JP2022093298A - Data prediction device and quartz glass crucible manufacturing system - Google Patents

Abstract

To provide a data prediction device capable of greatly shortening a cycle time required to adjust melting conditions in manufacturing a quartz glass crucible.SOLUTION: A data prediction device comprises: a storage part 12 which stores a data set generated by linking a measured value (input data) of a dimension at a predetermined position and a measured value (teacher data) of weight of a quartz glass crucible actually manufactured under melting conditions set to a crucible manufacturing device to each other; model generation means (control part 11) which generates a data prediction model for predicting the weight of the quartz glass crucible through machine learning using the data set; and data prediction means (control part 11) which receives as input data the measured value of the dimension at the predetermined position output from the crucible manufacturing device in a manufacturing process, and outputs a predicted weight of the quartz glass crucible manufactured in the manufacturing proces using the learnt data prediction model.SELECTED DRAWING: Figure 6

Description

The present invention relates to a data prediction device for predicting 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 a silicon melt formed in a quartz glass crucible (hereinafter, may be simply referred to as a "crucible"), the crucible is rotated, and the seed crystal is rotated in the opposite direction. A single crystal is formed at the lower end of the seed crystal by pulling it upward while rotating it.

The quartz glass rutsubo for accommodating this silicon melt has a transparent layer formed of high-purity synthetic silica glass on the inner surface side, and an opaque layer formed of natural silica glass having excellent heat on the outer surface side. It is common to use a quartz glass rut with a two-layer structure.

As an example of such a method for manufacturing a quartz glass crucible, a rotary molding method is known. In the rotary molding method, first, a raw material powder of natural silica glass (natural silica powder) is laminated on the surface of a rotating crucible molding mold, and then a raw material powder of synthetic silica glass (synthetic silica powder) is laminated on the surface of a layer of natural silica powder. ) Is laminated to form a raw material powder laminate (molding step). Next, the raw material powder laminate is heated and melted from the inside to the outside by arc discharge, and then cooled to form a synthetic silica glass layer (transparent layer) on the inner surface side and a natural silica glass layer (natural silica glass layer) on the outer front side. An opaque layer) is formed, and a crucible molded body having a two-layer structure is obtained (melting step). Finally, by cutting the upper end of the crucible molded body, a quartz glass crucible having a two-layer structure can be obtained (cutting step).

Further, the quartz glass crucible obtained by the above manufacturing method is required to have high accuracy in terms of weight. Therefore, after the cutting step, the weight is measured using the actually manufactured quartz glass crucible (measurement step). Then, the melting conditions are adjusted based on the actually measured value of the weight obtained from this measurement result. The melting conditions are usually adjusted based on the experience of the manufacturing operator.

On the other hand, regardless of the experience of the manufacturing operator, the manufacturing condition setting support device (computer) determines the manufacturing condition (including the melting condition) based on the measurement data of the actually manufactured quartz glass crucible. (See Patent Document 1).

For example, in Patent Document 1, the degree of coincidence between the data obtained by the simulation based on the predetermined physical property parameters and the manufacturing conditions and the measurement data of the rutsubo actually manufactured based on the manufacturing conditions is calculated and further matched. It is stated that the simulation is repeatedly executed while changing the physical characteristics parameters and manufacturing conditions until the degree reaches the specified level, and the manufacturing conditions when the degree of agreement exceeds the specified level is adopted as the improved manufacturing conditions. Has been done.

Japanese Unexamined Patent Publication No. 2017-128509

However, in the conventional method for manufacturing a quartz glass rutsubo, a step of measuring the weight (step of obtaining measurement data) is carried out after the rutsubo is manufactured, so that the adjustment of the melting conditions based on the measurement data is performed on the completed rutsubo. This is the work after the measurement data is fed back to the manufacturing operator.

