A wafer epitaxial growth control method based on machine learning
Through the machine learning-based wafer epitaxial growth control method, the feedforward neural network model is used to monitor and adjust process parameters and wafer characteristics in real time, which solves the problem of control error during SiC epitaxial growth and achieves higher process parameters Accuracy and consistency reduce the professional requirements of craftsmen.
Patent Information
- Application Number
- CN202410604850.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-15
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-05-15
AI Technical Summary
During SiC epitaxial growth, it is difficult for the prior art to accurately control process parameters, resulting in large control errors in the production process, and process personnel need to rely on manual experience to adjust parameters, which may not be accurate or most reasonable results.
Using machine learning-based wafer epitaxial growth control method, the device status and process process are processed through two processes and two display screens, and the feedforward neural network model is used to monitor and adjust process parameters and wafer characteristics in real time to reduce humans Data deviation, realize intelligent processing of equipment.
It significantly reduces the control error during wafer epitaxial growth, improves the accuracy and consistency of process parameters, reduces the professional requirements of process personnel, and realizes intelligent control of MOCVD epitaxial growth.
Smart Images

Figure CN118223119B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and specifically to a wafer epitaxial growth control method based on machine learning. Background Art
[0002] Thanks to the rapid improvement of computer hardware data processing level, the increasing maturity of new parallel computing frameworks, and the continuous emergence of learning frameworks that are keen to contribute high-quality machine learning software systems, the craze for artificial intelligence has swept the world. People have been amazed by various new artificial intelligence works again and again. How to use the theory and tools of machine learning to effectively explore and empower the current advanced material process is also a very meaningful direction in the current AI application. Silicon carbide (SiC) material is the third generation of wide bandgap semiconductor materials after the first generation of semiconductor materials silicon (Si) and the second generation of semiconductors (gallium arsenide GaAs). The crystal structure of silicon carbide has the characteristics of homogeneous polymorphism. Its basic structure is a tetrahedral structure of Si-C, which belongs to a densely packed structure. Its performance has superior properties such as higher bandgap width, high critical breakdown electric field, high thermal conductivity, and high carrier saturation drift velocity. It has great prospects in the application of semiconductor lighting, new energy vehicles, power electronic devices, lasers, detectors and other fields.
[0003] In the field of power semiconductors, especially in SiC epitaxial equipment, the sources used for MOCVD growth are flammable, explosive, and highly toxic substances, and multi-component, large-area, thin-layer and ultra-thin-layer heterogeneous materials need to be grown, and the process window for epitaxial growth is relatively limited. At the same time, wafer epitaxial growth is easily affected by the microenvironment in the chamber, such as the gas flow field, temperature field, and graphite accessories in the reaction chamber. Therefore, in the MOCVD process, it is first necessary to find the appropriate parameters for SiC epitaxial growth in the current epitaxial equipment. At the same time, in order to control the various wafer characterization indicators during the epitaxial batch growth process, it is necessary to manually intermittently adjust the process parameters to adapt to the various microenvironments of the current epitaxial growth. Due to the high temperature in the reaction chamber of MOCVD epitaxial growth, the mixed reaction gas raw materials enter the reaction chamber according to the process requirements, and chemical reactions occur in the chamber and on the surface of the wafer to form a silicon carbide film; the residual gas raw materials are pumped away by the exhaust subsystem to ensure that the gas pressure in the reaction chamber is constant. Therefore, we need to parameterize the process and adjust the parameters to adapt to the microenvironment in the reaction chamber, such as the temperature field and flow field required for MOCVD epitaxial growth. Since the microenvironment in the reaction chamber is gradually changing, and there are slight differences in the batch process of wafer substrates, it is necessary to regularly and quantitatively test the wafers in the production of wafer epitaxial growth. Based on field experience and wafer test content, process engineers make appropriate adjustments to the process parameters to adapt to the microenvironment of the reaction chamber. In the control structure of the MOCVD equipment, the process personnel use manual experience to synchronously modify the process parameters to be executed according to the various characteristics of the current wafer to adapt to the gradual microenvironment in the current reaction chamber. The existing technical process personnel use manual experience to modify and adjust the parameters according to the status and data displayed in real time on the UI end and the execution results combined with the various characteristics of the current wafer, but such a process is often not necessarily accurate or the most reasonable. Summary of the invention
[0004] In view of this, the main purpose of an embodiment of the present invention is to provide a wafer epitaxial growth control method based on machine learning, which can significantly reduce the control error in the production process of wafer epitaxial growth.
