Simulation device, simulation method, and program

Abstract

This simulation device comprises: a simulation unit provided with a sub-learning model corresponding to a different resolution for each stratum, the simulation unit simulating a phenomenon of a marine ecosystem using a learning model that inputs / outputs a simulation result to / from the sub-learning model; and a training unit that trains the sub-learning model by using training data pertaining to a corresponding resolution and training results from the preceding sub-learning model. The sub-learning model is provided with a fluid learning model for simulating a phenomenon of a fluid, and a biological learning model for simulating a phenomenon of an organism. The structure of the fluid learning model is the same irrespective of the stratum, and the structure of the biological learning model is different for each stratum.

Description

Simulation device, simulation method, and program

[0001] The present disclosure relates to a simulation device, a simulation method, and a program for simulating phenomena in marine ecosystems.

[0002] Simulations using physical models enable highly accurate predictions, but they tend to require large computational costs, such as computation time and resources. As a way to overcome this disadvantage, surrogate models that can replace physical models are attracting attention.

[0003] The surrogate model is realized by machine learning, for example, a neural network. Hereinafter, a surrogate model realized by machine learning is also referred to as a learning model. The learning model is trained using simulation results (hereinafter also referred to as training data) using a physical model. The trained (learned) learning model can perform simulations with lower computational cost than a physical model.

[0004] The U-Net model is known as a specific learning model. The U-Net model consists of multiple layers, and each layer symmetrically comprises an encoder that extracts features through convolution and a decoder that restores the resolution of the extracted features through deconvolution. However, the deepest layer only comprises a decoder. A distinctive feature of the U-Net model is the skip connection that inputs the extracted features from the encoder to the decoder in layers of the same depth. The U-Net model can improve the accuracy of simulations by using this skip connection.

[0005] There are various improved models of the U-Net model. For example, Non-Patent Document 1 combines the idea of ​​the multigrid method with the U-Net model. Specifically, the grid in each layer of the U-Net model is replaced with grids of three different resolutions, and max-pooling and sub-sampling are performed between these grids of different resolutions.

[0006] Le, Quang Tuyen, and Chinchun Ooi. "Surrogate modeling of fluid dynamics with a multigrid inspired neural network architecture." Machine Learning with Applications 6 (2021): 100176.

[0007] Non-Patent Document 1 utilizes such an improved model in fluid simulation. However, when simulating phenomena in marine ecosystems, it is necessary to simulate not only fluid phenomena such as eddies and tidal currents, but also biological phenomena that are coupled with fluid phenomena. Note that "biological phenomena" may be interpreted as phenomena including chemical substances exchanged through interactions between multiple organisms, and may also be called "biogeochemical phenomena."

[0008] In fluid simulation, even if the resolution at which the simulation is performed changes, there is no need to change the physical model (e.g., the Navier-Stokes equations, which are the governing equations), and therefore the structure of the learning model corresponding to the physical model (e.g., the number of layers, the number of nodes, etc.) does not change.

[0009] On the other hand, in biological simulations, the physical model must also be changed when the resolution at which the simulation is performed changes. For example, a corresponding physical model is required to represent large mammals or schools of fish, individual fish, or the movement of zooplankton. Therefore, the structure of the learning model corresponding to the physical model (number of layers, number of nodes, etc.) must also be changed depending on the resolution at which the simulation is performed.

[0010] Due to these circumstances, conventional techniques have had the problem of making it difficult to simulate phenomena in marine ecosystems.

[0011] Therefore, the present disclosure has been made in consideration of such circumstances, and aims to simulate phenomena in marine ecosystems.

[0012] One aspect of the present disclosure is a simulation device comprising: a simulation unit that simulates phenomena in a marine ecosystem using sub-learning models corresponding to different resolutions for each layer, the sub-learning models inputting and outputting simulation results between the sub-learning models; and a learning unit that causes the sub-learning models to learn using learning data of the corresponding resolution and the learning results of the previous sub-learning model, the sub-learning models comprising a fluid learning model that simulates fluid phenomena and an organism learning model that simulates organism phenomena, the structure of the fluid learning model being the same regardless of the layer, and the structure of the organism learning model being different for each layer.

