Silicon carbide crystal manufacturing apparatus, control device thereof, and method of generating learning model and controlling the same

Abstract

The control device has a learning model that outputs an estimated value of a second physical quantity that cannot be observed under conditions for manufacturing SiC crystals, from a first physical quantity that can be observed under the conditions for manufacturing SiC crystals. The control device generates a base learning model by machine learning using a simulation result of a simulation model based on structure data of the SiC crystal manufacturing apparatus as teacher data. The control device acquires the first physical quantity and a measured value of the second physical quantity measured under conditions in which SiC crystals cannot be manufactured but the second physical quantity can be observed, and generates a learning model that corrects an output of the base learning model based on the measured value.

Description

Technical Field

[0001] This disclosure relates to a silicon carbide (SiC) crystal manufacturing apparatus, a control device for the SiC crystal manufacturing apparatus, and a generative learning model and a method for controlling the SiC crystal manufacturing apparatus. Background Technology

[0002] In SiC crystal manufacturing, the chamber where the SiC crystal grows (especially the crystal growth surface) needs to be heated to very high temperatures (2000°C or higher on the crystal growth surface), and the inner wall of the chamber needs to be covered with an opaque heat-insulating material. Furthermore, due to interference with the path for supplying raw materials to the crystal growth surface, it is impossible to install sensors on the crystal growth surface or other surfaces within the chamber to measure the state during SiC crystal manufacturing. JP 2018-169818 A discloses a technique for generating a predictive model using machine learning that only uses simulation results based on theoretical formulas as teacher data and only sets manual command signals as input values, and using the predictive model to display an image of the theoretically predicted state within the chamber. In the technique disclosed in JP 2018-169818 A, a first state output from a semiconductor crystal manufacturing apparatus is input into the machine learning predictive model. An image generated based on a second state output from the predictive model is displayed outside the manufacturing apparatus. The first state is a physical quantity within the internal state of the manufacturing apparatus that can be directly measured by sensors or the like. The first state includes, for example, the position of components of the manufacturing apparatus, temperature or pressure at the measurement point, etc. The second state refers to the physical quantities within the manufacturing equipment that are extremely difficult to measure directly using sensors or similar means. The second state includes, for example, the temperature, feed concentration, flow rate, and vectors indicating the flow direction at each point in a specific region within the manufacturing equipment. Operators of the manufacturing equipment can understand the second state within the equipment by viewing images.

[0003] According to JP 2018-169818 A, the predictive model is obtained through machine learning. The results of simulating multiple states of the manufacturing equipment using the analytical simulation model are used as teacher data. The simulation model is generated from 3D models derived from CAD data of the manufacturing equipment, physical property data of the materials used in the reaction, etc.

[0004] JP 2020-181959 A discloses that data from various sensors in plasma processing equipment are used as teacher data for machine learning. Summary of the Invention

[0005] According to the technology disclosed in JP 2018-169818 A, the teacher data used to generate the predictive model is obtained through simulation. However, simulation may not reflect the exact characteristics of the actual semiconductor crystal manufacturing equipment, or it may not reflect the changes in the semiconductor crystal manufacturing equipment over time. That is, in the predictive model generated by simulation based on teacher data, there is a possibility that the prediction accuracy of the actual state in the furnace is insufficient. On the other hand, JP 2020-181959 A discloses the use of sensor data as teacher data. Although generating a high-accuracy learning model requires a large amount of teacher data, obtaining a large amount of teacher data through experiments increases both economic and time costs.

[0006] Furthermore, in semiconductor crystal manufacturing, SiC crystals, in particular, require high-temperature environments of 2000°C or higher. Therefore, it is necessary to cover the heated parts, such as the cavity space, with opaque thermal insulation materials, and it is impossible to measure the temperature around the crystal growth site, such as the susceptor, using pyrometers or similar instruments. Moreover, if sensors are installed to measure the temperature of the crystal growth site, the feed path and the sensors interfere with each other during the actual crystal growth process. For the same reason, it is difficult to directly observe the state within the cavity during the actual crystal growth process.

[0007] Simulations using methods such as the finite element method (FEM) can be used to estimate locations that cannot be directly observed. However, because simulations typically require tens of hours or longer of computation time, even when using measurements of constantly changing chamber states (temperature, flow rate, pressure, etc.) as input, obtaining simulation results takes much longer than the timescale (approximately 100 milliseconds to 1 second) required for feedback-controlled crystal growth. Therefore, it is difficult to estimate these locations during the growth process via simulation.

[0008] Furthermore, the physical property values ​​used in the simulations are those under predetermined ideal conditions, such as those described in the literature, and do not accurately reflect the components and conditions in the manufacturing equipment during the actual growth process, including changes over time. Machine learning derived from simulation data as teacher data cannot achieve sufficiently accurate estimates of the actual growth process. The machine learning model derived from simulation data as teacher data must be fine-tuned using measurement data obtained from the manufacturing equipment used for actual crystal growth.

[0009] This disclosure presents a novel technique for manufacturing SiC crystals, unlike any previously known. Specifically, it provides a method for generating a learning model capable of accurately predicting physical quantities that cannot be observed during crystal manufacturing, a SiC crystal manufacturing apparatus capable of using such a learning model to manufacture high-quality SiC crystals, and a control device for the SiC crystal manufacturing apparatus. Although the term "predictive model" is used in JP 2018-169818 A, in this disclosure, the model obtained through machine learning is referred to as a "learning model."

[0010] One aspect of this disclosure provides a SiC crystal manufacturing apparatus. The SiC crystal manufacturing apparatus includes a chamber, an actuator, a first sensor, and a control device. In the chamber, SiC crystals are manufactured from raw materials (e.g., silicon-containing gas and carbon-containing gas). The chamber may be referred to as a reactor. The actuator controls the temperature in the chamber and the flow rate, velocity, and concentration of the raw materials, etc. The first sensor measures a first physical quantity observable inside or outside the chamber under the conditions for manufacturing SiC crystals.

[0011] The control device has a learning model that outputs an estimate of a second physical quantity that cannot be observed under the conditions of SiC crystal manufacturing from the measurement value of the first sensor. The control device provides feedback control to the actuator, thereby reducing the difference between the estimate output by the learning model and the target value of the second physical quantity during SiC crystal manufacturing.

