Ion source beam intensity prediction method and prediction device

Abstract

Predict the beam intensity of the ion source. [Solution] In the method for predicting the beam intensity of an ion source exemplified in the present disclosure, image data of a plasma emission image of an ion source (10) acquired by an imaging device is input to a machine learning model (20) including a neural network, and a predicted value of the beam intensity of an ion beam output by the ion source is output based on the output of the trained neural network of the machine learning model. The machine learning model also receives control numerical data, which is numerical data of controllable quantities for operating the ion source, and observation numerical data, which is numerical data of information observable during operation of the ion source. Once the neural network of the machine learning model has been trained with a training dataset, a prediction dataset, which includes the numerical data of the controllable quantities to be applied to the ion source and image data of the plasma emission of the ion source, can be input to output a predicted value of the beam intensity of the ion beam.

Description

[Technical Field]

[0001] The present disclosure relates to a method and apparatus for predicting beam intensity in an ion source, and more particularly, to a method and apparatus for predicting beam intensity in an ion source with plasma emission. [Background technology]

[0002] Accelerators are used to generate various charged particle beams with various energies. An ion source is installed in front of the accelerator to ionize atoms in gases or solids and output them as an ion beam toward the accelerator. The ion source ionizes the various ions used in the accelerator to a predetermined valence, and outputs an ion beam with the required intensity. Stable operation of the ion source is essential to accelerate the ions through the accelerator and form the desired irradiation beam.

[0003] There are many types of ion sources, but to operate them properly, various parameters must be adjusted appropriately. In the case of an ECR (electron cyclotron resonance) ion source, the parameters can number more than 10, including detailed settings such as microwave input and main valve opening. Summary of the Invention [Problem to be solved by the invention]

[0004] Stable operation of ion sources has been a challenge. Stable operation of an ion source requires careful adjustment of numerous parameters. The relationship between each adjusted parameter and operational indicators (such as ion beam output intensity) is not direct and depends on the experience of skilled operating engineers. For example, even if the ion source is continued to operate with the parameters fixed after adjustment, it is not uncommon for the ion beam output to fluctuate over time.

[0005] The present disclosure provides a method for estimating beam intensity during operation of an ion source by solving the above problem. [Means for solving the problem]

[0006] The inventors have noticed that in many ion sources, the ion generation process is accompanied by plasma emission, a phenomenon in which ions generated by ionization emit electromagnetic waves in a wavelength range that includes the visible range.

[0007] The present inventors noticed that an experienced operating engineer operates the ion source by observing the plasma light emission, i.e., the operating engineer recognizes that the state of the ion source is reflected in the plasma light emission. The present inventors came up with the idea of ​​predicting the beam intensity of an ion source by applying a machine learning image recognition technique to image data of the plasma light emission. They actually constructed a system that predicts the beam intensity based on image data of the plasma light emission image, evaluated its performance, and confirmed its usefulness. In this way, they created a solution to the above-mentioned problem.

[0008] That is, in one embodiment of the present disclosure, there is provided a method for predicting the beam intensity of an ion source, the method including: acquiring a plasma emission image of an ion source operating with plasma emission by an imaging device; inputting image data of the plasma emission image into a machine learning model including a neural network; inputting control numerical data, which is numerical data of controllable quantities for operating the ion source, and observation numerical data, which is numerical data of information observable during operation of the ion source, into the machine learning model, provided that the observation numerical data does not include the image data; training the machine learning model with a training data set including the numerical data of the controllable quantities to be given to the ion source, the beam intensity of an ion beam output by the ion source, and the image data of plasma emission from the ion source; and inputting a prediction data set including the numerical data of the controllable quantities to be given to the ion source and the image data of plasma emission from the ion source into the trained neural network of the machine learning model, and outputting a predicted value of the beam intensity of the ion beam output by the ion source based on the output of the trained neural network.

[0009] In one embodiment of the present disclosure, there is provided an ion source beam intensity prediction device including an imaging device that acquires a plasma emission image of an ion source operating with plasma emission, and a computer device including at least an arithmetic unit and a recording unit, wherein the computer device implements a machine learning model including a convolutional neural network (CNN) and at least one fully connected layer to which an output of the CNN is input, and the machine learning model inputs image data of the plasma emission image to the CNN, and inputs control numerical data, which is numerical data of controllable amounts for operating the ion source, and observation numerical data, which is numerical data of information observable during operation of the ion source, to the at least one fully connected layer, where the observation numerical data does not include the image data, and and an output of the beam intensity of the ion beam output by the ion source from a fully connected layer of the CNN, wherein the machine learning model is trained by the CNN and the at least one fully connected layer using a training dataset including the numerical data of the controllable quantities to be given to the ion source, the beam intensity of the ion beam output by the ion source, and the image data of plasma light emission of the ion source, and the trained CNN and the trained at least one fully connected layer output a predicted value of the beam intensity of the ion beam output by the ion source in response to input of a prediction dataset to the CNN and the at least one fully connected layer, the prediction dataset including the numerical data of the controllable quantities to be given to the ion source and the image data of plasma light emission of the ion source.

