A critical dimension control method and system

Abstract

The disclosure provides a critical dimension control method and system, belonging to the technical field of semiconductor production and manufacturing, wherein the control method comprises: obtaining historical data; establishing and training a neural network prediction model through the historical data; obtaining initial previous layer factor data, inputting the initial previous layer factor data as an input variable into the trained neural network prediction model for prediction, and outputting predicted current layer measurement data; and according to a preset condition met by the predicted current layer measurement data, performing correction in a previous layer process to obtain target previous layer factor data. The disclosure realizes prediction and correction of the critical dimension of the current layer in the previous layer, thereby improving the problem of deviation of the critical dimension test result of the current layer caused by the existing adjustment lag and even scrapping. The disclosure separately trains and optimizes the data outside the standard and the data within the standard to establish a correction model, thereby avoiding interference.

Description

Technical Field

[0001] This disclosure pertains to the field of semiconductor manufacturing technology, and specifically relates to a method and system for controlling critical dimensions. Background Technology

[0002] Semiconductor manufacturing is expensive. Before the actual measurement results of the critical dimension (CD) of a wafer are available, for some wafers that are likely to be scrapped due to the impact of known problems, it is necessary to do everything possible to save them and reduce production costs.

[0003] In the dynamic random access memory (DRAM) manufacturing process, the dimensions of the previous layer A are passed down through the pattern, thereby affecting the critical dimensions of the current layer B. Other process parameters and other factors of the previous layer A will also affect the critical dimensions (Final CD) of the current layer B to some extent.

[0004] In existing technologies, R2R (Run-to-Run) systems provide a feedforward function in the manufacturing process, which uses the output results to appropriately adjust the input influencing factors such as process parameters in order to obtain more stable production process parameter control. However, this parameter control method currently limits the accurate correction of critical dimensions in advance.

[0005] like Figure 1 As shown, the feedforward algorithms currently available for R2R include the following: simple function control (univariate multivariate function), complex function control (multivariate function), and PID proportional-integral-derivative control.

[0006] The above-mentioned production process parameter control methods cannot fully consider complex production situations to achieve the goal of precise production guidance, which greatly limits the accuracy and applicability of the control method and makes it impossible to accurately correct the deviation of key dimensions in advance.

[0007] Public content

[0008] To address the above problems, this disclosure provides a critical dimension control method and system employing the following technical solution:

[0009] A critical dimension control method includes the following steps: acquiring historical data, including historical previous layer factor data and historical current layer measurement data; establishing and training a neural network prediction model using the historical data; acquiring initial previous layer factor data, using the initial previous layer factor data as input variables to the trained neural network prediction model for prediction, and outputting predicted current layer measurement data; and performing correction in the previous layer process according to preset conditions satisfied by the predicted current layer measurement data to obtain target previous layer factor data.

[0010] For example, the historical previous layer factor data has multiple sets, and the historical current layer measurement data includes the historical current layer key dimension measurement data corresponding to each set of the historical previous layer factor data.

[0011] For example, a neural network prediction model is built using the historical data, as follows:

[0012] The input layer of a neural network prediction model consists of n neurons, the hidden layer consists of m neurons, and the output layer consists of 1 neuron.

[0013] Each neuron in the n neurons of the input layer is defined as representing an input sub-variable, and each of the input sub-variables is defined as representing one of the historical previous layer factor data;

[0014] Define each of the m neurons in the hidden layer as a hidden layer output variable;

[0015] Define the output layer output variable as the predicted key dimension value of the current layer, and construct the prediction function.

[0016] For example, training a neural network prediction model using the historical data is specifically as follows:

[0017] The l sets of historical previous factor data are input into the input layer of the neural network prediction model. After being processed layer by layer by each hidden layer, l output variables are output.

[0018] Based on the deviation between each output variable and the corresponding historical measured data of the current layer's key dimensions, the weights between each neuron in the hidden layer and each neuron in the input layer are updated. When the error between the output variable and the corresponding historical measured data of the current layer's key dimensions is less than a set error, training is stopped, and the trained neural network prediction model is obtained.

