Epitaxial growth rate calculation method and system based on neural network

Abstract

The invention provides an epitaxial growth rate calculation method and system based on a neural network. The method comprises the following steps: collecting reflectivity data in a wafer epitaxial growth process; and preprocessing the reflectivity data to obtain input data. According to the method, the complex nonlinear relationship among the reflectivity data, the refractive index value (n) and the extinction coefficient value (k) is constructed through the neural network model, and the preprocessed reflectivity data is input into the neural network model, so that the refractive index value (n) and the extinction coefficient value (k) of the nonlinear relationship can be quickly and accurately obtained; the epitaxial growth rate of the wafer is calculated according to the refractive index value (n) and the extinction coefficient value (k). The method is high in processing efficiency, the epitaxial growth rate can be rapidly and accurately calculated in real time, the neural network model can adapt to multiple types of epitaxial growth by adjusting weight parameters in training, and the method has multi-process universality and is wide in application range.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a method and system for calculating epitaxial growth rate based on a neural network. Background Art

[0002] Epitaxial growth is the process of growing a single crystal film (single crystal layer) on a wafer substrate. This single crystal film is called an epitaxial layer. The wafer substrate does not necessarily have the same material and composition as the epitaxial layer. When the epitaxial layer and the wafer substrate are made of the same material, it is called homoepitaxial growth, and when they are made of different materials, it is called heteroepitaxial growth. Therefore, the crystal structure of the epitaxial layer is an extension of the crystal structure of the wafer substrate. Through epitaxial growth, the components can have a higher breakdown voltage or current handling capability. In the semiconductor epitaxial growth process, accurate calculation of the growth rate is crucial to optimizing the growth process and improving the film quality. Traditional methods have many limitations in calculating the growth rate, such as insufficient ability to handle complex nonlinear relationships, poor scalability, and poor accuracy. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for calculating epitaxial growth rate based on a neural network, which can quickly and accurately calculate the growth rate and solve the problems of insufficient ability to handle complex nonlinear relationships, poor scalability and poor accuracy of traditional methods.

[0004] To achieve the above-mentioned purpose, in a first aspect, the present invention provides a method for calculating epitaxial growth rate based on a neural network, comprising: collecting reflectivity data during wafer epitaxial growth; preprocessing the reflectivity data to obtain input data; inputting the input data into a neural network model, and performing calculations on the neural network model to obtain the refractive index value (n) and extinction coefficient value (k) of the epitaxial growth component; and calculating the wafer epitaxial growth rate based on the obtained refractive index value (n) and extinction coefficient value (k). The present invention constructs a complex nonlinear relationship between reflectivity data and a refractive index value (n) and an extinction coefficient value (k) through a neural network model. The preprocessed reflectivity data is input into the neural network model to quickly and accurately obtain the refractive index value (n) and the extinction coefficient value (k) of the nonlinear relationship. The epitaxial growth rate of the wafer is calculated based on the refractive index value (n) and the extinction coefficient value (k). The present invention has high processing efficiency and can calculate the epitaxial growth rate quickly and accurately in real time. Moreover, the neural network model can adapt to various types of epitaxial growth by adjusting the weight parameters in training, has multi-process versatility, and has a wide range of applications.

[0005] In some embodiments, the input data is input into a neural network model, including: the neural network model includes a network architecture design submodule and a model training submodule; the network architecture design submodule is used to establish a nonlinear mapping relationship between reflectivity data and a refractive index value (n) and an extinction coefficient value (k), wherein the mapping relationship is F(x), wherein the input data is R, and the mapping relationship about R is expressed as F(R)=(n, k); the model training submodule is used for training local data, annotating and preprocessing the reflectivity data, the refractive index value (n) and the extinction coefficient value (k) actually detected in the local data to form an annotated data set, wherein the annotated data set is formed into a training set and a validation set by a K-fold crossover method, wherein the training set inputs the mapping relationship F(x), the model training submodule adjusts the mapping relationship F(x) according to the input training set, the validation set inputs the mapping relationship F(x) for verification, and the adjusted mapping relationship F(x) is fed back to the network architecture design submodule. The beneficial effect is that the network architecture design submodule in the neural network model adjusts the network parameters according to different epitaxial growth conditions and different material systems, and can calculate the epitaxial growth rate of different epitaxial growth conditions and different material systems, and has a wide range of applications. The model training submodule continuously learns and adjusts the mapping relationship F(x) according to the reflectivity data, refractive index value (n) and extinction coefficient value (k) actually detected locally, optimizes the mapping relationship F(x), and feeds back to the network architecture design submodule to further improve the performance of the neural network model.

