Wafer Polishing Surface Profile Control Method, Its Model Training Method, and Related Devices

Abstract

The present application discloses a method for controlling the surface profile of a wafer and its model training method, as well as related devices. Among them, the wafer surface profile control model is a generative adversarial network model including a generator and a discriminator. The training method of the wafer surface profile control model includes: obtaining historical processing data; wherein, the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data; inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the generator to output predicted process control parameters; inputting the predicted process control parameters and the actual process control parameters into the discriminator to respectively output the true / false discrimination results for the predicted process control parameters and the actual process control parameters; and iteratively updating the parameters of the wafer surface profile control model based on the true / false discrimination results. By the above method, the control accuracy and production efficiency of the wafer surface profile can be improved.

Description

Technical Field

[0001] The present application relates to the field of semiconductor processing technologies, and particularly to a method for controlling the surface profile of a wafer during polishing, a method for training a model thereof, and related devices. Background Art

[0002] Wafer polishing is a key process in the semiconductor manufacturing process, aiming to remove minute impurities and unevenness on the wafer surface through mechanical polishing to ensure its flatness and smoothness. The surface profile of the wafer is an important indicator reflecting the surface morphology of the wafer. Among them, GBIR (TTV) refers to the total thickness variation of the silicon wafer surface profile, and SFQR refers to the local flatness of the silicon wafer, which can be used to judge the edge collapse situation after wafer polishing.

[0003] However, the polishing process is a time-varying and highly non-linear process. The product morphology is affected by various factors such as working conditions, types of raw materials and auxiliary materials. It is difficult for the solutions in the prior art to describe the relationship between various process parameters and surface profile changes through a mechanism model. And during the wafer production process, it is necessary to manually adjust the process parameters according to the predicted surface profile later. It is difficult to cope with the changes in different materials and different production batches of products, and has poor adaptability to environmental changes, seriously affecting the control accuracy of the wafer polishing surface profile and production efficiency. Summary of the Invention

[0004] The present application provides a method for controlling the surface profile of a wafer during polishing, a method for training a model thereof, and related devices to improve the control accuracy of the wafer polishing surface profile and production efficiency.

[0005] To solve the above technical problems, the present application provides a method for training a wafer polishing surface profile control model. The wafer polishing surface profile control model is a generative adversarial network model including a generator and a discriminator. The method includes: obtaining historical processing data; wherein the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data; inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the generator to output predicted process control parameters; inputting the predicted process control parameters and the actual process control parameters into the discriminator to respectively output true / false discrimination results for the predicted process control parameters and the actual process control parameters; and iteratively updating the parameters of the wafer polishing surface profile control model based on the true / false discrimination results.

[0006] Among them, the generator includes a connected convolutional neural network and a fully connected neural network, and the historical processing data further includes auxiliary material life data at different stages; the step of inputting the pre-polishing surface curve data and the desired surface curve data into the generator and outputting predicted process control parameters includes: inputting the pre-polishing surface curve data and the desired surface curve data into the convolutional neural network to extract polishing feature information; and inputting the polishing feature information and the auxiliary material life data into the fully connected neural network to map the polishing feature information and the auxiliary material life data to the predicted process control parameters.

[0007] Among them, the convolutional neural network includes a plurality of alternately connected convolutional layers and pooling layers; the convolutional layer is used for feature extraction of the input pre-polishing surface curve data and the desired surface curve data; the pooling layer is used for performing pooling operations on the features extracted by the convolutional layer;

[0008] And / or, the fully connected neural network includes a plurality of first fully connected layers connected in sequence; the first fully connected layer is used for performing non-linear mapping on the polishing feature information extracted by the convolutional neural network and finally outputting the predicted process control parameters.

[0009] Among them, the discriminator includes a deep neural network, and the deep neural network includes a plurality of second fully connected layers connected in sequence; each neuron in each layer of the second fully connected layer is connected to all neurons in the previous layer of the second fully connected layer, and the number of neurons in adjacent second fully connected layers gradually decreases; the second fully connected layer is used for performing non-linear mapping on the input predicted process control parameters and the actual process control parameters and finally outputting the true / false discrimination result.

[0010] Among them, the step of iteratively updating the parameters of the wafer polishing surface control model based on the true / false discrimination result includes: determining the loss function of the generator according to the true / false discrimination result of the discriminator on the predicted process control parameters; and determining the loss function of the discriminator according to the true / false discrimination result of the discriminator on the predicted process control parameters and the actual process control parameters; alternately optimizing the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator.

[0011] Among them, the step of alternately optimizing the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator includes: in the Nth round of training process, fixing the network parameters of the generator, minimizing the loss function of the discriminator, and updating the network parameters of the discriminator through gradient descent; in the (N + 1)th round of training process, fixing the network parameters of the discriminator, minimizing the loss function of the generator, and updating the network parameters of the generator through gradient descent.

[0012] To solve the above technical problems, the present application also provides a method for controlling the surface profile of a wafer polishing, which is applied to the process control system of a wafer polishing equipment. The process control system is integrated with a generator of a surface profile control model of the wafer polishing, and the surface profile control model of the wafer polishing is trained by the training method of the surface profile control model of the wafer polishing according to any one of the above; the method for controlling the surface profile of the wafer polishing includes: when polishing and processing the current batch of wafers, obtaining the processing data corresponding to the current batch of wafers; among them, the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers; inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator, and outputting the predicted process control parameters of the current batch of wafers; generating and sending an optimized control instruction to the process control system based on the predicted process control parameters of the current batch of wafers; and the process control system adjusts the processing parameters of the wafer polishing equipment according to the optimized control instruction to polish and process the current batch of wafers.