[0012] The control device includes a basic model generation module, an experimental result acquisition module, a correlation data acquisition module, a correlation determination module, and a model calibration module. The basic model generation module uses a simulation model based on structural data from a SiC crystal manufacturing equipment to generate multiple simulation results (first simulation results) of a first physical quantity and a second physical quantity, which are used as teacher data to generate a basic learning model through machine learning. The experimental result acquisition module acquires multiple experimental results when the actuator is driven under conditions where SiC crystals cannot be manufactured but the second physical quantity can be observed. These results include measurements from a first sensor (first sensor measurement) and a second sensor (second sensor measurement) attached to the SiC crystal manufacturing equipment that measures the second physical quantity. The correlation data acquisition module acquires the simulation result (second simulation result) of the second physical quantity output from the simulation model when the first physical quantity matches the first sensor measurement. The correlation determination module determines the correlation between the second simulation result and the second sensor measurement. The model calibration module generates a learning model that outputs the result of correcting the output of the basic learning model using the aforementioned correlation as an estimate of the second physical quantity.

[0013] This disclosure also provides a method for generating a learning model, which is used to obtain an estimate of a second, unobservable physical quantity from a first, observable physical quantity when manufacturing a SiC crystal using SiC crystal manufacturing equipment. The method includes a process for generating a basic model, a process for acquiring experimental results, a process for acquiring correlation data, a process for determining correlation, and a process for model calibration.

[0014] The basic model generation process includes using a simulation model generated from structural data of a SiC crystal manufacturing device to produce multiple simulation results (first simulation results) of a first physical quantity and a second physical quantity, which serve as teacher data. A basic learning model is then generated through machine learning. The experimental result acquisition process includes obtaining multiple experimental results, including the measurement values ​​of the first sensor (first sensor measurement value) and the measurement values ​​of the second sensor (second sensor measurement value), when the actuator is driven under conditions where SiC crystals cannot be manufactured in the SiC crystal manufacturing device (attached to a first sensor and a second sensor for measuring the second physical quantity), but the second physical quantity can be observed.

[0015] The correlation data acquisition process includes obtaining a simulation result (second simulation result) of a second physical quantity when the first physical quantity matches the measurement value of the first sensor using a simulation model. The correlation determination process includes determining the correlation between the second simulation result and the measurement value of the second sensor. The model calibration process includes generating a learning model, the output of which is used to calibrate the output of the basic learning model based on the aforementioned correlation as an estimate of the second physical quantity.

[0016] The technology disclosed in this disclosure enables the use of a learning model that can accurately predict a second physical quantity to provide feedback control for SiC crystal manufacturing equipment. Since the learning model, obtained through machine learning, does not require complex computations, obtaining the output—an estimate of the second physical quantity—from the input takes only a few milliseconds to one second. Therefore, the learning model can be used for real-time feedback control.

[0017] Specifically, the technique disclosed in this disclosure also uses a first sensor, employed in SiC crystal manufacturing, to obtain a first and a second measurement value to correct the output of the underlying learning model. Therefore, by utilizing the correlation obtained from the experimental results to correct the learning model used in SiC crystal manufacturing, a high-precision estimate of the second physical quantity can be obtained. Furthermore, the control device using the learning model controls the actuator of the SiC crystal manufacturing equipment, thereby reducing the difference between the target value and the estimated value of the second physical quantity. The control device operates as if feeding back the second physical quantity.

[0018] The control device enables the second physical quantity, which cannot be observed during SiC crystal fabrication, to approach a target value. An example of this second physical quantity is the surface temperature of the substrate. The surface temperature of the substrate is a crucial factor affecting the quality of the SiC crystal. If the surface temperature of the substrate can be brought close to the target value, a high-quality SiC crystal can be obtained. The substrate is the support plate that holds the substrate during crystal growth.

[0019] The details and further improvements of the technology disclosed herein will be described below. Attached Figure Description

[0020] The above and other objects, features, and advantages of this disclosure will become more apparent from the following detailed description with reference to the accompanying drawings. In the drawings:

[0021] Figure 1 This is a cross-sectional view of a SiC crystal manufacturing apparatus according to a first embodiment of the present disclosure;

[0022] Figure 2 This is a schematic diagram of the SiC crystal manufacturing equipment.

[0023] Figure 3 This is a schematic diagram used to explain the basic learning model;

[0024] Figure 4 It is a graph showing the relationship between the second simulation results and the second sensor measurements;

[0025] Figure 5 This is a schematic diagram used to explain the structure of the learning model;

[0026] Figure 6 This is a block diagram of the main control module; and

[0027] Figure 7 This is a flowchart of the learning model generation method. Detailed Implementation

[0028] (First Embodiment)

[0029] A SiC crystal manufacturing apparatus 10 according to a first embodiment of the present disclosure will be described with reference to the accompanying drawings. The SiC crystal manufacturing apparatus 10 is an apparatus for manufacturing SiC single crystals. Figure 1 A cross-sectional view of the SiC crystal manufacturing apparatus 10 is shown.

[0030] The SiC crystal manufacturing apparatus 10 includes a chamber 12, a heater 22, a base 13, a rotating device 21, and a gas supply device 23. These components are housed within a casing 11. The chamber 12 is a reactor in which high-temperature feed gases react to grow SiC single crystals. Because SiC is grown at temperatures of 2000°C or higher, the inner wall of the chamber 12 is covered with an opaque insulating material, although not shown. Optionally, the chamber 12 is made of an opaque insulating material.

[0031] Gas supply device 23 generates a first raw material gas containing silicon and a second raw material gas containing carbon, and supplies these raw material gases into chamber 12 through a hole located at the bottom of chamber 12. The first and second raw material gases are collectively referred to as raw material gases. Gas supply device 23 heats the raw material gases to a predetermined preparation temperature.

[0032] A heater 22 for heating the raw material gas in the furnace is disposed on the outer periphery of the chamber 12. The heater 22 includes a high-frequency coil 22a and a high-frequency generator 22b. The heater 22 heats the raw material gas inside the chamber 12 to 2000°C or higher. The raw material gas heated to 2000°C or higher crystallizes on the surface 13a of the substrate 13. The substrate 13 is a support plate that supports the substrate used for growing SiC crystals. The surface 13a of the substrate 13 corresponds to the crystal growth surface. The substrate 13 is supported by a rotating device 21. The rotating device 21 rotates the substrate 13. The rotating device 21 is also capable of moving the substrate 13 in the vertical direction.

[0033] Cooler 24 is disposed above chamber 12. Cooler 24 has a circulator 24a configured to circulate refrigerant and a cooling pipe 24b through which the refrigerant flows. Cooler 24 cools residual gas from the feed gas discharged from chamber 12.