That is, when the conventional method for manufacturing a quartz glass crucible is adopted, the melting conditions are adjusted by actually manufacturing the crucible and feeding back the measurement data regardless of whether the person performing the adjustment work is a person or a computer. It can only be done later. Therefore, for example, the time (cycle time) from the setting of the melting conditions in a specific manufacturing process to the setting of the next adjusted melting conditions is the time required for the rutsubo to be completed in the above manufacturing process. It will be a time that greatly exceeds (lead time). Such redundant cycle times need to be improved from the viewpoint of improving the efficiency of crucible manufacturing.

The present invention has been made in view of the above problems, and is a data prediction device and a quartz glass crucible manufacturing system capable of significantly shortening the cycle time required for adjusting the melting conditions in the manufacture of quartz glass crucibles. The purpose is to provide.

The data prediction device according to the present invention uses as input data the measured values of the dimensions of the quartz glass rutsubo actually manufactured under the melting conditions set in the rutsubo manufacturing device, and the quartz actually manufactured under the melting conditions. A storage means (corresponding to a storage unit 12 described later) for storing a data set created by associating the teacher data with the input data, and reading from the storage means, using the measured value of the weight of the glass rut as the teacher data. A model generation means (corresponding to the control unit 11 described later) that generates a data prediction model for predicting the weight of a quartz glass rut using a data set by machine learning, and the rut in the measurement process in the manufacturing process of the quartz glass rut. Data prediction means that accepts the measured values of the dimensions at a predetermined position output from the manufacturing equipment as input data and outputs the predicted weight of the quartz glass rutsubo manufactured in the manufacturing process using the trained data prediction model (described later). (Corresponding to the control unit 11), and the like.

According to the data prediction device according to the present invention, the weight of the quartz glass crucible to be completed is predicted by using the data prediction model learned in advance, instead of measuring the weight using the completed quartz glass crucible. As a result, the time until the weight is fed back to the manufacturing operator can be significantly shortened as compared with the case where the weight is measured using the completed quartz glass crucible, and the manufacturing operation is accompanied by this. It is possible to start the adjustment of the melting condition by the person as soon as possible. Therefore, for example, the cycle time from the setting of the melting conditions in a specific manufacturing process to the next setting of the adjusted melting conditions can be significantly shortened.

Further, it is desirable that the data prediction device according to the present invention includes a display means for displaying the predicted weight output by the data prediction means.

Further, the data prediction device according to the present invention measures the measured values of the dimensions at the predetermined positions as the outer diameter of the straight body portion, the wall thickness of the bottom portion, the wall thickness of the corner portion, and the wall thickness of the straight body portion in the quartz glass rut. It is desirable to make all or part of the thickness of the transparent layer at the bottom, the thickness of the transparent layer at the corners, and the thickness of the transparent layer at the straight body.

The quartz glass rutsubo manufacturing system according to the present invention includes the data prediction apparatus and a rutsubo manufacturing apparatus for manufacturing quartz glass rutsubo by a manufacturing process including a molding step, a melting step and a measuring step, and the rutsubo manufacturing apparatus includes the rutsubo manufacturing apparatus. It is characterized in that the melting process is carried out under the melting conditions adjusted based on the tendency of the predicted weight of the quartz glass rubbish output from the data prediction device that operates as a machine-learned data prediction model.

Further, in the quartz glass crucible manufacturing system according to the present invention, the melting conditions are set to the current value at the time of melting, the voltage value, the integrated power value, the electrode opening degree, the electrode position, the melting temperature, the pressure, the cooling water temperature, and the cooling water. It is desirable to use the data of the melting process including the flow rate and melting time.

The data prediction device and the quartz glass crucible manufacturing system according to the present invention have an effect that the cycle time required for adjusting the melting conditions can be significantly shortened in the manufacturing of the quartz glass crucible.

FIG. 1 is a diagram showing an example of a system configuration of a quartz 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 of 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 a quartz glass crucible obtained by the cutting step. FIG. 6 is a diagram showing a hardware configuration example of a computer operating as a data prediction device according to the present invention. FIG. 7 is a diagram showing an example of a data set for generating a data prediction model. FIG. 8 is a diagram showing an example of the configuration of the neural network. FIG. 9 is a diagram showing an example of the configuration of input data. FIG. 10 is a diagram showing the evaluation results of learning. FIG. 11 is a diagram showing a prediction result by the trained data prediction model.

