JP7756553B2 - Data prediction device and quartz glass crucible manufacturing system - Google Patents

Description

The present invention relates to a data prediction device that predicts the weight of a quartz glass crucible and a quartz glass crucible manufacturing system.

The Czochralski method (CZ method) is widely used for growing silicon single crystals. In this method, a seed crystal is brought into contact with the surface of molten silicon formed in a quartz glass crucible (hereinafter sometimes simply referred to as the "crucible"), and while the crucible is rotated, the seed crystal is pulled upward while rotating in the opposite direction, forming a single crystal at the bottom of the seed crystal.

The quartz glass crucible used to contain this silicon melt typically has a two-layer structure, with a transparent layer made of high-purity synthetic silica glass on the inner surface and an opaque layer made of natural silica glass with excellent thermal properties on the outer surface.

One example of a method for manufacturing such quartz glass crucibles is the rotating mold method. In the rotating mold method, natural silica glass raw powder (natural silica powder) is first layered on the surface of a rotating crucible mold, and then synthetic silica glass raw powder (synthetic silica powder) is layered on the surface of the natural silica powder layer to form a raw powder layered body (molding process). Next, this raw powder layered body is heated and melted from the inside to the outside using arc discharge, and then cooled to form a synthetic silica glass layer (transparent layer) on the inner surface and a natural silica glass layer (opaque layer) on the outer surface, resulting in a two-layered crucible molded body (melting process). Finally, the top end of the crucible molded body is cut off to obtain a two-layered quartz glass crucible (cutting process).

Furthermore, the quartz glass crucible obtained using the above manufacturing method requires high weight accuracy. Therefore, after the above cutting process, the weight of the actually manufactured quartz glass crucible is measured (measurement process). The melting conditions are then adjusted based on the actual weight value obtained from this measurement. Note that adjustment of the melting conditions is usually made based on the experience of the manufacturing worker.

Meanwhile, studies have been conducted on a method in which a manufacturing condition setting support device (computer) determines manufacturing conditions (including melting conditions) based on measurement data of actually manufactured quartz glass crucibles, without relying on the experience of the manufacturing worker (see Patent Document 1).

For example, Patent Document 1 describes a method of calculating the degree of agreement between data obtained by a simulation based on predetermined physical property parameters and manufacturing conditions and measurement data of a crucible actually manufactured based on those manufacturing conditions, and then repeatedly running the simulation while changing the physical property parameters and manufacturing conditions until the degree of agreement reaches a predetermined level, and adopting the manufacturing conditions when the degree of agreement reaches or exceeds the predetermined level as improved manufacturing conditions.

JP 2017-128509 A

However, in conventional methods for manufacturing quartz glass crucibles, a process for measuring the weight of the crucible (a process for obtaining measurement data) is carried out after the crucible is manufactured, so adjustments to the melting conditions based on the measurement data are made after the measurement data for the completed crucible has been fed back to the manufacturing workers.

In other words, when using conventional methods for manufacturing quartz glass crucibles, adjustments to the melting conditions can only be made after the crucible has actually been manufactured and the measurement data has been fed back, whether the adjustments are performed by a person or a computer. As a result, for example, the time (cycle time) from when the melting conditions are set in a particular manufacturing process until the next adjusted melting conditions are set can significantly exceed the time (lead time) required to complete the crucible in that manufacturing process. This lengthy cycle time needs to be improved from the perspective of improving the efficiency of crucible manufacturing.

The present invention was made in consideration of the above-mentioned problems, and aims to provide a data prediction device and a quartz glass crucible manufacturing system that can significantly shorten the cycle time required to adjust melting conditions in the manufacture of quartz glass crucibles.

The data prediction device of the present invention is characterized by comprising: storage means (corresponding to storage unit 12 described below) that stores a dataset created by linking the input data to the training data, where the input data is the measured value of the dimension at a predetermined position of a quartz glass crucible actually manufactured under melting conditions set in a crucible manufacturing apparatus, and the measured value of the weight of a quartz glass crucible actually manufactured under the melting conditions; model generation means (corresponding to control unit 11 described below) that generates a data prediction model by machine learning using the dataset read from the storage means; and data prediction means (corresponding to control unit 11 described below) that accepts as input data the measured value of the dimension at a predetermined position output from the crucible manufacturing apparatus during a measurement process in the quartz glass crucible manufacturing process, and uses the trained data prediction model to output a predicted weight of the quartz glass crucible manufactured in the manufacturing process.