[0019] For example, based on the preset conditions that the predicted layer measurement data meets, corrections are performed in the preceding process to obtain the target preceding layer factor data as follows:

[0020] If the predicted current layer measurement data exceeds the first preset range, the input variables of the neural network prediction model are adjusted according to the first correction model. The adjusted input variables are then input into the trained neural network prediction model for prediction to obtain the adjusted predicted current layer measurement data. This process continues until the adjusted predicted current layer measurement data meets the first preset range, and the corresponding current input variable is used as the target front-layer factor data.

[0021] For example, the input variables are adjusted according to the first correction model as follows:

[0022] The initial front-layer factor data are adjusted according to the first correction model, and additional correction process steps are added to the front-layer and / or current-layer process steps as new input variables.

[0023] For example, based on the preset conditions that the predicted current layer measurement data meets, to perform correction in the previous layer process to obtain the target previous layer factor data, the following further methods are also included:

[0024] If the predicted current layer measurement data is within the first preset range, the input variables of the neural network prediction model are adjusted according to the second correction model. The adjusted input variables are then input into the neural network prediction model for prediction to obtain the adjusted predicted current layer measurement data. This process continues until the adjusted predicted current layer measurement data meets the second preset range. The corresponding current input variable is then used as the target front-layer factor data. The second preset range is within the first preset range and includes the standard value.

[0025] For example, adjusting the input variables according to the second correction model includes:

[0026] The second corrective model only adjusts the initial front-layer factor data as new input variables.

[0027] For example, the adjustment range of the input variable in the second correction model is less than or equal to the adjustment range of the input variable in the first correction model.

[0028] For example, the initial front-layer factor data includes the initial front-layer critical dimension, the initial front-layer waiting time, the initial front-layer etching temperature, the initial front-layer etching pressure, and the initial front-layer gas flow rate;

[0029] The target front layer factor data includes the target front layer critical dimensions, target front layer waiting time, target front layer etching temperature, target front layer etching pressure, and target front layer gas flow rate.

[0030] For example, adding additional correction steps in the previous and / or current layer process steps is as follows:

[0031] Identify the single or multiple process actions included in the additional correction process steps;

[0032] Determine the sequential relationship between one or more of the process actions and the preceding and / or current process steps;

[0033] Determine the process parameters for one or more of the process actions.

[0034] For example, the process actions include deionized water rinsing, N2 purging, or O2 flushing.

[0035] This disclosure also provides a critical dimension control system, including:

[0036] The data acquisition module is used to acquire historical data, which includes historical previous layer factor data and historical current layer measurement data.

[0037] The model building and training module is used to build and train neural network prediction models using historical data;

[0038] The prediction module is used to acquire initial front-layer factor data, input the initial front-layer factor data as input variables into the trained neural network prediction model for prediction, and output the predicted measurement data of the current layer.

[0039] The correction module is used to correct deviations in the preceding process based on preset conditions that the predicted current layer measurement data meets, so as to obtain the target preceding layer factor data.

[0040] For example, the correction module is specifically used to: if the predicted current layer measurement data exceeds the first preset range, adjust the input variables of the neural network prediction model according to the first correction model, and then input the adjusted input variables into the neural network prediction model for prediction to obtain the adjusted predicted current layer measurement data, until the adjusted predicted current layer measurement data meets the first preset range, and use the corresponding current input variable as the target front layer factor data.

[0041] For example, adjusting the input variables according to the first correction model includes:

[0042] The initial front-layer factor data is adjusted according to the first correction model, and additional correction process steps are added to the front-layer and / or current-layer process steps as new input variables.

[0043] For example, the correction module is specifically used to: if the predicted current layer measurement data is within a first preset range, then adjust the input variables of the trained neural network prediction model according to the second correction model, and then input the adjusted input variables back into the neural network prediction model for prediction to obtain the adjusted predicted current layer measurement data, until the adjusted predicted current layer measurement data meets the second preset range, and use the corresponding current input variable as the target front-layer factor data, wherein the second preset range is within the first preset range and includes the standard value.

[0044] The beneficial effects of this disclosure are:

[0045] 1. This disclosure enables the prediction and correction of critical dimensions in the current layer at the previous layer, thereby improving the problem of deviation or even scrapping of critical dimension inspection results caused by the original adjustment lag.

[0046] 2. This disclosure establishes a correction model by separately training and optimizing data with key dimensions that exceed the standard and data that are within the standard, in order to avoid interference.