[0013] Among them, the generator includes a connected convolutional neural network and a fully connected neural network, and the processing data corresponding to the current batch of wafers further includes the auxiliary material life data corresponding to the current batch of wafers; the step of inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator and outputting the predicted process control parameters of the current batch of wafers includes: inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the convolutional neural network to extract polishing feature information; and inputting the polishing feature information and the auxiliary material life data corresponding to the current batch of wafers into the fully connected neural network to map the polishing feature information and the auxiliary material life data corresponding to the current batch of wafers to the predicted process control parameters of the current batch of wafers.

[0014] Among them, the wafer polishing surface profile control method further includes: before polishing the next batch of wafers, obtaining the historical processing data corresponding to the previous batch of wafers; updating the parameters of the wafer polishing surface profile control model according to the historical processing data corresponding to the previous batch of wafers; and the process control system calling the updated generator to control the polishing process of the next batch of wafers.

[0015] To solve the above technical problems, the present application also provides a training device for a wafer polishing surface profile control model. The wafer polishing surface profile control model is a generative adversarial network model including a generator and a discriminator. The training device includes: a first acquisition module for acquiring historical processing data, where the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data; a first processing module for inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the generator to output predicted process control parameters, inputting the predicted process control parameters and the actual process control parameters into the discriminator, and respectively outputting true / false discrimination results for the predicted process control parameters and the actual process control parameters; and a first update module for iteratively updating the parameters of the wafer polishing surface profile control model based on the true / false discrimination results.

[0016] To solve the above technical problems, the present application also provides a wafer polishing surface profile control device for controlling a wafer polishing device to polish wafers. The wafer polishing surface profile control device integrates a generator of a wafer polishing surface profile control model, and the wafer polishing surface profile control model is trained by the training method of the wafer polishing surface profile control model described in any one of the above. The wafer polishing surface profile control device includes: a second acquisition module for acquiring the processing data corresponding to the current batch of wafers when polishing the current batch of wafers, where the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers; a second processing module for inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator to output the predicted process control parameters of the current batch of wafers, generating an optimized control instruction based on the predicted process control parameters of the current batch of wafers; and a second update module for adjusting the processing parameters of the wafer polishing device according to the optimized control instruction to polish the current batch of wafers.

[0017] To solve the above technical problems, the present application also provides an electronic device, which includes: a memory and a processor coupled to each other, and the processor is configured to execute program instructions stored in the memory to implement the training method of the wafer polishing surface profile control model described in any one of the above, and / or the wafer polishing surface profile control method.

[0018] To solve the above technical problems, the present application also provides a computer-readable storage medium storing program instructions that can be executed to implement the training method of the wafer polishing surface profile control model described in any one of the above, and / or the wafer polishing surface profile control method.

[0019] The beneficial effects of the present application are as follows: Different from the prior art, the wafer polishing surface profile control model of the present application is a generative adversarial network model including a generator and a discriminator. During the training process of the wafer polishing surface profile control model, historical processing data is first obtained, where the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data; then the pre-polishing surface profile curve data and the desired surface profile curve data are input into the generator to output predicted process control parameters; then the predicted process control parameters and the actual process control parameters are input into the discriminator to respectively output the true / false discrimination results for the predicted process control parameters and the actual process control parameters; thus, the parameters of the wafer polishing surface profile control model can be iteratively updated based on the true / false discrimination results. By building a generative adversarial network model including a generator and a discriminator as the wafer polishing surface profile control model and training the generative adversarial network model with historical processing data, since the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data, the generator will generate predicted process control parameters according to the pre-polishing surface profile curve data and the desired surface profile curve data at different stages, and the discriminator is used to discriminate the gap between the generated predicted process control parameters and the actual process control parameters at the corresponding stages. Through the adversarial training between the generator and the discriminator, the discriminator can maximize its classification accuracy for the actual process control parameters and the generated predicted process control parameters, while the generator can minimize the difference between the generated predicted process control parameters and the actual process control parameters; thus, it can be understood that after the wafer polishing surface profile control model is trained, when polishing any batch of wafers, the pre-polishing surface profile curve data, the desired surface profile curve data of this batch of wafers, and the generator of the wafer polishing surface profile control model can be used to generate the process control parameters corresponding to this batch of wafers, and after processing with these process control parameters, the wafers of this batch can obtain processing results close to their desired surface profile curve data. Therefore, during the wafer polishing process, by using the trained wafer polishing surface profile control model of the present application, it is possible to accurately adjust its process control parameters according to the surface profile data of any batch of wafers, ensure that the final surface shape of the wafers meets the desired target, and thus improve the accuracy and production efficiency of the polishing process. Description of the Drawings

[0020] Figure 1 is a schematic flowchart of the first embodiment of the training method of the wafer polishing surface profile control model provided by the present application;

[0021] Figure 2 is a schematic flowchart of the second embodiment of the training method of the wafer polishing surface profile control model provided by the present application;

[0022] Figure 3It is a schematic structural diagram of a wafer polishing surface shape control model in an application scenario of the present application;

[0023] Figure 4 is Figure 2 a schematic flowchart of an embodiment of step S25 in

[0024] Figure 5 a schematic diagram of a training method for a wafer polishing surface shape control model in an application scenario of the present application;

[0025] Figure 6 a schematic flowchart of the first embodiment of the wafer polishing surface shape control method provided by the present application;

[0026] Figure 7 a schematic flowchart of the second embodiment of the wafer polishing surface shape control method provided by the present application;

[0027] Figure 8 a schematic diagram of a wafer polishing surface shape control method in an application scenario of the present application;

[0028] Figure 9 a schematic structural diagram of an embodiment of a training device for a wafer polishing surface shape control model provided by the present application;

[0029] Figure 10 a schematic structural diagram of an embodiment of a wafer polishing surface shape control device provided by the present application;

[0030] Figure 11 a schematic structural diagram of an embodiment of an electronic device provided by the present application;

[0031] Figure 12 a schematic structural diagram of an embodiment of a computer-readable storage medium provided by the present application. Detailed implementation manners

[0032] The solutions of the embodiments of the present application will be described in detail below with reference to the accompanying drawings of the specification.