The data prediction device of the present invention predicts the weight of a quartz glass crucible that is to be completed using a pre-trained data prediction model, rather than measuring the weight using a completed quartz glass crucible. This significantly reduces the time it takes for the weight to be fed back to the manufacturing worker compared to measuring the weight using a completed quartz glass crucible, allowing the manufacturing worker to begin adjusting the melting conditions earlier. Therefore, for example, the cycle time from when the melting conditions are set in a specific manufacturing process until the adjusted melting conditions are then set can be significantly reduced.

Furthermore, it is desirable that the data prediction device of the present invention be equipped with a display means for displaying the predicted weight output by the data prediction means.

Furthermore, it is desirable that the data prediction device of the present invention use as the actual measured values of the dimensions at the specified positions all or some of the outer diameter of the body portion of the quartz glass crucible, the thickness of the bottom portion, the thickness of the corner portions, the thickness of the body portion, the thickness of the transparent layer at the bottom portion, the thickness of the transparent layer at the corner portions, and the thickness of the transparent layer at the body portion.

The quartz glass crucible manufacturing system of the present invention comprises the data prediction device and a crucible manufacturing apparatus that manufactures quartz glass crucibles through a manufacturing process including a molding process, a melting process, and a measurement process, and is characterized in that the crucible manufacturing apparatus carries out the melting process under melting conditions that are adjusted based on the trend in the predicted weight of the quartz glass crucible output from the data prediction device, which operates as a machine-learned data prediction model.

Furthermore, in the silica glass crucible manufacturing system according to the present invention, it is desirable that the melting conditions be data on the melting process, including the current value, voltage value, integrated power value, electrode opening, electrode position, melting temperature, pressure, cooling water temperature, cooling water flow rate, and melting time during melting.

The data prediction device and quartz glass crucible manufacturing system of the present invention have the effect of significantly shortening the cycle time required to adjust melting conditions in the manufacture of quartz glass crucibles.

FIG. 1 is a diagram showing an example of the system configuration of a silica glass crucible manufacturing system according to the present invention. FIG. 2 is a diagram showing an example of the configuration of a crucible manufacturing apparatus. FIG. 3 is a flowchart showing an example of a manufacturing process for a quartz glass crucible. FIG. 4 is a cross-sectional view of the crucible molded body formed in the melting step. FIG. 5 is a cross-sectional view of the vitreous silica crucible obtained by the cutting process. FIG. 6 is a diagram showing an example of the hardware configuration of a computer that operates as a data prediction device according to the present invention. FIG. 7 is a diagram illustrating an example of a data set for generating a data prediction model. FIG. 8 is a diagram illustrating an example of the configuration of a neural network. FIG. 9 is a diagram illustrating an example of the configuration of input data. FIG. 10 is a diagram showing the results of learning evaluation. FIG. 11 is a diagram showing prediction results using a trained data prediction model.

Embodiments of a data prediction device and a quartz glass crucible manufacturing system according to the present invention will be described in detail below with reference to the drawings. Note that the present invention is not limited to these embodiments. Furthermore, in the specification and drawings of this application, elements that can be similarly described may be designated by the same reference numerals, and duplicate explanations may be omitted.

<System Configuration>
1 is a diagram showing an example of the system configuration of a vitreous silica crucible manufacturing system according to the present invention. The vitreous silica crucible manufacturing system of this embodiment comprises a data prediction device 1 that operates as a host computer for performing a process of predicting the weight of a vitreous silica crucible to be actually manufactured (hereinafter referred to as a data prediction process) and a process of generating a learning model for executing the weight prediction (hereinafter referred to as a learning model generation process), and a crucible manufacturing device 3 that manufactures vitreous silica crucibles in accordance with melting conditions adjusted and set by a manufacturing operator 2 based on the weight prediction results.

