JP7739722B2 - Growth condition determination support device, single crystal growth system, growth condition determination support method and program - Google Patents

Description

The present invention relates to a growth condition determination support device, a single crystal growth system, a growth condition determination support method, and a program.

LT and LN single crystals, which are used to make SAW filters, are primarily grown by the Czochralski method (hereafter referred to as the "Cz method"). The Cz method involves contacting a seed crystal, a single crystal piece, with the surface of the raw material melt in a crucible, and then pulling the seed crystal upward while rotating it to grow a cylindrical single crystal of the same orientation. In the Cz method, the rotation speed and pulling speed of the crystal, as well as the temperature environment during growth, must be appropriately selected depending on the type of crystal.

For example, Patent Document 1 discloses a method for evaluating the state of melting of a seed crystal based on the range of fluctuation in the weight measured as the seed crystal melts, with the aim of accurately evaluating the state of melting of the seed crystal.

Japanese Patent Application Laid-Open No. 2019-112264

Conventionally, growth equipment such as crucibles has been subject to significant deterioration due to use, which means that the quality of the grown crystals changes depending on the number of times they are used, making it difficult to accurately predict the quality.

The present invention was made in consideration of the above circumstances, and aims to assist in determining growth conditions according to the number of times the crystal growth device is used.

In order to achieve the above object, a cultivation condition determination support device according to one aspect of the present invention comprises:
An apparatus for assisting in determining crystal growth conditions, comprising:
an acquisition unit that acquires learning data in which growth conditions and crystal quality information are associated;
a prediction model learning unit that updates a prediction model for each number of uses by machine learning using the associated growth conditions as an input and the crystal quality information as an output, based on the learning data classified according to the number of uses of the growth device used to grow the crystal;
a quality prediction value calculation unit that calculates a predicted quality value of the crystal for each number of times the growth apparatus is used by inputting growth conditions into the prediction model for each number of times the growth apparatus is used;
a growth condition determination unit that determines the growth conditions to be output based on the calculated quality prediction value;
Equipped with
Each of the prediction models for each of the number of uses is a logistic regression model,
The prediction model learning unit updates the prediction model so that a loss function of the logistic regression model is minimized .

It can help determine growth conditions based on the number of times the crystal growth equipment is used.

FIG. 1 is a diagram showing an example of a system configuration of a single crystal growth system. FIG. 1 is a diagram illustrating an example of a hardware configuration of a single crystal growth apparatus. FIG. 2 is a diagram illustrating an example of a hardware configuration of a growth condition determination support device. FIG. 2 is a diagram illustrating an example of a functional configuration of a growth condition determination support device. 10 is a flowchart illustrating an example of the flow of a machine learning process. 10 is a flowchart showing an example of the flow of a cultivation condition determination support process. FIG. 10 is a diagram illustrating an example of learning data. FIG. 10 is a diagram showing experimental results.

The following describes an embodiment of the present invention (present embodiment) with reference to the drawings.

(System configuration of single crystal growth system 1)
FIG. 1 is a diagram showing an example of the system configuration of a single crystal growth system.

The single crystal growth system 1 according to this embodiment comprises a single crystal growth apparatus 2, a growth condition determination support device 3, and a growth management database 4. The single crystal growth apparatus 2, the growth condition determination support device 3, and the growth management database 4 are communicatively connected via a communication network 5. The communication network 5 may be the Internet or an in-house network.

The single crystal growth apparatus 2 is equipped with a crucible and other components, and is used to grow single crystals such as lithium niobate. Using the growth conditions determined by the growth condition determination support device 3, the single crystal growth apparatus 2 brings a seed crystal into contact with a molten raw material, and raises the seed crystal while rotating it to grow the crystal.

The development condition determination support device 3 receives development result data from the development management database 4 and analyzes the received development result data. Then, the development condition determination support device 3 outputs information indicating the development conditions based on the results of the analysis.