[0033] In the following description, specific details such as specific system architectures, interfaces, and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the present application.

[0034] The terms "system" and "network" are often used interchangeably herein. The term "and / or" herein merely describes an association relationship between associated objects and indicates that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after. In addition, "plurality" herein means two or more than two.

[0035] Please refer toFigure 1 , Figure 1 FIG. Figure 1 is a schematic flowchart of the first embodiment of the training method for the wafer polishing surface profile control model provided in this application. The wafer polishing surface profile control model in the embodiments of this application is a generative adversarial network model including a generator and a discriminator. The training method for the wafer polishing surface profile control model in this embodiment includes the following steps:

[0036] Step S11: Obtain historical processing data; where the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data.

[0037] The wafers in the embodiments of this application may include workpieces such as silicon wafers, sapphire wafers, and silicon carbide wafers. The wafer polishing surface profile control model can use the historical processing data collected from wafer polishing equipment as training samples. The historical processing data includes process control parameters at different stages during the wafer processing and wafer surface profile curve data. Among them, the process control parameters may include the rotation speed and feed speed of the polishing disc, the inclination angle of the polishing disc, the temperature parameter during grinding, the sensor thickness measurement parameter, the pressure parameter applied to the wafer, and the polishing time, etc. The process control parameters will all affect the surface profile of the wafer after polishing. The wafer surface profile curve data includes the pre-polishing surface profile curve data and the desired surface profile curve data of the wafer; among them, the pre-polishing surface profile curve data of the wafer refers to the surface shape of the wafer before polishing, expressed as a two-dimensional curve image, and the main information it contains is the change in the surface height or shape of the wafer, which can reflect the surface profile characteristics of the wafer before polishing; the desired surface profile curve data of the wafer refers to the target surface shape of the wafer, that is, the surface profile that the wafer is expected to achieve after ideal polishing treatment, expressed as a two-dimensional curve image.

[0038] Step S12: Input the pre-polishing surface profile curve data and the desired surface profile curve data into the generator, and output predicted process control parameters.

[0039] It can be understood that after inputting the pre-polishing surface profile curve data of the wafer and the corresponding desired surface profile curve data into the generator, the generator can process the pre-polishing surface profile curve data and the desired surface profile curve data simultaneously, and generate predicted process control parameters based on these inputs. That is, the generator infers that by using the predicted process control parameters it generates to polish the wafer with the pre-polishing surface profile curve data, the polished wafer can have the corresponding desired surface profile curve data.

[0040] Step S13: Input the predicted process control parameters and the actual process control parameters into the discriminator, and respectively output the true / false discrimination results for the predicted process control parameters and the actual process control parameters.

[0041] It can be understood that the discriminator needs to learn the features extracted from the real data and the generated data of the generator, and finally give a discrimination probability between the real data and the generated data. Therefore, after inputting the predicted process control parameters and the actual process control parameters into the discriminator, the predicted process control parameters are used as the generated data of the generator, and the discriminator can give a true / false discrimination result on whether the predicted process control parameters are real data. Similarly, the actual process control parameters are used as the real data, and the discriminator can also give a true / false discrimination result on whether the actual process control parameters are real data.

[0042] Step S14: Iteratively update the parameters of the wafer polishing surface profile control model based on the true / false discrimination result.

[0043] It can be understood that the predicted process control parameters generated by the generator of the generative adversarial network should try to confuse the discriminator as much as possible, while the discriminator should try to distinguish between the predicted process control parameters generated by the generator and the actual process control parameters. During the adversarial optimization training, it is necessary to continuously update and optimize the parameters of the wafer polishing surface profile control model according to the true / false discrimination result of the discriminator on whether the predicted process control parameters and the actual process control parameters are real data, so as to minimize the loss between the predicted value of the model and the true output value as much as possible. The smaller the loss value between the predicted value and the true output value, the closer the model prediction is to the true value.

[0044] In the above solution, a generative adversarial network model including a generator and a discriminator is built as the wafer polishing surface shape control model, and historical processing data is used to train the generative adversarial network model. Since the historical processing data at least includes the actual process control parameters at different stages, the pre-polishing surface shape curve data, and the desired surface shape curve data, the generator will generate predicted process control parameters based on the pre-polishing surface shape curve data and the desired surface shape curve data at different stages. The discriminator is used to determine the gap between the generated predicted process control parameters and the actual process control parameters at the corresponding stages. Through the adversarial training of the generator and the discriminator, the discriminator can maximize its classification accuracy for the actual process control parameters and the generated predicted process control parameters, while the generator can minimize the difference between the generated predicted process control parameters and the actual process control parameters. Therefore, it can be understood that after the wafer polishing surface shape control model is trained, when polishing any batch of wafers, the pre-polishing surface shape curve data, the desired surface shape curve data of this batch of wafers, and the generator of the wafer polishing surface shape control model can be used to generate the process control parameters corresponding to this batch of wafers. After processing with these process control parameters, the wafers of this batch can obtain processing results close to their desired surface shape curve data. The wafer polishing surface shape control model of the present application adopts an end-to-end design solution, which inputs the pre-polishing surface shape curve data and the desired surface shape curve data and outputs the process control parameters that can obtain the desired surface shape curve data, that is, directly obtains the control quantity from the current quantity and the desired quantity. Since the whole process is a unified model, it can better capture the non-linear relationship between the input and the output, thereby improving the accuracy of the model. Therefore, during the wafer polishing process, using the trained wafer polishing surface shape control model of the present application can achieve precise adjustment of its process control parameters according to the surface shape data of any batch of wafers, ensuring that the final surface shape of the wafers meets the desired target, thereby improving the accuracy and production efficiency of the polishing process.