<Crucible manufacturing equipment>
As an example, the crucible manufacturing apparatus 3 of this embodiment uses a rotating mold method to manufacture a two-layer quartz glass crucible having a transparent layer formed of high-purity synthetic silica glass on the inner surface and an opaque layer formed of natural silica glass with excellent thermal properties on the outer surface.

Figure 2 is a cross-sectional view showing an example of the configuration of a crucible manufacturing apparatus 3 according to this embodiment. This crucible manufacturing apparatus 3 includes a crucible molding mold 34, which is composed of an inner member 31 formed from a metal mold having multiple through-holes (not shown) drilled therein, and a holder 33 that holds the inner member 31 and has a ventilation portion 32 formed on its outer periphery.

A rotating shaft 35 connected to a rotating means (not shown) is fixed to the bottom of the holder 33, rotatably supporting the crucible mold 34. The ventilation section 32 is connected to an exhaust port 36 that penetrates from approximately the center of the bottom of the holder 33 in the axial direction of the rotating shaft 35, and this exhaust port 36 is further connected to a pressure reduction mechanism 37.

In addition, an arc electrode 38 for arc discharge, as well as a raw material supply nozzle and a nozzle for spraying gas (nitrogen gas, oxygen gas, etc.) onto a specified portion of the crucible (not shown) are provided on the upper part facing the inner member 31.

Here, we will explain the method for manufacturing a quartz glass crucible using the crucible manufacturing apparatus 3. Figure 3 is a flowchart showing an example of the manufacturing process for a quartz glass crucible.

To manufacture a vitreous silica crucible using the crucible manufacturing apparatus 3, first, a rotary drive source (not shown) is used to rotate the rotary shaft 35 in the direction of the arrow (see FIG. 2), thereby rotating the crucible mold 34 at a predetermined speed. The pressure in the ventilating section 32 is then reduced by driving the pressure reducing mechanism 37, and vitreous silica raw material powder (natural silica powder, synthetic silica powder) is supplied into the inner member 31 from the raw material supply nozzle while suction is applied to the inner surface of the inner member 31 through the numerous through-holes formed in the inner member 31. Specifically, coarse-grained natural silica powder is first supplied, and a layer of natural silica powder is formed on the surface of the inner member 31 by suction and centrifugal pressure. Fine-grained synthetic silica powder is then supplied, and another layer of synthetic silica powder is formed on the surface of the natural silica powder by suction and centrifugal pressure, forming a two-layered raw material powder laminate (Step S1: Molding Process).

After the raw material powder stack is formed, while continuing to reduce the pressure using the pressure reduction mechanism 37, electricity is passed through the arc electrode 38 to heat and melt the raw material powder stack from the inside, first vitrifying the surface layer (forming a synthetic silica glass layer). The pressure reduction using the pressure reduction mechanism 37 and the heating and melting using the arc electrode 38 continue, and the raw material powder stack is vitrified all the way to the outer surface (forming a natural silica glass layer), forming a two-layered crucible molded body (Step S2: Melting Process).

Figure 4 is a cross-sectional view of the crucible molded body formed in the melting process. The crucible molded body 40 shown in Figure 4 is composed of a transparent layer 40a that forms the synthetic silica glass layer and an opaque layer 40b that forms the natural silica glass layer.

After the crucible molded body 40 is formed, an appearance inspection and dimensional measurements are performed using this crucible molded body 40 (Step S3: Measurement process). Dimensional measurements include, for example, measuring the outer diameter, wall thickness, and transparent layer thickness at predetermined positions on the crucible molded body 40.

Finally, the upper end of the crucible molded body 40 is cut at a predetermined height to obtain a two-layered quartz glass crucible (Step S4: Cutting process). Specifically, a two-layered quartz glass crucible is obtained by cutting the crucible molded body 40 shown in Figure 4 along the dotted lines. Generally, the crucible molded body 40 produced by the rotational mold method is manufactured to be taller than the required dimensions of the finished quartz glass crucible, and the required dimensions are met by cutting the upper opening of the crucible molded body 40.