Hereinafter, embodiments of the data prediction device and the quartz glass crucible manufacturing system according to the present invention will be described in detail with reference to the drawings. The present invention is not limited to this embodiment. Further, in the specification and drawings of the present application, elements that can be similarly described may be designated by the same reference numerals to omit duplicate description.

<System configuration>
FIG. 1 is a diagram showing an example of a system configuration of a quartz glass crucible manufacturing system according to the present invention. The quartz glass crucible manufacturing system of the present embodiment is a process of predicting the weight of the quartz glass crucible actually manufactured (hereinafter referred to as a data prediction process) and a process of generating a learning model for executing the weight prediction (hereinafter, learning). A data prediction device 1 that operates as a host computer that performs model generation processing), and a crucible manufacturing device 3 that manufactures quartz glass crucibles according to melting conditions adjusted and set by the manufacturing operator 2 based on the weight prediction results. And.

<Crucible manufacturing equipment>
As an example, the crucible manufacturing apparatus 3 of the present embodiment has a transparent layer formed of high-purity synthetic silica glass on the inner surface side and is formed of natural silica glass having excellent heat on the outer front side by a rotary molding method. A two-layer structure quartz glass crucible having an opaque layer is manufactured.

FIG. 2 is a cross-sectional view showing an example of the configuration of the crucible manufacturing apparatus 3 of the present embodiment. The crucible manufacturing apparatus 3 is, for example, an inner member 31 composed of a mold having a plurality of through holes (not shown), and a holding body provided with a ventilation portion 32 on the outer periphery thereof to hold the inner member 31. A mold 34 for crucible molding, which is composed of 33 and 33, is provided.

Further, a rotating shaft 35 connected to a rotating means (not shown) is fixed to the lower portion of the holding body 33 to rotatably support the crucible molding mold 34. Further, the ventilation portion 32 is connected to an exhaust port 36 penetrating in the axial direction of the rotating shaft 35 from substantially the center of the bottom of the holding body 33, and the exhaust port 36 is further connected to the decompression mechanism 37.

Further, on the upper portion facing the inner member 31, a gas (nitrogen gas, oxygen gas, etc.) is sprayed on an arc electrode 38 for arc discharge, a raw material supply nozzle (not shown), and a predetermined portion of the rutsubo. Nozzles and the like are provided.

Here, a method for manufacturing a quartz glass crucible by the crucible manufacturing apparatus 3 will be described. FIG. 3 is a flowchart showing an example of a manufacturing process of a quartz glass crucible.

To manufacture a quartz glass crucible using the crucible manufacturing apparatus 3, first, a rotation drive source (not shown) is used to rotate the rotation shaft 35 in the direction of the arrow (see FIG. 2). The molding mold 34 is rotated at a predetermined speed. Then, the ventilation portion 32 is depressurized by driving the depressurizing mechanism 37, and while sucking the inner surface side of the inner member 31 through a large number of through holes formed in the inner member 31, quartz glass is introduced into the inner member 31 from the raw material supply nozzle. Supply raw material powder (natural silica powder, synthetic silica powder). 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 pressing with suction force and centrifugal force. After that, fine synthetic silica powder is supplied, and a layer of synthetic silica powder is further formed on the surface of the layer of natural silica powder by pressing with attractive force and centrifugal force to form a raw material powder laminate having a two-layer structure (step). S1: Molding process).

After forming the raw material powder laminate, the arc electrode 38 is energized to heat and melt the raw material powder laminate from the inside while continuing the depressurization by the depressurizing mechanism 37, and the surface layer is first vitrified (a synthetic silica glass layer is formed). Ru). After that, the depressurization by the decompression mechanism 37 and the heating and melting by the arc electrode 38 are continued, and the raw material powder laminate is vitrified to the outer front side (a natural silica glass layer is formed) to form a crucible molded body having a two-layer structure (a natural silica glass layer is formed). Step S2: melting step).