[0034] Gas supply device 23, heater 22, rotating device 21, and cooler 24 are controlled by control device (CTRL) 30. Gas supply device 23, heater 22, rotating device 21, and cooler 24 are actuators for operating SiC crystal manufacturing equipment 10, and for convenience, they are collectively referred to as actuator 20 below. Actuator 20 controls the manufacturing conditions as shown in the following examples. Examples of manufacturing conditions include the temperature and pressure in chamber 12; the flow rate, velocity, gas concentration, gas temperature, gas pressure, and gas inflow location of the raw material gas supplied from gas supply device 23; the current value, control voltage, and frequency of the current flowing through high-frequency coil 22a; the vertical position and moving speed of high-frequency coil 22a; the vertical position, moving speed, and rotational speed of base 13; the temperature of the outer wall of the chamber; the amount of refrigerant; the temperature of the refrigerant; the pressure of the refrigerant; and the pressure, flow rate, and temperature of the exhaust gas.

[0035] Figure 1 The thick arrows in the diagram illustrate the flow of reactant gases during SiC crystal growth. As described above, the feed gas reaching the surface 13a of the substrate 13 has a temperature of 2000°C or higher.

[0036] The SiC crystal manufacturing apparatus 10 also includes multiple sensors 17a-17e. Sensor 17a measures the temperature of the outer wall of chamber 12. Sensor 17b measures the temperature of heater 22 (or the current flowing through high-frequency coil 22a). Sensor 17c measures the temperature, pressure, flow rate, velocity, and concentration of the feed gas supplied to chamber 12. Sensor 17d measures the temperature, pressure, flow rate, velocity, and concentration of the residual gas discharged from chamber 12. Sensor 17e measures the temperature of cooler 24. These sensors 17a-17e are used to know the state of chamber 12 and the state of the feed gas during SiC manufacturing. Hereinafter, these sensors 17a-17e are collectively referred to as first sensors 17. First sensors 17 function during SiC crystal manufacturing. That is, the physical quantities (temperature, flow rate, etc.) measured by first sensors 17 correspond to first physical quantities that can be observed under the conditions of SiC crystal manufacturing. Each of the first sensors 17 is arranged at a location where the temperature is below the sensor's heat resistance temperature even under the conditions of SiC crystal manufacturing. The first sensor 17 also includes a sensor that measures the output of the actuator 20. When the actuator 20 operates to follow a target value, the target value is an approximation of the output of the actuator 20, and therefore the target value can be considered a pseudo-measurement of the first sensor 17. The SiC crystal manufacturing apparatus 10 has a number of other first sensors, but their illustrations are omitted.

[0037] The measurement value of the first sensor 17 is sent to the control device 30. The control device 30 controls the actuators 20, such as the rotating device 21, heater 22, gas supply device 23, and cooler 24, based on the measurement value of the first sensor 17. The control device 30 determines a target value for each actuator 20 and controls each actuator 20 such that its output follows the target value. For example, the target value and output of the rotating device 21 is the rotational speed of the base 13, and the control device 30 controls the rotating device 21 such that its actual rotational speed follows the target rotational speed. The target values ​​for the gas supply device 23 are the temperature, pressure, flow rate, flow rate, and concentration of the raw material gas, and the control device 30 controls the gas supply device 23 such that the temperature, pressure, flow rate, flow rate, and concentration of the raw material gas generated by the gas supply device 23 follow the target temperature, target pressure, target flow rate, target flow rate, and target concentration.

[0038] When the control device 30 controls the actuator 20, a SiC single crystal SC is grown on the surface 13a of the substrate 13. That is, a SiC crystal is manufactured using the SiC crystal manufacturing apparatus 10. The temperature and pressure inside the chamber 12, especially the temperature of the surface 13a (crystal growth surface) of the substrate 13, are important for manufacturing crystals of good quality. In particular, the temperature at the center of the surface 13a of the substrate 13 is important. In other words, if the temperature and pressure of each part inside the chamber 12 can be matched with the temperature of the surface 13a of the substrate 13 (the temperature at the center of the surface) to the target value, a single crystal of good quality can be obtained. For convenience, the temperature of the surface 13a of the substrate 13 (e.g., the temperature at the center of the surface 13a) will be simply referred to as the surface temperature in the following text.

[0039] However, under the conditions of SiC crystal manufacturing, the temperature inside chamber 12 (especially the surface temperature) exceeds 2000°C. During SiC crystal manufacturing, the temperature inside chamber 12 exceeds the heat resistance temperature of the sensor, therefore the sensor cannot be placed inside chamber 12. Alternatively, if the sensor is placed in the flow path of the feed gas, the flow of the feed gas is obstructed, therefore the sensor cannot be placed in the flow path of the feed gas. However, the SiC crystal manufacturing equipment 10 can be operated such that, although SiC crystals cannot be manufactured, the temperature inside chamber 12 is lower than the heat resistance temperature of the sensor. When the SiC crystal manufacturing equipment 10 is operated such that the temperature inside chamber 12 is lower than the heat resistance temperature of the sensor, the temperature and pressure (especially the surface temperature) inside chamber 12 can be measured. Physical quantities that cannot be observed under the conditions of SiC crystal manufacturing, such as the temperature and pressure inside chamber 12, and the flow rate and velocity of the feed gas on the surface of the substrate, are called second physical quantities. When the actuator 20 is operated such that the temperature inside chamber 12 is lower than the heat resistance temperature of the sensor, although SiC crystals cannot be manufactured, the second physical quantity can be observed. The second physical quantity is closely related to the quality of the SiC crystal to be manufactured. If the second physical quantity can be made to follow the target value, high-quality SiC crystals can be manufactured.

[0040] As will be described in detail later, the SiC crystal manufacturing apparatus 10 includes sensor mounting portions 18a-18d for measuring a second physical quantity when the temperature inside the chamber 12 is lower than the heat resistance temperature of the sensor by controlling the actuator 20. Sensor mounting portion 18a is disposed on the surface of the base 13. Sensor mounting portion 18b is disposed on the inner surface of the chamber 12. Sensor mounting portion 18c is disposed on the inner bottom surface of the chamber 12. Sensor mounting portion 18d is disposed in the raw material gas supply path at the bottom of the chamber 12. By arranging sensors (sensors for measuring temperature and pressure) and the control actuator 20 on the sensor mounting portions 18a-18d under conditions not exceeding the heat resistance temperature of the sensor, it is possible to measure the second physical quantity (temperature and pressure) in various parts inside the chamber 12. The sensors for measuring the second physical quantity are collectively referred to as second sensors 19. Since the second sensors 19 cannot be attached during SiC crystal manufacturing, therefore... Figure 1 The second sensor 19 is drawn with a dashed line. The sensor fixing parts 18a-18d are examples, and the sensor fixing parts 18a-18d can be set in other positions.

[0041] To make the explanation easier to understand, Figure 2 A schematic diagram of a SiC crystal manufacturing apparatus 10 is shown. Figure 2In the diagram, multiple actuators (ACTRs) 20 are drawn as a rectangle. A first sensor 17 is positioned outside the chamber 12, specifically at a location where the temperature during SiC crystal fabrication does not exceed the sensor's heat resistance temperature. A sensor for detecting the output of the actuators 20 is also included in the first sensor 17.