Figure 5 is a cross-sectional view of the quartz glass crucible obtained through the cutting process. The quartz glass crucible 41 shown in Figure 5 is composed of a transparent layer 41a that forms the synthetic silica glass layer and an opaque layer 41b that forms the natural silica glass layer.

While this embodiment has been described with reference to an example in which a two-layer quartz glass crucible is manufactured, the present invention is not limited to this. The data prediction device and quartz glass crucible manufacturing system of the present invention can also be applied to the manufacture of, for example, a three-layer or more quartz glass crucible. That is, they can also be applied to the manufacture of a three-layer or more quartz glass crucible formed by an inner layer, an intermediate layer, and an outer layer, or even a four-layer or more quartz glass crucible formed by multiple intermediate layers. For example, when manufacturing a three-layer or more quartz glass crucible, a three-layer or more quartz glass crucible can be obtained by forming a raw material powder stack and a crucible molded body having a three-layer or more structure composed of an inner layer, one or more intermediate layers, and an outer layer in the molding and melting processes described above.

<Configuration of Data Prediction Device>
FIG. 6 is a diagram illustrating an example of the hardware configuration of a computer operating as a data prediction device according to the present invention. In FIG. 6, the data prediction device 1 includes a control unit 11 including a central processing unit (CPU) and a field programmable gate array (FPGA), a storage unit 12 including various memories such as read-only memory (ROM) and random access memory (RAM), an input unit 13 including a user interface such as a keyboard and a mouse, an interface unit 14 performing input/output processing such as printing and scanning, a display unit 15 serving as a display, and a communication unit 16 for communicating with the outside via a predetermined network. While FIG. 6 illustrates the data prediction device 1 including the input unit 13 including a user interface such as a keyboard and a mouse, the present invention is not limited to this. The display unit 15 may have touch panel functionality, eliminating the input unit 13 or using the input unit 13 in combination.

In FIG. 6, the control unit 11 executes, for example, a data prediction program for predicting the weight of a quartz glass crucible and a learning model generation program for generating a learning model for weight prediction, in order to realize the data prediction process and learning model generation process by the data prediction device 1 of this embodiment. The memory unit 12 stores programs (data prediction program, learning model generation program) related to the data prediction process and learning model generation process of this embodiment, various information (datasets for learning, etc.), and various data obtained during the process (dimensions of specified positions, predicted weight, etc.). The control unit 11 executes the data prediction process and learning model generation process of this embodiment by reading out the various programs stored in the memory unit 12.

Note that the storage unit 12 is not limited to internal memory, but may be an external storage medium such as a DVD (Digital Versatile Disc) or SD memory, or may be composed of both internal memory and external storage medium (such as a DVD or SD memory). Furthermore, for the sake of convenience, the hardware configuration of the data prediction device 1 of this embodiment lists the configuration related to the data prediction processing and learning model generation processing of this embodiment, and does not represent all of the functions of the computer that makes up the data prediction device 1.

Furthermore, the data prediction device 1 of this embodiment is intended to be a general-purpose PC such as a desktop personal computer or laptop computer, but is not limited to these and may also be a mobile terminal such as a smartphone or tablet terminal.

<Learning model generation process>
Next, before describing the data prediction process in the data prediction device 1 of this embodiment, the learning model generation process, which is a prerequisite for the data prediction process, will be described in detail. Note that the input data used in the learning model generation process and the data prediction process is assumed to be standardized in advance by the control unit 11. That is, as a preprocessing step for machine learning and weight prediction, the control unit 11 performs a standardization process on each of the actual measurement values of dimensions at predetermined positions that make up the input data, setting the average value to 0 and the standard deviation to 1.

In the data prediction device 1 of this embodiment, the control unit 11 operates as a learning model that predicts the weight of a quartz glass crucible, i.e., a data prediction model. For example, it accepts as input data the actual measured values of dimensions at a predetermined position output from the crucible manufacturing device 3 in the measurement step (step S3) within the manufacturing process shown in Figure 3, and outputs the weight (predicted weight) of the quartz glass crucible. This data prediction model is generated using a known machine learning algorithm, such as a neural network.