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

After forming the crucible molded body 40, visual inspection and dimensional measurement are performed using the crucible forming body 40 (step S3: measurement step). In the dimensional measurement, for example, the outer diameter, the wall thickness, the thickness of the transparent layer, and the like at a predetermined position in the crucible forming body 40 are measured.

Finally, by cutting the upper end portion of the crucible molded body 40 at a predetermined height, a quartz glass crucible having a two-layer structure can be obtained (step S4: cutting step). Specifically, a quartz glass crucible having a two-layer structure can be obtained by cutting the crucible molded body 40 shown in FIG. 4 at a dotted line portion. Generally, the crucible molded body 40 manufactured by the rotary molding method is manufactured so as to be higher than the required size of the finished quartz glass crucible, and the required size is obtained by cutting the upper end opening portion of the crucible molded body 40. Meet.

FIG. 5 is a cross-sectional view of a quartz glass crucible obtained by the cutting step. The quartz glass crucible 41 shown in FIG. 5 is composed of a transparent layer 41a forming a synthetic silica glass layer and an opaque layer 41b forming a natural silica glass layer.

In this embodiment, the case of manufacturing a quartz glass crucible having a two-layer structure has been described as an example, but the present invention is not limited to this. The data prediction device and the quartz glass crucible manufacturing system according to the present invention can be similarly applied to, for example, the case of manufacturing a quartz glass crucible having a three-layer structure or more. That is, it can be applied even in the case of manufacturing a quartz glass crucible having a three-layer structure formed by an inner layer, an intermediate layer, and an outer layer, and further, a quartz glass crucible having a four-layer structure or more in which a plurality of intermediate layers are formed. Is. For example, in the case of producing a quartz glass crucible having a three-layer structure or more, in the molding step and the melting step described above, a raw material powder laminate having a three-layer structure or more composed of an inner layer, a single or a plurality of intermediate layers, and an outer layer. By forming the body and the crucible molded body, a quartz glass crucible having a three-layer structure or more can be obtained.

<Configuration of data prediction device>
FIG. 6 is a diagram showing a hardware configuration example 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 composed of a CPU (Central Processing Unit) and an FPGA (Field Programmable Gate Array), and various types such as a ROM (Read Only Memory) and a RAM (Random Access Memory). A storage unit 12 including a memory, an input unit 13 including a user interface such as a keyboard and a mouse, an interface unit 14 that performs input / output processing such as printing and scanning, a display unit 15 that is a display, and a predetermined network. A communication unit 16 that communicates with the outside is provided. Although it was decided to include an input unit 13 including a user interface such as a keyboard and a mouse in FIG. 6, the data prediction device 1 of the present embodiment is not limited to this, and the display unit 15 is provided with a touch panel function. By having the input unit 13, the input unit 13 may not be provided or may be used in combination with the input unit 13.

In FIG. 6, in order to realize the data prediction process and the learning model generation process by the data prediction device 1 of the present embodiment, the control unit 11 performs, for example, a data prediction program for predicting the weight of a quartz glass rut and a weight prediction. Execute a learning model generation program for generating a learning model to be performed. The storage unit 12 is obtained in the process of processing the program (data prediction program, learning model generation program) and various information (data set for learning, etc.) related to the data prediction process and the learning model generation process of the present embodiment. Various data (dimensions at predetermined positions, predicted weight, etc.) are stored. The control unit 11 executes the data prediction process and the learning model generation process of the present embodiment by reading out various programs stored in the storage unit 12.

The storage unit 12 is not limited to the internal memory, and may be an external storage medium such as a DVD (Digital Versatile Disc) or an SD memory, and the internal memory and the external storage medium (DVD or SD). It may be composed of both (memory, etc.). Further, the hardware configuration of the data prediction device 1 of the present embodiment enumerates the configurations related to the data prediction processing and the learning model generation processing of the present embodiment for convenience of explanation, and is a computer constituting the data prediction device 1. It does not represent all the functions of.