[0042] The sensor mounting portion 18 is located at a position where the temperature exceeds the sensor's heat resistance temperature during SiC crystal manufacturing (typically the surface of the base 13). A second sensor 19 can be attached to the sensor mounting portion 18. However, the second sensor 19 is removed during SiC crystal manufacturing and is attached when the actuator 20 of the SiC crystal manufacturing apparatus 10 is controlled to operate at a temperature below the sensor's heat resistance temperature. The sensor mounting portion 18 may have attachment portions for attaching the second sensor 19. These attachment portions include, for example, a bracket, a screw hole for fastening to the second sensor 19, or a engagement portion for engaging with the second sensor 19.

[0043] The control device 30 in the SiC crystal manufacturing apparatus 10 is capable of generating a learning model that outputs an estimated value of a second physical quantity from the measurements of the first sensor 17 and the second sensor 19. The learning model is capable of estimating the second physical quantity (typically, surface temperature) during SiC crystal manufacturing with high accuracy. In generating the learning model, the measured values ​​are used when the actuator 20 is controlled to bring the interior of the chamber 12 below the heat resistance temperature of the sensors. Therefore, a highly accurate estimate can be obtained.

[0044] The control device controls the actuator 20 to operate the SiC crystal manufacturing equipment 10, and obtains an estimate of a second physical quantity by using the measurement value of the first sensor 17 and the learning model. The control device 30 controls the actuator 20 to reduce the difference between the target value and the estimated value of the second physical quantity. Because the learning model can output a high-precision estimate, high-quality single crystals can be manufactured.

[0045] The control device 30 will be described. The control device 30 includes a basic model generation module (BMG) 31, an experimental results acquisition module (ERA) 32, a correlation data acquisition module (CDA) 33, a correlation determination module (CD) 34, a model calibration module (MC) 35, a learning model (LM) 36, and a main control module (MCTRL) 37. Each module will be described below.

[0046] The basic model generation module 31 generates a basic learning model that serves as the foundation for the high-precision learning model 36. Specifically, the basic model generation module 31 uses a simulation model generated based on the structural data of the SiC crystal manufacturing equipment 10 to generate multiple simulation results (first simulation results) for multiple sets of first and second physical quantities. Then, the basic model generation module 31 uses the first simulation results as teacher data to generate a basic learning model through machine learning. The structural data of the SiC crystal manufacturing equipment 10 typically includes physical property values ​​(thermal conductivity, stiffness, etc.) of the materials constituting the SiC crystal manufacturing equipment 10. The structural data also includes physical property values ​​(viscosity, coefficient of thermal expansion, boiling point, melting point, etc.) of the feed gas. The simulation model can be obtained analytically using, for example, the finite element method (FEM). Computational fluid dynamics (CFD) techniques can also be used to simulate the feed gas. In the FEM, the SiC crystal manufacturing equipment 10 and its internal space are divided into a large number of grid points, each grid point is given structural characteristics (thermal conductivity, characteristics of the feed gas, etc.), and energy equations (heat conduction equations and fluid motion equations) between adjacent grid points are given. The grid points corresponding to each actuator 20 are assigned values ​​corresponding to the operation of the respective actuator 20, and the energy equation described above is solved to obtain the change of the physical quantity (temperature, pressure, etc.) of the corresponding grid point over time. The physical quantity of the grid point corresponding to the mounting position of the first sensor 17 corresponds to a first physical quantity, and the physical quantity of the grid point corresponding to the mounting position of the second sensor 19 (sensor fixing part 18) corresponds to a second physical quantity. A typical example of the second physical quantity is the surface temperature of the base 13.

[0047] The basic model generation module 31 simulates the state of the SiC crystal manufacturing equipment 10 during SiC crystal manufacturing using the aforementioned simulation model. In other words, the basic model generation module 31 simulates the state of the SiC crystal manufacturing equipment 10 when the actuator 20 is controlled under SiC crystal manufacturing conditions. The basic model generation module 31 acquires multiple sets of first and second physical quantities at different times during the simulation as first simulation results.

[0048] The basic model generation module 31 uses multiple first simulation results as teacher data to perform machine learning and generate a learning model for the SiC crystal manufacturing device 10. For ease of explanation, the learning model generated in this process is called the basic learning model. One first simulation result (a set of first and second physical quantities) corresponds to one set of teacher data. For machine learning, for example, neural network deep learning methods are used. Figure 3 A schematic diagram of the basic learning model is shown. The input layer of the neural network has the same number of nodes as the number of the first physical quantities in the first simulation result. Figure 3In the example, the output layer of the neural network has one node. The weights of the neural network are learned so that when a first physical quantity contained in the first simulation result is input into the input layer, a second physical quantity contained in the first simulation result can be obtained from the output layer. Therefore, a basic learning model can be obtained. In this example, the value of the node in the output layer is the surface temperature of base 13.

[0049] If a new first physical quantity is input into the resulting basic learning model, a new estimate of the second physical quantity (the surface temperature of the base 13) corresponding to the new first physical quantity can be obtained. However, the basic learning model cannot accurately reflect the physical performance values ​​of each part of the SiC crystal manufacturing apparatus 10 and the effects of degradation over time. That is, considering degradation over time, the output of the basic learning model cannot be expected to have high accuracy. Therefore, in the SiC crystal manufacturing apparatus 10 disclosed in this disclosure, a correlation between the behavior of the actual SiC crystal manufacturing apparatus 10 and the simulation results is specified, and this correlation is used to correct the output of the basic learning model.

[0050] The experimental result acquisition module 32 uses a SiC crystal manufacturing device 10 with a first sensor 17 and a second sensor 19 attached thereto, and the actuator 20 is driven under conditions where SiC crystals cannot be manufactured but a second physical quantity can be observed. Figure 1 and Figure 2 The second sensor 19, shown by a dashed line, is actually attached to the SiC crystal manufacturing apparatus 10 for experimentation. The control device 30 drives the actuator 20 to ensure that the temperature inside the chamber 12 does not exceed the sensor's heat resistance temperature. The experimental result acquisition module 32 then acquires multiple sets of experimental results, including measurements from the first sensor 17 (first sensor measurements) and the second sensor 19 (second sensor measurements). As described above, the second sensor 19 measures a second physical quantity. For ease of explanation, the experiment performed by the experimental result acquisition module 32 to acquire data is referred to below as a preliminary experiment.