Growth conditions include, for example, the crystal pulling speed, the crystal rotation speed, and conditions related to the temperature control of a heater, such as a high-frequency heating device. Growth conditions may also include various measurement results during the crystal growth process. Growth conditions include, for example, the time at which a shoulder begins to form after the crystal melt and seed crystal come into contact, the time at which the crystal diameter reaches the diameter of the straight body, the temperature change in the crucible that stores the melt, the ambient temperature inside the device, and the temperature change in the cooling section. These do not have to be directly related to the crystal pulling speed or rotation speed or the temperature control of the heating device, and may be statistical values that combine these. More specifically, growth conditions include the start or end time of the shoulder of the crystal, the start or end of the body, the crucible bottom temperature or temperature change, the time or amount of heat generated by solidification and the weight of the melt at that time, the start or end time of power output, the power rise time, duration, or power rise amount, the start time of seeding, the average temperature of the chamber, the average temperature of the cooler, the average temperature of the crucible stand, the average temperature of the window, the average temperature of the inflow or outflow into the chamber, the average temperature of the high-frequency power source, the average temperature of the tube or cable output from the cooler, the average temperature of the seal, the average temperature of the chamber door, the average temperature of the chamber body, the flow rate of the inflow or outflow into the chamber, the average inflow and outflow rate of the water cooler, the flow rate of oil used for sealing, the average flow rate of the water-cooled tubes in the chamber, the weight at the time of convergence of hunting, the average growth rate of the body, the average difference between the shoulder and diameter, and the time of bottoming out. Note that the values exemplified as averages can be changed to maximum or minimum values as appropriate.

The growth management database 4 is a device that stores information for managing the growth status of single crystals. The growth management database 4 transmits growth result data to the growth condition determination support device 3. The growth result data is data indicating the results of single crystal growth, and is data that associates the single crystal growth conditions with quality information of the grown single crystal (specifically, the result of determining whether the crystal is good or bad). The growth management database 4 also transmits information indicating the growth conditions to the single crystal growth device 2. The information indicating the growth conditions and the single crystal quality information are associated with each other using an initially set ID or the like. The growth management database 4 may store information indicating the growth conditions and the quality information in association with each other.

Here, quality information on single crystals refers to information on the polycrystallization of crystals, information on the occurrence of cracks or breaks in crystals pulled using the Cz method, information on the falling of crystals, etc. Single crystals mean that they are composed of a single crystallite, while polycrystals mean that there are multiple crystallites. Therefore, if a visual inspection reveals that the crystal contains something resembling an amorphous pattern, it is assumed that the crystal has become polycrystallized. This can also be confirmed by observing the cross section using electron backscatter diffraction (EBSD). In the case of polycrystalline bodies, it can be confirmed that there are crystallites with multiple orientations.

In this embodiment, the quality information for the single crystal is determined by visually inspecting the growth results, with those that have not become polycrystallized being judged as "good" and those that have become polycrystallized being judged as "bad."

(Hardware configuration of each device)
Next, the hardware configuration of each device will be described. Fig. 2 is a diagram showing an example of the hardware configuration of a single crystal growth device. Fig. 2 is a diagram conceptually showing the device for growing a single crystal.

The single crystal growth apparatus 2 is an apparatus that brings a seed crystal SC into contact with a raw material melt L, which is a melted raw material contained in a crucible 10, and grows a single crystal CR by rotating and raising the seed crystal SC. The single crystal growth apparatus 2 includes, for example, a crucible 10, a crucible stand 20, a crucible lifting mechanism 30, a pulling shaft 40, a heater 50, a refractory inner cylinder 60, a refractory shield plate 70, a refractory material 80, and a control device 90. Also shown as related components are a seed crystal 100, a raw material melt 110, and a single crystal 120.

The crucible 10 is placed on the crucible stand 20. A refractory material 80 is arranged in the chamber so as to surround the crucible 10. A heater 50 is arranged between the crucible 10 and the refractory material 80. If necessary, the heaters 50 may be arranged in multiple stages. Figure 1 shows heaters 50 arranged in two stages. The crucible 10 is heated by the heaters 50.