Further, the data prediction device 1 of the present embodiment is assumed to be a general-purpose PC such as a desktop personal computer and a notebook personal computer, but is not limited to these, and may be, for example, a mobile terminal such as a smartphone or a tablet terminal. ..

<Learning model generation process>
Next, before explaining the data prediction process in the data prediction device 1 of the present embodiment, the learning model generation process which is a premise thereof will be described in detail. The input data used in the learning model generation process and the data prediction process are pre-standardized by the control unit 11. That is, as a preprocessing for performing machine learning and weight prediction, the control unit 11 individually performs standardization processing in which the average value is 0 and the standard deviation is 1 for the actually measured values of the dimensions of the predetermined positions constituting the input data. ..

In the data prediction device 1 of the present embodiment, the control unit 11 operates as a learning model for predicting the weight of the quartz glass rut, that is, a data prediction model, and for example, a measurement step (step S3) in the manufacturing process shown in FIG. ), The measured value of the dimension of the predetermined position output from the rutsubo manufacturing apparatus 3 is accepted as input data, and the weight (predicted weight) of the quartz glass rutsubo is output. Then, this data prediction model is generated by using a known machine learning algorithm such as a neural network.

Further, the data set used in the machine learning algorithm is, for example, the predetermined position of the quartz glass crucible actually manufactured under the melting conditions (current value at the time of melting, melting time, etc., which will be described later) set in the crucible manufacturing apparatus 3. The measured value of the dimensions is used as the input data, the measured value of the weight of the quartz glass crucible actually manufactured under the above-mentioned melting conditions is used as the output label (correct answer label), and the output label is associated with this input data.

As described above, the input data of the data set is an actually measured value of the dimension corresponding to the predetermined position of the actually manufactured quartz glass rutsubo, but in this embodiment, as an example, the outer diameter and the position of the rutsubo position G Outer diameter of F, outer diameter of position E, wall thickness of position A, wall thickness of position D, wall thickness of position F, thickness of transparent layer at position A, thickness of transparent layer at position D, thickness of position F The thickness of the transparent layer is 9 dimensions (mm) in total. The positions (A, D, E, F, G) are, for example, the bottom portion (A), the corner portion (D), and the straight body portion (as described in the cross-sectional view of the quartz glass crucible 41 shown in FIG. It shall refer to each part of E, F, G).

Further, in the present embodiment, as described above, the input data is set to the above nine dimensions, but the present invention is not limited to this, and the input data may be, for example, a part of the above nine dimensions. Alternatively, other dimensions may be added to the above nine dimensions, or the dimensions other than the above nine dimensions may be used. Further, the input data may include environmental conditions such as melting data (current value at the time of melting, melting time, etc.) and climatic conditions as parameters other than the dimensions. In addition, when parameters other than dimensions are included in the input data, for example, when the parameters include non-numerical qualitative data (for example, equipment such as electrodes and molds), the qualitative data is quantitatively included. It will be converted into data (numerical value) and input. The method of digitization is not particularly limited, but for example, qualitative data is converted into quantitative data by using a dummy variable (One-Hot encoding) or the like.

FIG. 7 is a diagram showing an example of a data set for generating a data prediction model. In FIG. 7, the data set for causing the neural network to perform machine learning is the above-mentioned input data (the outer diameter of the position G of the actually manufactured lutsubo, the outer diameter of the position F, the outer diameter of the position E, and the position A). Measured values of wall thickness, wall thickness at position D, wall thickness at position F, thickness of transparent layer at position A, thickness of transparent layer at position D, thickness of transparent layer at position F, etc.) and described above. It consists of a combination with an output label (actually measured value of the weight of the actually manufactured rutsubo).

