Information processing device, plasma magnetic flux calculation method, and program

Abstract

This information processing device comprises an input unit to which a signal acquired by a magnetic sensor in a fusion reactor is input as input data, and an inference unit that uses a model having a plurality of neural networks to calculate a plasma magnetic flux from the input data.

Description

Information Processing Apparatus, Plasma Flux Calculation Method, and Program

[0001] The present invention relates to a technique for calculating the magnetic flux (plasma magnetic flux) created outside the plasma by the current flowing in the fusion plasma.

[0002] In the prior art, the magnetic flux (plasma magnetic flux) created outside the plasma by the current flowing in the fusion plasma in a fusion experimental reactor is calculated by solving an inverse problem for a boundary integral equation (Non-Patent Documents 1 and 2) based on information obtained by analyzing signals acquired by magnetic sensors installed in the experimental reactor.

[0003] On the other hand, the plasma magnetic flux varies greatly depending on the current flowing in the plasma and the shape of the plasma. For example, in the case of current, it has a problem of having a dynamic range of about 100 to 1000 times. Conventionally, the Cauchy Condition Surface method has been developed as a method for solving the inverse problem for the boundary integral equation (Non-Patent Document 2), but the application of neural networks (NNs) has not been made so far.

[0004] K. Kurihara, 「A new shape reproduction method based on the Cauchy-condition surface for real-time tokamak reactor control」, Fusion Engineering and Design, Vol 51-52, 1049-1057 (2000). https: / / doi.org / 10.1016 / s0920-3796(00)00174-5.V. D. Pustovitov,「Decoupling in the problem of tokamak plasma response to asymmetric magnetic perturbations」, Plasma Physics and Controlled Fusion, Vol.50, 105001 (2008). https: / / doi.org / 10.1088 / 0741-3335 / 50 / 10 / 105001.Y. Miyata, T. Suzuki, M. Takechi, H. Urano, and S. Ide, “Assessment of the accuracy of plasma shape reconstruction by the Cauchy condition surface method in JT-60SA,” Review of Scientific Instruments, Vol. 86(7), 073511 (2015). https: / / doi.org / 10.1063 / 1.4926997S. Joung, et al., 「Deep neural network Grad-Shafranov solver constrained with measured magnetic signals」, Nuclear Fusion, Vol.60, 2020, 016034, https: / / doi.org / 10.1088 / 1741-4326 / ab555fJ. Lu, et al., 「Fast equilibrium reconstruction by deep learning on EAST tokamak」, AIP Advances, Vol.13, 2023, 075007, https: / / doi.org / 10.1063 / 5.0152318.

[0005] In conventional computations that do not use neural networks, the main approach employed is to limit the number of conditions for representing the inverse problem in order to perform high-speed computation in real time. The determination of the number of conditions is usually empirically determined through simulations or experiments (Non-Patent Document 3).

[0006] When training with a neural network (NN), it is common practice to use general mean squared error or gradient error as the loss function, as described in Non-Patent Documents 4 and 5. However, it is difficult to train a single NN on training data that includes various plasma states. For example, training works well during periods when the plasma flux is large and the output signal of the magnetic sensor is large (flat top), but training does not progress for plasmas during periods when the output signal is weak (rise / fall), resulting in low prediction accuracy.

[0007] This invention has been made in view of the above points, and aims to provide a technology that enables the calculation of plasma magnetic flux using a neural network.

[0008] According to the disclosed technology, an information processing device is provided that includes an input unit that receives signals acquired by a magnetic sensor in a nuclear fusion reactor as input data, and an inference unit that calculates plasma magnetic flux from the input data using a model having multiple neural networks.

[0009] The disclosed technology provides a method for calculating plasma magnetic flux using a neural network.

[0010] This is a diagram showing an example configuration of the information processing device 100. This is a flowchart explaining the operation of the information processing device 100. This is a diagram showing an example configuration of the information processing device 200. This is a diagram explaining the arrangement of magnetic sensors. This is a flowchart explaining the operation of the information processing device 200. This is a diagram showing an example configuration of Model 300 (parallel NN). This is a diagram showing an example of plasma magnetic flux output. This is a diagram showing an image of the time change of the weights of each expert NN output by the weighted NN. This is a diagram showing an example of the hardware configuration of the device.