[0047] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 Several algorithmic functions for the feedforward of existing R2R technologies are shown;

[0050] Figure 2 A schematic flowchart of a critical dimension control method according to an embodiment of the present disclosure is shown;

[0051] Figure 3 A schematic diagram of the neural network prediction model structure according to an embodiment of the present disclosure is shown;

[0052] Figure 4 A schematic diagram of the process for performing bias correction at the front layer according to an embodiment of the present disclosure is shown;

[0053] Figure 5 A schematic diagram comparing the predicted values ​​and actual values ​​of the control function in the prior art is shown.

[0054] Figure 6 A schematic diagram showing a comparison between the predicted value and the actual value of the prediction model function according to an embodiment of the present disclosure is provided.

[0055] Figure 7 A schematic diagram of a critical dimension control system is shown. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0057] It should be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0058] To facilitate understanding of this disclosure, a brief introduction to critical dimensions (CD) and neural network models is provided below:

[0059] Critical dimensions are the physical dimensions of the microstructures on a chip, reflecting the patterning precision of the semiconductor manufacturing process. With advancements in semiconductor process nodes and photolithography technology, the dimensions of lines and gaps in a single exposure are becoming increasingly smaller, allowing for the integration of more transistors onto a single chip, resulting in more powerful chip performance. In a sense, advancements in semiconductor process technology aim to achieve smaller critical dimensions. When the change in critical dimensions exceeds a certain threshold, device performance will be affected, or even rendered unusable.

[0060] A neural network model consists of an input layer, an output layer, and hidden layers. The output layer and hidden layers contain neurons with activation functions. The neurons in each layer are connected layer by layer through weighted connections (connection weights).

[0061] Each input layer node receives feature parameters that reflect the nature of the problem under investigation. Their values ​​are propagated to each neuron in the first hidden layer, where the value of each neuron is the weighted sum of the values ​​of the individual nodes in the input layer multiplied by their connection weights. This process is repeated in each hidden layer until a prediction is produced.

[0062] The learning process consists of two stages: forward propagation of the signal and backward propagation of the error. During forward propagation, the input parameters are introduced from the input layer, processed layer by layer by the hidden layers, and then transmitted to the output layer. If the actual output of the input layer does not match the expected output, the process transitions to backward propagation of the error. Backpropagation involves transmitting the output error in some form back through the hidden layers to the input layer, distributing the error to all units in each layer, thereby obtaining the error signal of each unit. This error signal serves as the basis for adjusting the weights of each node. It is this backward propagation of the error that continuously adjusts the weights of each layer to achieve the purpose of network learning, thus determining the connection weights between nodes.

[0063] In existing semiconductor manufacturing processes, there may be dozens or even hundreds of input variables, and most of these parameters have a normal rather than a nonlinear effect on the output variables.

[0064] The ultra-high precision control of various parameters in semiconductor manufacturing makes it feasible to correct input variables in a timely manner.

[0065] Based on the above theory, this disclosure provides a critical dimension control method and system. It utilizes a neural network model trained on a large amount of online actual measurement data with multiple factors affecting the preceding layers. This allows the neural network to predict the measurement data of the current layer using factors from the preceding layers. Based on the degree of deviation between the predicted value and the target value, the prediction and correction of the critical dimension of the current layer are achieved at the preceding layers, thereby improving the problem of out-of-specification (OOS) or even scrapping of the critical dimension inspection results caused by the existing adjustment lag.

[0066] like Figure 2 As shown, a critical dimension control method includes the following steps: acquiring historical data, establishing and training a neural network prediction model using the historical data; acquiring initial preceding layer factor data, using the initial preceding layer factor data as input variables to the trained neural network prediction model for prediction, and outputting the predicted current layer measurement data; and correcting deviations in the preceding process according to the preset conditions satisfied by the predicted current layer measurement data to obtain the target preceding layer factor data.

[0067] The historical data includes historical preceding factor data and historical current layer measurement data. The historical preceding factor data consists of multiple sets, and the historical current layer measurement data includes the historical current layer key dimension measured data corresponding to each set of preceding factor data.

[0068] This disclosure combines intelligent neural network control with semiconductor manufacturing to achieve "intelligent" semiconductor manufacturing through the "Internet+" model. The neural network control used in this disclosure actively improves itself based on online data training, taking into full account the interference of continuous process changes on function accuracy.