The dataset used in the machine learning algorithm is created by, for example, using the measured values of the dimensions at a specified position of a quartz glass crucible actually manufactured under the melting conditions (such as the current value and melting time during melting, as described below) set in the crucible manufacturing device 3 as input data, using the measured value of the weight of a quartz glass crucible actually manufactured under the aforementioned melting conditions as the output label (correct answer label), and linking this output label to this input data.

As described above, the input data for the dataset are actual measurement values of dimensions corresponding to specific positions on an actually manufactured quartz glass crucible. In this embodiment, as an example, nine dimensions (mm) are used: the outer diameter of the crucible at position G, the outer diameter at position F, the outer diameter at position E, the wall thickness at position A, the wall thickness at position D, the wall thickness at position F, the transparent layer thickness at position A, the transparent layer thickness at position D, and the transparent layer thickness at position F. Note that each position (A, D, E, F, G) refers to, for example, the bottom (A), corner (D), and straight body (E, F, G), as shown in the cross-sectional view of the quartz glass crucible 41 in Figure 5.

In addition, in this embodiment, as described above, the input data is the nine dimensions, but this is not limited to this. The input data may be, for example, a portion of the nine dimensions, or may include other dimensions in addition to the nine dimensions, or may be composed of dimensions other than the nine dimensions. Furthermore, the input data may include parameters other than dimensions, such as melting data (current value during melting, melting time, etc.) and environmental conditions such as climatic conditions. When parameters other than dimensions are included in the input data, if the parameters include non-numeric qualitative data (for example, equipment such as electrodes or molds), the qualitative data is converted into quantitative data (numeric values) and then input. The method of quantification is not particularly limited, but may, for example, be to convert qualitative data into quantitative data using dummy variables (one-hot encoding) or the like.

Figure 7 shows an example dataset for generating a data prediction model. In Figure 7, the dataset for performing machine learning on a neural network is composed of a combination of the above-mentioned input data (actually measured values of the outer diameter at position G, outer diameter at position F, outer diameter at position E, wall thickness at position A, wall thickness at position D, wall thickness at position F, transparent layer thickness at position A, transparent layer thickness at position D, transparent layer thickness at position F, etc. of an actually manufactured crucible) and the above-mentioned output label (actually measured value of the weight of an actually manufactured crucible).

In this embodiment, for example, a dataset for training the data prediction model (the dataset shown in FIG. 7) is stored in advance in the storage unit 12. Approximately 80% of the dataset stored in the storage unit 12 is used as training data, and approximately 20% is used as evaluation data. Here, the training data is a dataset used to train the data prediction model, and the evaluation data is a dataset that is not used to train the model but is used to verify whether the trained data prediction model is versatile. The data prediction model was verified using the holdout method. It is desirable to prepare a number of training data sets sufficient to achieve the desired accuracy in the data prediction model. To evaluate the performance of the data prediction model in this embodiment, the mean squared error (MES), one of the evaluation indices, is used. That is, the difference between the actual measured value and the predicted value is squared, and the sum is divided by the number of data points to be evaluated. Therefore, for example, the smaller this value, the lower the model's error. Note that the evaluation indices are not limited to this.

Then, with the prepared data set (training data, evaluation data) as described above stored in the memory unit 12, the control unit 11 reads the training data from the memory unit 12 and generates a data prediction model optimized for predicting the weight of the crucible by performing supervised learning using the training data, for example, using a neural network, which is a machine learning algorithm.

<Example of a machine learning algorithm>
FIG. 8 is a diagram showing an example of the configuration of a neural network for performing machine learning. The machine learning algorithm used in this embodiment is a fully connected neural network, which is composed of, for example, an input layer, an intermediate layer having one or more layers, and an output layer. Each unit in the input layer of this neural network corresponds one-to-one to the dimensions ( _x1 , _x2 , _x3 , ...) of a predetermined position that constitutes the input data in the above-mentioned dataset. In other words, the number of each dimension that constitutes the input data is the number of input units. Meanwhile, the number of units in the output layer corresponds to the weight of the crucible (output label: z), and in this embodiment, it is one.