[0045] Please refer to Figure 2 , Figure 2 which is a schematic flowchart of the second embodiment of the training method of the wafer polishing surface shape control model provided by the present application. The training method of the wafer polishing surface shape control model in this embodiment includes the following steps:

[0046] Step S21: Obtain historical processing data; wherein, the historical processing data at least includes the actual process control parameters at different stages, the pre-polishing surface shape curve data, the desired surface shape curve data, and the auxiliary material life data.

[0047] It is understandable that the service life of auxiliary materials such as polishing liquid and polishing pad also has a certain impact on the wafer polishing surface profile. Therefore, the difference from the previous embodiment is that the historical processing data in this embodiment further includes the service life data of auxiliary materials at different stages. The service life data of auxiliary materials is the life information of polishing liquid, polishing pad and other auxiliary materials, which contains time series data, such as the usage situation and wear life of each auxiliary material. The service life data of auxiliary materials is represented in the form of a one-dimensional vector.

[0048] In other embodiments, the historical processing data may further include randomly generated noise data having the same dimension as the surface profile curve. It is understandable that appropriately adding noise to the model can make the model more challenging and have better generalization ability.

[0049] Step S22: Input the pre-polishing surface profile curve data and the desired surface profile curve data into the convolutional neural network to extract polishing feature information.

[0050] Step S23: Input the polishing feature information and the service life data of auxiliary materials into the fully connected neural network to map the polishing feature information and the service life data of auxiliary materials to the predicted process control parameters.

[0051] Please combine Figure 3 , the generator of the wafer polishing surface profile control model in this embodiment includes a connected convolutional neural network and a fully connected neural network. The structure combining the convolutional neural network and the fully connected neural network can simultaneously process the pre-polishing surface profile curve data, the desired surface profile curve data and the service life data of auxiliary materials, and generate predicted process control parameters based on these inputs.

[0052] In one embodiment, the convolutional neural network includes a plurality of alternately connected convolutional layers and pooling layers; the convolutional layers are used to extract features from the input pre-polishing surface profile curve data and the desired surface profile curve data; the pooling layers are used to perform pooling operations on the features extracted by the convolutional layers.

[0053] Specifically, the convolutional layer extracts features from the input image data (i.e., the pre-polishing surface profile curve data and the desired surface profile curve data in this application) through a sliding window. Each convolutional layer performs a convolution operation on the input image data through a set of convolutional kernels (filters) to extract local features. Let the input image data of the l-th layer be , the convolutional kernel be , and the bias term be , then the convolution operation can be expressed as:

[0054] .

[0055] Among them, * represents the convolution operation, ReLU is the activation function, and the form of the ReLU function is .

[0056] The pooling layer is used for dimensionality reduction and ensuring the smoothness of extracting local features. In the embodiments of the present application, pooling methods such as max pooling and average pooling can be adopted. For example, the pooling operation can slide a window and take the maximum or average value of each local area to reduce the size of the extracted feature map. The pooling operation can be expressed as:

[0057] .

[0058] It can be understood that after multiple convolution and pooling operations on the polished front surface curve data and the expected surface curve data, high-dimensional polished feature information can be extracted. As Figure 3 shown, between the convolutional neural network and the fully connected neural network, a flattening layer can also be set as a transition layer. The flattening layer can flatten the high-dimensional features extracted by the convolutional neural network into one-dimensional data, that is, obtain one-dimensional polished feature information, and then input the one-dimensional polished feature information together with the auxiliary material life data represented in the form of a one-dimensional vector into the fully connected neural network. The fully connected neural network maps the polished feature information extracted by the convolutional neural network and the directly input auxiliary material life data together to the generated predicted process control parameters.

[0059] In one embodiment, the fully connected neural network includes a plurality of first fully connected layers connected in sequence; the first fully connected layer is used for performing non-linear mapping on the polished feature information extracted by the convolutional neural network and finally outputting the predicted process control parameters.

[0060] The fully connected neural network performs non-linear mapping on the high-dimensional polished feature information extracted by the convolutional neural network through the first fully connected layer, and finally can output the generated predicted process control parameters. As Figure 3 shown, each first fully connected layer of the fully connected neural network is composed of a plurality of neurons, and each neuron is connected to all neurons of the previous layer. Assume that the input of the k-th first fully connected layer is , the weight is , the bias is , then the output of the k-th layer is calculated by the formula:

[0061] .

[0062] The final output layer of the fully connected neural network is also a first fully connected layer, which is used for generating the predicted process control parameters. Assume that the final output is , representing the generated predicted process control parameters:

[0063] .

[0064] Among them, is the weight of the output layer, is the output of the previous layer.

[0065] Step S24: Input the predicted process control parameter and the actual process control parameter into the discriminator, and respectively output the true / false discrimination results for the predicted process control parameter and the actual process control parameter.

[0066] Please combine Figure 3 , in one embodiment, the discriminator includes a deep neural network, and the deep neural network includes a plurality of second fully connected layers connected in sequence; each neuron of each layer of the second fully connected layer is connected to all neurons of the previous layer of the second fully connected layer, and the number of neurons in adjacent second fully connected layers gradually decreases; the second fully connected layer is used to perform a non-linear mapping on the input predicted process control parameter and actual process control parameter, and finally output the true / false discrimination result.

[0067] It can be understood that the deep neural network is the core part of the discriminator, and each neuron of each layer of the second fully connected layer in the deep neural network is connected to all neurons of the previous layer. The discriminator gradually extracts the high-level features of the input data through a plurality of second fully connected layers and finally performs true / false classification. Similar to the convolutional neural network in the foregoing generator, let the input of the l-th layer of the deep neural network be , the weight be , the bias term be , then the output of the l-th layer is calculated by the following formula:

[0068] .

[0069] Among them, the ReLU activation function is used to introduce non-linear transformation to enhance the expression ability of the network.