[0069] like Figure 3 As shown, the neural network prediction model built using historical data is as follows:

[0070] S101. Define a neural network prediction model with an input layer consisting of n neurons, a hidden layer consisting of m neurons, and an output layer consisting of 1 neuron.

[0071] S102. Define each neuron in the n neurons of the input layer as an input sub-variable, and define each input variable x(i) as one of the historical previous layer factor data, where i = 1, 2, 3... n.

[0072] The front-layer factor data includes the front-layer critical dimension (CD), front-layer waiting time, front-layer etching temperature, front-layer etching pressure, and front-layer gas flow rate. For example, the gas can be O2.

[0073] It should be noted that all the selected front-layer factor data above affect the critical dimension values. Taking the formation process of the channel hole as an example, the influence of the front-layer factor data on the critical dimension values ​​is explained:

[0074] The specific formation process of the via is as follows: First, a patterned photoresist layer is formed on a semiconductor device including a substrate, a stacked structure, and a hard mask layer. The photoresist layer has photoresist layer openings, and the photoresist layer openings correspond to the vias. Then, based on the photoresist layer openings, the pattern of the photoresist layer is transferred to the hard mask layer through an etching process, that is, the hard mask layer has hard mask layer openings corresponding to the photoresist layer openings. Finally, based on the hard mask layer openings, a stacked structure including alternating interlayer sacrificial layers and interlayer insulating layers is formed through an etching process to form the via.

[0075] Currently, high aspect ratio channel etching uses a time-controlled method (also known as the by-time method). A set time is used as the condition to trigger and stop the etching reaction. For example, if the set time is 60 minutes, the etching reaction will be triggered to stop after 60 minutes. Therefore, controlling the waiting time is also an important process parameter of the current control method.

[0076] Because the process steps before forming the via (i.e., the front layer), such as the size of the photoresist layer opening on the patterned photoresist layer, the thickness of the hard mask layer, or the size of the hard mask layer opening on the hard mask layer, will have certain differences, the differences brought about by the front layer cannot be compensated when forming a high aspect ratio via using the by-time etching method. This results in unstable dimensions of the formed via, and some via dimensions may exceed the specification standard. Therefore, it can be seen that the critical dimension (CD) control of the front layer is particularly important.

[0077] During the etching process, chemical and physical reactions cause particles to generate heat as they collide with and impact the sample surface. Excessive heat can damage the silica sample surface, affecting critical dimensions. Choosing the appropriate flow rate and gas introduction time ratio significantly influences the sidewall morphology and reaction rate of the etched surface, thus affecting critical dimensions. Therefore, O2 flow rate and etching temperature also affect critical dimension values.

[0078] For example, input variable x(1) represents the critical dimension of the previous layer, input variable x(2) represents the waiting time of the previous layer, input variable x(3) represents the etching temperature of the previous layer, and input variable x(4) represents the O2 flow rate of the previous layer.

[0079] S103. Define each of the m neurons in the hidden layer as a hidden layer output variable. Define the hidden layer output variable function as follows:

[0080]

[0081] In the formula, u j Let f represent the output variable of the j-th hidden layer, where j = 1, 2, 3, ..., m, and f represent the mapping relationship of the activation function (introducing nonlinearity). This represents the hidden layer bias term (equivalent to a constant term), v ij This represents the weights of the i-th input variable and the j-th hidden layer output variable.

[0082] S104. Define the output layer's output variable as the predicted key dimension value of the current layer, and construct the prediction function as follows:

[0083]

[0084] In the formula, y(x) represents the output variable of the output layer, i.e., the predicted value, and ω j θ represents the weights of the j-th hidden layer output variable and the output variable. y This represents the output layer bias term (equivalent to a constant term).

[0085] For example, training a neural network prediction model using historical data is as follows:

[0086] Forward propagation of the neural network prediction model: l sets of historical previous factor data are input into the input layer of the neural network prediction model. After being processed layer by layer by each hidden layer, the neural network prediction model outputs l output variables.

