drawings may be used as appropriate.

[0088] The prediction computer 2 of the fifth embodiment (Figure 1) executes a correction process 43 when performing the first transition process 41 and the second transition process 42. The correction process 43 includes a process of correcting the data output from the numerical model N and inputting it into the trained model 20 (Figure 2) when coupling the boundary G data, and a process of correcting the data output from the trained model 20 and inputting it into the numerical model N.

[0089] The simulations using the trained model 20 (Figure 2) and the numerical model N repeatedly recalculate their own domains using the boundary G data calculated by the other model, and are evaluated until a condition for smooth connection is met.

[0090] In this case, instead of using the boundary G data calculated by the other party as is, the number of iterations can sometimes be reduced by performing correction processing 43. For example, one method is to take the average of the boundary G data calculated by the other party and the data calculated by oneself and use it as the boundary G data for one's own domain. Another method is to take the average of the boundary G data calculated by the other party and the previous value of the boundary G data calculated by the other party and use it as the boundary G data for one's own domain.

[0091] In the fifth embodiment described above, the prediction computer 2 (Figure 1) performs a correction process 43, which allows for the smooth connection and coupling of the spatial K simulation and the dynamic unit 30 simulation.

[0092] (Sixth Embodiment) Next, the sixth embodiment will be described with reference to Figure 9. Note that components identical to those shown in the previously described embodiments are denoted by the same reference numerals, and redundant descriptions are omitted. Referencing the previously mentioned drawings may be used as appropriate.

[0093] In the sixth embodiment, a shape factor S is used to calculate the boundary G data. For example, the prediction computer 2 (Figure 1) uses the shape factor S when performing the first transition process 41 and the second transition process 42.

[0094] The connection points between the trained model 20 (Figure 2) and the numerical model N, such as the connection point between the tank and the pipe 32, have local shapes, such as the nozzle shape of the pipe 32, which can affect the flow.

[0095] To simulate the nozzle shape in the spatial K side of the fluid R simulation, it is necessary to increase the number of meshes used to represent the nozzle shape. Furthermore, if the nozzle shapes differ, prior machine learning using a large amount of data corresponding to each nozzle shape is required.

[0096] The sixth embodiment makes it possible to evaluate the effects of the nozzle shape on the flow in the simulation without directly representing the nozzle shape.

[0097] Shape factors S are considered and represented for two regions: space K and the dynamic part 30. Shape factors S define the relative magnitude of the fluid flow R based on the position within the two regions. Furthermore, by using 'Ashape' as the boundary G data, shape-considered boundary G data can be used.

[0098] Expansion into space K: Σ(A·B·S)·xs=f (calculate xs that satisfies the above) Ashape'=〈A·S〉·x

[0099] Note that the connection point between the trained model 20 (Figure 2) and the numerical model N does not have to be a nozzle shape; for example, it may be a mesh shape, or a shape in which part of the opening of the pipe 32 is closed.

[0100] In summary, the sixth embodiment allows for the smooth connection and coupling of the spatial K simulation and the dynamic part 30 simulation, taking into account the shape of the connection point between the trained model 20 (Figure 2) and the numerical model N.

[0101] Although the present invention has been described above based on the first to sixth embodiments, a configuration applied in any one embodiment may be applied to another embodiment, or the configurations applied in each embodiment may be combined.

[0102] The aforementioned simulator 1 comprises a control device, a memory device, an output device, an input device, and a communication interface. Here, the control device includes a highly integrated processor such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), or a dedicated chip. The memory device includes ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), SSD (Solid State Drive), etc. The output device includes a display panel, a head-mounted display, a projector, a printer, etc. The input device includes a mouse, a keyboard, a touch panel, etc. This simulator 1 can be implemented using a hardware configuration that utilizes a standard computer.

[0103] The program or trained model 20 executed by the aforementioned simulator 1 is provided pre-loaded into ROM or similar media. Alternatively, this program or trained model 20 may be provided as an installable or executable file stored on a computer-readable, non-temporary storage medium. This storage medium may include CD-ROMs, CD-Rs, memory cards, DVDs, flexible disks (FDs), etc.

[0104] Furthermore, the program or trained model 20 executed in this simulator 1 may be stored on a computer connected to a network such as the Internet and provided for download via the network. In other words, the program or trained model 20 may be provided from cloud computing resources. Alternatively, a server on the cloud may execute the program or trained model 20, and only the processing results may be provided via the cloud. In addition, this simulator 1 can also be configured by combining separate modules, each independently performing the functions of its components, which are interconnected via a network or dedicated line.

