Method for predicting photoresist film thickness through convolutional neural network based on FDC data

Abstract

The invention discloses a method for predicting photoresist film thickness by using a convolutional neural network based on FDC data, which comprises the following steps of: firstly, acquiring machine data and corresponding film thickness data recorded by a Track machine of a fault detection and classification FDC system in a wafer rotating and spin coating process, the machine data comprising a plurality of process parameter variables; and screening the process parameter variables, inputting the screened machine data into a regression prediction model based on a one-dimensional convolutional neural network for training, testing and optimizing, and obtaining predicted photoresist film thickness data by using the optimized regression prediction model. According to the method, the prediction of the photoresist film thickness is realized, the determination of a process window is facilitated, namely on the premise of ensuring the photoresist performance and the pattern quality, the process parameter variation range is allowed, the process parameter variation range can be used as a reference for the maintenance and calibration of photoetching machine equipment, the equipment is ensured to be in an optimal state, and the improvement of the reliability of batch production is facilitated.

Description

Technical Field

[0001] The invention belongs to the field of semiconductor manufacturing measurement and relates to a method for predicting photoresist film thickness by using a convolutional neural network based on FDC data. Background Art

[0002] In the photolithography process, using a track machine for photoresist development to evenly coat the photoresist on the wafer surface is one of the important steps before exposure. This step ensures that the photoresist forms a uniform and consistent film on the wafer surface, which is crucial for subsequent photolithography exposure and pattern transfer, because the thickness and uniformity of the photoresist will directly affect the quality of pattern transfer, and thus affect the resolution of photolithography. If the photoresist is unevenly distributed on the wafer surface, it will lead to inconsistent uniformity of the photoresist line width (CD), which will also affect the consistency of the etching rate in the subsequent process and the dose required during the exposure process. Uneven thickness may even cause pattern deformation or breakage, which will affect the subsequent pattern transfer and the performance of the integrated circuit, and ultimately affect the yield of the product. Therefore, it is very important to predict the thickness of the photoresist film.

[0003] Normally, the thickness of the photoresist film needs to be measured on a measuring machine, and there is currently no method that can predict the thickness of the photoresist film. If the parameters of the photoresist film thickness can be predicted without consuming the cost of measuring equipment and a lot of time, it can not only reduce the loss of measuring equipment and materials, shorten the production cycle, but also improve economic benefits.

[0004] In the semiconductor industry, the Fault Detection and Classification (FDC) system is a key system used in the semiconductor manufacturing field to monitor and analyze equipment operation data in real time to improve product yield and production efficiency. Its role is to collect, store and analyze fault or defective product data detected during the semiconductor manufacturing process. The technology commonly used by the current leading FDC systems is to develop statistical fault models for different production operation points based on training sets of "good" and "bad" operations, and then use key parameters collected from the equipment during the processing to evaluate these models in real time to detect process deviations and predict impending equipment failures. However, no one has used it for product parameter prediction in the semiconductor manufacturing process. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies in the prior art and provide a method for predicting photoresist film thickness using a convolutional neural network based on FDC data, thereby solving the problem of predicting film thickness in the field of rotary coating in integrated circuit manufacturing.

[0006] The objective of the present invention is achieved through the following technical solutions: A method for predicting photoresist film thickness using a convolutional neural network based on FDC data, comprising the following steps: Step S1, during the photoresist coating process on the wafer surface, the machine data of the Track machine during the rotating coating and the photoresist film thickness are collected by using the fault detection and classification FDC system, and the machine data are preprocessed; wherein each set of machine data after preprocessing contains m process parameter variables, where m is an integer greater than 1; The machine data collected by the FDC system is preprocessed in the following ways, including: The FDC system stores the data recorded by the Track machine in time sequence during the spinning process. s will record all the parameters of the machine once; After the spin coating is completed The various machine operation parameters recorded during the time period of s are extracted for preprocessing, and the FDC data samples (i.e., machine data) when the same photoresist is spin-coated with the same film thickness are extracted.

[0007] Data cleaning is performed on machine data containing missing values ​​and errors in the intercepted data samples, including: The incomplete data with only FDC recorded machine data but no corresponding film thickness data is deleted. The machine data recorded by FDC is compared with the data of the set recipe of the Track machine. If the following formula (1) is satisfied, it means that the data recorded by the Track machine is wrong and is deleted; otherwise, it is retained. Formula (1) in, Indicates the machine data recorded by the FDC system. Indicates the data for setting the process recipe. , Indicates the set threshold range.

[0008] The data set consisting of the FDC data samples obtained after deletion is normalized, and the formula of the conversion function is: Formula (2) In the formula, represents the normalized result; Indicates the machine data after deletion. Indicates the maximum value of the machine data after deletion. Indicates the minimum value of the machine data after deletion.