A refractory inner cylinder 60 and a refractory shield plate 70 may also be installed to surround the upper part of the crucible 10. The refractory inner cylinder 60 and the refractory shield plate 70 are provided to keep the pulled single crystal 120 warm. A pulling shaft 40 is provided above the refractory material 80 so as to be rotatable and movable up and down. A seed crystal holder 41 is provided at the tip of the lower end of the pulling shaft 40, and holds the seed crystal 100. The crucible 10 and crucible stand 20 can be moved up and down by a crucible lifting mechanism 30.

Crucible 10 is a container for storing and holding the raw material melt and growing a single crystal. Because oxide crystals such as LN are grown in an oxygen-containing atmosphere, it is preferable for the crucible to be made of a heat-resistant, oxygen-insensitive precious metal such as Pt (platinum), Rh (rhodium), or Ir (iridium), or an alloy of these metals. Crucible 10 is an example of a growth device used to grow crystals.

The crucible stand 20 is provided as a support stand that supports the crucible 10 from below. The crucible stand 20 may be made of various materials as long as it has sufficient heat resistance to withstand the heat of the heater 50 and durability to support the crucible 10.

The crucible lifting mechanism 30 is a drive mechanism for raising and lowering the crucible 10. In this embodiment, the crucible 10 is placed on the crucible stand 20, so the crucible 10 may be raised and lowered by raising and lowering the lifting plate 31 that supports the crucible stand 20. In other words, it is sufficient that the crucible 10 can be raised and lowered directly or indirectly.

The crucible lifting mechanism 30 is configured to be able to lift and lower the crucible 10 without rotating it. The crucible lifting mechanism 30 can be any mechanism as long as it can lift and lower the crucible 10 without rotating it. For example, a ball screw may be used to lift the lifting plate 31, and an intermediate jig may be provided to prevent the lifting plate 31 from rotating.

The pulling shaft 40 holds the seed crystal 100 on the seed crystal holder 41 at its lower end, brings the seed crystal 100 into contact with the surface of the raw material melt 110 held in the crucible 10, and rotates to pull up a single crystal. The pulling shaft 40 has the seed crystal holder 41 at its lower end that holds the seed crystal 100, and is equipped with a pulling shaft drive motor (not shown). The pulling shaft drive motor is a rotational drive mechanism that rotates the crystal while pulling it up.

The heater 50 is a means for heating the crucible 10 and is arranged to surround the periphery of the crucible 10. It may be a resistance heating heater or a high-frequency induction heating heater that uses an induction coil for heating.

The refractory inner tube 60, refractory shield plate 70, and refractory material 80 are made of sintered refractories such as alumina, zirconia, magnesia, and calcia, and function to prevent heat from leaking to the outside.

The control device 90 controls the overall operation of the single crystal growth apparatus 2. The control device 90 is composed of, for example, a CPU, main memory, storage devices such as a hard disk, a communication device for wired or wireless communication, an input device such as a keyboard, touch panel, or mouse, a computer equipped with a display device such as a monitor, a sequencer, etc. The control device 90 executes various processes according to the programs stored in the storage device. The control device 90 may also be used (or supplied) alone as a control device for controlling a growth apparatus that grows a single crystal by the rotational pulling method. In other words, the control device 90 can be applied alone to known single crystal growth apparatuses.

The control device 90 controls the temperature of the heater 50. The control device 90 also controls the rotation and elevation of the pulling shaft 40 by controlling the pulling shaft drive motor.

Next, we will explain the hardware configuration of the development condition determination support device 3. Figure 3 shows an example of the hardware configuration of the development condition determination support device.

The growth condition determination support device 3 includes a CPU (Central Processing Unit) 101, a main memory device 102, an auxiliary memory device 103, an input device 104, a display device 105, a communication interface device 106, and a drive device 107. Each of these devices is connected via a bus.

The CPU 101 is the main control unit that controls the operation of the growth condition determination support device 3, and realizes various functions described below by reading and executing programs stored in the main memory device 102.

When the development condition determination support device 3 is started, the main memory device 102 reads and stores the program from the auxiliary memory device 103. The auxiliary memory device 103 stores the installed program, as well as files, data, etc. required for the various functions described below.