[0013] One aspect of the present disclosure is a simulation method comprising: a simulation step of simulating a phenomenon in a marine ecosystem using a learning model that has sub-learning models corresponding to different resolutions for each layer and that inputs and outputs simulation results between the sub-learning models; and a learning step of having the sub-learning models learn using learning data of the corresponding resolution and the learning results of the previous sub-learning model, wherein the sub-learning models comprise a fluid learning model that simulates a fluid phenomenon and an organism learning model that simulates an organism phenomenon, and the structure of the fluid learning model is the same regardless of the layer, and the structure of the organism learning model is different for each layer.

[0014] According to the present disclosure, it is possible to simulate phenomena in marine ecosystems.

[0015] Fig. 1 is a diagram illustrating an example of a learning model according to the present disclosure. Fig. 2 is a functional block diagram of a simulation device according to the present disclosure. Fig. 3 is a flowchart illustrating an example of learning according to the present disclosure. Fig. 4 is a flowchart illustrating an example of a simulation according to the present disclosure. Fig. 5 is a diagram illustrating an example of a hardware configuration of a simulation device according to the present disclosure.

[0016] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0017] The learning model of the present disclosure will be described with reference to FIG. 1. The learning model of the present disclosure is based on a U-Net model, which consists of multiple layers and symmetrically includes, in each layer, an encoder that extracts features by convolution and a decoder that restores the resolution of the extracted features by deconvolution. However, as described above, the deepest layer includes only a decoder. Although not shown, the decoder performs deconvolution while referring to features extracted by the encoder in the layer at the same depth. Note that, for convenience of explanation, hereinafter, a layer in the learning model will also be referred to as a "learning layer."

[0018] The learning model of the present disclosure includes a learning model in each layer of multiple learning layers. Hereinafter, the learning model in each layer is also referred to as a sub-learning model. That is, the learning model of the present disclosure includes a sub-learning model in each layer of multiple learning layers. In a simulation using the learning model of the present disclosure, the simulation results of each sub-learning model are input and output between the sub-learning models continuously from the previous stage to the next stage, and a final simulation result is output. Furthermore, the sub-learning model constitutes the above-mentioned encoder or decoder.

[0019] As shown in FIG. 1, the sub-learning model includes, for each learning layer, a fluid learning model (represented by "Fluid" in the figure) that simulates fluid phenomena and a biological learning model (represented by "Biology" in the figure) that simulates biological phenomena. The fluid learning model is, for example, a surrogate model of a fluid physics model described by the Navier-Stokes equations. The biological learning model is, for example, a surrogate model of a biophysical model such as Biogeochemistry with Light, Iron, Nutrients and Gas (BLING), PlankTOM5, HAMOCC, PISCES, REcom2, or COBALT.

[0020] The structure of the fluid learning model (number of layers, number of nodes, etc.) is the same regardless of the depth of the learning layer. On the other hand, the fluid learning model can output simulation results by specifying a resolution. In the above-mentioned learning model, the fluid learning model outputs simulation results by specifying a different resolution for each learning layer.

[0021] The structure of the biological learning model (number of layers, number of nodes, etc.) varies depending on the depth of the learning layer. Specifically, a biological learning model with shallow layers is suitable for simulations with low resolution, while a biological learning model with deep layers is suitable for simulations with high resolution. For example, the biological learning model with shallow layers may be BLING, and the biological learning model with deep layers may be PlankTOM5. Therefore, the biological learning model differs for each learning layer, and outputs simulation results with the corresponding resolution. Using the example of Figure 1, the shallowest layer (first layer) on the input / output side has N 2 A biological learning model corresponding to the resolution is used. Similarly, in the next deep layer (layer 2), (N / 2) 2 A biological learning model corresponding to the resolution is used, and at the deepest layer (nth layer), (N / n) 2 A biological learning model that corresponds to the resolution is used. 2 The biological learning model corresponding to the resolution is N 2 It is a surrogate model of the biophysical model corresponding to the resolution. Similarly, (N / 2) 2 The biological learning model corresponding to the resolution is (N / 2) 2 A surrogate model of the biophysical model corresponding to the resolution (N / n) 2 The biological learning model corresponding to the resolution is (N / n) 2 The resolution is a surrogate model of the biophysical model corresponding to the resolution. Thus, the resolution in the present disclosure is proportional to the depth of the learning layer. Furthermore, the structure of the biological learning model and the corresponding biophysical model differ depending on the resolution.