[0010] In this disclosure, the term "ion source" includes any type of ion source, particularly any ion source that operates with plasma emission, as further details are provided in the Detailed Description of the Invention section below. [Effects of the Invention]

[0011] According to any one of the embodiments of the present disclosure, the operation of the ion source can be stabilized by improving the prediction of the beam intensity of the ion source, which has previously required skill and depended on the individual. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a plan view illustrating the arrangement of an exemplary ion source according to an embodiment of the present invention. [Figure 2A-B] 2A and 2B are a perspective view (FIG. 2A) and a plan view (FIG. 2B) showing a magnet configuration in an ECR ion source, which is an exemplary ion source to which an embodiment of the present invention is applied. [Figure 3A-B] 3A and 3B are explanatory diagrams for explaining operation schemes of an ECR ion source, showing a conventional operation method (FIG. 3A) and an embodiment of the present invention (FIG. 3B). [Figure 4] Figure 4 shows an example of an image of the plasma emission from an operating ECR ion source. [Figure 5] FIG. 5 is a configuration diagram showing the configuration of an example machine learning model in an embodiment of the present invention. [Figure 6] FIG. 6 is a graph of the actual beam intensity used for machine learning in an embodiment of the present invention. [Figure 7] FIG. 7 shows a learning curve that examines how the loss defined by Equation 1 changes during the training and validation stages in an embodiment of the present invention. [Figure 8A-B] 8A and 8B are graphs confirming the reproducibility of actual measurement data by a trained machine learning model in an embodiment of the present invention, for training (FIG. 8A) and validation (FIG. 8B), respectively. [Figure 9A-B] Figures 9A and 9B are graphs confirming the reproducibility of actual measurement data by a trained machine learning model in an image-only ablation study for an embodiment of the present invention, for training (Figure 9A) and validation (Figure 9B), respectively. [Figure 10A-B] Figures 10A-B are graphs confirming the reproducibility of actual measurement data by a trained machine learning model in an ablation study using only numerical data, for training (Figure 10A) and validation (Figure 10B), respectively. [Figure 11A-B] Figures 11A-B are graphs confirming the reproducibility of actual measurement data by a trained machine learning model in an ablation study using grayscale images and numerical data for an embodiment of the present invention, for training (Figure 11A) and validation (Figure 11B), respectively. [Figure 12A-B] Figures 12A-B are graphs confirming the reproducibility of actual measurement data by a trained machine learning model in an ablation study using blurred images and numerical data for an embodiment of the present invention, for training (Figure 12A) and validation (Figure 12B), respectively. [Figure 13] FIG. 13 is a graph plotting measured and predicted beam intensities for an ion source operated under different conditions to verify an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, a method for predicting beam intensity of an ion source according to an embodiment of the present disclosure will be described. In the description, unless otherwise specified, common parts or elements are designated by common reference numerals. In addition, in the drawings, elements of each embodiment are not necessarily shown to scale.

[0014] 1. Ion source configuration FIG. 1 is a plan view illustrating the arrangement of an exemplary ion source according to this embodiment, and FIGS. 2A and 2B are a perspective view (FIG. 2A) and a plan view (FIG. 2B) showing the magnet configuration of an ECR ion source, which is an exemplary ion source. A typical accelerator experiment using ions generally uses an ion source 10 and an accelerator that accelerates ions from the ion source 10 to a required velocity. To obtain a high-energy beam required for the experiment at the accelerator exit, it is necessary to increase the intensity of the ion source 10 and stabilize its operation.

[0015] The ECR ion source, a typical ion source, can extract a highly charged ion beam by sequentially ionizing ions in a space confined by the magnetic field of a magnet. However, the ECR ion source has a wide range of operating parameters (described below), making it difficult to operate the beam intensity stably for long periods of time.

[0016] 2. Implementation Example (Ion Source Beam Intensity Prediction) When an experienced operator operates the ion source 10, approximately ten controllable values ​​are adjusted. The operator searches for a set of control values ​​that will produce a stable and as intense a beam as possible. In doing so, the operator adjusts the controllable values ​​using experience and intuition while looking at observable information. However, because there are many different types of ion sources with different designs, the operator must gain experience with a variety of ion sources. In particular, for those in which ions are prepared from solid raw materials, operating the ion source can be difficult because the process of converting the solid into vapor is the target of control.