[0051] The correlation data acquisition module 33 uses a simulation model to obtain the simulation result of the second physical quantity when the first physical quantity matches the measurement value of the first sensor. The simulation model is the model used by the basic model generation module 31. The correlation data acquisition module 33 uses the simulation model to perform a simulation under the same conditions as the preliminary experiment and obtains the simulation result of the second physical quantity as the output. The simulation output when the first physical quantity matches the measurement value of the first sensor is called the second simulation result.

[0052] The correlation determination module 34 determines the correlation between the second simulation result obtained by the correlation data acquisition module 33 and the second sensor measurement value obtained by the second sensor 19 in the preliminary experiment. For ease of interpretation, the second simulation result is represented by the symbol Te, and the experimental result (second sensor measurement value) is represented by the symbol Ts. Figure 4 An example is shown where multiple sets of curves representing the second simulation results Te and the second sensor measurement Ts are plotted. Figure 4 In the diagram, the horizontal axis represents the second simulation result Te, and the vertical axis represents the second sensor measurement value Ts.

[0053] When the correlation between multiple sets of second simulation results Te and second sensor measurements Ts is obtained through linear regression, the regression line is obtained as [second sensor measurement Ts] = Ca × [second simulation result Te] + Cb. Here, Ca represents the slope of the regression line, and Cb represents the intercept of the regression line. Figure 4 As just an example, the correlation between the second simulation result Te and the second sensor measurement Ts can be represented by a polynomial or nonlinear function.

[0054] The second sensor measurements, as a result of the preliminary experiment, reflect characteristics of the actual SiC crystal manufacturing apparatus 10, such as the structure and physical properties of the materials of chamber 12, and their degradation over time. The second simulation results are data obtained from the simulation model and do not reflect all characteristics of the actual SiC crystal manufacturing apparatus 10. The correlation between the second simulation results and the second sensor measurements is useful for correcting the simulation model results based on accurate physical property values ​​of the structure of the SiC crystal manufacturing apparatus 10.

[0055] The model calibration module 35 generates a learning model 36, which outputs the result of the basic learning model being calibrated by the correlation determined by the correlation determination module 34 as an estimate (an estimate of the second physical quantity). Figure 5 A diagram illustrating the final learning model 36 is shown. Learning model 36 consists of a basic learning model and a correlation (i.e., a correlation formula obtained by the correlation determination module 34) regarding the output of the basic learning model. When a new first physical quantity, adjusted for manufacturing SiC crystals, is input into the basic learning model, an output (second simulation result Te) is obtained. The second simulation result Te is an estimate of a second physical quantity (e.g., the temperature at surface 13a of the base 13) obtained through the basic learning model during crystal production. Learning model 36 corrects the output (second simulation result Te) of the basic learning model using the correlation formula (final estimate Tx = Ca × Te + Cb). Learning model 36 outputs the corrected estimate Tx using the correlation formula.

[0056] As described above, the learning model 36 reflects the results of preliminary experiments using the actual SiC crystal manufacturing equipment 10. That is, the learning model 36 reflects the precise physical performance values ​​of each part of the SiC crystal manufacturing equipment 10 and the effects of degradation over time.

[0057] The advantages of the SiC crystal manufacturing apparatus 10 of this embodiment (especially the advantages of the technology for generating the learning model 36) will be described. Generally, generating a learning model through machine learning requires a massive amount of teacher data. In the SiC crystal manufacturing apparatus 10, teacher data is obtained through simulation using the analysis simulation model of the SiC crystal manufacturing apparatus 10. Since simulation is performed using a computer, a large amount of teacher data can be automatically obtained through batch processing or the like.

[0058] When experimental results are used as teacher data to generate a learning model, numerous experiments must be conducted to obtain a large amount of teacher data, which increases both economic and time costs. The SiC crystal manufacturing apparatus 10 disclosed in this disclosure uses simulation results to generate a basic learning model, and the results of preliminary experiments are used to fine-tune the basic learning model. Since the experimental results are only used for fine-tuning, the sample size may be smaller than when generating teacher data for the learning model. The SiC crystal manufacturing apparatus 10 of this embodiment is capable of generating a learning model 36 that can accurately predict furnace conditions in a shorter time compared to conventional techniques.

[0059] The SiC crystal manufacturing apparatus 10 also uses the first sensor 17 used in crystal manufacturing to obtain data for correcting the output of the basic learning model (experimental results acquired by the experimental results acquisition module 32). Since the first sensor 17 used in actual crystal manufacturing is also used when acquiring data for correction, the above correlation has high accuracy.

[0060] The basic model generation module 31, experimental result acquisition module 32, correlation data acquisition module 33, correlation determination module 34, and model correction module 35 can achieve the acquisition of a high-precision learning model. The high-precision learning model obtains the estimated value of the second physical quantity from the first physical quantity and reflects the characteristics of the actual SiC crystal manufacturing equipment 10.

[0061] The main control module 37 performs feedback control on the actuator 20 based on the measurement value of the first sensor 17 during the actual manufacturing of SiC crystals. Figure 6 A block diagram of the main control module 37 is shown. Figure 6 In the diagram, the SiC crystal manufacturing equipment (SCMA) 10 is simplified and drawn using a rectangle.

[0062] The main control module 37 includes a command generation unit (CG) 38 and a differential device 39. The main control module 37 controls the actuator 20 of the SiC crystal manufacturing equipment 10 so that a second physical quantity (e.g., the surface temperature of the substrate 13) follows a predetermined target value. As described above, the target value is preset to an optimal value for SiC crystal growth.

[0063] The command generation unit 38 generates a command value for the actuator 20 based on the target value of the second physical quantity. Figure 6 The block remains inactive until the temperature inside chamber 12 reaches a predetermined initial temperature, and the actuator 20 is controlled in a predetermined sequence. When the temperature inside chamber 12 reaches the initial temperature, Figure 6 The box is active, and actuator 20 is controlled to make the second physical quantity follow the target value.

[0064] The SiC crystal manufacturing apparatus 10 includes a first sensor 17, and the measurement value of the first sensor 17 is input into a learning model 36. The learning model 36 outputs an estimate of a second physical quantity corresponding to the measurement value of the first sensor 17.

[0065] The target value is directly input to the command generation unit 38, along with the difference between the target value and the estimated value. This difference is generated by the difference device 39. The command generation unit 38 implements a control rule that combines feedforward and feedback. The command generation unit 38 applies a first control rule to the target value and generates a feedforward command. Simultaneously, the command generation unit 38 applies a second control rule to the difference between the target value and the estimated value and generates a feedback command. The final command is the sum of the feedforward and feedback commands. The first and second control rules can be well-known control rules, such as PID (proportional-integral-differential) control rules, PI (proportional-integral) control rules, PD (proportional-differential) control rules, etc.