[0011] Hereinafter, embodiments of the present invention (this embodiment) will be described with reference to the drawings. The embodiments described below are merely examples, and the embodiments to which the present invention is applied are not limited to the embodiments described below.

[0012] The following describes the configuration and operation of the apparatus for calculating (or reconstructing) plasma magnetic flux using a neural network model.

[0013] In this embodiment, separate devices are used for the learning phase and the inference phase, but the same device may be used for both phases.

[0014] Although the following explanation uses a fusion experimental reactor, this is merely an example, and the technology according to the present invention is applicable to any fusion reactor. A fusion reactor may be an experimental reactor, a prototype reactor, or a commercial reactor.

[0015] (Device configuration for the learning phase) Figure 1 shows an example configuration of the information processing device 100 that performs learning. The information processing device 100 that performs learning may also be called a learning device.

[0016] As shown in Figure 1, the information processing device 100 comprises an input unit 110, a learning unit 120, and an output unit 130. The learning unit 120 holds a model 300 to be learned. The model is a parallel neural network (NN) model, and its structure will be described later.

[0017] The operation of the information processing device 100 will be explained according to the steps in the flowchart in Figure 2. In S101, learning data is input from the input unit 110. The learning data is a set of pairs of input data to the model and correct output data from the model. As learning data, for example, measured values ​​from a plasma experiment or calculation results from a physical simulation that simulates it can be used. The input data is the signal from the magnetic sensor, and the correct data is the plasma magnetic flux corresponding to that signal.

[0018] In S102, the learning unit 120 inputs the input data to the model 300 and learns the model 300 by updating the parameters of the model 300 so that the error between the output from the model 300 and the correct data is minimized. For learning, gradient methods such as gradient descent and backpropagation are used.

[0019] When training is complete, in S103, the output unit 130 outputs the trained model 300 (specifically, model parameters, hyperparameters, etc.). The trained model 300 is used in the inference phase.

[0020] (Device Configuration for Inference Phase) Figure 3 shows an example configuration of the information processing device 200 that performs inference. The information processing device 200 that performs inference may also be called an inference device. Alternatively, the information processing device 200 may also be called a plasma magnetic flux reconstruction device. The information processing device 200 is a device that reconstructs the plasma magnetic flux, which is the magnetic flux created outside the plasma by the electric current flowing through the fusion plasma.

[0021] As shown in Figure 3, the information processing device 200 includes an input unit 210, an inference unit 220, an output unit 230, and an analysis unit 240. Alternatively, it may include only one of the output unit 230 or the analysis unit 240.

[0022] The inference unit 220 holds the trained model 300. Alternatively, the inference unit 220 can be considered as the model 300 itself. The trained model 300 is the model trained by the information processing unit 100 described above. Furthermore, if the information processing unit 200 also trains the model 300, the trained model 300 may be the model trained by the information processing unit 200.

[0023] In this embodiment, the fusion experimental reactor (and the prototype and commercial reactors as well) has a mechanism for confining plasma by creating a magnetic field within a torus-shaped vacuum vessel. Multiple magnetic sensors, described later, are arranged outside the vacuum vessel. Figure 4 shows a schematic cross-section of the torus-shaped vacuum vessel at the location of the magnetic sensors. Although Figure 4 shows an example where the magnetic sensors are arranged on the outer wall of the vacuum vessel, the configuration is not limited to this, and the magnetic sensors may also be installed on the inner wall of the vacuum vessel.

[0024] The operation of the information processing device 200 will be explained according to the procedure in the flowchart of Figure 5. In S201, input data is received from the input unit 210. The input data is a signal acquired by the magnetic sensor of the nuclear fusion experimental reactor. Here, the input data used is the total magnetic sensor signal minus the external magnetic flux (magnetic flux generated by coils used for control, etc.). The subtraction process may be performed before input to the input unit 210, or it may be performed by the input unit 210. The above-mentioned coil is, for example, a poloidal field coil (PF coil). The external magnetic flux can be determined by measuring the current flowing through the coil and obtaining the value of that current.

[0025] The input data is time-series data. For example, the total magnetic sensor signals at time t (units are Weber or Tesla) are φ ｍｅａｓ Let (t) be the external magnetic flux (in Weber or Tesla units) and φ ｅｘｔ If (t), then the input data for time t is φ ｍｅａｓ (t) - φ ｅｘｔ (t)

[0026] In S202, the inference unit 220 inputs input data to the model 300 and obtains output data from the model 300. The output data from the model 300 consists of the plasma magnetic flux and the weights of each parallel NN output from the weighted NN described later.