[0017] The present inventors came up with the idea of ​​applying a machine learning technique to image data captured of the plasma light emission. It is easy for those skilled in the art to perform machine learning on a computer device including at least an arithmetic unit and a recording unit. Figures 2A and 2B are a perspective view (Figure 2A) and a plan view (Figure 2B) showing the magnet configuration of an ECR ion source, an exemplary ion source to which this embodiment is applied. Experienced operating engineers recognize that the state of the ion source is reflected in the plasma light emission, and have long operated the ion source while incorporating the plasma light emission into their observations. Therefore, the present inventors came up with the idea of ​​a method for predicting the beam intensity of an ion source by applying a machine learning image recognition technique to image data captured of the plasma light emission.

[0018] 3A and 3B are explanatory diagrams for explaining the operation scheme of the ECR ion source, showing a conventional operation method (FIG. 3A) and the operation method of this embodiment (FIG. 3B). Also, FIG. 4 shows an example of an image taken of the plasma emission of the ECR ion source during operation. [Table 1] [Table 2]

[0019] In ion source operation, the controllable values ​​shown in Table 1 and the observable information shown in Table 2 are the operating parameters that change the operating conditions. These can be directly manipulated or obtained during operation of the ion source 10. Furthermore, although the detailed state of the physical phenomena occurring in the plasma for ion beam generation at each point in time can be predicted based on physical models if physical quantities such as plasma electron density, plasma electron temperature, and ion confinement time are known, it is difficult to actually measure or estimate these physical quantities. Given that this is a phenomenon occurring in a vacuum, the information that can be directly obtained from the plasma where ionization of ions occurs is limited. The observable quantities shown in Table 2 are merely examples. As listed at the end of Table 2, images of the plasma emission state inside the ion source taken from the extraction port (ion outlet) of the ion source 10 contain information such as the brightness and color distribution of the plasma and the discharge status, as shown in Figure 4. Traditionally, operating engineers have empirically inferred the plasma state from images of plasma emission and controlled operation. In contrast, in this embodiment, as shown in Figure 3B, a machine learning model is used in which controllable values ​​and observable information are input (x) and beam quality, i.e., beam intensity and beam stability, are output (y), and the observable information explicitly includes an image of this plasma emission.

[0020] FIG. 5 is a diagram illustrating the configuration of an example machine learning model 20 according to this embodiment. The image data may be, for example, a 224 × 224 pixel visible video image captured by camera C ( FIG. 1 ) from the outlet of the ion source 10 ( FIG. 1 ) of the plasma light emission inside the ion source. In the following description, "image data" refers to this image data unless otherwise specified. This image data 210 has typical pixel values ​​of R (red), G (green), and B (blue). In this machine learning model 20, this image data 210 is first input to an image recognition neural network 220, which is a convolutional neural network (CNN) such as ResNet50. The outputs of the 2,048 nodes are input to a first fully connected layer 230, which has 512 node outputs. The outputs, along with the 17 pieces of numerical data in Tables 1 and 2 (excluding the image data), are input to a second fully connected layer 240, which has 64 node outputs, and then to a third fully connected layer 250, which has one node output. In the following explanation, "numerical data" refers to these 17 data unless otherwise specified. The output node of the third fully connected layer is assigned to the beam intensity (beam intensity), which is a predicted value of 290.

[0021] Using the above machine learning model 20, machine learning (training) was actually carried out and its performance was verified. Figure 6 is a graph of the actual beam intensity used for machine learning in this embodiment, with the horizontal axis representing real time and the vertical axis representing beam intensity (unit: eμA), which are actual measured values ​​using a Faraday cup FC (Figure 1). For the neural network for image recognition, ResNet50, a 50-layer CNN, was used, and the initial weights were values ​​learned using ImageNet. The ions generated by the ion source are 56 Fe 15+A beam was used. First, to train the machine learning model, an operating engineer operated the ion source while measuring the beam intensity. The curves shown in the graph in Figure 6 are from the measurement data. In this machine learning model, training and validation were performed using this measurement data. As shown in the graph in Figure 6, the measurement data is divided into a first half, an unused period, and a second half. Of these, the measurement data in the second half, which has a larger amount of data, was used for training, and the measurement data in the first half was used for validating the machine learning model. Specifically, the machine learning model was trained using the second half of the graph in Figure 6 (measurement time: 18 minutes 31 seconds, number of data points: 6,850). Next, the machine learning model was validated using the first half of the graph in Figure 6 (measurement time: 13 minutes 16 seconds, number of data points: 4,900).