[0022] As described above, the sub-learning models correspond to different resolutions for each learning layer. Therefore, the learning data used to train the sub-learning models also has a corresponding resolution for each learning layer. In the example of Figure 1, the learning data in the first layer is N 2 N output to a fluid physics model with specified resolution 2 Simulation results of resolution and N 2 N output to a biophysical model corresponding to the resolution 2 N coupled with the simulation results of resolution 2 The simulation results show the resolution. Similarly, the training data in the second layer is (N / 2) 2 Specify the resolution and output to the fluid physics model (N / 2) 2 Simulation results for resolution and (N / 2) 2 Output to a biophysical model corresponding to the resolution (N / 2) 2 Coupling with the simulation results of resolution (N / 2) 2 This is a simulation result of the resolution, and the learning data in the nth layer is (N / n) 2 Specify the resolution and output to the fluid physics model (N / n) 2 Simulation results of resolution and (N / n) 2 Output to a biophysical model corresponding to the resolution (N / n) 2 Coupling with the simulation results of resolution (N / n) 2 10 shows the results of a simulation of resolution. As described above, the training data used in this disclosure has different resolutions for each training layer.

[0023] Furthermore, a sub-learning model that has been trained using training data may input its training results to a subsequent sub-learning model. That is, in its training, the sub-learning model may use the training results of a previous sub-learning model in addition to training data at the corresponding resolution. For example, the parameters of a sub-learning model optimized through training may be used as the initial values ​​of the parameters of a subsequent sub-learning model. In other words, a sub-learning model may be trained using training data after using the parameters of a previous sub-learning model as its initial parameters. The parameters of a training model are, for example, weighting coefficients between nodes in a neural network that realizes the training model. Note that the terms "subsequent stage" and "previous stage" used here indicate the order of input and output, regardless of the depth of the learning layer. The deeper layer in the learning layer does not necessarily mean the "subsequent stage." As is clear from Figure 1, in a decoder, the shallower layer is the "subsequent stage."

[0024] A simulation device 10 according to the present disclosure will be described with reference to FIG. 2 . The simulation device 10 simulates a phenomenon in a marine ecosystem. For example, when a phenomenon in a marine ecosystem at a certain time is input to the simulation device 10, the simulation device 10 outputs the phenomenon in the marine ecosystem after a predetermined time has elapsed as a simulation result. In other words, the simulation device 10 simulates how the phenomenon in the marine ecosystem changes after the predetermined time has elapsed. As shown in FIG. 2 , the simulation device 10 includes a generation unit 11, a learning unit 12, a simulation unit 13, and a storage unit 14.

[0025] The generator 11 generates training data using a fluid physics model with the same structure regardless of resolution and a biophysical model with a different structure depending on the resolution. First, the generation of training data for the sub-training models in the first layer of the above-mentioned training model will be described, followed by the generation of training data for the sub-training models in the second to nth layers.

[0026] First, the generation unit 11 generates training data for the fluid physics model using a fluid physics model that expresses a fluid phenomenon.2 Specify the resolution and add N to the fluid physics model. 2 The simulation results of the resolution are output, and the simulation results are used as learning data for the fluid learning model.

[0027] Furthermore, the generation unit 11 generates training data for the biophysical model using a biophysical model that expresses a biological (biogeochemical) phenomenon. 2 N in biophysical models corresponding to resolution 2 The results of the simulation of the resolution are output, and the results of the simulation are used as learning data for the biological learning model.