The input device 104 is a device for inputting various types of information and is realized, for example, by a keyboard or pointing device. The display device 105 is for displaying various types of information and is realized, for example, by a display. The communication interface device 106 includes a LAN card, etc., and is used for connecting to the development management database 20.

The program according to this embodiment is at least a part of the various programs that control the cultivation condition determination support device 3. The program is provided, for example, by distributing a storage medium 108 or by downloading it from a network. The storage medium 108 on which the program is recorded can be of various types, including storage media that record information optically, electrically, or magnetically, such as CD-ROMs, flexible disks, and magneto-optical disks, and semiconductor memories that record information electrically, such as ROMs and flash memories.

In addition, when the storage medium 108 on which the program is recorded is inserted into the drive device 107, the program is installed from the storage medium 108 into the auxiliary storage device 103 via the drive device 107. Programs downloaded from the network are installed into the auxiliary storage device 103 via the communication interface device 106.

(Functional configuration of the development condition determination support device 3)
Next, a description will be given of the functions of the cultivation condition determination support device 3. Fig. 4 is a diagram showing an example of the functional configuration of the cultivation condition determination support device.

The growth condition determination support device 3 includes an acquisition unit 31, a learning unit 32, a quality prediction value calculation unit 33, a growth condition determination unit 34, and an output unit 35.

During the learning process, the acquisition unit 31 acquires learning data that associates growth conditions with crystal quality information.

The learning unit 32 updates the prediction model by learning based on the learning data. Specifically, the learning unit 32 includes a learning data classification unit 321 and a prediction model learning unit 322. The learning data classification unit 321 classifies the learning data according to the number of times the crucible 10 has been used. Note that the crucible 10 is an example of a growth device and may be other devices.

The prediction model learning unit 322 updates the prediction model for each number of uses through machine learning based on the classified learning data. That is, a prediction model is constructed for each number of uses of the crucible 10. If the upper limit for the number of uses of the crucible 10 is 20, 20 prediction models ranging from 0 to 19 are constructed. Then, for example, the prediction model learning unit 322 updates the prediction model for the Nth number of uses based on the learning data for the Nth number of uses. The prediction model is, for example, a logistic regression model. However, other prediction models may also be used, such as a support vector machine or a neural network.

The quality prediction value calculation unit 33 calculates a predicted quality value according to the given growth conditions. These growth conditions include the number of times the crucible 10 has been used. The quality prediction value calculation unit 33 then calculates a predicted quality value based on a prediction model according to the number of times it has been used. The predicted quality value is, for example, a value that indicates the probability that the quality is good, expressed as a number between 0 and 1. In this case, if the probability of good quality is 100%, the predicted value is 1, and if the probability of good quality is 50%, the predicted value is 0.5.

The growth condition determination unit 34 determines the growth conditions to be output based on the calculated threshold. The output unit 35 outputs information indicating the determined growth conditions.

(Operation of the growth condition determination support device 3)
Next, the operation of the breeding condition determination support device 3 will be described with reference to the drawings. Fig. 5 is a flowchart showing an example of the flow of machine learning processing.

The growth condition determination support device 3 performs machine learning processing periodically or at times in response to user operation. When the growth condition determination support device 3 starts machine learning processing, the acquisition unit 31 acquires learning data that associates growth conditions with quality information from the growth management database 4 (step S11). Next, the learning data classification unit 321 classifies the learning data by the number of times the crucible 10 has been used (step S12).

The prediction model training unit 322 updates the prediction model for each use count of the crucible 10 by learning based on the classified training data (step S13). Specifically, each prediction model is a logistic function that represents the probability y of a single crystal being of good quality depending on the growth conditions x.

It is calculated using the following formula.