[0022] FIG. 7 shows a learning curve obtained by examining how the loss defined by Equation 1 changes during the training and validation stages in this embodiment.

number

[0023] Figures 8A and 8B are graphs confirming the reproducibility of actual measurement data using the trained machine learning model of this embodiment, showing training (Figure 8A) and validation (Figure 8B), respectively. Each graph plots the same measurement data as shown in Figure 6, with values ​​predicted by the machine learning model overlaid. As can be seen from Figures 8A and 8B, the trained machine learning model reproduces the actual measurement data used for training extremely well. In contrast, the validation values ​​predicted by the trained machine learning model do not necessarily match the actual measurement data. However, the trained machine learning model successfully predicts the overall trend and characteristic features such as plasma discharge. Figures 8A and 8B each show the loss values ​​at epoch number 100.

[0024] Next, we performed an ablation study to investigate the impact of intentionally restricting the training data on the accuracy of the machine learning model. The data was restricted to contrast with the training data shown in Figure 8A and B, which used both images and numerical data: (1) images only, (2) numerical data only, (3) images and numerical data but with the images in grayscale, and (4) images and numerical data but with the images blurred.

[0025] 9A and 9B show the results using only images. Compared with the results shown in FIGS. 8A and 8B, which use both images and numerical data, the results using only images show that the loss does not increase significantly, indicating that image data is important for training a machine learning model for operating the ion source 10 and for predicting the beam intensity during operation in the trained machine learning model.

[0026] Figures 10A and 10B show the results using only numerical data. Compared to Figures 8A and 8B, which use both image and numerical data, the results using only numerical data show that the training itself was insufficient, and as a result, even the machine learning model after training was unable to make predictions.

[0027] From these results, it was confirmed that useful machine learning can be performed not only by using both images and numerical data (Figures 8A and 8B) but also by using only images (Figures 9A and 9B), and that image data is useful observation data for machine learning for operating the ion source 10.

[0028] Figures 11A and 11B show the results using both image and numerical data, but with the images converted to grayscale. As is clear from a comparison with Figures 8A and 8B, even with grayscale images containing only brightness information, it is possible to perform useful machine learning, and it was confirmed that the image data is useful observation data for machine learning for the operation of the ion source 10. However, it was also confirmed that the loss of beam intensity data estimated by the trained machine learning model was greater than when RGB color information was used, resulting in a decrease in the accuracy of the machine learning.

[0029] Figures 12A and 12B show the results for a blurred image using both image and numerical data. A comparison with Figures 8A and 8B and Figures 11A and 11B reveals that although training is possible, the prediction loss of the trained machine learning model is greater when the image is blurred than when it is grayscale (Figures 11A and 11B). This confirms the importance of subtle brightness information, or fidelity, present in numerous pixels in image data for machine learning. This demonstrates the usefulness of using image data and demonstrates that physical reality is reflected in image data, with information about the patterns (shapes) of brightness and darkness being particularly important. Furthermore, the focus of the optical system used to capture the image is important.

[0030] Furthermore, to confirm the usefulness of the machine learning model of this embodiment, training and prediction were performed using data acquired when the ion source 10 was operated under different conditions. Here, an ECR ion source was actually operated as the ion source 10 under operating conditions different from those shown in FIGS. 12A and 12B. Specifically, the RF source (not shown) that provides microwave power to the ECR ion source was changed to a different device from the previous one. In addition, beam intensity data obtained by intentionally scanning parameters by changing the currents in the mirror coils MC1 and MC2 (FIGS. 1 and 2) was used. The machine learning model of this embodiment was trained using image data and numerical data obtained during these operations.

[0031] Figure 13 is a graph plotting the measured and predicted beam intensity values ​​for an ion source operated under different conditions to verify this embodiment. The actual measurement is obtained from the readout value of the Faraday cup FC (Figure 1), and the measured value is 0 eμA when the Faraday cup FC is retracted. As shown in Figure 13, the measured value fluctuates due to fluctuations in the current of the mirror coils MC1 and MC2, and is observed except when the Faraday cup FC is retracted. The predicted value does not necessarily show the same value as the actual measurement, but fluctuations in the actual measurement, such as increases, decreases, or maintenance of the value, are similarly reproduced in the predicted value.

[0032] From the results obtained so far, it was confirmed that there was a certain degree of reproducibility in the prediction of the operating state of the ion source 10, particularly the beam intensity.