[0005] The embodiment of the present invention is implemented as follows: a wafer epitaxial growth control method based on machine learning comprises the following steps:
[0006] S1, process one:
[0007] S11, in process one, issuing a process execution instruction through the first UI end process, wherein the process execution instruction includes process parameters for wafer epitaxial growth;
[0008] S12, executing the process business corresponding to the process parameters, writing the process parameters into the equipment, and returning the actual state of the process engineering executed by the equipment to the first UI end process;
[0009] S13, real-time monitoring of equipment status and process execution process through process one;
[0010] S2, process two:
[0011] S21, manually modifying the process parameters in the second UI end process;
[0012] S22, the device executes the modified process parameters and returns the wafer characteristics to the second UI end process;
[0013] S23, storing the process parameters input in process one, the process parameters modified in process two, and the wafer features returned in process two into a process parameter database;
[0014] S24, repeatedly performing steps S21, S22 and S23 to obtain process parameters and wafer feature sample data;
[0015] S3, machine learning: obtaining a wafer epitaxial growth model through training and learning according to the sample data, wherein the wafer epitaxial growth model is used to fit the functional relationship between the output process parameters and wafer characteristics;
[0016] S4, epitaxial growth process: adjusting the input process parameters to the wafer characteristics in real time according to the wafer epitaxial growth model to perform wafer epitaxial growth.
[0017] Furthermore, in step S3, obtaining a wafer epitaxial growth model through training and learning according to the sample data includes:
[0018] Regularize the process parameters as input parameters of the feedforward neural network At the same time, the epitaxial indicators of the wafer: uneven thickness distribution, uneven concentration distribution, and defect feature distribution are used as independent and unrelated variables as the output of the neural network. .
[0019] Furthermore, the process parameters at least include: carrier gas amount, process gas amount, pressure in the reaction chamber, process temperature, process time, process sequence, and process control strategy.
[0020] Furthermore, the method further comprises: given a training sample , ,in represents the data after the process recipe is regularized, is the attribute of process recipe data, Indicates wafer index, It is the property of wafer index.
[0021] Furthermore, the The attributes of the process recipe data include at least: temperature, pressure, and flow rate; The properties of wafer indicators include at least: thickness distribution, concentration distribution, and various types of defects.
[0022] Furthermore, the method further comprises: setting the output layer The threshold value of each neuron is It means that the hidden layer The threshold value of each neuron is It means that the input layer neurons and hidden layers The connection weight of each neuron is , hidden layer The neurons in the output layer The connection weight of each neuron is , hidden layer The input received by a neuron is: ; Output layer The input received by a neuron is: ;
[0023] If the discriminant function of the hidden layer and the output layer uses the Sigmoid function , for the training set, assuming that the output of the current neural network is ,Right now: ; Then the entire neural network is Mean square error over the dataset for: ;
[0024] In the neural network model, the total Parameters need to be determined: the output layer The threshold of a neuron , hidden layer The threshold of a neuron , input layer neurons and hidden layers The connection weights of neurons , Tibetan layer The neurons in the output layer The connection weights of neurons ; In the process of neural network learning, an iterative algorithm is used, and the estimation formula for any parameter is: ;
[0025] Using the gradient descent strategy, the target The negative gradient direction adjusts the parameters, for the error , given a learning rate , after one round of learning:
[0026]
[0027]
[0028]
[0029]
[0030] in: , ; Learning rate Control the update step size and learning rate of each learning iteration of the algorithm The value is (0,1).
[0031] Furthermore, the process one and the process two are displayed through two different displays.
[0032] Furthermore, the judgment process of S4 includes: in automatic mode, the trained learning model predicts and outputs the results of the process parameters to determine whether the epitaxial results of the wafer meet the growth requirements; if so, the process parameters are saved to the process parameter database for epitaxial growth; if not, the process parameters are corrected and the learning model is used again for judgment.