[0070] Therefore, the discriminator includes a plurality of second fully connected layers. After each second fully connected layer, the ReLU activation function will be applied to further capture the non-linear relationship in the input data. The number of neurons in each second fully connected layer will gradually decrease until the last layer outputs a single probability value. Assume that the discriminator has L layers, and the output of the last layer is , and the value of the output layer obtained through linear transformation is:

[0071] .

[0072] The output layer of the discriminator uses the sigmoid activation function to map the final output to the probability space, representing the probability value that the input data is real data. Assume that the final output is , then the output formula is:

[0073] 。

[0074] Among them, is the sigmoid activation function, which compresses the output value into the interval [0, 1], indicating the probability that the data is real.

[0075] Step S25: Iteratively update the parameters of the wafer polishing surface profile control model based on the true / false discrimination result.

[0076] Please combine Figure 4 , Figure 4 is Figure 2 a schematic flowchart of an embodiment of step S25 in

[0077] Step S251: Determine the loss function of the generator according to the true / false discrimination result of the discriminator for the predicted process control parameters; and determine the loss function of the discriminator according to the true / false discrimination result of the discriminator for the predicted process control parameters and the actual process control parameters.

[0078] It can be understood that the core of the generative adversarial network model lies in its loss function, which usually includes two parts: the generator loss function and the discriminator loss function. Let the predicted process control parameters generated by the generator be G(z), the actual process control parameters be x, and the outputs of the discriminator be D(x) and D(G(z)), where D(x) represents the output probability of the discriminator for judging the actual process control parameters, and D(G(z)) represents the output probability of the discriminator for judging the predicted process control parameters generated by the generator. Among them, the generator loss function LG can be expressed as:

[0079] 。

[0080] The optimization goal of the discriminator is to maximize its classification accuracy for the actual process control parameters and the predicted process control parameters. For this purpose, the discriminator can use the binary cross-entropy loss function to calculate the loss, and the formula of the discriminator's loss function is as follows:

[0081] 。

[0082] Among them, pdata represents the real data distribution, and pz is the noise distribution input to the generator.

[0083] Step S252: Alternately optimize the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator.

[0084] In this application, the training of the generative adversarial network model is a process of adversarial training between the generator and the discriminator. Its optimization goal is to optimize the loss functions of the generator and the discriminator through gradient descent. During the training process, the generator and the discriminator continuously adjust their network parameters through alternating optimization.

[0085] In one embodiment, during the Nth round of training, the network parameters of the generator are fixed, the loss function of the discriminator is minimized, and the network parameters of the discriminator are updated through gradient descent; during the (N + 1)th round of training, the network parameters of the discriminator are fixed, the loss function of the generator is minimized, and the network parameters of the generator are updated through gradient descent.

[0086] Please refer to Figure 5 , Figure 5 which is a schematic diagram of the training method of the wafer polishing surface shape control model in an application scenario of this application. As shown in the figure, during one round of training, after the historical processing data is input into the generator, the generator can generate predicted process control parameters, and then the predicted process control parameters and the actual process control parameters are input into the discriminator for training. When training the discriminator, it is necessary to fix the parameters of the generator and use the backpropagation algorithm to update the parameters of the discriminator to minimize the loss function of the discriminator; during the next round of training, it is necessary to fix the parameters of the discriminator and also use the backpropagation algorithm to update the parameters of the generator to minimize the loss function of the generator. During the optimization process of the discriminator, it is necessary to minimize the loss function LD of the discriminator and update the parameters of the discriminator through gradient descent: . During the optimization process of the generator, it is necessary to minimize the loss function LG of the generator and update the parameters of the generator through gradient descent: . Through the adversarial training between the generator and the discriminator, the predicted process control parameters generated by the generator can enable the wafer to obtain a processing result close to its desired surface shape curve data.

[0087] It can be understood that the training goal of the wafer polishing surface shape control model is to minimize the difference between the generated predicted process control parameters and the actual process control parameters. It can be measured by the following loss function:

[0088] .

[0089] Among them, is the generated predicted process control parameter, ytarget is the real data, and the loss function Lfinal evaluates the difference between the generated result and the desired result by minimizing the Euclidean distance.

[0090] Please refer to Figure 6 , Figure 6It is a schematic flowchart of the first embodiment of the wafer polishing surface profile control method provided by this application. The wafer polishing surface profile control method in the embodiments of this application is applied to the process control system of a wafer polishing device. The process control system is integrated with a generator for generating a wafer polishing surface profile control model, and the wafer polishing surface profile control model is obtained by training through the wafer polishing surface profile control model training method in any of the above embodiments. Specifically, the wafer polishing surface profile control method in this embodiment includes the following steps:

[0091] Step S61: When polishing the current batch of wafers, obtain the processing data corresponding to the current batch of wafers; wherein, the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers.

[0092] Step S62: Input the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator, and output the predicted process control parameters of the current batch of wafers.

[0093] Step S63: Based on the predicted process control parameters of the current batch of wafers, generate and send an optimized control instruction to the process control system.

[0094] Step S64: The process control system adjusts the processing parameters of the wafer polishing device according to the optimized control instruction to polish the current batch of wafers.

[0095] It can be understood that the wafer polishing surface profile control model after training can be used for the control of the wafer polishing surface profile. Among them, the generator is integrated into the process control system of the wafer polishing device. At this time, the generator no longer simply relies on historical processing data, but can generate dynamically optimized control instructions according to the latest input processing data corresponding to the current batch of wafers. The generator can receive input signals from the process control system in real time. The input signals of the process control system contain the processing data corresponding to the current batch of wafers, such as the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers. Then the generator can output the predicted process control parameters of the current batch of wafers in real time according to these signals, generate optimized control instructions and send them to the process control system. For example, the optimized control instruction can be to adjust the parameters of the wafer polishing device to the predicted process control parameters of the current batch of wafers in real time. The process control system can then adjust the processing parameters of the wafer polishing device according to the optimized control instruction, and polish the current batch of wafers using the predicted process control parameters to make the surface profile of the current batch of wafers approach the desired surface profile curve data and meet the desired target.