The prediction model learning unit 322 determines the parameters w and b so that the loss function L is minimized, with the probability y of the single crystal being good being 0.5 or greater as good and the probability y being poor being poor if it is less than 0.5.

where N is the number of samples of data on growth conditions, M is the number of growth conditions, _yj is the quality of the single crystal predicted from the jth growth condition, and _tj is the quality of the single crystal grown under the jth growth condition.

is a regularization (L2 regularization) term to avoid overfitting. For example, the prediction model training unit 322 determines the parameter α using the following procedure. That is, the prediction model training unit 322 randomly divides the data for each growth condition into five parts, trains a prediction model using four of the parts, and evaluates the performance of the prediction model using the remaining data. Furthermore, the prediction model training unit 322 repeats the process of training the prediction model and evaluating the performance five times, and evaluates the performance of each α by averaging the performance index obtained five times. In this way, the prediction model training unit 322 updates the value of α to the value that maximizes the average performance index.

Figure 6 is a flowchart showing an example of the flow of the growth condition determination support process. The growth condition determination support device 3 executes the growth condition determination support process at a timing corresponding to a user operation, such as before starting single crystal growth.

When the growth condition determination support device 3 starts the growth condition determination support process, the quality prediction value calculation unit 33 receives input of the growth conditions (step S21). The growth conditions include the number of times the crucible 10 has been used. For example, when starting growth using a crucible 10 that has been used N times, the quality prediction value calculation unit 33 receives input of the growth conditions including the number N of times the crucible 10 has been used.

Next, the quality prediction value calculation unit 33 calculates a quality prediction value under the input growth conditions using the prediction model for the number of uses N that has been learned through the above-mentioned machine learning process (step S22). The quality prediction value is, for example, a value between 0 and 1 that indicates the probability that the quality is good.

Next, the growth condition determination unit 34 determines whether the quality prediction value is equal to or greater than a threshold value (step S23). The threshold value is a preset value, for example, 0.5.

If the growth condition determination unit 34 determines that the quality prediction value is not greater than or equal to the threshold value (step S23: No), the process returns to step S21.

If the growth condition determination unit 34 determines that the quality prediction value is equal to or greater than the threshold value (step S23: Yes), it determines the input growth conditions and terminates the growth condition determination support process.

Figure 7 shows an example of learning data. Learning data is data that associates growth conditions with quality information (good or bad) determined under those growth conditions. In the example of Figure 7, for example, Lot ID represents the number of times the crucible 10 has been used.

(Experimental results)
The prediction model by the above-described cultivation condition determination support device 3 was evaluated. The performance of the prediction model for each number of uses was evaluated by F1 score. The F1 score was calculated from the confusion matrix shown in Table 1.

As shown in Table 1, the confusion matrix is classified into four categories: true positive (TP), true negative (TN), false positive (FP), and false negative (FN), and summarizes the number of prediction results that fall into each category. From these, the recall, precision, and F1 score that are judged to be "good" or "bad" can be expressed as follows:

Good recall Recall = TP/(TP + FN)
Good precision Precision = TP/(TP + FP)
Good F-score = 2 x Recall x Precision / (Recall + Precision)
Defect reproducibility rate TNR = TN/(TN + FP)
Defective conformance rate = TN/(TN+FN)
F-score of defects = 2 × TNR × defective precision / (TNR + defective precision)
F1 score = (number of good samples x F-score of good samples + bad samples x F-score of bad samples) / total number of samples

Figure 8 shows the experimental results. For comparison, Figure 8 shows the results 902 predicted using a single prediction model regardless of the number of uses, and the results 901 predicted using the prediction model according to this embodiment. Note that the results 902 show the average F1 score because calculations were not performed according to the number of uses of the crucible 10.

In the prediction model according to this embodiment, the F1 score differs depending on the number of times the crucible 10 is used, but the F1 score for each number of uses exceeded the F1 score (approximately 0.4) in result 902. Note that in result 901 predicted by the prediction model according to this embodiment, analysis was not possible for the 4th, 10th, 11th, and 12th uses due to the small number of defective products and the small number of samples. Note that the average F1 score that could be analyzed was approximately 0.6.

The growth condition determination support device 3 according to this embodiment can assist in determining growth conditions according to the number of times the crucible 10, an example of a growth device, has been used, through machine learning according to the number of times it has been used.

In the above-described embodiments, examples have been shown in which the judgment results included in the cultivation result data 30 are binary values, either good or bad. Depending on the prediction model used, it may not be binary. For example, if an unnecessary element of the product, such as a scratch or stain, is found, but it does not affect the product's performance, it may be indicated by a "△" or similar symbol, indicating a gap between good "○" and bad "×".