[0027] The output data described above is time-series data. The output plasma magnetic flux at time t, corresponding to the input data at time t, is, for example, the two-dimensional spatial distribution of plasma magnetic flux in the cross-section of the vacuum vessel that confines the plasma (e.g., the plasma magnetic flux value at each point in a mesh divided into planes).

[0028] The plasma magnetic flux is passed to the output unit 230, and the weights are passed to the analysis unit 240. In S203, the output unit 230 displays the plasma magnetic flux as, for example, a two-dimensional spatial distribution. The analysis unit 240 performs an analysis on the weights of the parallel NN (e.g., plasma state diagnosis) and outputs the analysis results.

[0029] (Configuration of Model 300) Figure 6 shows an example configuration of a parallel NN used as Model 300. The parallel NN has multiple expert NNs arranged in parallel and a weighted NN.

[0030] The input data is fed into each expert NN and the weighted NN. The weighted NN calculates the activation level (weight) of each expert for each input data.

[0031] In this parallel NN, for each input data, each expert NN in all expert NNs predicts the plasma magnetic flux, and the weighted NN calculates the weights of each expert NN's prediction, thereby outputting a weighted average of the multiple predictions.

[0032] The structure of the NN in each expert NN and weighted NN is not limited to a specific structure. Furthermore, the NN structures in each expert NN and weighted NN may be the same or different. Examples of NNs that can be used in each expert NN and weighted NN include fully connected layers, Transformers, Transformer encoders, or Transformer decoders.

[0033] For example, suppose we have expert NNs 1 through 5, and their respective predicted values ​​are p1, p2, p3, p4, and p5. Also, let w1, w2, w3, w4, and w5 be the weights calculated by the weighted NN for each of the expert NNs 1 through 5. Furthermore, let w1 + w2 + w3 + w4 + w5 = 1. In this case, the output value will be "w1 × p1 + w2 × p2 + w3 × p3 + w4 × p4 + w5 × p5". This output value corresponds to the weighted average of the output values ​​of expert NNs 1 through 5.

[0034] Furthermore, since each expert NN and weighted NN are trained simultaneously using gradient descent, the weighted NN automatically divides the data according to its characteristics, and each expert NN continues to train on the divided data. However, in order to efficiently divide the data, it is also possible to analyze the characteristics of the data using clustering or other methods and devise ways to adjust hyperparameters such as the total number of expert NNs and the number of expert NNs selected at output.

[0035] (Output Example) Figure 7 shows the correct value of the plasma magnetic flux, the output value from the information processing device 200 (weighted average of expert NNs 1 to 5), and the output values ​​from each expert NN. Figure 7 is based on the input data when the system is flat-topped. In the example shown in Figure 7, the output values ​​of expert NN1 and expert NN5 are mainly weighted and averaged to produce the output.

[0036] Figure 8 shows an image of the time evolution of the weights of each expert NN output by the weighted NN.

[0037] For example, by the analysis unit 240 pre-holding the changes in the weights of each expert NN according to the passage of time as shown in FIG. 8, the analysis unit 240 compares the data of the time change with the weights output from the inference unit 220 (model 300) simultaneously with the reconstruction of the plasma flux, thereby inferring the state of the plasma at that time (e.g., startup / flat top / shutdown).

[0038] As described above, according to the technique of the present embodiment, by weighting according to various plasma states and selectively using the expert NNs, the prediction accuracy can be improved. For example, by training expert NNs specialized for flat top and startup / shutdown respectively, and training the weighting NN to change the weighting according to the input, the prediction accuracy of startup / shutdown can be improved.

[0039] Also, as described above, by the analysis unit 240 analyzing the weight distribution, the difference in the plasma state can be grasped. For example, as shown in FIG. 8, by analyzing the weights and the NNs used for each time period, the plasma state change for each time period can be inferred.

[0040] Also, the analysis unit 240 may investigate the number of expert NNs required for prediction. The "expert NNs required for prediction" is, for example, an expert NN whose weight is equal to or greater than a predetermined threshold.

[0041] Investigating the number of expert NNs required for prediction corresponds to classifying the entire data, so from this investigation, it can be inferred how many different plasma states exist in the data.