[0028] Finally, the generation unit 11 couples these two sets of training data to generate N 2 Generate training data for resolution.

[0029] Similarly, the generation unit 11 generates training data for the sub-training models in the second to n-th layers. 2 Specify the resolution and output to the fluid physics model (N / 2) 2 Simulation results for resolution and (N / 2) 2 Output to a biophysical model corresponding to the resolution (N / 2) 2 Coupling with the simulation results of resolution (N / 2) 2 Generate training data with a resolution of (N / n) 2 Specify the resolution and output to the fluid physics model (N / n) 2 Simulation results of resolution and (N / n) 2 Output to a biophysical model corresponding to the resolution (N / n) 2 Coupling with the simulation results of resolution, (N / n) 2 Generate training data for resolution.

[0030] As described above, the generation unit 11 generates the fluid learning model having the same structure regardless of the resolution and the N 2 Resolution ~ (N / n) 2 Using biophysical models with different structures corresponding to different resolutions, 2Resolution ~ (N / n) 2 Generate training data for resolution.

[0031] The fluid physics model is, for example, a fluid physics model described by the Navier-Stokes equations, and the biophysical model is, for example, Biogeochemistry with Light, Iron, Nutrients and Gas (BLING), PlankTOM5, HAMOCC, PISCES, REcom2, or COBALT.

[0032] The learning unit 12 trains the sub-learning models in each layer of the above-described learning model using the learning data with different resolutions for each learning layer generated by the generation unit 11. In other words, the learning unit 12 trains the sub-learning models for each learning layer using the learning data with a resolution corresponding to the sub-learning model. First, the training of the sub-learning models in the first layer of the learning model will be described, followed by the training of the sub-learning models in the second to n-th layers. Note that the training data trained on the sub-learning models is not limited to that generated by the generation unit 11 and may be generated separately from the simulation device 10. In this case, the simulation device 10 may omit the generation unit 11 from its configuration.

[0033] First, the learning unit 12 2 The sub-learning model in the first layer is trained using the training data of resolution. The sub-learning model in the first layer is trained using the fluid learning model with the same structure in each layer and the N 2 and a biological learning model corresponding to the resolution. Note that the fluid learning model is a surrogate model of the above-mentioned fluid physics model, and the biological learning model is a surrogate model of the above-mentioned biophysical model.

[0034] Next, the learning unit 12 calculates (N / 2) 2 Using the training data of resolution, the sub-training model in the second layer is trained, and... (N / n) 2 The sub-learning model in the nth layer is trained using the training data of resolution. The sub-learning model in the second layer is trained using the fluid learning model with the same structure in each layer and (N / 2) 2The sub-learning model in the nth layer is a fluid learning model with the same structure in each layer and (N / n) 2 The biological learning model has a structure that differs in each layer depending on the resolution.

[0035] Here, the learning unit 12 may set the parameters of the sub-learning model of the first-layer encoder that has been trained as the initial values ​​of the parameters of the sub-learning model of the second-layer encoder. Similarly, the learning unit 12 may set the parameters of the sub-learning model of the second-layer decoder that has been trained as the initial values ​​of the parameters of the sub-learning model of the first-layer decoder.

[0036] That is, the learning unit 12 may set the parameters of the sub-learning model of the encoder of the k-1th layer that has been trained as the initial values ​​of the parameters of the sub-learning model of the encoder of the kth layer. Similarly, the learning unit 12 may set the parameters of the sub-learning model of the decoder of the kth layer that has been trained as the initial values ​​of the parameters of the sub-learning model of the decoder of the k-1th layer. Note that k is a positive integer between 2 and n, and when k=n, the "sub-learning model of the encoder of the kth layer" can be read as the "sub-learning model of the decoder of the kth layer."