[0006] In some embodiments, the network architecture design submodule includes an input layer, a hidden layer, and an output layer, wherein the input layer, the hidden layer, and the output layer all include neurons; a fully connected network model is formed between the input layer, the hidden layer, and the output layer for data feature mapping; the hidden layer includes a hidden single layer, the hidden single layer is greater than or equal to 3 layers, the hidden single layer includes a plurality of neurons, the plurality of neurons are fully connected, and the neurons are used for data feature extraction. The beneficial effect is that the neurons in the plurality of hidden single layers can perform more advanced feature extraction on the information and obtain a more advanced feature representation.

[0007] In some embodiments, the hidden single layer adopts an activation function, and the activation function introduces a nonlinear factor to enhance the expression ability of the mapping relationship F(x). Its beneficial effect is that the activation function in the hidden single layer introduces a nonlinear factor into the linear model, so that the linear model is expressed as a nonlinear relationship through the activation function, and the refractive index value (n) and the extinction coefficient value (k) expressed as a nonlinear relationship from the input data are realized, and the nonlinear operation is completed.

[0008] In some embodiments, the reflectivity data of the validation set is input into the network architecture design submodule, and a plurality of predicted refractive index values ​​(n) and extinction coefficient values ​​(k) are obtained through the mapping relationship F(x); the network architecture design submodule includes a loss function, which is a root mean square error function, and the loss function quantifies the difference between the predicted value obtained by the mapping relationship F(x) and the actual detected value in the local data, and the loss function is used to optimize the mapping relationship F(x); a plurality of predicted refractive index values ​​(n) and extinction coefficient values ​​(k) are obtained through the optimized mapping relationship F(x). Refractive index value (n) and extinction coefficient value (k), multiple predicted refractive index values ​​(n) and extinction coefficient values ​​(k) are made into a scatter plot for observing robustness; multiple predicted refractive index values ​​(n) and extinction coefficient values ​​(k) are compared with the actual detected refractive index values ​​(n) and extinction coefficient values ​​(k) in local data through the root mean square error function to calculate the root mean square error, and obtain the root mean square error calculation value, and reflect the stability and accuracy of the corresponding mapping relationship F (x) according to the root mean square error calculation value and robustness. Its beneficial effect is that the robustness and root mean square error calculation value obtained by observation can reflect the stability and accuracy of the mapping relationship F (x).

[0009] In some embodiments, the collecting of reflectivity data during the epitaxial growth of the wafer includes: collecting the reflectivity data monitored in real time during the epitaxial growth process and making it into a data set. The beneficial effect is that the reflectivity data is collected and made into a data set, which facilitates the centralized preprocessing of the data in the data set.

[0010] In some embodiments, the preprocessing of the reflectivity data to obtain input data includes: performing sliding average filtering on a data set made from the reflectivity data, wherein the window in the sliding average filtering is used to remove abnormal values ​​and obviously erroneous data values, and denoising the reflectivity data; performing normalization processing on the denoised reflectivity data to obtain input data, and the input data is used to input the neural network model. The beneficial effects are: performing sliding average filtering on the data set, performing rapid calculation on the data set, being able to achieve real-time data processing, and having a fast data processing response speed, performing normalization processing on the data after the sliding average filtering processing, converting the numerical range of the denoised reflectivity data to a uniform scale, and obtaining input data for the neural network model.

[0011] In some embodiments, the window size is b×1, where b is the data length.

[0012] In some embodiments, the normalization process converts the de-noised reflectivity data into a distribution with a mean of 0 and a standard deviation of 1, which is used to convert the de-noised reflectivity data into a uniform scale to obtain input data. The beneficial effect is that the normalization process of the de-noised reflectivity data simplifies the calculation process, reduces the amount of calculation, and improves the model convergence speed during the gradient descent process.