[0033] Furthermore, it is characterized in that the method further includes: S5, manual mode: in specific working conditions, when the machine learning capability is insufficient, resulting in large deviations in the prediction results that cannot meet the extended requirements, manual adjustments are performed through human experience intervention.
[0034] According to the above technical scheme, the embodiment of the present invention has the following effects: the present application adopts two processes and two display screens, which are respectively used for real-time monitoring of equipment status and process, modification or editing of process parameters and recipes, and learning and result processing of process data samples, so that the human data deviation caused by insufficient reliance on real-time monitoring data of equipment by process personnel can be reduced, and the mining and analysis of equipment data can be realized, and the equipment is given the ability to intelligently process the process, which reduces the professional requirements of the process for process personnel. Therefore, this method has very practical economic significance in the actual epitaxial growth of MOCVD. The present application processes the equipment status and data, and the real-time monitoring events of the process in different processes and different screens, so as to minimize the data deviation of process parameters caused by insufficient grasp of real-time data of equipment, and also proposes a feasible structural method for equipment intelligence. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:
[0036] Figure 1 A process schematic diagram of a wafer epitaxial growth control method based on machine learning provided by the present application is shown;
[0037] Figure 2 A schematic diagram of a feedforward neural network for sample learning in the circular epitaxial growth control method provided by the present application is shown;
[0038] Figure 3 A schematic diagram of process 2 in the circular epitaxial growth control method provided in the present application is shown. DETAILED DESCRIPTION
[0039] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.
[0040] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments. The step sequence or circuit diagram in the following embodiments is only an example and can be adjusted without conflict.
[0041] like Figure 1 As shown, a process diagram of a wafer epitaxial growth control method based on machine learning provided by an embodiment of the present invention includes the following steps:
[0042] S1, process one:
[0043] S11, in process one, issuing a process execution instruction through the first UI end process, wherein the process execution instruction includes process parameters for wafer epitaxial growth;
[0044] S12, executing the process business corresponding to the process parameters, writing the process parameters into the equipment, and returning the actual state of the process engineering executed by the equipment to the first UI end process;
[0045] S13, real-time monitoring of equipment status and process execution process through process one;
[0046] S2, process two:
[0047] S21, manually modifying the process parameters in the second UI end process;
[0048] S22, the device executes the modified process parameters and returns the wafer characteristics to the second UI end process;
[0049] S23, storing the process parameters input in process one, the process parameters modified in process two, and the wafer features returned in process two into a process parameter database;
[0050] S24, repeatedly performing steps S21, S22 and S23 to obtain process parameters and wafer feature sample data;
[0051] S3, machine learning: obtaining a wafer epitaxial growth model through training and learning according to the sample data, wherein the wafer epitaxial growth model is used to fit the functional relationship between the output process parameters and wafer characteristics;
[0052] S4, epitaxial growth process: adjusting the input process parameters to the wafer characteristics in real time according to the wafer epitaxial growth model to perform wafer epitaxial growth.
[0053] Specifically, in step S3, obtaining the wafer epitaxial growth model through training and learning according to the sample data includes: regularizing the process parameters as input parameters of the feedforward neural network At the same time, the epitaxial indicators of the wafer: uneven thickness distribution, uneven concentration distribution, and defect feature distribution are used as independent and unrelated variables as the output of the neural network. The process parameters at least include: carrier gas volume, process gas volume, pressure in the reaction chamber, process temperature, process time, process sequence, and process control strategy.
[0054] In conventional control structure methods, the input or correction of process parameters cannot monitor the equipment and process in real time, and the input or correction of process parameters cannot be performed when the equipment and process are being monitored in real time. The two cannot be taken into account at the same time. At the same time, in conventional control structure methods, there is no targeted machine training and learning of the process set by the current process parameters and the process results generated by the process parameters, and the process data and results are not fully mined. The equipment cannot intelligently and automatically correct the process parameters to adapt to the microenvironment in the reaction chamber.