[0096] Therefore, deploying the wafer polishing surface profile control model after training in the process control system of the wafer polishing equipment can achieve real-time control of the polishing process control parameters.

[0097] Please refer to Figure 7 , Figure 7 which is a schematic flowchart of the second embodiment of the wafer polishing surface profile control method provided by this application. The wafer polishing surface profile control method in this embodiment includes the following steps:

[0098] Step S71: When polishing the current batch of wafers, obtain the processing data corresponding to the current batch of wafers; wherein, the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data, the desired surface profile curve data, and the consumable life data of the current batch of wafers.

[0099] It can be understood that the life of consumables such as polishing liquid and polishing pad also has a certain impact on the wafer polishing surface profile. Therefore, the difference from the previous embodiment is that the processing data corresponding to the current batch of wafers in this embodiment further includes the consumable life data corresponding to the current batch of wafers.

[0100] Step S72: Input the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the convolutional neural network to extract polishing feature information.

[0101] Step S73: Input the polishing feature information and the consumable life data corresponding to the current batch of wafers into the fully connected neural network to map the polishing feature information and the consumable life data corresponding to the current batch of wafers to the predicted process control parameters of the current batch of wafers.

[0102] The generator of the wafer polishing surface profile control model in this embodiment includes a connected convolutional neural network and a fully connected neural network. The combined structure of the convolutional neural network and the fully connected neural network can simultaneously process the pre-polishing surface profile curve data, the desired surface profile curve data, and the consumable life data corresponding to the current batch of wafers, and generate the predicted process control parameters of the current batch of wafers based on these inputs.

[0103] Step S74: Based on the predicted process control parameters of the current batch of wafers, generate and send an optimized control instruction to the process control system.

[0104] Step S75: The process control system adjusts the processing parameters of the wafer polishing equipment according to the optimized control instruction to polish the current batch of wafers.

[0105] Steps S74 and S75 in this embodiment are basically the same as steps S63 and S64 in the above embodiment, and will not be elaborated here.

[0106] In one embodiment, the above wafer polishing surface profile control method further includes the following steps:

[0107] Step S76: Before polishing the next batch of wafers, obtain the historical processing data corresponding to the previous batch of wafers.

[0108] Step S77: Update the parameters of the wafer polishing surface profile control model according to the historical processing data corresponding to the previous batch of wafers.

[0109] Step S78: The process control system calls the generator with updated parameters to control the polishing of the next batch of wafers.

[0110] Please refer to Figure 8 , Figure 8 FIG. is a schematic diagram of a wafer polishing surface profile control method in an application scenario of the present application. When the trained wafer polishing surface profile control model is deployed in the process control system of the wafer polishing equipment, before each polishing, the wafer polishing surface profile control model can learn the processing data of the previous polishing; for example, before the (K + 1)-th polishing, the wafer polishing surface profile control model can learn the latest processing data of the K-th time, and update the parameters of the wafer polishing surface profile control model; then, when the (K + 1)-th polishing is performed, the process control system calls the generator with updated parameters to generate the predicted process control parameters of the wafer corresponding to the (K + 1)-th polishing, generate and send the optimized control instruction to the process control system, so as to ensure that the generated optimized control instruction can adapt to the continuous change of the production environment and ensure the accuracy of the process control parameters of the (K + 1)-th polishing. This closed-loop feedback mechanism enables the polishing control strategy of each production process to be continuously optimized, and realizes the update of efficient and precise control strategies in a dynamic and complex production environment.

[0111] In the embodiment of the present application, in order to ensure the continuous optimization of the generator, the feedback of the discriminator needs to be transmitted to the generator in a timely manner for gradient update to form an R2R control, that is, to adjust and optimize the future control strategy through historical and real-time feedback information. The generator not only depends on static process control parameters, but can flexibly adjust the control signal by learning historical and real-time data to adapt to the fluctuations of the production environment. This method can achieve precise regulation of the production process and improve the consistency and stability of product quality.

[0112] Please refer to Figure 9 , Figure 9It is a schematic structural diagram of an embodiment of a training device for a wafer polishing surface profile control model provided by this application. The wafer polishing surface profile control model is a generative adversarial network model including a generator and a discriminator. The training device 90 of the wafer polishing surface profile control model in this embodiment includes a first acquisition module 900, a first processing module 902, and a first update module 904 that are interconnected. The first acquisition module 900 is used to acquire historical processing data; wherein, the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data. The first processing module 902 is used to input the pre-polishing surface profile curve data and the desired surface profile curve data into the generator, and output predicted process control parameters; input the predicted process control parameters and the actual process control parameters into the discriminator, and respectively output true / false discrimination results for the predicted process control parameters and the actual process control parameters. The first update module 904 is used to iteratively update the parameters of the wafer polishing surface profile control model based on the true / false discrimination results.

[0113] In one embodiment, the generator includes a connected convolutional neural network and a fully connected neural network, and the historical processing data further includes auxiliary material life data at different stages. The step that the first processing module 902 executes to input the pre-polishing surface profile curve data and the desired surface profile curve data into the generator and output predicted process control parameters includes: inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the convolutional neural network to extract polishing feature information; inputting the polishing feature information and the auxiliary material life data into the fully connected neural network to map the polishing feature information and the auxiliary material life data to the predicted process control parameters.

[0114] In one embodiment, the step that the first update module 904 executes to iteratively update the parameters of the wafer polishing surface profile control model based on the true / false discrimination results includes: determining the loss function of the generator according to the true / false discrimination result of the discriminator for the predicted process control parameters; and determining the loss function of the discriminator according to the true / false discrimination results of the discriminator for the predicted process control parameters and the actual process control parameters; alternately optimizing the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator.