[0009] Step S2, screening the pre-processed machine data, then constructing a data set, and dividing it into a training set and a test set; the label of the data set adopts the photoresist film thickness collected in step S1; The process parameter variables contained in the normalized machine data sample are screened according to predetermined feature screening rules in the following ways, including: Delete the values ​​whose standard deviation is less than the threshold range The standard deviation is the square root of the variance and is calculated as: Formula (3) In the formula, represents the standard deviation, Represents the data points in the machine data, It is a data point The average value of It is a data point The number of

[0010] Step S3: construct a regression prediction model based on a one-dimensional convolutional neural network, and perform training and testing using a training set and a test set; according to The data set is randomly divided into a training set and a test set according to the ratio of; the regression prediction model is trained by the training set; and the prediction performance of the regression prediction model after training is evaluated by the test set. In the data set formed by the data samples, the process parameter variables related to the pretreatment, spin coating, baking and cooling steps in the spin coating after screening in step S2 are used as the input parameters of the regression prediction model; the output parameter of the regression prediction model is the prediction result of the photoresist film thickness.

[0011] Step S4, collecting the machine data recorded by the Track machine during the rotating coating, pre-processing the machine data and screening the process parameter variables, and then inputting the data into the trained and tested regression prediction model based on the one-dimensional convolutional neural network to obtain the photoresist film thickness prediction data.

[0012] The second object of the present invention is to provide a photoresist film thickness prediction system for implementing the above method, comprising the following modules: Data acquisition module, used to obtain machine data during wafer manufacturing process; Data processing module, used for preprocessing machine data and screening process parameter variables; The film thickness prediction module inputs the processed machine data into the optimized regression prediction model based on the one-dimensional convolutional neural network to obtain the prediction result of the photoresist film thickness.

[0013] A third object of the present invention is to provide an electronic device, characterized in that it includes a processor and a memory, the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the above method.

[0014] The fourth object of the present invention is to provide a machine-readable storage medium, characterized in that the machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the above method.

[0015] Compared with the prior art, the photoresist film thickness prediction method provided by the present invention has the following beneficial effects: The present invention constructs a regression model of photoresist film thickness based on a one-dimensional convolutional neural network, and by introducing feature screening rules, screens features with little correlation in pre-processed FDC data, and eliminates features with little correlation and that may cause noise interference; and uses the one-dimensional convolutional neural network to input properties of multiple features to optimize the regression model, thereby improving the effectiveness and accuracy of the regression prediction model, realizing the prediction of photoresist film thickness, reducing the workload of semiconductor measurement staff, reducing the amount of photoresist used, reducing production costs, and being greatly beneficial for maintaining film thickness consistency and stability; the film thickness prediction based on the model not only helps to stabilize the process window, that is, the allowable range of process parameter variations under the premise of ensuring photoresist performance and pattern quality, but also can be used as a reference for maintenance and calibration of photolithography equipment, ensuring that the equipment is in an optimal state, and helping to improve the reliability of mass production. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the steps of the method for predicting the thickness of the photoresist film provided by the present invention.

[0017] Figure 2 It is a flow chart of the Track machine glue spreading provided by the present invention.

[0018] Figure 3 It is a schematic diagram of the actual meaning of various features in the COT cavity provided by an embodiment of the present invention.

[0019] Figure 4 It is a schematic diagram of the normalization of three characteristic variables contained in a FDC data sample after screening provided by an embodiment of the present invention, wherein A is Spin_rotation_speed (rotation speed), B is Spin_Exh_Pressure (exhaust pressure during rotation and glue spreading), and C is RRC_Nozzle_flow (RRC (Reduced Resist Consumption) flow).

[0020] Figure 5 It is a schematic diagram of the structure of a one-dimensional convolutional neural network provided in an embodiment of the present invention.

[0021] Figure 6 This is a box plot of the root mean square error (RMSE) obtained by using four models to predict 100 test samples using four feature combinations provided in an embodiment of the present invention. S, P, and R represent the three features Spin_rotation_speed (rotation speed), Spin_Exh_Pressure (exhaust pressure during rotary spraying), and RRC_Nozzle_flow (RRC (Reduced Resist Consumption) flow), respectively.

[0022] Figure 7 The coefficient of determination (R) obtained by using the four models to predict 100 test samples using the four feature combinations provided in the embodiment of the present invention is 2 ) box plot. Among them, there are three features: Spin_rotation_speed (rotation speed), Spin_Exh_Pressure (exhaust pressure during rotary spraying), and RRC_Nozzle_flow (RRC (Reduced ResistConsumption) flow).

[0023] Figure 8 It is the prediction result of the model trained by considering only the two features of Spin_rotation_speed (rotation speed) and RRC_Nozzle_flow (RRC (Reduced Resist Consumption) flow) provided by the embodiment of the present invention under a new set of data. DETAILED DESCRIPTION

[0024] The specific technical scheme of the present invention is further described in detail below with reference to specific examples.