Further, in the present embodiment, for example, a data set (data set shown in FIG. 7) for training a data prediction model is stored in the storage unit 12 in advance. Then, about 80% of the data set stored in the storage unit 12 is used as training data, and about 20% is used as evaluation data. Here, the training data is a data set used for training the data prediction model, and the evaluation data is not used for training the model, and it is verified whether the trained data prediction model is versatile. The data set used to do this. The data prediction model was verified by the holdout method. Further, it is desirable that the number of training data is prepared to the extent that the desired accuracy can be obtained in the data prediction model, for example. Further, in the performance evaluation of the data prediction model in the present embodiment, a mean squared error (MES: Mean Squared Error), which is one of the evaluation indexes, is adopted. That is, since the calculation result obtained by squaring the difference between the measured value and the predicted value and dividing the sum by the number of data is used as the value to be evaluated, for example, the smaller the value, the smaller the error. The evaluation index is not limited to this.

Then, in a state where the data set (training data, evaluation data) prepared as described above is stored in the storage unit 12, the control unit 11 reads the training data from the storage unit 12, for example, a machine learning algorithm. A data prediction model optimized for the weight prediction of the rutsubo is generated by performing supervised learning with training data using one neural network.

<Example of 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, and is composed of, for example, an input layer, an intermediate layer having one or more layers, and an output layer. Each unit of the input layer of this neural network has a one-to-one correspondence with the dimensions (x ₁ , x ₂ , x ₃ , ...) At predetermined positions constituting the input data in the above-mentioned data set. That is, the number of each dimension constituting the input data is the number of input units. On the other hand, the number of units in the output layer corresponds to the weight of the crucible (output label: z), and is one in the present embodiment.

Further, in the present 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. The number of layers and the number of units of the intermediate layer are not specified and can be set arbitrarily. Therefore, the number of layers and the number of units of the intermediate layer are appropriately determined according to the amount of data, the type of data, the required accuracy of output, and the like. For example, as the number of layers and units in the middle layer increases, the performance of the neural network such as analysis flexibility and output accuracy improves, but on the other hand, the amount of data, memory usage area, calculation amount, etc. increase. do.

Specifically, first, the control unit 11 reads one training data from the storage unit 12, and inputs the input data (x ₁ , x ₂ , x ₃ , ..., X _n ) constituting the training data to the input layer. Input to each unit of. Then, weights (w ₁₁ to w _{1 m} , w ₂₁ to w _{2 m} , w ₃₁ to w _{3 m} , ..., w _n1 to w _nm ) are added to the input data (x ₁ , x ₂ , x ₃ , ..., X _n ), respectively. Multiply and output the multiplication result to each unit of the next layer (intermediate layer). As a result, the input values (y ₁ , y ₂ , y ₃ , ..., _ym ) of each unit in the intermediate layer are as follows. Note that n is the number of units in the input layer, m is the number of units in the intermediate layer, and b ₁ , b ₂ , b ₃ , ..., B _m are biases. In addition, weights (w ₁₁ to w _{1 m} , w ₂₁ to w _{2 m} , w ₃₁ to w _{3 m} , ..., w _n1 to w _nm ) and biases (b ₁ , b ₂ , b ₃ , b _m ) are used in the learning process. It is a numerical value to be updated.
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, in each unit in the intermediate layer, the ReLU function described in the following equation is calculated for each of the multiple input values (y ₁ , y ₂ , y ₃ , ..., _ym ) from each unit in the previous layer. , The result is output to each unit in the next layer. In addition, M is 1 to m.

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

Then, the control unit 11 calculates the mean square error between the predicted value which is the output of the output layer and the weight of the teacher data, and repeatedly learns over a predetermined number of epochs so that this error is minimized. By doing so, the data prediction model is trained (weight and bias updates).

In the present embodiment, the learning model is generated using the data set shown in FIG. 7, but the input data of the data set is not limited to this. For example, some dimensions are selected from the dimensions constituting the input data shown in FIG. 7, and the selected dimensions (input data) and the actually measured value (output label) of the weight of the actually manufactured rutsubo are used. Supervised learning may be performed using the linked data set. Alternatively, for example, supervised learning can be performed using a data set in which the input data in which other parameters are newly added to the above input data and the measured value (output label) of the actually manufactured rutsubo weight are linked. It is also possible to do it.