[0033] 3. Conclusion The embodiments of the present disclosure have been specifically described above. The above-described embodiments, modifications, and experimental confirmations have been described to explain the invention, and the scope of the present disclosure should be determined based on the claims. Furthermore, modifications within the scope of the present disclosure, including other combinations of the embodiments, are also included in the scope of the claims of the present application. In other words, a person skilled in the art may make various modifications, combinations, subcombinations, and substitutions of the components of the above-described embodiments within the technical scope of the present disclosure or its equivalents. [Explanation of symbols]

[0034] 10. Ion Source 20 Machine Learning Models 210 Image Data 220 Neural Networks for Image Recognition 230 First fully connected layer 240 Second fully connected layer 250 Third fully connected layer 290 predicted values C Camera MC1, MC2 mirror coils

Claims

1. acquiring a plasma emission image of the ion source operating with plasma emission by an imaging device; inputting image data of the plasma emission image into a machine learning model including a neural network; inputting control numerical data, which is numerical data of controllable quantities for operating the ion source, and observation numerical data, which is numerical data of information observable during operation of the ion source, into the machine learning model; provided that the observation numerical data does not include the image data; Including, training the machine learning model with a training data set including the numerical data of the controllable quantity applied to the ion source, the beam intensity of an ion beam output by the ion source, and the image data of plasma emission from the ion source; a step of inputting a prediction data set including the numerical data of the controllable quantity to be applied to the ion source and the image data of plasma emission of the ion source into an input of the trained neural network of the machine learning model, and outputting a predicted value of the beam intensity of the ion beam output by the ion source from an output of the trained neural network; A method for predicting beam intensity of an ion source, comprising:

2. the neural network of the machine learning model comprises a convolutional neural network (CNN) and at least one fully connected layer; the control numerical data and the observation numerical data are input to the at least one fully connected layer; Image data of the plasma emission image is input to the convolutional neural network (CNN). The method for predicting beam intensity of an ion source according to claim 1 .

3. acquiring a plasma emission image of the ion source operating with plasma emission by an imaging device; inputting image data of the plasma emission image into a convolutional neural network (CNN); inputting the output of the CNN into at least one fully connected layer; inputting control numerical data, which is numerical data of controllable quantities for operating the ion source, and observation numerical data, which is numerical data of information observable during operation of the ion source, into the at least one fully connected layer; wherein the observation numerical data does not include the image data; Including, training the CNN and the at least one fully connected layer with a training dataset including the numerical data of the controllable quantity applied to the ion source, the beam intensity of an ion beam output by the ion source, and the image data of a plasma emission of the ion source; a prediction data set including the numerical data of the controllable quantity to be applied to the ion source and the image data of plasma emission of the ion source as inputs to the trained CNN and the trained at least one fully connected layer, and outputting a predicted value of the beam intensity of the ion beam output by the ion source as an output of the trained at least one fully connected layer; A method for predicting beam intensity of an ion source, comprising:

4. the at least one fully connected layer A first fully connected layer to which the output of the CNN is input; a second fully connected layer to which the output of the first fully connected layer is input; It is equipped with The control numerical data and the observation numerical data are input to the second fully connected layer.

4. The method for predicting beam intensity of an ion source according to claim 2 or 3.

5. an imaging device for acquiring a plasma emission image of the ion source operating with plasma emission; a computer device including at least a computing device and a recording device; 1. An apparatus for predicting beam intensity of an ion source, comprising: the computer device Convolutional Neural Networks (CNNs) At least one fully connected layer to which the output of the CNN is input; We have implemented machine learning models including The machine learning model is inputting image data of the plasma emission image into the CNN; inputting, into the at least one fully coupled layer, control numerical data, which is numerical data of a controllable amount for operating the ion source, and observation numerical data, which is numerical data of information observable during operation of the ion source; wherein the observation numerical data does not include the image data; an output of a beam intensity of an ion beam output by the ion source from the at least one fully coupled layer; Equipped with The machine learning model: the CNN and the at least one fully connected layer are trained with a training dataset including the numerical data of the controllable quantity to be applied to the ion source, the beam intensity of an ion beam output by the ion source, and the image data of plasma emission from the ion source; The trained CNN and the trained at least one fully connected layer output a predicted value of the beam intensity of the ion beam output by the ion source in response to input to the CNN and the at least one fully connected layer of a prediction data set including the numerical data of the controllable quantity to be applied to the ion source and the image data of plasma emission of the ion source. A device for predicting beam intensity in an ion source.

6. the at least one fully connected layer A first fully connected layer to which the output of the CNN is input; a second fully connected layer to which the output of the first fully connected layer is input; It is equipped with The control numerical data and the observation numerical data are input to the second fully connected layer. The apparatus for predicting beam intensity of an ion source according to claim 5.

Images