[0087] Backpropagation of the neural network prediction model: l output variables and l sets of corresponding historical measured key dimensions of the current layer are merged into p sets of experimental data. Based on the deviation between each output variable and the corresponding historical measured key dimensions of the current layer, the total error is calculated using a loss function. The weights between each neuron in the hidden layer and the output layer are updated using the partial derivative of the total error with the weights of each neuron in the hidden layer and the weights of each neuron in the output layer, and the first learning rate. The weights between each neuron in the hidden layer and each neuron in the input layer are updated using the partial derivative of the error value of each neuron in the hidden layer with the weights of each neuron in the output layer, and the second learning rate. Training stops when the error value between the output variable and the corresponding historical measured key dimensions of the current layer is less than a set error, thus obtaining the trained neural network prediction model.

[0088] It's important to note that during the training of a deep neural network, because we want the network's output to be as close as possible to the desired predicted value, we compare the network's current prediction with the target value. Then, based on the difference, we update the weight vector of each layer (of course, there's usually an initialization process before the first update, where parameters are pre-configured for each layer). For example, if the network's prediction is too high, the weight vector is adjusted to predict a lower value. This adjustment continues until the deep neural network can predict the target value or a value very close to it. Therefore, we need to predefine "how to compare the difference between the predicted and target values," which is the loss function or objective function. These are important equations used to measure the difference between the predicted and target values. Taking the loss function as an example, a higher output value (loss) indicates a greater difference, so training the deep neural network becomes a process of minimizing this loss. Common loss functions include mean squared error, cross-entropy, logarithmic, and exponential loss functions, and the specific loss function can be chosen based on the actual application scenario.

[0089] For example, the mean square error can be used as the loss function, specifically:

[0090]

[0091] In the formula, E represents the total error, and y k Y represents the output variable of the preceding factor data in the k-th group of actual measurement data, i.e., the predicted key dimension value of the current layer, k = 1, 2, 3...p. k This represents the actual value of the critical dimension of the current layer in the k-th group of actual measurement data.

[0092] For example, the weights between each neuron in the hidden layer and the neurons in the output layer are updated as follows:

[0093]

[0094] newω m =Δω m +ω m

[0095] In the formula, Δω m The weights added between the output variables of the m-th hidden layer represent the weights added between the output variables, l1 represents the first learning rate, and E represents the weights added between the output variables. yk ω represents the error between the output variable of the preceding factor data and the actual value of the key dimension of the current layer in the k-th group of actual measurement data. m Newω represents the weights of the output variables of the m-th hidden layer. m This represents the weight between the updated m-th hidden layer output variable and the output variable.

[0096] For example, the specific steps for updating the weights between each neuron in the hidden layer and each neuron in the input layer are as follows:

[0097]

[0098] newv ij =Δv ij +v ij

[0099] In the formula, Δv ij L1 represents the weight added between the i-th input variable and the j-th hidden layer output variable, L2 represents the second learning rate, and E represents the second learning rate. yk newv represents the error between the output variable of the preceding factor data and the actual value of the key dimension of the current layer in the k-th group of actual measurement data. ij This represents the weight between the updated i-th input variable and the j-th hidden layer output variable.

[0100] This disclosure improves the training accuracy of the data model by using online actual measurement data to train the neural network prediction model, thereby increasing the modeling accuracy and computational efficiency, and making the prediction results more accurate.

[0101] The preceding factor data in this disclosure includes process parameters for one or more steps, and the process parameters include the sequence of operations between different steps.

[0102] like Figure 4 As shown, for example, based on the preset conditions satisfied by the predicted current layer measurement data, correction is performed in the previous layer process to obtain the target previous layer factor data as follows:

[0103] S201. If the predicted measurement data of the current layer exceeds the first preset range, the input variables of the neural network prediction model are adjusted according to the first correction model. The adjusted input variables are then input into the trained neural network prediction model for prediction to obtain the adjusted predicted measurement data of the current layer. This process continues until the adjusted predicted measurement data of the current layer meets the first preset range, and the corresponding current input variables are used as the target front-layer factor data.

[0104] The first correction model is as follows:

[0105]

[0106] In the formula, y(x) o f represents the input variable after correction by the first correction model. o This represents the mapping relationship of the activation functions in the first correction model. This represents the first bias term of the first bias correction model. For example, adjusting the input variables of the neural network prediction model according to the first bias correction model specifically involves:

[0107] The initial preceding factor data are adjusted based on the first corrective model, and additional corrective process steps are added to the preceding and / or current process steps as new input variables.