[0013] In a second aspect, an embodiment of the present invention provides an epitaxial growth rate calculation system based on a neural network, comprising: a data acquisition module, used to collect reflectivity data during the epitaxial growth process of a wafer; a data preprocessing module, which preprocesses the received reflectivity data to obtain input data; a neural network model module, which calculates the received input data to obtain a refractive index value (n) and an extinction coefficient value (k) of the epitaxial growth component; and a growth rate calculation module, which calculates the epitaxial growth rate of the wafer based on the obtained refractive index value (n) and extinction coefficient value (k). BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flow chart of the method for calculating epitaxial growth rate based on neural network provided by the present invention; Figure 2 A schematic diagram of a neural network model in the method for calculating epitaxial growth rate based on a neural network provided by the present invention; Figure 3 A schematic diagram of a network architecture design submodule in the neural network-based epitaxial growth rate calculation method provided by the present invention; Figure 4 A flow chart of the neural network-based epitaxial growth rate calculation system provided by the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0016] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0017] In this application, unless otherwise clearly specified and limited, the terms "connection", "fixation", etc. should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0018] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present application, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing in the full text includes three parallel schemes. Taking "A and / or B" as an example, it includes scheme A, or scheme B, or a scheme that satisfies both A and B. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0019] Traditional methods have many limitations in growth rate calculation, such as insufficient ability to handle complex nonlinear relationships, poor scalability, and poor accuracy.

[0020] In view of the above problems, the embodiments of the present invention provide the following technical solutions to overcome the above problems.

[0021] Figure 1 A flow chart of the method for calculating epitaxial growth rate based on neural network provided by the present invention; Figure 2 A schematic diagram of a neural network model in the method for calculating epitaxial growth rate based on a neural network provided by the present invention; Figure 3 A schematic diagram of a network architecture design submodule in the neural network-based epitaxial growth rate calculation method provided by the present invention; Figure 4 A flow chart of the neural network-based epitaxial growth rate calculation system provided by the present invention.

[0022] like Figure 1-Figure 3 As shown, in a specific implementation of the present invention, the method for calculating the epitaxial growth rate based on a neural network includes: S101: Collecting reflectivity data during wafer epitaxial growth; S102: preprocessing the reflectivity data to obtain input data; S103: inputting the input data into a neural network model, and the neural network model performs calculation to obtain a refractive index value (n) and an extinction coefficient value (k) of the epitaxial growth component; S104: Calculate the wafer epitaxial growth rate according to the obtained refractive index value (n) and extinction coefficient value (k).

[0023] The above-mentioned process of calculating the epitaxial growth rate of the wafer based on the refractive index value (n) and the extinction coefficient value (k) is as follows: the epitaxial layer growth process of the wafer substrate is carried out in the chamber of the semiconductor epitaxial equipment, and the epitaxial growth process is then carried out by adjusting the process parameters in advance. Therefore, in this process, it is necessary to accurately monitor the growth rate in real time to ensure the quality of the epitaxial layer. In addition, according to different epitaxial processes involving different epitaxial layer materials and corresponding reflectivity parameters, the growth rate calculation process expression is: growth rate = G (n, k), and the present invention can calculate the epitaxial growth rate in the process in real time.

[0024] The present invention constructs a complex nonlinear relationship between reflectivity data and a refractive index value (n) and an extinction coefficient value (k) through a neural network model. The preprocessed reflectivity data is input into the neural network model to quickly and accurately obtain the refractive index value (n) and the extinction coefficient value (k) of the nonlinear relationship. The epitaxial growth rate of the wafer is calculated based on the refractive index value (n) and the extinction coefficient value (k). The present invention has high processing efficiency and can calculate the epitaxial growth rate quickly and accurately in real time. Moreover, the neural network model can adapt to various types of epitaxial growth by adjusting the weight parameters in training, has multi-process versatility, and has a wide range of applications.

[0025] In a specific embodiment of the present invention, Figure 2 and Figure 3 As shown, the input data is input into the neural network model, including: the neural network model includes a network architecture design submodule and a model training submodule; the network architecture design submodule establishes a nonlinear mapping relationship between reflectivity data and refractive index value (n) and extinction coefficient value (k), the mapping relationship is F(x), wherein the input data is R, and the mapping relationship about R is expressed as F(R)=(n, k); the model training submodule is used for training local data, and the reflectivity data, refractive index value (n) and extinction coefficient value (k) actually detected in the local data are annotated and preprocessed to form an annotated data set, the annotated data set is formed into a training set and a verification set by a K-fold crossover method, the training set inputs the mapping relationship F(x), the model training submodule adjusts the mapping relationship F(x) according to the input training set, the verification set inputs the mapping relationship F(x) for verification, and the adjusted mapping relationship F(x) is fed back to the network architecture design submodule.