Further, in the present embodiment, a neural network is used as an example of the 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. do not have. For example, it is conceivable to use an algorithm that combines a decision tree and gradient boosting, such as LightGBM.

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

In the present embodiment, the control unit 11 that operates as a trained data prediction model is, for example, an actually measured value of the dimension of a predetermined position output by the rutsubo manufacturing apparatus 3 in the measurement step (step S3) in the manufacturing process shown in FIG. (Outer diameter of position G, outer diameter of position F, outer diameter of position E, wall thickness of position A, wall thickness of position D, wall thickness of position F, thickness of transparent layer at position A, transparency of position D The thickness of the layer and the thickness of the transparent layer at position F) are accepted as input data, and the weight (estimated weight) of the quartz glass rubbing pot being manufactured is output. That is, the control unit 11 immediately predicts the weight of the quartz glass crucible to be completed by inputting the measured value of the dimension at the predetermined position obtained in the measurement step (step S3) without waiting for the completion of the quartz glass crucible. do.

Specifically, first, the operator (manufacturing worker 2) of the data prediction device 1 uses the user interface (input unit 13 including the keyboard and mouse) to obtain a predetermined position in the measurement step (step S3). The measured value of the dimension is input to the data prediction device 1 as input data (step S11 in FIG. 3). At this time, as the input data, the measured value of the dimension at the same position as when the data prediction model was generated is input. FIG. 9 is a diagram showing an example of the configuration of input data. When parameters other than dimensions are included in the input data, for example, when the parameters include non-numerical qualitative data, the amount of qualitative data is measured by the same method as when generating the data prediction model described above. Input after converting to target data. Then, the control unit 11 that operates as a trained data prediction model 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). Then, the weight predicted here (predicted weight) is displayed on the display unit 15 of the data prediction device 1.

<Reflections in the manufacturing process>
By confirming the predicted weight displayed on the display unit 15 of the data prediction device 1, the manufacturing worker 2 tends to manufacture the quartz glass crucible when the melting process is performed under the current melting conditions (large size). ， Small) can be known. Therefore, based on the predicted weight displayed on the display unit 15, the manufacturing worker 2 assumes that the tendency of the weight value becomes smaller when the tendency of the weight value is larger than the assumed value, for example. If it is smaller than the value, the current melting condition is adjusted so that it becomes larger (step S13 in FIG. 3) and reflected in the melting step of the crucible (step S14 in FIG. 3).

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

<Effect>
The data prediction device 1 of the present embodiment is actually manufactured under the melting conditions, using the actually measured values of the dimensions of the quartz glass ruts actually manufactured under the melting conditions set in the rutsubo manufacturing device 3 as input data. Using the measured value of the weight of the quartz glass rut as the teacher data, the storage unit 12 that stores the data set created by associating the teacher data with the input data, and the data set read from the storage unit 12 are used. A model generation means (control unit 11) that generates a data prediction model for predicting the weight of a quartz glass rutsubo by machine learning, and dimensions at a predetermined position output from the rutsubo manufacturing apparatus 3 in a measurement step in the manufacturing process of the quartz glass rutsubo. It is decided to provide a data prediction means (control unit 11) that accepts the measured value of the above as input data and outputs the predicted weight of the quartz glass rut pot manufactured in the manufacturing process by using the trained data prediction model. ..

Then, in the present embodiment, instead of measuring the weight using the completed quartz glass rutsubo (without waiting for the completion of the quartz glass rutsubo), the quartz to be completed is to be completed using a pre-trained data prediction model. Predict the weight of the quartz pot. As a result, the time until the weight is fed back to the manufacturing operator 2 can be significantly shortened as compared with the case where the weight is measured using the completed quartz glass crucible. It is possible to start the adjustment of the melting conditions by the operator 2 as soon as possible. Therefore, for example, the cycle time from the setting of the melting conditions in a specific manufacturing process to the next setting of the adjusted melting conditions can be significantly shortened.