[0115] In one embodiment, the first update module 904 performs the step of alternately optimizing the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator, which specifically includes: in the Nth round of training process, fixing the network parameters of the generator, minimizing the loss function of the discriminator, and updating the network parameters of the discriminator by gradient descent; in the (N + 1)th round of training process, fixing the network parameters of the discriminator, minimizing the loss function of the generator, and updating the network parameters of the generator by gradient descent.

[0116] Please refer to Figure 10 , Figure 10 FIG. is a schematic structural diagram of an embodiment of a wafer polishing surface profile control device provided by the present application. The wafer polishing surface profile control device 100 in this embodiment is used to control a wafer polishing device to polish a wafer. The wafer polishing surface profile control device 100 is integrated with a generator of a wafer polishing surface profile control model, and the wafer polishing surface profile control model is obtained by training with the training method of the wafer polishing surface profile control model in any of the above embodiments.

[0117] The wafer polishing surface profile control device 100 includes a second acquisition module 1000, a second processing module 1002, and a second update module 1004 that are connected to each other. The second acquisition module 1000 is configured to acquire processing data corresponding to the current batch of wafers when polishing the current batch of wafers; wherein, the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers. The second processing module 1002 is configured to input the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator, output the predicted process control parameters of the current batch of wafers, and generate an optimized control instruction based on the predicted process control parameters of the current batch of wafers. The second update module 1004 is configured to adjust the processing parameters of the wafer polishing device according to the optimized control instruction to polish the current batch of wafers.

[0118] In one embodiment, the generator includes a connected convolutional neural network and a fully connected neural network, and the processing data corresponding to the current batch of wafers further includes the auxiliary material life data corresponding to the current batch of wafers. The second processing module 1002 executes the step of inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator and outputting the predicted process control parameters of the current batch of wafers, including: inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the convolutional neural network to extract polishing feature information; inputting the polishing feature information and the auxiliary material life data corresponding to the current batch of wafers into the fully connected neural network to map the polishing feature information and the auxiliary material life data corresponding to the current batch of wafers to the predicted process control parameters of the current batch of wafers.

[0119] In one embodiment, the second processing module 1002 is further configured to obtain the historical processing data corresponding to the previous batch of wafers before polishing the next batch of wafers; update the parameters of the wafer polishing surface profile control model according to the historical processing data corresponding to the previous batch of wafers; and the process control system calls the generator with updated parameters to control the polishing process of the next batch of wafers.

[0120] Please refer to Figure 11 , Figure 11 FIG. is a schematic structural diagram of an embodiment of the electronic device of the present application. The electronic device 110 in this embodiment includes a processor 1102 and a memory 1101 connected to each other; the memory 1101 is used to store program instructions, and the processor 1102 is used to execute the program instructions stored in the memory 1101 to implement the steps of any of the above embodiments of the training method of the wafer polishing surface profile control model, and / or, the steps of any of the above embodiments of the wafer polishing control method. In a specific implementation scenario, the electronic device 110 may include, but is not limited to: a microcomputer, a server.

[0121] Specifically, the processor 1102 is used to control itself and the memory 1101 to implement the steps of any of the above embodiments of the training method of the wafer polishing surface type control model, and / or, the steps of any of the above embodiments of the wafer polishing surface type control method. The processor 1102 may also be referred to as a CPU (Central Processing Unit). The processor 1102 may be an integrated circuit chip with signal processing capabilities. The processor 1102 may also be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. Additionally, the processor 1102 may be implemented jointly by integrated circuit chips.

[0122] Please refer to Figure 12 , Figure 12 FIG. is a schematic structural diagram of an embodiment of the computer-readable storage medium of the present application. The computer-readable storage medium 120 of the present application stores program instructions 1200 thereon, and when the program instructions 1200 are executed by a processor, the steps of any of the above embodiments of the training method of the wafer polishing surface type control model, and / or, the steps of any of the above embodiments of the wafer polishing surface type control method are implemented.

[0123] The computer-readable storage medium 120 may specifically be a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc, etc., which can store the program instructions 1200, or it may also be a server storing the program instructions 1200. The server can send the stored program instructions 1200 to other devices for operation, or it can also run the stored program instructions 1200 by itself.

[0124] In several embodiments provided by the present application, it should be understood that the disclosed methods, devices, and apparatuses can be implemented in other ways. For example, the device and apparatus embodiments described above are merely illustrative. For example, the division of modules or units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0126] In addition, each functional unit in various embodiments of the present application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0127] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods in various embodiments of the present application. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs and other various media that can store program codes.

[0128] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A training method for a wafer polishing surface profile control model, characterized in that The wafer polishing surface profile control model is a generative adversarial network model including a generator and a discriminator. The method includes: Obtain historical processing data; wherein, the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data; Input the pre-polishing surface profile curve data and the desired surface profile curve data into the generator, and output predicted process control parameters; Input the predicted process control parameters and the actual process control parameters into the discriminator, and respectively output true / false discrimination results for the predicted process control parameters and the actual process control parameters; Iteratively update the parameters of the wafer polishing surface profile control model based on the true / false discrimination results; Wherein, the generator includes a connected convolutional neural network and a fully connected neural network, and the historical processing data further includes auxiliary material life data at different stages; The step of inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the generator and outputting predicted process control parameters includes: Input the pre-polishing surface profile curve data and the desired surface profile curve data into the convolutional neural network to extract polishing feature information; Input the polishing feature information and the auxiliary material life data into the fully connected neural network to map the polishing feature information and the auxiliary material life data to the predicted process control parameters.

2. The training method of the wafer polishing surface type control model according to claim 1, wherein The convolutional neural network includes a plurality of convolutional layers and pooling layers connected alternately; the convolutional layer is used to extract features from the input pre-polishing surface profile curve data and the desired surface profile curve data; the pooling layer is used to perform pooling operations on the features extracted by the convolutional layer; And / or, the fully connected neural network includes a plurality of first fully connected layers connected in sequence; the first fully connected layer is used to perform non-linear mapping on the polishing feature information extracted by the convolutional neural network and finally output the predicted process control parameters.