[0055] Process one of the wafer epitaxial growth control method uses display screen 1. Real-time monitoring of the equipment status and process execution process, as well as the issuance of process execution instructions are realized through the UI end process of process one, and human-computer interaction is realized in display screen 1. The process business execution process of process one executes the process business corresponding to the process parameters from the UI end, writes the process parameters into the equipment, and returns the process engineering status to the UI end. The data service process of process one realizes the writing of data and status in the process business process into the equipment, and real-time reading of the equipment status and data, and sends the equipment status and data to the UI end process and the process business execution process. The database business process of process one mainly realizes the storage of other data edited by the UI end into the database, or reads the process parameter library in the database to the UI end.
[0056] The process and results of process 2 of the wafer epitaxial growth control method are displayed on display screen 2. The UI end process of process 2 is used for human-computer interaction, and the process parameters can be manually edited and modified. At the same time, the process parameters and the wafer characteristics of the process parameters are used as sample data for machine learning; a model with high accuracy is obtained by training and learning the generated numerous samples to fit the functional relationship between input and output. The trained model can predict and judge the output of the process parameters: whether it meets the expectations of process adjustment.
[0057] like Figure 2 As shown, it is a feedforward neural network for sample learning provided by this application. Since the process parameters (recipe) include: carrier gas volume, process gas volume, reaction chamber pressure, process temperature, process time, process sequence, process control strategy and other parameters are independent and unrelated, these parameters can be regularized as input parameters of the feedforward neural network. At the same time, the epitaxial indicators of the wafer: uneven thickness distribution, uneven concentration distribution, and defect feature distribution can be approximated as independent and unrelated variables as the output of the neural network. Due to the intrinsic relationship between the epitaxial index of the wafer and the process parameters (recipe), the parameters can be optimized through neural network training, thereby obtaining a learning model with higher accuracy.
[0058] Given a training sample , . represents the data after regularization of our process recipe, It is the attribute of process recipe data, i.e. temperature, pressure, flow, etc. Represents our wafer index, are the properties of wafer indicators, namely thickness distribution, concentration distribution, various defects, etc. Figure 2 Describes a person who has input neurons, output neurons, A multilayer feedforward neural network with hidden layers. If the output layer is set The threshold value of each neuron is It means that the hidden layer The threshold value of each neuron is It means that the input layer neurons and hidden layers The connection weight of each neuron is , hidden layer The neurons in the output layer The connection weight of each neuron is Then the hidden layer The input received by a neuron is:
[0059]
[0060] Output layer The input received by a neuron is:
[0061]
[0062] If the discriminant function of the hidden layer and the output layer uses the Sigmoid function , for the training set, assuming that the output of the current neural network is ,Right now:
[0063] Then the entire neural network Mean square error over the dataset for:
[0064]
[0065] In the neural network model, the total Parameters need to be determined: the output layer The threshold of a neuron , hidden layer The threshold of a neuron , input layer neurons and hidden layers The connection weights of neurons , Tibetan layer The neurons in the output layer The connection weights of neurons In the process of neural network learning, an iterative algorithm is used, and the estimation formula for any parameter is:
[0066]
[0067] We use the gradient descent strategy to The negative gradient direction adjusts the parameters, for the error , given a learning rate , after one round of learning:
[0068]
[0069]
[0070]
[0071]
[0072] in: , Learning rate The update step size of each learning iteration of the control algorithm, generally the learning rate The value is (0,1). The iteration process continues until certain set conditions are met.
[0073] Based on the above embodiment, the judgment process of S4 includes: in the automatic mode, the trained learning model predicts and outputs the results of the process parameters to determine whether the epitaxial growth results of the wafer meet the growth requirements; if they meet the requirements, the process parameters are saved to the process parameter database for epitaxial growth; if they do not meet the requirements, the process parameters are corrected and the learning model is used again for judgment. Figure 3 As shown, it is a specific schematic diagram of process 2 provided by the present application. In process 2, in the automatic mode, the trained learning model predicts and outputs the results of the process parameters to determine whether the current process parameters meet the adjustment expectations: that is, whether the epitaxial growth results of the wafer meet the growth requirements. If it meets the adjustment expectations, the process parameters are saved to the process parameter database for epitaxial growth. If the currently input process parameters do not meet the adjustment expectations, the process parameters are corrected and the learning model is used again for judgment.