Next, an example of the data prediction model optimized by the above learning model generation process will be described.

In this embodiment, for example, with the generation of a data prediction model for predicting the weight of the crucible, 6000 data sets used for learning were prepared. Then, this data set was randomly divided into training data and evaluation data. The breakdown was 4800 training data and 1200 evaluation data. In addition, a mean squared error (MES: Mean Squared Error), which is one of the loss functions, was adopted for the performance evaluation of the model.

In addition, the optimization algorithm uses "Adam", the activation function in the intermediate layer is the ReLU function, the activation function in the output layer is the linear function, and the learning rate is learned with the initial value set to "0.01". The learning rate is automatically lowered according to the progress. The batch size was 32. In addition, the initial value of the number of epochs was set to 1000, and for the purpose of preventing overfitting, learning was terminated if the error was not improved as compared with the case of the previous epoch.

Then, using the training data, the neural network was made to perform machine learning according to the above specifications, and the trained data prediction model was made to perform weight prediction using nine test data prepared in advance.

FIG. 10 is a diagram showing the evaluation results of learning, and shows, for example, the transition of the mean squared error with respect to the degree of learning (Epoches). As shown in FIG. 10, in the data prediction model of this embodiment, the mean square error becomes smaller as the learning progresses, so that it can be confirmed that the learning is proceeding smoothly. In addition, since the mean square error is sufficiently small from about 4 epochs and the mean square error does not change thereafter, it is said that about 4 epochs are suitable for learning the data prediction model in this embodiment. I can say.

Further, FIG. 11 is a diagram showing a prediction result by a trained data prediction model using test data. Here, the predicted value output from the data prediction model for predicting the weight of the rutsubo and the weight of the teacher data (actual measurement value) were plotted and the results were observed. As a result of observing the plot data of FIG. 11, the variation of the predicted value was small with respect to the teacher data, and the error was about 5% at the maximum.

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

Claims (5)

The measured value of the dimensions of the quartz glass crucible actually manufactured under the melting conditions set in the crucible manufacturing equipment is used as input data, and the measured value of the weight of the quartz glass crucible actually manufactured under the melting conditions is used as the teacher. As data, a storage means for storing a data set created by associating the teacher data with the input data, and
Using the data set read from the storage means, a model generation means for generating a data prediction model for predicting the weight of the quartz glass crucible by machine learning, and a model generation means.
In the measurement process in the manufacturing process of quartz glass rutsubo, the measured value of the dimension of the predetermined position output from the rutsubo manufacturing device is accepted as input data, and the quartz manufactured in the manufacturing process is used by using the trained data prediction model. A data prediction means that outputs the predicted weight of the glass rutsubo,
To prepare
A data prediction device characterized by that. moreover,
A display means for displaying the predicted weight output by the data prediction means,
To prepare
The data prediction device according to claim 1. The measured values of the dimensions at the predetermined positions are the outer diameter of the straight body portion, the wall thickness of the bottom portion, the wall thickness of the corner portion, the wall thickness of the straight body portion, the thickness of the transparent layer at the bottom, and the thickness of the corner portion in the quartz glass rut. All or part of the thickness of the transparent layer and the thickness of the transparent layer on the straight body,
The data prediction apparatus according to claim 1 or 2. The data prediction device according to claim 1, 2 or 3,
A rutsubo manufacturing device that manufactures quartz glass rutsubo by a manufacturing process including a molding process, a melting process, and a measuring process.
Equipped with
The crucible manufacturing apparatus carries out the melting process under melting conditions adjusted based on the tendency of the predicted weight of the quartz glass crucible output from the data prediction apparatus that operates as a machine-learned data prediction model.
A quartz glass crucible manufacturing system characterized by this. The melting conditions are used as data of the melting process including the current value, the voltage value, the integrated power value, the electrode opening degree, the electrode position, the melting temperature, the pressure, the cooling water temperature, the cooling water flow rate and the melting time at the time of melting.
The quartz glass crucible manufacturing system according to claim 4.

Images