[0026] The network architecture design submodule in the neural network model of this embodiment adjusts the network parameters according to different epitaxial growth conditions and different material systems, establishes the mapping relationship between different epitaxial growth conditions and different material systems, and obtains the refractive index value (n) and extinction coefficient value (k) corresponding to different epitaxial growth conditions and different material systems. The epitaxial growth rate of different epitaxial growth conditions and different material systems can be calculated, which has a wide range of applications and strong versatility. The model training submodule in the neural network model forms a training set and a verification set with the annotated data set through the K-fold crossover method. The annotated data set is randomly or hierarchically divided into k subsets, and k-1 subsets are selected each time to form a training set for training the mapping relationship F (x), and a more accurate mapping relationship F (x) is adjusted. The remaining subset is used as a verification set, and the verification set inputs the adjusted mapping relationship F (x) to verify the adjusted mapping relationship F (x). This process is repeated k times, and each time the test set and training set combination are different, which can make full use of all samples in the annotated data set, and reduce the deviation and uncertainty caused by a single data division through multiple training and testing. Specifically, the K value is set to 5 to ensure the accuracy of the mapping relationship F(x) evaluated by the labeled data set and reduce the training calculation cost. The model training submodule continuously learns and adjusts the mapping relationship F(x) based on the reflectivity data, refractive index value (n) and extinction coefficient value (k) actually detected locally, optimizes the mapping relationship F(x), and feeds back to the network architecture design submodule to further improve the performance of the neural network model.

[0027] Specifically, the neural network model includes hyperparameters, which are used to control the learning process and structure of the neural network model. The hyperparameters are set before the neural network model is trained. The learning rate in the hyperparameters is set to 0.0001, and the batch size in the hyperparameters is set to 64.

[0028] In a specific embodiment of the present invention, Figure 3 As shown, the network architecture design submodule includes an input layer, a hidden layer and an output layer, and the input layer, the hidden layer and the output layer all include neurons; a fully connected network model is formed between the input layer, the hidden layer and the output layer for data feature mapping; the hidden layer includes a hidden single layer, the hidden single layer is greater than or equal to 3 layers, the hidden single layer includes a plurality of the neurons, the plurality of neurons are fully connected, and the neurons are used for extracting data features.

[0029] In this embodiment, the network architecture design submodule includes an input layer, a hidden layer and an output layer. A fully connected network model is formed between the input layer, the hidden layer and the output layer to perform data transmission and data feature mapping. The input layer uses a single neuron as the input of the input data. The hidden layer after the input layer is used to extract data features. The hidden layer includes a hidden single layer greater than or equal to 3 layers for data feature extraction. Specifically, three hidden single layers are set after the input layer. The three hidden single layers after the input layer are the first hidden single layer, the second hidden single layer and the third hidden single layer, respectively. The first hidden single layer, the second hidden single layer and the third hidden single layer contain 8 Neurons, 16 neurons and 8 neurons. The 8 neurons in the first hidden layer are used to perform preliminary nonlinear transformation on the input data and extract the basic features of the data. The data of the preliminary nonlinear transformation is input into the second hidden layer. The 16 neurons in the second hidden layer further perform feature extraction on the data of the preliminary nonlinear transformation. The data of the further extracted features is input into the third hidden layer. The 8 neurons in the third hidden layer perform more advanced feature extraction on the input data of the further extracted features. The more advanced feature data is input into the output layer. The two neurons in the output layer output the refractive index value (n) and the extinction coefficient value (k). When the number of hidden layers in the hidden layer is greater than 3, the neurons in multiple hidden layers can perform more advanced feature extraction on the information and obtain more advanced feature representation, which has a wide range of applications.

[0030] In a specific embodiment of the present invention (not shown), the hidden single layer adopts an activation function, and the activation function introduces a nonlinear factor to enhance the expression ability of the mapping relationship F (x). The activation function in the hidden single layer introduces a nonlinear factor into the linear model, so that the linear model is expressed as a nonlinear relationship through the activation function, and the refractive index value (n) and the extinction coefficient value (k) expressed as a nonlinear relationship from the input data are realized, and the nonlinear operation is completed. The activation function includes a sigmoid function, a tanh function, a ReLU function, and the like. The specific sigmoid function and the tanh function are effective when processing small data, and the ReLU function is effective when processing large data. When selecting the activation function, the appropriate activation function can be selected according to the specific situation of the data.