[0066] Command generation unit 38 provides the generated command to actuator 20 and drives SiC crystal manufacturing equipment 10 through feedback control. When actuator 20 operates according to the command value, the measurement value of first sensor 17 changes. Since the measurement value of first sensor 17 is input into learning model 36, the estimated value of the second physical quantity output by learning model 36 also changes. Although the difference between the target value and the estimated value changes, main control module 37 controls actuator 20 of SiC crystal manufacturing equipment 10 to reduce the difference between the target value and the estimated value. The time required to receive the measurement value from the sensor and control the actuator is approximately the same as the time required for mechanical model input and output (a few milliseconds to 1 second).

[0067] As described above, the first physical quantity may include the output of actuator 20. The command value to actuator 20 corresponds to an approximation of the output of actuator 20. Therefore, when the output of actuator 20 is included in the first physical quantity, the command value to actuator 20 can be input into the learning model 36 as part of the first physical quantity. Figure 6 The dashed arrows in the diagram indicate cases where command values ​​for executor 20 are input into learning model 36.

[0068] In summary, the main control module 37 executes the following process. The main control module 37 uses a learning model 36. The learning model 36 outputs an estimate of a second physical quantity that cannot be observed under the conditions for manufacturing the SiC crystal, based on the measurement value from the first sensor 17. The main control module 37 controls the actuator 20 to reduce the difference between the estimate output by the learning model 36 and the target value of the second physical quantity during SiC crystal manufacturing. Because the second physical quantity, which deeply affects crystal growth, can be made close to the target value, high-quality SiC crystal manufacturing is achieved.

[0069] The advantages of the control device 30, including the learning model 36, will be described. The learning model 36 consists of a machine learning-based basic learning model and correlations. The basic learning model is a machine learning-based model, and because it does not require solving complex equations, results can be obtained quickly. The correlation is a relational expression representing the relationship between experimental results (measured values) and the corresponding simulated output (second simulation results). The correlation is represented by a regression line or a relational expression such as an approximate polynomial, and can be calculated quickly. That is, when using the learning model 36, estimates of the second physical quantity can be obtained quickly.

[0070] Furthermore, the output of the basic learning model is corrected based on the correlation of the results of preliminary experiments using the actual SiC crystal manufacturing equipment 10. This correction improves the accuracy of the estimate. By using the learning model 36, accurate estimates of a second physical quantity that cannot be observed during crystal manufacturing can be obtained at high speed. Therefore, the control device 30 using the learning model 36 is suitable for feedback control. When the control device 30, which performs feedback control using the learning model 36, controls the actuator 20 of the SiC crystal manufacturing equipment 10, high-quality SiC crystals can be manufactured.

[0071] The output of the learning model 36 is an estimate of the second physical quantity, not a measured value. Feedback control that uses an estimate instead of a measured value can be called pseudo-feedback control.

[0072] The control device 30 in this embodiment is effective in itself. That is, this disclosure provides a novel SiC crystal manufacturing apparatus 10 and a novel control device 30 for controlling the SiC crystal manufacturing apparatus.

[0073] (Second Embodiment)

[0074] This disclosure also provides a method for generating a learning model suitable for controlling a SiC crystal manufacturing apparatus. The method generates a learning model that obtains estimates of a second, unobservable physical quantity from a first observable physical quantity during the manufacturing of SiC crystals using the SiC crystal manufacturing apparatus. Figure 7 A flowchart of the learning model generation method is shown. This method includes a basic model generation process (S2), an experimental results acquisition process (S3), a correlation data acquisition process (S4), a correlation determination process (S5), and a model calibration process (S6).

[0075] The basic model generation process (S2) involves using multiple simulation results (first simulation results) of multiple sets of first and second physical quantities as teacher data to generate a basic learning model through machine learning. These simulation results are generated using a simulation model based on structural data from a SiC crystal manufacturing device. The details of the basic model generation process are similar to those of the basic model generation module 31 described above.

[0076] The experimental result acquisition process (S3) includes using a SiC crystal manufacturing device with a first sensor attached for measuring a first physical quantity and a second sensor attached for measuring a second physical quantity, and driving the actuator of the SiC crystal manufacturing device under the condition that SiC crystals cannot be manufactured but the second physical quantity can be observed. This experiment is called a preliminary experiment. The experimental result acquisition process includes acquiring multiple sets of experimental results in the preliminary experiment, including multiple sets of measurement values ​​from the first sensor (first sensor measurement value) and measurement values ​​from the second sensor (second sensor measurement value). The details of the experimental result acquisition process are similar to those of the experimental result acquisition module 32 described above.

[0077] The correlation data acquisition process (S4) includes acquiring the simulation result (second simulation result) of the second physical quantity when the first physical quantity matches the measurement value of the first sensor using a simulation model. The details of the correlation data acquisition process are similar to those of the correlation data acquisition module 33 described above. The correlation determination process (S5) includes determining the correlation between the second simulation result and the measurement value of the second sensor. The details of the correlation determination process are similar to those of the correlation determination module 34 described above.

[0078] The model calibration process (S6) includes generating a learning model 36, which outputs an estimate by correcting the output of the base learning model using the aforementioned correlation. The details of the model calibration process are similar to those of the model calibration module 35 described above.

[0079] This disclosure also provides a control method using the learning model 36 obtained by the above-described generation method. In this control method, firstly, a measurement value from a first sensor 17 is acquired from a SiC crystal manufacturing apparatus 10 operating under SiC crystal manufacturing conditions. Next, the measurement value from the first sensor is input into the learning model to obtain an estimate of a second physical quantity. Then, the SiC crystal manufacturing apparatus is controlled such that the difference between the target value and the estimated value of the second physical quantity is reduced. According to this control method, high-quality SiC crystals can be manufactured.

[0080] Next, typical examples of learning model generation methods will be described. The generation methods described below are examples, and the techniques disclosed in this disclosure are not limited to the following examples.

[0081] The learning model outputs an estimate of a second physical quantity at a predetermined location where the temperature exceeds the sensor's heat resistance temperature, derived from a first physical quantity observable during SiC crystal fabrication using SiC crystal manufacturing equipment. The method includes a basic model generation process, an experimental results acquisition process, a correlation data acquisition process, a correlation determination process, and a model calibration process. In the following text, for ease of explanation, the predetermined location where the temperature exceeds the sensor's heat resistance temperature during SiC crystal fabrication will be referred to as the location of interest. A typical location of interest is the surface of the substrate, and a typical second physical quantity is temperature (or pressure).