[0108] For example, adding an additional correction process step specifically involves: determining the single or multiple process actions included in the additional correction process step; determining the sequential relationship between the single or multiple process actions and the previous and / or current process steps; and determining the process parameters of the single or multiple process actions.

[0109] It should be noted that one or more process actions can be interspersed in other process steps. For example, additional correction process steps include deionized water cleaning and N2 purging. Deionized water cleaning can be performed before the previous process step, or N2 purging can be performed after the previous process step.

[0110] For example, if the wafer's predicted measurement data in the previous layer exceeds a first preset range, the following parameters are modified using the first correction model: the critical dimension value x(1) in the previous layer step A is modified. o And waiting time x(2) o Modify the etching temperature x(3) in step B of the previous layer. o And O2 flow rate x(4) o After step B, add a correction process step including deionized water cleaning, N2 purging, and O2 rinsing. Modify the other front-layer factor data x(a) in step C after the correction process step. o and x(b) o .

[0111] The new input variable x(1)o x(2) o x(3) o x(4) o Deionized water cleaning, N2 purging, O2 rinsing, x(a) o and x(b) o The trained neural network prediction model is input for prediction. If the neural network prediction model predicts that the key size value of the current layer meets the first preset range, then the previous layer is determined to produce based on the previous layer factor data represented by the current input variable, thus completing the correction.

[0112] The pre-treatment factors for deionized water cleaning include cleaning time and deionized water concentration; the pre-treatment factors for N2 purging include N2 concentration, N2 flow rate and purging time; and the pre-treatment factors for O2 flushing include O2 concentration, O2 flow rate and flushing time.

[0113] For example, the following parameters can also be modified using the first correction model: modify the etching temperature x(3) in step A of the previous layer. o And O2 flow rate x(4) o After step A, a correction process step is added, including deionized water cleaning, N2 purging, and O2 rinsing. In step B after the correction process step, the critical dimension value x(1) of the previous layer is modified. o And waiting time x(2) o In step C, modify the data of other preceding factors x(a). o and x(b) o .

[0114] As can be seen, the above correction process involves continuously adjusting new input variables into the first correction model for prediction until the predicted critical dimension value of the current layer meets the first preset range. At this point, target front-layer factor data is obtained, and wafers that meet the standards are produced based on the target front-layer factor data, thus avoiding wafer scrap and reducing production costs.

[0115] S202. If the predicted measurement data of the current layer is within the first preset range, the input variables of the neural network prediction model are adjusted according to the second correction model. The adjusted input variables are then input into the trained neural network prediction model for prediction to obtain the adjusted predicted measurement data of the current layer. This process continues until the adjusted predicted measurement data of the current layer meets the second preset range. The corresponding current input variable is used as the target front-layer factor data. The second preset range is within the first preset range and includes the standard value.

[0116] The second correction model is as follows:

[0117]

[0118] In the formula, y(x) if represents the input variable after correction by the second correction model. i This represents the mapping relationship of the activation functions in the second correction model. This represents the bias term of the second correction model.

[0119] For example, adjusting the input variables of the neural network prediction model according to the second correction model is as follows: the second correction model only adjusts the initial front-layer factor data as new input variables. The adjustment range of the input variables of the second correction model is less than or equal to the adjustment range of the input variables of the first correction model.

[0120] For example, if the wafer's predicted measurement data in the previous layer is within the first preset range, the following parameters are modified using the second correction model: In step A of the previous layer, the critical dimension value x(1) of the previous layer is modified. i And waiting time x(2) i In step B of the previous layer, the etching temperature x(3) is modified. i And O2 flow rate x(4) i In step C, modify the data of other preceding factors x(a). i and x(b) i If the critical dimension value predicted by the second correction model in this cycle meets the second preset range, then the previous layer will be determined to produce based on the previous layer factor data represented by the input variables in this cycle.

[0121] The new input variable x(1) i x(2) i x(3) i x(4) i x(a) i and x(b) i The trained neural network prediction model is input for prediction. If the neural network prediction model predicts that the key size value of the current layer meets the second preset range, then the previous layer factor data represented by the current input variable is determined as the target previous layer factor data, and the correction is completed.