[0031] In a specific embodiment of the present invention, (not shown), the reflectivity data of the validation set is input into the network architecture design submodule, and a plurality of predicted refractive index values ​​(n) and extinction coefficient values ​​(k) are obtained through the mapping relationship F(x); the network architecture design submodule includes a loss function, which is a root mean square error function, and the loss function quantifies the difference between the predicted value obtained by the mapping relationship F(x) and the actual detected value in the local data, and the loss function is used to optimize the mapping relationship F(x); the optimized mapping relationship F(x) is obtained A plurality of predicted refractive index values ​​(n) and extinction coefficient values ​​(k), and the plurality of predicted refractive index values ​​(n) and the extinction coefficient values ​​(k) are made into a scatter plot for observing robustness; a plurality of predicted refractive index values ​​(n) and the extinction coefficient values ​​(k) are subjected to root mean square error calculation with the refractive index values ​​(n) and the extinction coefficient values ​​(k) actually detected in the local data through the root mean square error function, and a root mean square error calculation value is obtained, and the stability and accuracy of the corresponding mapping relationship F (x) are reflected according to the root mean square error calculation value and the robustness.

[0032] In this embodiment, the loss function in the network architecture design submodule is a key indicator for measuring model performance, thereby guiding model training. The loss function uses a root mean square error function, which provides a smooth and convex optimization landscape, which helps the mapping relationship F (x) to be efficiently optimized based on the gradient algorithm, so that the refractive index value (n) and the extinction coefficient value (k) calculated by the mapping relationship F (x) are closer to the actual value, wherein the gradient algorithm includes any one of batch gradient descent, random gradient descent, small batch gradient descent, momentum method, and adaptive gradient algorithm. The data in the validation set obtains multiple corresponding refractive index values ​​(n) and extinction coefficient values ​​(k) through the optimized mapping relationship F (x), and the obtained multiple corresponding refractive index values ​​(n) and extinction coefficient values ​​(k) are made into a scatter plot, wherein the scatter plot visualizes the multiple corresponding refractive index values ​​(n) and extinction coefficient values ​​(k), thereby evaluating the robustness of the overall system, and the robustness can reflect the stability of the mapping relationship F (x).

[0033] in: Indicates the actual detection value in the local data; : represents the value predicted by the mapping relationship F(x); n: is the total number of data; The root mean square error function provides a quantitative prediction accuracy evaluation value. The smaller the root mean square error function value, the better the prediction performance of the mapping relationship F (x) and the higher the degree of fit with the actual value. When the root mean square error function is used as a loss function, when there are outliers in the data, the root mean square error function reduces the impact of outliers on the stability of the model by reducing the impact of the square root of the error. The robustness and root mean square error calculation values ​​obtained through observation can reflect the stability and accuracy of the mapping relationship F (x), with strong visualization, which is convenient for direct observation by personnel.

[0034] In a specific embodiment of the present invention (not shown), the collection of reflectivity data during the epitaxial growth of the wafer includes: collecting the reflectivity data monitored in real time during the epitaxial growth process and making it into a data set. In this embodiment, the sensor and the measuring device in the epitaxial growth device collect the reflectivity data during the epitaxial growth process in real time, and the reflectivity data is collected and made into a data set, so that the data in the data set can be centrally preprocessed.

[0035] In a specific embodiment of the present invention (not shown), the reflectivity data is preprocessed to obtain input data, including: performing sliding average filtering on a data set made from the reflectivity data, wherein the window in the sliding average filtering is used to remove abnormal values ​​and obviously erroneous data values, and denoise the reflectivity data; the denoised reflectivity data is normalized to obtain input data, and the input data is used to input the neural network model. In this embodiment, the data set is subjected to sliding average filtering, and the data set is quickly calculated, so that real-time data processing can be achieved, and the data processing response speed is fast. The sliding average filtering process has a window, and the window is used to calculate the average value of the segment of data. The average value is used to smooth the data sequence, remove abnormal values ​​and obviously erroneous data values, and can effectively reduce noise interference. The data after the sliding average filtering process is normalized, and the numerical range of the denoised reflection data is converted to a unified scale to obtain the input data of the neural network model, and the preprocessing of the reflectivity data is completed.