[0082] In the process of generating the basic learning model, a simulation model based on structural data from SiC crystal manufacturing equipment is used to simulate the conditions under which SiC crystals are manufactured, obtaining multiple sets of simulation results for the first and second physical quantities, i.e., multiple first simulation results. Next, a basic learning model is generated, in which these multiple first simulation results serve as teacher data for machine learning. When the first physical quantity is input, the basic learning model outputs an estimated value for the second physical quantity.

[0083] During the acquisition of experimental results, an auxiliary sensor (corresponding to the second sensor 19 mentioned above) for measuring the second physical quantity is attached to the location of interest. The SiC crystal manufacturing equipment operates under conditions where the temperature at the location of interest is lower than the heat resistance temperature of the sensor, although SiC crystals cannot be manufactured. Operating the SiC crystal manufacturing equipment under these conditions, even though SiC crystals cannot be manufactured, is referred to as a preliminary experiment.

[0084] During the correlation data acquisition process, a simulation model is used to perform simulations under the same conditions as the preliminary experiment, and a second simulation result corresponding to the auxiliary sensor measurements obtained in the preliminary experiment is obtained. The second simulation result corresponds to an estimate of the second physical quantity at the location of interest (the estimate before correction, described later). During the correlation determination process, the correlation between the second simulation result and the second sensor measurement is determined. An example of this correlation is the regression line described above. During the model correction process, a learning model is generated. In the learning model, a first physical quantity obtained when operating the SiC crystal manufacturing equipment under the conditions for manufacturing SiC crystals is input into the basic learning model, and the output of the basic learning model corrected using the aforementioned correlation is output as an estimate of the second physical quantity.

[0085] The first physical quantity includes the temperature (or pressure) at multiple locations within the SiC crystal fabrication apparatus. In preliminary experiments, the SiC crystal fabrication apparatus is controlled such that the temperature (or pressure) at these multiple locations varies. By conducting the preliminary experiments as described above, the structural characteristics of the various parts of the SiC crystal fabrication apparatus and the locations of interest are well reflected.

[0086] For convenience, the temperature at the aforementioned location of interest will be referred to as the interest temperature hereinafter. The relationship between the temperature at each location and the interest temperature depends on the structure of the SiC crystal manufacturing apparatus 10, and is approximately the same regardless of whether the interest temperature is high or low. Therefore, in the generation method of the above embodiment, the temperature at each location and the interest temperature are measured in the SiC crystal manufacturing apparatus 10, which operates in a manner not exceeding the heat resistance temperature of the sensor, through a preliminary experiment. Furthermore, the interest temperature is obtained when simulation is performed under the same conditions as the preliminary experiment. The relationship between the interest temperature obtained in the preliminary experiment (the second sensor measurement value) and the interest temperature obtained through simulation (the second simulation result) (the aforementioned correlation) generally holds and is independent of the absolute value of the temperature. Even at high temperatures, the correlation between the simulation result and the preliminary experiment result is maintained when the temperature is low. By using such a correlation, the temperature at the location of interest exceeding the heat resistance temperature of the sensor during crystal manufacturing can be accurately estimated.

[0087] Regarding the techniques described so far, please remember the following points. The learning models, namely the basic learning model and learning model 36 in the above embodiments, are both multi-input single-output models. The techniques disclosed in this disclosure can also use multi-input multi-output learning models. That is, the number of second physical quantities can also be two or more. Furthermore, in the techniques disclosed in this disclosure, learning models with different numbers of nodes in each layer (neural network learning models) or other learning algorithms (e.g., Gaussian process regression) can be employed.

[0088] In the technology disclosed in this disclosure, a basic learning model is generated using teacher data obtained through simulation. All teacher data used to obtain the basic learning model is obtained through simulation. Therefore, a large amount of teacher data can be obtained in a relatively short time. On the other hand, the correlation between the results of experiments using actual SiC crystal manufacturing equipment and the simulation results corresponding to those experiments is determined. Since the experiments that obtain sensor measurements are for obtaining the correlation between simulation results and experimental results (results reflecting the state of the actual SiC crystal manufacturing equipment), the amount of data from the experimental results may be less than the amount of teacher data used to obtain the learning model.

[0089] The second physical quantity refers to a physical quantity that cannot be observed under the conditions for manufacturing SiC crystals. Here, "cannot be observed under the conditions for manufacturing SiC crystals" includes situations where it could be observed even if it were costly, but it is not practically suitable to incur such a cost.

[0090] After the SiC crystal manufacturing equipment 10 has been operating for a considerable period of time following the generation of learning model 36, the experimental results acquisition module 32, the correlation data acquisition module 33, the correlation determination module 34, and the model calibration module 35 can restart. The learning model 36 can be updated based on the latest state of the SiC crystal manufacturing equipment 10.

[0091] The technology disclosed in this specification is not limited to Figure 1 The SiC crystal manufacturing apparatus 10 shown can also be applied to other crystal manufacturing apparatuses (e.g., crystal manufacturing apparatuses based on the sublimation method, solution method, HT-CVD method, or CVD method).

[0092] Although specific examples of this disclosure have been described in detail above, these are merely examples and do not limit the scope of the claims. The technology described in this specification includes various modifications and variations to the specific examples illustrated above. Furthermore, the technical elements described in this specification or drawings are technically useful individually or in various combinations, and are not limited to the combinations described in this specification at the time of application. Moreover, the technology illustrated in this specification or drawings can achieve multiple purposes simultaneously, but achieving one of these purposes is itself technically useful.

Claims

1. A silicon carbide (SiC) crystal manufacturing apparatus, comprising: A chamber in which SiC crystals are formed from raw materials; An actuator configured to control the manufacturing conditions used to generate the SiC crystal; A first sensor is configured to measure a first physical quantity observable under the conditions of manufacturing the SiC crystal in the chamber; and A control device having a learning model that outputs an estimated value of a second physical quantity from measurements of a first sensor, the second physical quantity being unobservable under the conditions of manufacturing the SiC crystal, the control device being configured to provide feedback control to the actuator such that the difference between the estimated value output from the learning model and a target value of the second physical quantity during the manufacturing of the SiC crystal is reduced. The control device includes The basic model generation module is configured to use the first simulation results as teacher data to generate a basic learning model through machine learning. The first simulation results are multiple simulation results of a set of first and second physical quantities generated using a simulation model based on the structural data of the SiC crystal manufacturing equipment. The experimental result acquisition module is configured to acquire multiple experimental results, including a set of measurements from a first sensor and a second sensor. The first sensor measurement is the value measured by the first sensor, and the second sensor measurement is the value measured by the second sensor. The second sensor is configured to measure the second physical quantity in a SiC crystal manufacturing apparatus equipped with both the first and second sensors, when the actuator is driven under conditions where SiC crystal fabrication is not possible but the second physical quantity can be observed. The correlation data acquisition module is configured to acquire a second simulation result, which is the simulation result of the second physical quantity output from the simulation model when the simulation result of the first physical quantity generated using the simulation model matches the measurement value of the first sensor. A correlation determination module, configured to determine the correlation between the second simulation result and the second sensor measurement, and The model calibration module is configured to generate the learning model, which outputs the result of calibrating the base learning model using the correlation as an estimate. The conditions for manufacturing the SiC crystal in the chamber include a temperature within the chamber exceeding the heat resistance temperature of the second sensor. The conditions under which SiC crystals cannot be manufactured but the second physical quantity can be observed include a temperature within the chamber that is lower than the heat resistance temperature of the second sensor. The first sensor is positioned at a location where the temperature is still lower than the heat resistance temperature of the first sensor, even under the conditions of manufacturing the SiC crystal in the chamber. When the actuator is driven under the condition that SiC crystals cannot be manufactured but the second physical quantity can be observed, the second sensor is attached to the cavity, and when the second sensor is attached to the cavity, the experimental result acquisition module acquires the plurality of experimental results.