[0042] (Hardware configuration example) Any of the devices (information processing devices 100, 200, learning device, inference device) described in the present embodiment can be realized, for example, by causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0043] In other words, the device can be realized by using hardware resources such as the CPU and memory built into a computer to execute a program corresponding to the processing performed by the device. The program can be recorded on a computer-readable recording medium (such as portable memory), saved, and distributed. It can also be provided via a network, such as the Internet or email.

[0044] Figure 9 shows an example of the hardware configuration of the computer described above. The computer in Figure 9 has a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., all of which are interconnected by bus B. The computer may also be equipped with a GPU.

[0045] The program that enables processing on the computer is provided on a recording medium 1001, such as a CD-ROM or memory card. When the recording medium 1001 containing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001; it may also be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files and data.

[0046] When there is an instruction to start a program, the memory device 1003 reads the program from the auxiliary storage device 1002 and stores it. The CPU 1004 realizes the functions related to the device according to the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network or the like. The display device 1006 displays a GUI (Graphical User Interface) or the like according to the program. The input device 1007 is composed of a keyboard, a mouse, buttons, or a touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the calculation result.

[0047] The functions of the elements disclosed in this specification can be implemented using circuitry or processing circuitry including combinations thereof, which are programmed using a general-purpose processor, a special-purpose processor, an integrated circuit, ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), conventional circuits, and / or one or more programs stored in one or more memories, or configured in other ways to execute the disclosed functions. Since a processor includes transistors and other circuits, it is regarded as processing circuitry or circuitry. The processor may be a programmed processor that executes a program stored in a memory. In the present disclosure, a circuit, unit, or means is hardware that executes the recited functions or hardware programmed to execute the recited functions. The hardware may be any hardware as long as it is the hardware disclosed in this specification and is programmed or configured to execute the recited functions.

[0048] (Effect of the Embodiment) As described above, with the technology according to the present embodiment, it becomes possible to calculate the plasma magnetic flux using the NN.

[0049] Furthermore, the technology according to this embodiment allows for the real-time calculation (reconstruction) of the plasma magnetic flux in a fusion reactor, thereby enabling accurate determination of the plasma's position within the reactor. This contributes to the safe control and operational stabilization of the fusion reactor.

[0050] More specifically, this enables the reconstruction of plasma magnetic flux with high precision for various plasma states. Furthermore, by analyzing the weight distribution, differences in plasma states can be identified, allowing for the prediction of plasma state changes. Additionally, by analyzing the optimal number of expert neural networks, the entire dataset can be classified, allowing for the estimation of how many different plasma states exist within the data.

[0051] The following additional information is disclosed regarding the embodiments described above.

[0052] <Note> (Note 1) An information processing device comprising an input unit that inputs signals acquired by a magnetic sensor in a nuclear fusion reactor as input data, and an inference unit that calculates plasma magnetic flux from the input data using a model having multiple neural networks.

[0053] (Appendix 2) An information processing device, wherein the information processing device is an information processing device that reconstructs plasma magnetic flux, which is the magnetic flux created outside the plasma by an electric current flowing through a fusion plasma, and comprises an input unit that inputs a signal obtained by a magnetic sensor in a fusion reactor, which has been processed by subtracting the external magnetic flux from the total magnetic sensor signal, as input data, and an inference unit that holds a model having multiple neural networks that corresponds to multiple plasma states and performs calculations of the plasma magnetic flux, and uses the model to reconstruct the plasma magnetic flux from the input data.

[0054] (Appendix 3) The information processing device according to Appendix 1 or 2, wherein the plurality of neural networks comprises a plurality of expert neural networks arranged in parallel and a weighting neural network that calculates weights for the output values ​​from each expert neural network.

[0055] (Appendix 4) The inference unit outputs a weighted average of the output values ​​from the multiple expert neural networks, using the weights calculated by the weighted neural network, as the plasma magnetic flux value, as described in Appendix 3.

[0056] (Appendix 5) The inference unit is the information processing device described in Appendix 3, which outputs weights for each expert neural network.

[0057] (Appendix 6) The information processing apparatus according to Appendix 5, further comprising an analysis unit for estimating the state of the plasma based on the weights.

[0058] (Appendix 7) The information processing apparatus according to Appendix 6, wherein the analysis unit estimates the state of the plasma based on data showing the time change of weights for each expert neural network and the weights output from the inference unit.