[0036] In a specific embodiment of the present invention (not shown), the window size is b×1, where b is the data length. In this embodiment, when the reflectivity data is subjected to sliding average filtering, since the reflectivity data is one-dimensional data, the window size is set to b×1 to be applicable to one-dimensional reflectivity data. Specifically, the window size in the sliding average filter affects the filtering effect. b=5 is selected, and the window size is 5×1. The sliding average filter processing of the reflectivity data is performed to avoid excessive smoothing of the reflectivity data after processing due to a larger window. A smaller window cannot effectively remove outliers and obviously erroneous data values ​​to retain more reflectivity data features.

[0037] In a specific embodiment of the present invention (not shown), the normalization process converts the denoised reflectivity data into a distribution with a mean of 0 and a standard deviation of 1, which is used to convert the denoised reflectivity data into a uniform scale to obtain input data. In this embodiment, the normalization process converts the denoised reflectivity data into a distribution with a mean of 0 and a standard deviation of 1, eliminates the differences in the dimensions and value ranges between different evaluation indicators, and makes all data in a uniform scale to obtain input data, which is convenient for inputting into the neural network model for calculation. The normalization process of the denoised reflectivity data simplifies the calculation process, reduces the amount of calculation, and improves the model convergence speed during the gradient descent process.

[0038] Based on the neural network-based epitaxial growth rate calculation method provided in the above embodiment, Figure 4 As shown, the present invention provides an epitaxial growth rate calculation system based on a neural network, comprising: a data acquisition module 201, used to collect reflectivity data during the epitaxial growth process of a wafer; a data preprocessing module 202, which preprocesses the received reflectivity data to obtain input data; a neural network model module 203, which calculates the received input data to obtain a refractive index value (n) and an extinction coefficient value (k) of an epitaxial growth component; and a growth rate calculation module 204, which calculates the epitaxial growth rate of the wafer according to the obtained refractive index value (n) and extinction coefficient value (k).

[0039] All relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding unit module. Specifically, first, the data acquisition module 201 receives the reflectivity data during the epitaxial growth of the wafer, and then transmits the reflectivity data to the data preprocessing module 202. The data preprocessing module 202 preprocesses the received reflectivity data to obtain input data, and then the data preprocessing module 202 transmits the input data to the neural network model module 203. The neural network model module 203 performs nonlinear calculation on the input data to obtain the refractive index value (n) and extinction coefficient value (k) of the required epitaxial growth component. Finally, the neural network model module 203 inputs the refractive index value (n) and the extinction coefficient value (k) into the growth rate calculation module 204. The growth rate calculation module 204 calculates the epitaxial growth rate based on the refractive index value (n) and the extinction coefficient value (k). The process of the growth rate calculation module 204 calculating the epitaxial growth rate is as follows: The growth rate calculation module 204 calculates the epitaxial growth rate in real time according to the refractive index value (n) and the extinction coefficient value (k) during the epitaxial process according to the trained model.

[0040] The growth rate calculation module 204 calculates the absorption coefficient and reflection coefficient of the epitaxial growth component according to the refractive index value (n) and the extinction coefficient value (k), infers the electronic structure and band structure of the epitaxial growth component according to the absorption coefficient and reflection coefficient, infers the growth dynamics of the epitaxial growth component according to the electronic structure and band structure, and calculates the epitaxial growth rate according to the growth dynamics.

[0041] In some embodiments of the present application, the present application embodiment discloses an electronic device (not shown), which is used for training local data of a neural network model, and the electronic device may include: one or more processors (CPU); memory; display; one or more graphics processing units (GPU); and a system running on the electronic device. The system is stored in the memory, and the system is configured to run in one or more processors (CPU) and one or more graphics processing units (GPU), and the memory stores local data, and the local data is configured to one or more processors (CPU) and one or more graphics processing units (GPU) for training.

[0042] The above is only a specific implementation of the embodiment of the present application, but the protection scope of the embodiment of the present application is not limited thereto, and any changes or replacements within the technical scope disclosed in the embodiment of the present application should be included in the protection scope of the embodiment of the present application. Therefore, the protection scope of the embodiment of the present application should be based on the protection scope of the claims.