3. The training method of the wafer polishing surface type control model according to claim 1, characterized in that The discriminator includes a deep neural network, and the deep neural network includes a plurality of second fully connected layers connected in sequence; each neuron in each layer of the second fully connected layer is connected to all neurons in the previous layer of the second fully connected layer, and the number of neurons in adjacent second fully connected layers gradually decreases; the second fully connected layer is used to perform non-linear mapping on the input predicted process control parameters and the actual process control parameters and finally output the true / false discrimination results.

4. The training method of the wafer polishing surface type control model according to claim 1, wherein, The step of iteratively updating the parameters of the wafer polishing surface profile control model based on the true / false discrimination results includes: Determine the loss function of the generator according to the true / false discrimination result of the discriminator for the predicted process control parameters; and determine the loss function of the discriminator according to the true / false discrimination results of the discriminator for the predicted process control parameters and the actual process control parameters; Alternately optimize the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator.

5. The training method of the wafer polishing surface profile control model according to claim 4, characterized in that, The step of alternately optimizing the network parameters of the generator and the discriminator according to the loss function of the generator and the loss function of the discriminator includes: In the Nth round of training, fix the network parameters of the generator, minimize the loss function of the discriminator, and update the network parameters of the discriminator by gradient descent; In the (N + 1)th round of training, fix the network parameters of the discriminator, minimize the loss function of the generator, and update the network parameters of the generator by gradient descent.

6. A method for controlling the surface profile of a wafer, characterized in that, A process control system applied to a wafer polishing device, the process control system integrating a generator of a wafer polishing surface profile control model, and the wafer polishing surface profile control model is trained by the training method of the wafer polishing surface profile control model according to any one of claims 1 to 5; The wafer polishing surface profile control method includes: When polishing the current batch of wafers, obtain the processing data corresponding to the current batch of wafers; wherein, the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers; Input the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator, and output the predicted process control parameters of the current batch of wafers; Based on the predicted process control parameters of the current batch of wafers, generate and send an optimized control instruction to the process control system; The process control system adjusts the processing parameters of the wafer polishing device according to the optimized control instruction to polish the current batch of wafers.

7. The wafer polishing surface type control method according to claim 6, characterized in that The generator includes a connected convolutional neural network and a fully connected neural network, and the processing data corresponding to the current batch of wafers further includes the auxiliary material life data corresponding to the current batch of wafers; The step of inputting the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator and outputting the predicted process control parameters of the current batch of wafers includes: Input the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the convolutional neural network to extract polishing feature information; Input the polishing feature information and the auxiliary material life data corresponding to the current batch of wafers into the fully connected neural network to map the polishing feature information and the auxiliary material life data corresponding to the current batch of wafers to the predicted process control parameters of the current batch of wafers.

8. The wafer polishing surface profile control method according to claim 6 or 7, characterized in that The wafer polishing surface profile control method further includes: Before polishing the next batch of wafers, obtain the historical processing data corresponding to the previous batch of wafers; Update the parameters of the wafer polishing surface profile control model according to the historical processing data corresponding to the previous batch of wafers; The process control system calls the generator with updated parameters to control the polishing of the next batch of wafers.

9. A training device for a wafer polishing surface shape control model, characterized in that, The wafer polishing surface profile control model is a generative adversarial network model including a generator and a discriminator, and the training device includes: The first acquisition module is configured to acquire historical processing data, where the historical processing data at least includes actual process control parameters at different stages, pre-polishing surface profile curve data, and desired surface profile curve data. The first processing module is configured to input the pre-polishing surface profile curve data and the desired surface profile curve data into the generator, and output predicted process control parameters; input the predicted process control parameters and the actual process control parameters into the discriminator, and respectively output true / false discrimination results for the predicted process control parameters and the actual process control parameters. The first update module is configured to iteratively update the parameters of the wafer polishing surface profile control model based on the true / false discrimination results. The generator includes a connected convolutional neural network and a fully connected neural network, and the historical processing data further includes auxiliary material life data at different stages. When the first processing module executes the step of inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the generator and outputting predicted process control parameters, it includes: inputting the pre-polishing surface profile curve data and the desired surface profile curve data into the convolutional neural network to extract polishing feature information; inputting the polishing feature information and the auxiliary material life data into the fully connected neural network to map the polishing feature information and the auxiliary material life data to the predicted process control parameters.

10. A wafer polishing surface profile control device, characterized in that, The wafer polishing surface profile control device is used to control a wafer polishing device to polish a wafer. The wafer polishing surface profile control device integrates a generator of a wafer polishing surface profile control model, and the wafer polishing surface profile control model is trained by the training method of the wafer polishing surface profile control model according to any one of claims 1 to 5. The wafer polishing surface profile control device includes: The second acquisition module is configured to acquire the processing data corresponding to the current batch of wafers when polishing the current batch of wafers, where the processing data corresponding to the current batch of wafers at least includes the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers. The second processing module is configured to input the pre-polishing surface profile curve data and the desired surface profile curve data of the current batch of wafers into the generator, and output the predicted process control parameters of the current batch of wafers; generate an optimized control instruction based on the predicted process control parameters of the current batch of wafers. The second update module is configured to adjust the processing parameters of the wafer polishing device according to the optimized control instruction to polish the current batch of wafers.

11. An electronic device, characterized in that, The electronic device includes a memory and a processor coupled to each other. The processor is configured to execute program instructions stored in the memory to implement the training method of the wafer polishing surface profile control model according to any one of claims 1 to 5, and / or the wafer polishing surface profile control method according to any one of claims 6 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that can be executed to implement the training method of the wafer polishing surface profile control model according to any one of claims 1 to 5, and / or the wafer polishing surface profile control method according to any one of claims 6 to 8.

Images