In addition, in this embodiment, the activation function in the intermediate layer is a ReLU (Rectified Linear Unit) function, and the activation function in the output layer is a linear function. There are no regulations regarding the number of layers and units in the intermediate layer, and they can be set arbitrarily. Therefore, the number of layers and units in the intermediate layer is determined appropriately depending on the amount of data, type of data, required output accuracy, and other factors. For example, as the number of layers and units in the intermediate layer increases, the performance of the neural network, such as analytical flexibility and output accuracy, improves; however, the amount of data, memory usage, and calculation volume also increase.

Specifically, the control unit 11 first reads one training data from the memory unit 12 and inputs the input data ( _x1 , _x2 , _x3 , ..., _xn ) constituting this training data to each unit in the input layer. Then, the control unit 11 multiplies the input data ( _x1 , _x2 , _x3 , ..., _xn ) by weights ( _w11 to _w1m , _w21 to _w2m , _w31 to _w3m , ..., _wn1 to _wnm ) and outputs the multiplication results to each unit in the next layer (hidden layer). As a result, the input values ( _y1 , _y2 , _y3 , ..., _ym ) of each unit in the hidden layer are as follows: where n is the number of units in the input layer, m is the number of units in the hidden layer, and _b1 , _b2 , _b3 , ..., _bm are biases. Furthermore, the weights (w ₁₁ to w _1m , w ₂₁ to w _2m , w ₃₁ to w _3m , . . . , w _n1 to w _nm ) and biases (b ₁ , b ₂ , b ₃ , . . . b _m ) are values that are updated in the learning process.
y ₁ =x ₁ w ₁₁ +x ₂ w ₂₁ +x ₃ w ₃₁ +...+x _n w _n1 +b ₁
y ₂ =x ₁ w ₁₂ +x ₂ w ₂₂ +x ₃ w ₃₂ +...+x _n w _n2 +b ₂
y ₃ =x ₁ w ₁₃ +x ₂ w ₂₃ +x ₃ w ₃₃ +…+x _n w _n3 +b ₃
…
y _m =x ₁ w _1m +x ₂ w _2m +x ₃ w _3m +...+x _n w _nm +b _m

Next, each unit in the intermediate layer calculates the ReLU function given by the following equation for each of the multiple input values (y ₁ , y ₂ , y ₃ , ..., y _m ) from each unit in the previous layer, and outputs the results to each unit in the next layer, where M ranges from 1 to m.

In the output layer, for example, a linear function is applied, the input values received from each unit in the middle layer are summed, and the sum is output as the predicted value of the crucible weight (predicted weight).

The control unit 11 then calculates the mean squared error between the predicted value, which is the output of the output layer, and the weight of the training data, and trains the data prediction model (updating the weights and biases) by repeatedly learning over a predefined number of epochs so as to minimize this error.

In this embodiment, a learning model is generated using the dataset shown in FIG. 7, but the input data for the dataset is not limited to this. For example, some dimensions may be selected from the dimensions that make up the input data shown in FIG. 7, and supervised learning may be performed using a dataset in which the selected dimensions (input data) are linked to the measured values (output labels) of the weights of crucibles that have actually been manufactured. Alternatively, supervised learning may be performed using a dataset in which input data in which other parameters have been newly added to the above input data is linked to the measured values (output labels) of the weights of crucibles that have actually been manufactured.

In addition, in this embodiment, a neural network is used as an example of a machine learning algorithm to generate the above-mentioned data prediction model, but the machine learning algorithm for generating the data prediction model is not limited to this. For example, an algorithm that combines a decision tree and gradient boosting, such as LightGBM, can be used.

<Data prediction processing>
Next, the data prediction process in the data prediction device 1 of this embodiment will be described in detail.