[0059] (Appendix 8) A plasma magnetic flux calculation method executed by an information processing device, comprising: an input step of inputting a signal acquired by a magnetic sensor in a nuclear fusion reactor as input data; and an inference step of calculating the plasma magnetic flux from the input data using a model having a plurality of neural networks.

[0060] (Appendix 9) A plasma magnetic flux calculation method performed by an information processing device, wherein the information processing device is an information processing device that reconstructs plasma magnetic flux, which is the magnetic flux created outside the plasma by an electric current flowing through a fusion plasma, and comprises an input step of inputting a signal obtained by a magnetic sensor in a fusion reactor, which has been processed by subtracting the external magnetic flux from the total magnetic sensor signal, as input data, and an inference step of reconstructing the plasma magnetic flux from the input data using a model having multiple neural networks that corresponds to multiple plasma states and performs calculations of the plasma magnetic flux.

[0061] (Appendix 10) A program for causing a computer to function as one of the components of an information processing device described in any one of Appendix 1 to 7.

[0062] (Appendix 11) A non-temporary storage medium storing a program for causing a computer to function as a part of an information processing device described in any one of the appendix items 1 through 7.

[0063] Although this embodiment has been described above, the present invention is not limited to this specific embodiment, and various modifications and changes are possible within the scope of the gist of the invention as described in the claims.

[0064] This patent application claims priority based on Japanese Patent Application No. 2024-190696, filed on 30 October 2024, and the entire contents of Japanese Patent Application No. 2024-190696 are incorporated herein by reference.

[0065] 100, 200 Information processing device 110, 210 Input unit 120 Learning unit 130, 230 Output unit 220 Inference unit 240 Analysis unit 300 Model 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims

1. An information processing device comprising an input unit that receives signals acquired by a magnetic sensor in a nuclear fusion reactor as input data, and an inference unit that calculates plasma magnetic flux from the input data using a model having multiple neural networks.

2. An information processing device, wherein the information processing device is an information processing device that reconstructs plasma magnetic flux, which is the magnetic flux created outside the plasma by an electric current flowing through a fusion plasma, and comprises an input unit that inputs a signal obtained by a magnetic sensor in a fusion reactor, which has been processed by subtracting the external magnetic flux from the total magnetic sensor signal, as input data, and an inference unit that holds a model having multiple neural networks that corresponds to multiple plasma states and performs calculations of the plasma magnetic flux, and uses the model to reconstruct the plasma magnetic flux from the input data.

3. The information processing apparatus according to claim 1, wherein the plurality of neural networks comprises a plurality of expert neural networks arranged in parallel and a weighted neural network that calculates weights for the output values ​​from each expert neural network.

4. The information processing apparatus according to claim 3, wherein the inference unit outputs a weighted average of the output values ​​from the plurality of expert neural networks using the weights calculated by the weighted neural network as the plasma magnetic flux value.

5. The information processing apparatus according to claim 3, wherein the inference unit outputs weights for each expert neural network.

6. The information processing apparatus according to claim 5, further comprising an analysis unit for estimating the state of the plasma based on the weights.

7. The information processing apparatus according to claim 6, wherein the analysis unit estimates the state of the plasma based on data showing the time change of weights for each expert neural network and the weights output from the inference unit.

8. A plasma magnetic flux calculation method executed by an information processing device, comprising: an input step of inputting signals acquired by a magnetic sensor in a nuclear fusion reactor as input data; and an inference step of calculating plasma magnetic flux from the input data using a model having multiple neural networks.

9. A plasma magnetic flux calculation method executed by an information processing device, wherein the information processing device is an information processing device that reconstructs plasma magnetic flux, which is the magnetic flux created outside the plasma by an electric current flowing through a fusion plasma, and comprises an input step of inputting a signal obtained by a magnetic sensor in a fusion reactor, which has been processed by subtracting the external magnetic flux from the total magnetic sensor signal, as input data, and an inference step of holding a model having multiple neural networks that corresponds to multiple plasma states and performs calculations of plasma magnetic flux, and using the model to reconstruct the plasma magnetic flux from the input data.

10. A program for causing a computer to function as a component of an information processing device described in any one of claims 1 to 7.

11. A non-temporary storage medium storing a program for causing a computer to function as a component of an information processing device according to any one of claims 1 to 7.

Images