The present invention has been described above based on various embodiments, but the present invention is not limited to the requirements set forth in the above embodiments. These aspects can be modified without departing from the spirit of the present invention, and can be determined appropriately depending on the application form.

The present invention can be applied to crystal growth.

REFERENCE SIGNS LIST 1 Single crystal growth system 2 Single crystal growth device 3 Growth condition determination support device 4 Growth management database 5 Communication network 31 Acquisition unit 32 Learning unit 33 Quality prediction value calculation unit 34 Growth condition determination unit 35 Output unit 321 Learning data classification unit 322 Prediction model learning unit

Claims (5)

An apparatus for assisting in determining crystal growth conditions, comprising:
an acquisition unit that acquires learning data in which growth conditions and crystal quality information are associated;
a prediction model learning unit that updates a prediction model for each number of uses by machine learning using the associated growth conditions as an input and the crystal quality information as an output, based on the learning data classified according to the number of uses of the growth device used to grow the crystal;
a quality prediction value calculation unit that calculates a predicted quality value of the crystal for each number of times the growth apparatus is used by inputting growth conditions into the prediction model for each number of times the growth apparatus is used;
a growth condition determination unit that determines the growth conditions to be output based on the calculated quality prediction value;
Equipped with
Each of the prediction models for each of the number of uses is a logistic regression model,
the prediction model learning unit updates the prediction model so as to minimize a loss function of the logistic regression model.
A support device for determining cultivation conditions. a training data classification unit that classifies the training data acquired by the acquisition unit according to the number of times the growth device used to grow the crystal has been used;
The cultivation condition determination support device according to claim 1. A single crystal growth system comprising a single crystal growth apparatus and a growth condition determination support apparatus,
The cultivation condition determination support device comprises:
An apparatus for assisting in determining crystal growth conditions, comprising:
an acquisition unit that acquires learning data in which growth conditions and crystal quality information are associated;
a prediction model learning unit that updates a prediction model for each number of uses by machine learning using the associated growth conditions as an input and the crystal quality information as an output, based on the learning data classified according to the number of uses of the growth device used to grow the crystal;
a quality prediction value calculation unit that calculates a predicted quality value of the crystal for each number of times the growth apparatus is used by inputting growth conditions into the prediction model for each number of times the growth apparatus is used;
a growth condition determination unit that determines the growth conditions to be output based on the calculated quality prediction value,
Each of the prediction models for each of the number of uses is a logistic regression model,
the prediction model learning unit updates the prediction model so as to minimize a loss function of the logistic regression model;
The single crystal growth apparatus is
growing a single crystal according to the output growth conditions;
Single crystal growth system. 1. A computer-implemented method comprising:
acquiring learning data that associates growth conditions with crystal quality information;
updating a prediction model for each number of uses by machine learning using the learning data classified according to the number of uses of the growth device used to grow the crystal, with the associated growth conditions as input and the quality information of the crystal as output;
calculating a predicted value of crystal quality for each number of times the growth apparatus is used by inputting growth conditions into the prediction model for each number of times the growth apparatus is used;
determining growth conditions to be output based on the calculated quality prediction value;
Equipped with
Each of the prediction models for each of the number of uses is a logistic regression model,
the updating step updates the prediction model so as to minimize a loss function of the logistic regression model.
A method for supporting the determination of development conditions. On the computer,
acquiring learning data that associates growth conditions with crystal quality information;
updating a prediction model for each number of uses by machine learning using the learning data classified according to the number of uses of the growth device used to grow the crystal, with the associated growth conditions as input and the quality information of the crystal as output;
calculating a predicted value of crystal quality for each number of times the growth apparatus is used by inputting growth conditions into the prediction model for each number of times the growth apparatus is used;
determining growth conditions to be output based on the calculated quality prediction value;
Execute
Each of the prediction models for each of the number of uses is a logistic regression model,
the updating step updates the prediction model so as to minimize a loss function of the logistic regression model.
program.