In this embodiment, the control unit 11, which operates as a trained data prediction model, accepts as input data, for example, the actual measurement values of dimensions at predetermined positions (outer diameter at position G, outer diameter at position F, outer diameter at position E, wall thickness at position A, wall thickness at position D, wall thickness at position F, transparent layer thickness at position A, transparent layer thickness at position D, transparent layer thickness at position F) output by the crucible manufacturing apparatus 3 in the measurement step (step S3) within the manufacturing process shown in FIG. 3 , and outputs the weight (predicted weight) of the quartz glass crucible being manufactured. In other words, the control unit 11 instantly predicts the weight of the quartz glass crucible to be completed by inputting the actual measurement values of dimensions at predetermined positions obtained in the measurement step (step S3), without waiting for the quartz glass crucible to be completed.

Specifically, first, the operator of the data prediction device 1 (manufacturing worker 2) uses the user interface (input unit 13 including a keyboard and mouse) to input the actual measurement values of the dimensions at the predetermined positions obtained in the measurement process (step S3) as input data to the data prediction device 1 (step S11 in FIG. 3). At this time, the input data input is the actual measurement values of the dimensions at the same positions as when the data prediction model was generated. FIG. 9 shows an example of the configuration of the input data. Note that when parameters other than dimensions are included in the input data, for example, if the parameters include non-numerical qualitative data, the qualitative data is converted to quantitative data in the same manner as when the data prediction model was generated. The control unit 11, operating as a trained data prediction model, then receives the input data shown in FIG. 9 and outputs the weight of the quartz glass crucible to be completed as the predicted weight (step S12 in FIG. 3). The weight predicted here (predicted weight) is then displayed on the display unit 15 of the data prediction device 1.

<Reflection in manufacturing processes>
The manufacturing worker 2 can know the tendency (large or small dimensions) of the quartz glass crucible to be manufactured when the melting process is carried out under the current melting conditions by checking the predicted weight displayed on the display unit 15 of the data prediction device 1. Therefore, based on the predicted weight displayed on the display unit 15, the manufacturing worker 2 adjusts the current melting conditions, for example, so that the weight value becomes smaller if the trend of the weight value is larger than the assumed value, or so that the weight value becomes larger if the trend of the weight value is smaller than the assumed value (step S13 in FIG. 3), and reflects this in the crucible melting process (step S14 in FIG. 3).

In adjusting the melting conditions, for example, data on the melting process (melting data) is adjusted. The melting data to be adjusted includes, for example, the current value (current value of the arc electrode 38) during melting, the voltage value, the integrated power value, the electrode opening, the electrode position, the melting temperature, the pressure, the cooling water temperature, the cooling water flow rate, and the melting time. These numerical data during melting include the set value, average value, maximum value, minimum value, integrated value, etc.

<Effects>
The data prediction device 1 of this embodiment is equipped with a memory unit 12 that stores a dataset created by linking the input data to the teacher data, using the actual measured values of the dimensions at a predetermined position of a quartz glass crucible actually manufactured under the melting conditions set in the crucible manufacturing apparatus 3 as input data and the actual measured values of the weight of the quartz glass crucible actually manufactured under the melting conditions as input data, a model generation means (control unit 11) that generates a data prediction model by machine learning using the dataset read from the memory unit 12, and a data prediction means (control unit 11) that accepts the actual measured values of the dimensions at a predetermined position output from the crucible manufacturing apparatus 3 during the measurement process in the quartz glass crucible manufacturing process as input data, and uses the trained data prediction model to output a predicted weight of the quartz glass crucible manufactured in the manufacturing process.

In this embodiment, rather than measuring the weight using a completed quartz glass crucible (without waiting for the quartz glass crucible to be completed), a pre-trained data prediction model is used to predict the weight of the quartz glass crucible that is to be completed. This significantly reduces the time it takes for the weight to be fed back to manufacturing worker 2 compared to measuring the weight using a completed quartz glass crucible, allowing manufacturing worker 2 to begin adjusting the melting conditions earlier. Therefore, for example, the cycle time from when the melting conditions are set in a particular manufacturing process until the adjusted melting conditions are then set can be significantly reduced.

Next, we will explain an example of a data prediction model optimized using the above learning model generation process.