[0037] As described above, the learning unit 12 2 Resolution ~ (N / n) 2 The learning unit 12 uses the learning data of the resolution to train the sub-learning models in the first to n-th layers. 2 Resolution ~ (N / n) 2 Using the training data of resolution, 2 Resolution ~ (N / n) 2 The learning unit 12 trains the above-described learning model by using a different biological learning model corresponding to the resolution.

[0038] The simulation unit 13 performs a simulation using the learned learning model trained by the learning unit 12. For example, when the simulation unit 13 receives a marine ecosystem phenomenon at a certain time, it outputs the marine ecosystem phenomenon after a predetermined time has elapsed as a simulation result. In other words, the simulation unit 13 simulates how the marine ecosystem phenomenon changes after a predetermined time has elapsed. Specifically, when the simulation unit 13 uses a learning model such as that shown in FIG. 1, the encoder sub-learning models in each layer sequentially convolve the simulation results for input data input to the encoder sub-learning model in the first layer. When the deepest layer is reached, the decoder sub-learning models in each layer sequentially deconvolve the simulation results and output them as the final simulation results of the learning model. Note that the above-mentioned skip connection may be used in this deconvolution.

[0039] The storage unit 14 stores the above-mentioned learning model. The learning unit 12 causes the learning model stored in the storage unit 14 to learn, and the simulation unit 13 causes the learning model stored in the storage unit 14 to perform simulation.

[0040] The storage unit 14 may also store the above-described physical model. The generation unit 11 may generate training data using the physical model stored in the storage unit 14. The storage unit 14 may also store training data.

[0041] The learning steps of the learning model of the present disclosure will be described with reference to FIG. 3 . First, the generator 11 generates learning data at different resolutions using a fluid physics model and a biophysical model (step S11). Specifically, the generator 11 couples learning data output to the fluid physics model at a specified resolution with learning data output to the biophysical model corresponding to the same resolution to generate learning data at the specified resolution. The generator 11 generates learning data at different resolutions as learning data for sub-learning models in the first to nth layers of the above-mentioned learning model.

[0042] Next, the learning unit 12 trains the sub-learning models using training data with different resolutions for each training layer (step S12). Here, the order in which the learning unit 12 trains the sub-learning models will be described with reference to FIG.

[0043] First, the learning unit 12 trains the sub-learning model of the encoder in the first layer. If there is a sub-learning model in a subsequent stage (YES in step S13), the learning unit 12 sets the parameters of the trained sub-learning model as the initial values ​​of the parameters of the subsequent sub-learning model (step S14). Here, in the example of FIG. 1, the sub-learning model of the encoder in the second layer is located subsequent to the sub-learning model of the encoder in the first layer. Therefore, the learning unit 12 sets the parameters of the sub-learning model of the encoder in the first layer as the initial values ​​of the parameters of the sub-learning model of the encoder in the second layer.

[0044] After this, the learning unit 12 repeats steps S12 and S13 (YES) until there are no further sub-learning models (NO in step S13). Returning to the example in Fig. 1, there are no further sub-learning models after the decoder sub-learning model in the first layer. This completes the learning step of the learning model, as shown in Fig. 3.

[0045] Finally, the simulation steps of the simulation device 10 (simulation unit 13) of the present disclosure will be described with reference to Fig. 4. It is assumed that the learning model has already been learned by the learning unit 12.

[0046] When using a learning model such as that shown in Figure 1, the encoder sub-learning models in each layer sequentially perform convolution on input data input to the encoder sub-learning model in the first layer (step S21). When the deepest layer is reached, the decoder sub-learning models in each layer sequentially perform deconvolution (step S22). Note that the above-mentioned skip connection may be used in this deconvolution.

[0047] As described above, the simulation device 10 of the present disclosure includes a simulation unit 13 that simulates a phenomenon in a marine ecosystem using a learning model that includes a sub-learning model for each learning layer, and a learning unit 12 that causes the sub-learning models to learn using learning data with different resolutions for each learning layer. Also, in the simulation device 10 of the present disclosure, the sub-learning models include a fluid learning model that simulates a phenomenon in a fluid and an organism learning model that simulates a phenomenon in an organism, and the organism learning model differs for each learning layer.