Claims

1. A method for calculating epitaxial growth rate based on a neural network, characterized in that: include: Collect reflectivity data during wafer epitaxial growth; Preprocessing the reflectivity data to obtain input data; Inputting the input data into a neural network model, the neural network model performs calculations to obtain a refractive index value (n) and an extinction coefficient value (k) of the epitaxial growth component; The wafer epitaxial growth rate is calculated according to the obtained refractive index value (n) and the extinction coefficient value (k).

2. The method according to claim 1, characterized in that: Inputting the input data into a neural network model comprises: The neural network model includes a network architecture design submodule and a model training submodule; The network architecture design submodule is used to establish a nonlinear mapping relationship between the reflectivity data and the refractive index value (n) and the extinction coefficient value (k), wherein the mapping relationship is F(x), wherein the input data is R, and the mapping relationship about R is expressed as F(R)=(n, k); The model training submodule is used for training local data. The reflectivity data, the refractive index value (n) and the extinction coefficient value (k) actually detected in the local data are labeled and preprocessed to form a labeled data set. The labeled data set is formed into a training set and a verification set through a K-fold crossover method. The training set inputs the mapping relationship F(x). The model training submodule adjusts the mapping relationship F(x) according to the input training set, and the verification set inputs the mapping relationship F(x) for verification, and the adjusted mapping relationship F(x) is fed back to the network architecture design submodule.

3. The method according to claim 2, characterized in that: The network architecture design submodule includes an input layer, a hidden layer and an output layer, and the input layer, the hidden layer and the output layer all include neurons; A fully connected network model is formed between the input layer, the hidden layer and the output layer for data feature mapping; The hidden layer includes a hidden single layer, the hidden single layer is greater than or equal to 3 layers, the hidden single layer includes a plurality of neurons, the plurality of neurons are fully connected, and the neurons are used for extracting data features.

4. The method according to claim 3, characterized in that: The hidden single layers all use an activation function, and the activation function introduces a nonlinear factor to enhance the expression ability of the mapping relationship F(x).

5. The method according to claim 3, characterized in that: Inputting the reflectivity data of the validation set into the network architecture design submodule, and obtaining a plurality of predicted refractive index values ​​(n) and extinction coefficient values ​​(k) through the mapping relationship F(x); The network architecture design submodule includes a loss function, which is a root mean square error function, which quantifies the difference between the predicted value obtained by the mapping relationship F(x) and the actual detected value in the local data, and the loss function is used to optimize the mapping relationship F(x); Obtaining a plurality of predicted refractive index values ​​(n) and extinction coefficient values ​​(k) through the optimized mapping relationship F(x), and making a scatter plot of the plurality of predicted refractive index values ​​(n) and the extinction coefficient values ​​(k) for observing robustness; The root mean square error (RMS) of the multiple predicted refractive index values ​​(n) and the extinction coefficient values ​​(k) and the actually detected refractive index values ​​(n) and the extinction coefficient values ​​(k) in the local data are calculated through the RMS error function, and a RMS error calculation value is obtained. The stability and accuracy of the corresponding mapping relationship F(x) are reflected based on the RMS error calculation value and robustness.

6. The method according to claim 1, characterized in that The collecting of reflectivity data during the wafer epitaxial growth process includes: The reflectivity data monitored in real time during the epitaxial growth process are collected and made into a data set.

7. The method according to claim 6, characterized in that The preprocessing of the reflectivity data to obtain input data comprises: Performing a sliding average filter process on the data set made from the reflectivity data, wherein the window in the sliding average filter is used to remove abnormal values ​​and obviously erroneous data values, and to reduce noise on the data set made from the reflectivity data; The data set made from the de-noised reflectivity data is normalized to obtain input data, which is used to input into a neural network model.

8. The method according to claim 7, characterized in that: The window size is b×1, where b is the data length.

9. The method according to claim 7, characterized in that: The normalization process converts the de-noised reflectivity data into a distribution with a mean of 0 and a standard deviation of 1, and is used to convert the de-noised reflectivity data into a uniform scale to obtain input data.

10. A neural network-based epitaxial growth rate calculation system, characterized in that: include: A data acquisition module is used to collect reflectivity data during the epitaxial growth of the wafer; A data preprocessing module, which preprocesses the received reflectivity data to obtain input data; A neural network model module calculates the received input data to obtain a refractive index value (n) and an extinction coefficient value (k) of the epitaxial growth component; The growth rate calculation module calculates the wafer epitaxial growth rate according to the obtained refractive index value (n) and extinction coefficient value (k).

Images