[0122] For example, the following parameters can also be modified using the second correction model: modify the etching temperature x(3) in step A of the previous layer. i And O2 flow rate x(4) i In step B of the previous layer, modify the critical dimension value x(1) of the previous layer. i And waiting time x(2) i In step C, modify the data of other preceding factors x(a). i and x(b) i .

[0123] It should be noted that there is a strict failure and scrapping procedure for the key dimension values predicted by the neural network for each layer. For example, it is defined that the key dimension value x predicted for each layer needs to satisfy: a ≤ x ≤ b. If x b occurs, the wafer needs to be scrapped.

[0124] In the present disclosure, a correction model is established and optimized through separate training for data with key dimensions exceeding the standard (Out spec data) and data within the standard (Inspec data) to avoid interference. Specifically, the data with key dimensions exceeding the standard (Outspec data) is relatively few and has a large degree of dispersion. To avoid the data with key dimensions exceeding the standard (Out spec data) interfering with the second correction model for data within the standard.

[0125] It should be noted that during the prediction process of the neural network prediction model, it will actively self-improve based on online data training. The specific process should be understood by those skilled in the art and will not be elaborated here.

[0126] As Figure 5 shown, Figure 5 shows a comparison diagram of the predicted values and actual values of the control function prediction in the prior art. By comparing the deviations between the predicted values and actual values of 10 sets of current control function predictions, it can be seen that the deviations between the predicted values and actual values of the current control function predictions are relatively large and the control accuracy is low; as Figure 6 shown, Figure 6 shows a comparison diagram of the predicted values and actual values of the prediction model function according to the embodiments of the present disclosure. By comparing the deviations between the predicted values and actual values of 10 sets of prediction model functions of the present disclosure, it can be seen that the prediction of the prediction model function of the present disclosure is more accurate.

[0127] As Figure 7 shown, the present disclosure also provides a key dimension control system, including a data acquisition module, a model establishment and training module, a prediction module, and a correction module. Among them, the data acquisition module is used to acquire historical data, and the model establishment and training module is used to establish and train a neural network prediction model through historical data; the prediction module is used to acquire initial pre-layer factor data, input the initial pre-layer factor data as input variables into the trained neural network prediction model for prediction, and output the predicted current layer measurement data; the correction module is used to perform correction in the pre-layer process according to the preset conditions satisfied by the predicted current layer measurement data to obtain the target pre-layer factor data.

[0128] In the present disclosure, data on the influencing factors of the dimensions in the pre-layer are collected, and the deviation degree between the real-time predicted value and the target value of the neural network prediction model in the pre-layer is used to realize the prediction of the measurement data of the current layer using the pre-layer factors.

[0129] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A critical dimension control method applied in the field of semiconductor production and manufacturing, characterized in that, The method comprises the following steps: acquiring historical data, the historical data comprising historical pre-layer factor data and historical in-layer measurement data; establishing and training a neural network prediction model based on the historical data; acquiring initial pre-layer factor data, inputting the initial pre-layer factor data as an input variable into the trained neural network prediction model for prediction, and outputting predicted in-layer measurement data; adjusting the input variable according to a first correction model to adjust the input variable, inputting the adjusted input variable into the trained neural network prediction model for prediction, and obtaining adjusted predicted in-layer measurement data until the adjusted predicted in-layer measurement data meets a first preset range, so as to take the current input variable as the target pre-layer factor data.

2. The critical dimension control method of claim 1, wherein The historical pre-layer factor data has multiple groups, and the historical in-layer measurement data comprises historical in-layer critical dimension measurement data corresponding to each group of historical pre-layer factor data.

3. The critical dimension control method of claim 2, wherein, The neural network prediction model is established based on the historical data, and the establishment is specifically as follows: defining that the input layer of the neural network prediction model comprises n neurons, the hidden layer comprises m neurons, and the output layer comprises one neuron; defining that each neuron in the n neurons of the input layer represents an input sub-variable, and each input sub-variable represents one of the historical pre-layer factor data; defining that each neuron in the m neurons of the hidden layer represents a hidden layer output variable; defining that the output variable of the output layer is a predicted in-layer critical dimension value, and constructing a prediction function.