[0048] The simulation device 10 of the present disclosure focuses on the fact that, unlike fluid phenomena, biological (biogeochemical) phenomena vary depending on the resolution at which the simulation is performed when simulating phenomena in a marine ecosystem. Specifically, the simulation device 10 uses a learning model that includes sub-learning models, each of which includes a fluid learning model with the same structure and an organism learning model with a different structure corresponding to the resolution, for each learning layer, and further trains the sub-learning models using learning data at the corresponding resolution. This enables the simulation device 10 to accurately simulate phenomena in a marine ecosystem, which consist of fluid phenomena that remain constant regardless of resolution and organism phenomena that vary depending on the resolution.

[0049] In particular, the learning model disclosed herein increases the resolution used for learning the biological learning model in proportion to the depth of the learning layer, and the biological learning model also corresponds to the resolution. This allows the simulation device 10 to take advantage of the multigrid method, which calculates from a low resolution and then uses the low-resolution solution for calculations at a high resolution. Specifically, the simulation device 10 can reduce learning costs and improve calculation accuracy by using the coefficients and outputs of a neural network trained at a low resolution as initial values ​​and inputs for learning at a high resolution.

[0050] Note that a learning device including the above-described learning unit 12, the learning steps of the learning unit 12, and a learning model (trained model) that has been trained by the learning steps of the learning unit 12 are also aspects of the present disclosure.

[0051] In the above disclosure, the resolution of a sub-learning model is proportional to the depth of the learning layer in which the sub-learning model is provided, but this is not limited to this. To obtain a desired simulation result, the resolution may be adjusted so that it is not partially proportional.

[0052] The simulation device 10 described above can be, for example, a general-purpose computer system including a central processing unit (CPU) 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Fig. 5. In this computer system, the ocean carbon flux uncertainty quantification system is realized by the CPU 901 executing a predetermined program loaded onto the memory 902. This program can be recorded on a computer-readable non-transitory recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or can be distributed via a network.

[0053] 10 Simulation device 11 Generation unit 12 Learning unit 13 Simulation unit 14 Storage unit

Claims

1. A simulation device comprising: a simulation unit which has sub-learning models corresponding to different resolutions for each layer, and which simulates phenomena in a marine ecosystem using a learning model which inputs and outputs simulation results between the sub-learning models; and a learning unit which trains the sub-learning models using learning data of the corresponding resolution and the learning results of the previous sub-learning model; wherein the sub-learning models comprise: a fluid learning model which simulates fluid phenomena; and an organism learning model which simulates organism phenomena; wherein the structure of the fluid learning model is the same regardless of the layer, and the structure of the organism learning model is different for each layer.

2. The simulation device according to claim 1, wherein the resolution is proportional to the depth of the hierarchy.

3. The simulation device according to claim 2, wherein the learning unit trains the sub-learning model using the biological learning model that differs in accordance with the resolution.

4. The simulation device according to claim 3, further comprising a generation unit that generates the learning data using a fluid physics model that represents the fluid phenomenon and a biophysical model that represents the biological phenomenon, wherein the generation unit generates the learning data using different biophysical models corresponding to the resolution.

5. The simulation device according to claim 1, wherein the learning unit uses parameters of the sub-learning model in a previous stage as initial values ​​of parameters of the sub-learning model.

6. A simulation method comprising: a simulation step of simulating a phenomenon in a marine ecosystem using a learning model which has sub-learning models corresponding to different resolutions for each hierarchy and which inputs and outputs simulation results between the sub-learning models; and a learning step of having the sub-learning models learn using learning data of the corresponding resolution and the learning results of the previous sub-learning model, wherein the sub-learning models comprise a fluid learning model which simulates a fluid phenomenon, and an organism learning model which simulates an organism phenomenon, wherein the structure of the fluid learning model is the same regardless of the hierarchy, and the structure of the organism learning model is different for each hierarchy.

7. A program for causing a computer to operate as each part of the simulation device according to any one of claims 1 to 5.

Images