[0105] According to at least one embodiment described above, when the prediction computer 2 simulates the physical quantities of the fluid R present in space K, it couples the data of the boundary G between space K and the dynamic part 30 with the trained model 20 and the numerical model N. This allows for a smooth coupling of prediction by the neural network and a normal simulation. Furthermore, it enables high-speed prediction by the neural network while taking into account the results of other simulations.

[0106] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be carried out in a variety of other forms, and various omissions, substitutions, modifications, and combinations are possible without departing from the spirit of the invention. These embodiments or their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of symbols]

[0107] 1...Simulator, 2...Prediction computer, 3...Input unit, 4...Output unit, 5...Communication unit, 6...Processing circuit, 7...Memory unit, 20...Trained model, 21...Input layer, 22...Hidden layer, 23...Output layer, 30...Dynamic unit, 31...Pump, 32...Piping, 41...First transition processing, 42...Second transition processing, 43...Correction processing, 50...Reaction model, 51...Inflow processing, 52...Reaction processing, 53...Outflow processing, D1...First boundary data, D2...Second boundary data, G...Boundary, H...Wall, K...Space, K1...First space, K2...Second space, L...Streamlines, N...Numerical model, Q...Object, R...Fluid, S...Shape coefficient.

Claims

1. It stores a trained model corresponding to the space to be simulated, and a numerical model configured to perform numerical simulations of dynamic parts that perform dynamic actions affecting the fluid present in the space. The trained model is machine-learned to output data from a second time point, which represents the physical quantities of the fluid in a future time, when it receives data from a first time point that represents the physical quantities of the fluid present in the space. When simulating the physical quantities of the fluid present in the space, the trained model and the numerical model are coupled with data of the boundary between the space and the dynamic part. A system comprising one or more computers that perform processing, Simulator.

2. The boundary data calculated by the trained model includes the fluid data present in the region adjacent to the boundary of the dynamic part. The boundary data calculated by the numerical model includes the fluid data present in the region adjacent to the boundary of the space. The simulator according to claim 1.

3. The process of coupling the boundary data is as follows: A first transition process involves converting two-dimensional or more data output from the trained model into one-dimensional data and inputting it into the numerical model. A second transition process involves converting the one-dimensional data output from the numerical model into two-dimensional or higher data and inputting it into the trained model. including, The simulator according to claim 1 or claim 2.

4. The boundary is set at the end of the space, The simulator according to claim 1 or claim 2.

5. The aforementioned boundary is set within the space, The simulator according to claim 1 or claim 2.

6. The numerical model includes a reaction model involving at least one of the reactions of consumption or generation of the substances constituting the fluid. The simulator according to claim 1 or claim 2.

7. The computer, when coupling the boundary data, performs a correction process that corrects the data output from the numerical model and inputs it into the trained model, and corrects the data output from the trained model and inputs it into the numerical model. The simulator according to claim 1 or claim 2.

8. Shape factors are used to calculate the data for the aforementioned boundary. The simulator according to claim 1 or claim 2.

9. It stores a trained model corresponding to the space to be simulated, and a numerical model configured to perform numerical simulations of dynamic parts that perform dynamic actions affecting the fluid present in the space. The trained model is machine-learned to output data from a second time point, which represents the physical quantities of the fluid in a future time, when it receives data from a first time point that represents the physical quantities of the fluid present in the space. When simulating the physical quantities of the fluid present in the space, the trained model and the numerical model are coupled with data of the boundary between the space and the dynamic part. The process is performed by one or more computers. Simulation method.

10. It stores a trained model corresponding to the space to be simulated, and a numerical model configured to perform numerical simulations of dynamic parts that perform dynamic actions affecting the fluid present in the space. When simulating the physical quantities of the fluid present in the space, the trained model and the numerical model couple the data of the boundary between the space and the dynamic part. The trained model is used in one or more computers that perform the processing, An input layer into which data at a first time point indicating the physical quantity of the fluid present in the space is input, An output layer that outputs data for a second time point showing the physical quantities of the fluid at a time point in the future than the first time point, An intermediate layer whose parameters have been machine-trained using training data where the data from the first time point is input and the data from the second time point is output, including, Pre-trained model for simulator use.

Images