In this example, for example, when generating a data prediction model to predict the weight of a crucible, 6,000 datasets were prepared for use in learning. These datasets were then randomly divided into training data and evaluation data. The breakdown was 4,800 training data and 1,200 evaluation data. Furthermore, mean squared error (MES), a loss function, was used to evaluate the performance of the model.

The optimization algorithm used was "Adam," with the activation function in the intermediate layer being a ReLU function and the activation function in the output layer being a linear function. Furthermore, the initial learning rate was set to "0.01," and the learning rate was automatically decreased as the learning progressed. The batch size was set to 32. Furthermore, the initial number of epochs was set to 1000, and in order to prevent overfitting, learning was terminated if the error had not improved compared to the previous epoch.

Then, we used the training data to perform machine learning on the neural network with the above specifications, and had the trained data prediction model predict weights using nine sets of test data that had been prepared in advance.

Figure 10 shows the results of the learning evaluation, showing, for example, the progress of the mean squared error (Mean Squared Error) against the degree of learning (Epochs). As shown in Figure 10, the mean squared error of the data prediction model in this example decreases as learning progresses, confirming that learning is progressing smoothly. Furthermore, the mean squared error becomes sufficiently small from about epoch number 4 onwards, and there is no change in the mean squared error thereafter. Therefore, in this example, it can be said that an epoch number of around 4 is suitable for training the data prediction model.

Figure 11 shows the prediction results of a trained data prediction model using test data. Here, the predicted values output from the data prediction model that predicts the crucible weight were plotted against the weight (actual measured value) of the training data, and the results were observed. Observation of the plotted data in Figure 11 revealed that there was little variation in the predicted values compared to the training data, with a maximum error of around 5%.

REFERENCE SIGNS LIST 1 Data prediction device 2 Manufacturing worker 3 Crucible manufacturing device 11 Control unit 12 Memory unit 13 Input unit 14 Output unit 15 Display unit 16 Communication unit 31 Inner member 32 Ventilation unit 33 Holder 34 Crucible molding mold 35 Rotating shaft 36 Exhaust port 37 Pressure reduction mechanism 38 Arc electrode 40 Crucible molding body 40a, 41a Transparent layer 40b, 41b Opaque layer 41 Quartz glass crucible

Claims (5)

A storage means for storing a data set created by using the measured values of the dimensions of a predetermined position of a quartz glass crucible actually manufactured under the melting conditions set in the crucible manufacturing device as input data and the measured values of the weight of the quartz glass crucible actually manufactured under the melting conditions as training data and linking the training data to the input data;
A model generation means for generating a data prediction model for predicting the weight of a quartz glass crucible by machine learning using the data set read from the storage means;
A data prediction means for receiving as input data the actual measurement values of the dimensions at a predetermined position output from the crucible manufacturing apparatus in a measurement step within the manufacturing process of the quartz glass crucible, and outputting a predicted weight of the quartz glass crucible manufactured in the manufacturing process using a trained data prediction model;
Equipped with
A data prediction device comprising: moreover,
a display means for displaying the predicted weight output by the data prediction means;
Equipped with
2. The data prediction device according to claim 1. The actual measured values of the dimensions at the predetermined positions are all or part of the outer diameter of the straight body part of the quartz glass crucible, the thickness of the bottom part, the thickness of the corner parts, the thickness of the straight body part, the thickness of the transparent layer at the bottom part, the thickness of the transparent layer at the corner parts, and the thickness of the transparent layer at the straight body part.
3. The data prediction device according to claim 1 or 2. A data prediction device according to claim 1, 2 or 3;
A crucible manufacturing apparatus for manufacturing a quartz glass crucible through a manufacturing process including a molding process, a melting process, and a measuring process;
Equipped with
The crucible manufacturing apparatus carries out a melting process under melting conditions adjusted based on the trend of the predicted weight of the quartz glass crucible output from the data prediction device operating as a machine-learned data prediction model.
A quartz glass crucible manufacturing system characterized by: The melting conditions are melting process data including a current value, a voltage value, an integrated power value, an electrode opening degree, an electrode position, a melting temperature, a pressure, a cooling water temperature, a cooling water flow rate, and a melting time during melting.
5. The quartz glass crucible manufacturing system according to claim 4.