2. The silicon carbide (SiC) crystal manufacturing equipment according to claim 1, further comprising: A sensor mounting part is arranged in the cavity and configured to be attached to the second sensor when the actuator is driven and the experimental result acquisition module acquires the plurality of experimental results, provided that SiC crystals cannot be manufactured but the second physical quantity can be observed.

3. The silicon carbide (SiC) crystal manufacturing apparatus according to claim 2, further comprising: A base, disposed within the cavity and configured to support a substrate for growing the SiC crystal, wherein The sensor fixing part is arranged on the surface of the base, and The second physical quantity includes the surface temperature of the base.

4. A control device for controlling a silicon carbide (SiC) crystal manufacturing equipment, the SiC crystal manufacturing equipment comprising: A chamber in which SiC crystals are generated from raw materials; an actuator configured to control the manufacturing conditions for generating the SiC crystals; and a first sensor configured to measure a first physical quantity observable under conditions for manufacturing the SiC crystal in the chamber, the control device comprising: The main control module has a learning model that outputs an estimated value of a second physical quantity from the measurement value of the first sensor. The second physical quantity cannot be observed under the conditions of manufacturing the SiC crystal. The main control module is configured to perform feedback control on the actuator so that the difference between the estimated value output from the learning model and the target value of the second physical quantity during the manufacturing of the SiC crystal becomes smaller. The basic model generation module is configured to use the first simulation results as teacher data to generate a basic learning model through machine learning. The first simulation results are a set of simulation results of the first physical quantity and the second physical quantity generated by a simulation model based on the structural data of the SiC crystal manufacturing equipment. The experimental result acquisition module is configured to acquire multiple experimental results of a set of first sensor measurement values ​​and second sensor measurement values, wherein the first sensor measurement value is the measurement value of the first sensor, and the second sensor measurement value is the measurement value of the second sensor, wherein the second sensor is configured to measure the second physical quantity in the SiC crystal manufacturing equipment with the first sensor and the second sensor attached when the actuator is driven under the condition that SiC crystal cannot be manufactured but the second physical quantity can be observed; A correlation data acquisition module is configured to acquire a second simulation result, wherein the second simulation result is the simulation result of the second physical quantity output from the simulation model when the simulation result of the first physical quantity generated using the simulation model matches the measurement value of the first sensor; A correlation determination module is configured to determine the correlation between the second simulation result and the second sensor measurement; and The model calibration module is configured to generate the learning model, wherein the learning model uses the correlation to calibrate the output of the base learning model as the estimated value output. The conditions for manufacturing the SiC crystal in the chamber include a temperature within the chamber exceeding the heat resistance temperature of the second sensor. The conditions under which SiC crystals cannot be manufactured but the second physical quantity can be observed include a temperature within the chamber that is lower than the heat resistance temperature of the second sensor. The first sensor is positioned at a location where the temperature, even under the conditions of manufacturing the SiC crystal, is lower than the heat resistance temperature of the first sensor. When the actuator is driven under the condition that SiC crystals cannot be manufactured but the second physical quantity can be observed, the second sensor is attached to the cavity, and when the second sensor is attached to the cavity, the experimental result acquisition module acquires the plurality of experimental results.

5. A method for generating learning models, comprising: The first simulation result is used as teacher data to generate a basic learning model through machine learning. The first simulation result is a set of simulation results of a first physical quantity and a second physical quantity generated by a simulation model based on the structural data of the silicon carbide (SiC) crystal manufacturing equipment according to any one of claims 1-3. The first physical quantity is a physical quantity that can be observed by a first sensor under the condition of manufacturing SiC crystals in the chamber of the SiC crystal manufacturing equipment. The second physical quantity is a physical quantity that cannot be observed under the condition of manufacturing SiC crystals. Multiple experimental results are obtained from a set of first sensor measurements and second sensor measurements, wherein the first sensor measurement is the measurement value of the first sensor and the second sensor measurement is the measurement value of the second sensor, wherein the second sensor is configured to measure the second physical quantity in the SiC crystal manufacturing equipment with the first sensor and the second sensor attached when the actuator of the SiC crystal manufacturing equipment is driven under the condition that the SiC crystal cannot be manufactured but the second physical quantity can be observed; Obtain a second simulation result, which is the simulation result of the second physical quantity output from the simulation model when the simulation result of the first physical quantity generated using the simulation model matches the measurement value of the first sensor; Determine the correlation between the second simulation result and the second sensor measurement; and A learning model is generated, and the output of the learning model is used to correct the output of the basic learning model using the correlation, as an estimate of the second physical quantity. The conditions for manufacturing the SiC crystal in the chamber include a temperature within the chamber exceeding the heat resistance temperature of the second sensor. The conditions under which the SiC crystal cannot be manufactured but the second physical quantity can be observed include a temperature within the chamber that is lower than the heat resistance temperature of the second sensor. The first sensor is positioned at a location where the temperature is still lower than the heat resistance temperature of the first sensor, even under the conditions of manufacturing the SiC crystal in the chamber. When the actuator of the SiC crystal manufacturing equipment is driven under conditions where the SiC crystal cannot be manufactured but the second physical quantity can be observed, the second sensor is attached to the chamber, and when the second sensor is attached to the chamber, the plurality of experimental results are acquired.

6. The method of claim 5, further comprising: The measurement value of the first sensor is obtained from a SiC crystal manufacturing apparatus operating under the conditions for manufacturing the SiC crystal; The estimated value of the second physical quantity is obtained by inputting the measurement value of the first sensor into the learning model; and The SiC crystal manufacturing equipment is controlled such that the difference between the target value and the estimated value of the second physical quantity is reduced.

Images