[0074] The present invention provides a preferred embodiment, the method further comprises: S5, manual mode: in a specific working condition, when the general lack of ability of machine learning is insufficient, resulting in a large deviation of the prediction result that cannot meet the extension requirements, manual adjustment is performed through human experience intervention. At the same time, the manual mode of process 2 retains the mechanism that in certain specific working conditions, due to the insufficient general lack of ability of machine learning, the prediction result has a large deviation and cannot meet the extension requirements, and human experience intervention is required.
[0075] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other.
Claims
1. A wafer epitaxial growth control method based on machine learning, characterized in that: The steps include: S1, process one: S11, in process one, issuing a process execution instruction through the first UI end process, wherein the process execution instruction includes process parameters for wafer epitaxial growth; S12, executing the process business corresponding to the process parameters, writing the process parameters into the equipment, and returning the actual state of the process engineering executed by the equipment to the first UI end process; S13, real-time monitoring of equipment status and process execution process through process one; S2, process two: S21, manually modifying the process parameters in the second UI end process; S22, the device executes the modified process parameters and returns the wafer characteristics to the second UI end process; S23, storing the process parameters input in process one, the process parameters modified in process two, and the wafer features returned in process two into a process parameter database; S24, repeatedly performing steps S21, S22 and S23 to obtain process parameters and wafer feature sample data; S3, machine learning: obtaining a wafer epitaxial growth model through training and learning according to the sample data, wherein the wafer epitaxial growth model is used to fit the functional relationship between the output process parameters and wafer characteristics; S4, epitaxial growth process: adjusting the input process parameters to the wafer characteristics in real time according to the wafer epitaxial growth model to perform wafer epitaxial growth; The step S3 of obtaining the wafer epitaxial growth model through training and learning according to the sample data includes: Regularize the process parameters as input parameters of the feedforward neural network At the same time, the epitaxial indicators of the wafer: uneven thickness distribution, uneven concentration distribution, and defect feature distribution are used as independent and unrelated variables as the output of the neural network. ; The process parameters at least include: carrier gas volume, process gas volume, reaction chamber pressure, process temperature, process time, process sequence, and process control strategy; The attributes of the process recipe data include at least: temperature, pressure, and flow rate; The attributes of wafer indicators include at least: thickness distribution, concentration distribution, and various defects; The method further comprises: given a training sample , ,in represents the data after the process recipe is regularized, is the attribute of process recipe data, Indicates wafer index, is the attribute of wafer index; set the output layer The threshold value of each neuron is It means that the hidden layer The threshold value of each neuron is It means that the input layer neurons and hidden layers The connection weight of each neuron is , hidden layer The neurons in the output layer The connection weight of each neuron is , hidden layer The input received by a neuron is: ; Output layer The input received by a neuron is: ; If the discriminant function of the hidden layer and the output layer uses the Sigmoid function , for the training set, assuming that the output of the current neural network is ,Right now: ; Then the entire neural network is Mean square error over the dataset for: ; In the neural network model, the total Parameters need to be determined: the output layer The threshold of a neuron , hidden layer The threshold of a neuron , input layer neurons and hidden layers The connection weights of neurons , Tibetan layer The neurons in the output layer The connection weights of neurons ; In the process of neural network learning, an iterative algorithm is used, and the estimation formula for any parameter is: ; Using the gradient descent strategy, the target The negative gradient direction adjusts the parameters, for the error , given a learning rate , after one round of learning: in: , ; Learning rate Control the update step size and learning rate of each learning iteration of the algorithm The value is (0,1).
2. The wafer epitaxial growth control method according to claim 1, characterized in that: The process one and the process two are displayed through two different displays.
3. The wafer epitaxial growth control method according to claim 2, characterized in that: The judgment process of S4 includes: in automatic mode, the trained learning model predicts and outputs the results of the process parameters to determine whether the epitaxial results of the wafer meet the growth requirements; if so, the process parameters are saved to the process parameter database for epitaxial growth; if not, the process parameters are corrected and the learning model is used again for judgment.
4. The wafer epitaxial growth control method according to claim 1, characterized in that: The method further comprises: S5, manual mode: In specific working conditions, when the machine learning capability is insufficient, resulting in large deviations in the prediction results that cannot meet the external requirements, manual adjustments are made through human experience intervention.