4. The critical dimension control method of claim 3, wherein The neural network prediction model is trained based on the historical data, and the training is specifically as follows: inputting l groups of historical pre-layer factor data into the input layer of the neural network prediction model, processing layer by layer through each hidden layer, and outputting l output variables; updating the weight between each neuron of the hidden layer and each neuron of the input layer based on the deviation value between each output variable and the corresponding historical in-layer critical dimension measurement data, and stopping training when the error value between the output variable and the corresponding historical in-layer critical dimension measurement data is less than a set error, thereby obtaining the trained neural network prediction model.

5. The critical dimension control method of claim 1, wherein The adjustment of the input variable according to the first correction model is specifically as follows: adjusting the initial pre-layer factor data according to the first correction model and adding an additional correction process step in the pre-layer and / or in-layer process step as a new input variable.

6. The critical dimension control method of claim 1, wherein The adjustment of the input variable according to the first correction model is specifically as follows: adjusting the initial pre-layer factor data according to the first correction model and adding an additional correction process step in the pre-layer and / or in-layer process step as a new input variable. If the predicted on-layer measurement data is within a first preset range, adjusting the input variable of the neural network prediction model according to a second correction model, inputting the adjusted input variable into the neural network prediction model for prediction to obtain adjusted predicted on-layer measurement data, until the adjusted predicted on-layer measurement data meets a second preset range, and taking the corresponding current input variable as the target pre-layer factor data, the second preset range being within the first preset range and containing a standard value.

7. The critical dimension control method of claim 6, wherein Adjusting the input variable according to the second correction model comprises: Adjusting only the initial pre-layer factor data as a new input variable according to the second correction model.

8. The critical dimension control method according to claim 6 or 7, wherein, The adjustment range of the input variable of the second correction model is less than or equal to the adjustment range of the input variable of the first correction model.

9. The critical dimension control method of claim 1, wherein, The initial pre-layer factor data comprises initial pre-layer critical dimensions, initial pre-layer waiting time, initial pre-layer etching temperature, initial pre-layer etching pressure and initial pre-layer gas flow. The target pre-layer factor data comprises target pre-layer critical dimensions, target pre-layer waiting time, target pre-layer etching temperature, target pre-layer etching pressure and target pre-layer gas flow.

10. The critical dimension control method of claim 5, wherein, The additional correction process step in the pre-layer and / or on-layer process step is specifically as follows: Determining a single or multiple process actions included in the additional correction process step; Determining the front-back relationship of the single or multiple process actions with the pre-layer and / or on-layer process step; Determining the process parameters of the single or multiple process actions.

11. The critical dimension control method of claim 10, wherein, The process actions comprise deionized water cleaning, N2purging or O2flushing.

12. A critical dimension control system applied in the field of semiconductor production and manufacturing, characterized in that, Comprise: A data acquisition module for acquiring historical data, the historical data comprising historical pre-layer factor data and historical on-layer measurement data; A model establishment and training module for establishing and training a neural network prediction model through the historical data; A prediction module for acquiring initial pre-layer factor data, inputting the initial pre-layer factor data as an input variable into the trained neural network prediction model for prediction, and outputting to obtain predicted on-layer measurement data; A correction module for correcting in the pre-layer process according to a preset condition met by the predicted on-layer measurement data to obtain target pre-layer factor data, comprising: if the predicted on-layer measurement data exceeds a first preset range, adjusting the input variable of the neural network prediction model according to a first correction model, inputting the adjusted input variable into the neural network prediction model for prediction to obtain adjusted predicted on-layer measurement data, until the adjusted predicted on-layer measurement data meets the first preset range, and taking the corresponding current input variable as the target pre-layer factor data.

13. The critical dimension control system of claim 12, wherein, Adjusting the input variable according to the first correction model comprises: Adjusting the initial pre-layer factor data and adding an additional correction process step in the pre-layer and / or on-layer process step as a new input variable according to the first correction model.

14. The critical dimension control system of claim 12, wherein, The correction module is specifically used for: If the predicted in-layer measurement data is within a first preset range, input variables of the trained neural network prediction model are adjusted according to a second correction model, the input variables are input to the neural network prediction model again for prediction to obtain adjusted predicted in-layer measurement data, until the adjusted predicted in-layer measurement data meets a second preset range, the corresponding current input variable is taken as the target previous layer factor data, the second preset range is within the first preset range and contains a standard value.

Images