[0025] like Figure 1 As shown, the present invention provides a method for predicting photoresist film thickness by using a convolutional neural network based on FDC data. In this embodiment, training is performed based on sample data of M91Y6 photoresist with a spin coating thickness of 6900 Å. The actual operation is as follows: 1) Data preprocessing, each time the machine is running The FDC system will record all the parameters of the machine every s; After the spin coating is completed The various parameters of the machine operation recorded during the time period of s are extracted for preprocessing, including deleting missing values ​​and erroneous data, normalizing the data, etc., to obtain 100 complete FDC data samples and corresponding film thickness data; The data recorded by the FDC system on the coating process include: The process of glue coating by Track machine is shown in Figure 2 In the process, the wafer mainly passes through four chambers: ADH (Adhesion Process Station), SCPL (Slim Chill Plate Process Station), COT (Coat Process Station), and CGHP (Chilling General-Purpose Hot Process Station). The specific coating operations in the four chambers are: ADH pre-treats the wafer and applies the adhesion promoter HMDS; SCPL cools the wafer; COT performs rotary coating; CGHP bakes the wafer; The FDC system will record the temperature, humidity, rotation speed, liquid flow rate and other data of the wafer passing through the four chambers ADH, SCPL, COT and CGHP of the Track machine during the coating operation, which are collectively referred to as FDC data; in this embodiment, the FDC data includes the following characteristic variables (process parameter variables): Resist_temp (photoresist temperature), Motor_flange_temp (tray temperature), Cup_temp (cavity temperature), Cup_humidity (cavity humidity), Spin_rotation_speed (rotation speed), Spin_Exh_Pressure (exhaust pressure during rotation coating), RRC_Nozzle_flow (photoresist flow of RRC process, where RRC is Reduced Resist Consumption, a new type of coating process).

[0026] Delete the machine data containing missing values ​​and errors in the intercepted data sample, including: Delete the incomplete data that only has FDC data but lacks the corresponding film thickness data. Compare the data recorded by FDC with the data of the set recipe of the Track machine. If the following formula is satisfied, it means that the data recorded by FDC is wrong and deleted, otherwise it is retained: Formula (1) in, Indicates the data recorded by the FDC system. Indicates the data for setting the recipe. Indicates the set threshold range. In this embodiment, .

[0027] The data set consisting of the FDC data samples obtained after deletion is normalized, and the formula of the conversion function is: In the formula, represents the normalized result; represents the sample data of FDC data, Indicates the maximum value of the sample data of the FDC data, Indicates the minimum value of the sample data of the FDC data.

[0028] 2) Feature selection: Consider the standard deviation of different feature variables in the FDC data to decide whether to retain the feature variable, as shown in Table 1: Table 1 Feature Name Standard Deviation Resist_temp (resist temperature) <![CDATA[4.4×10 -5 ]]> Motor_flange_temp (tray temperature) <![CDATA[4.4×10 -5 ]]> Cup_temp (cavity temperature) <![CDATA[4.9×10 -5 ]]> Cup_humidity <![CDATA[3.3×10 -4 ]]> Spin_rotation_speed 0.2316 Spin_Exh_Pressure (exhaust pressure during spin coating) 0.3152 RRC_Nozzle_flow (RRC traffic) 0.2290 By analyzing Table 1, the standard deviation of the features is less than After feature screening, it is found that the standard deviation is less than There are four features: Resist_temp (photoresist temperature), Cup_humidity (chamber humidity), Motor_flange_temp (rotating tray temperature), and Cup_temp (chamber temperature). For the effective modeling, these four features can be removed. The actual meaning of each feature can be found in Figure 3 ; The normalized result of the three characteristic variables contained in a screened FDC data sample, see Figure 4 .

[0029] In order to compare the results after feature screening, the one-dimensional convolutional neural network with training times epoch = 1000 and learning rate lr = 0.001 is used as the measurement basis, and the root mean square error is used as the and the coefficient of determination As the evaluation criteria, we compared the features before and after the feature selection, retained all features, and removed features with a standard deviation less than The root mean square error of the model trained in the two cases after the features are predicted for 100 groups of samples and the coefficient of determination , see Table 2, it is found that after removing the standard deviation less than better results can be obtained in the case of features.

[0030] Table 2 Feature variables Model trained with seven feature variables before screening Model trained with three feature variables after screening The proportion of RMSE less than 50 0.55 0.76 <![CDATA[R 2 The proportion greater than 0.8]]> 0.21 0.36 3) Establish a regression prediction model based on a one-dimensional convolutional neural network. The data set formed by the features obtained in step S2 is randomly divided into a training set and a test set. The structure of the model is shown in Figure 5 The regression prediction model based on the one-dimensional convolutional neural network includes an input layer, a first convolution block, a second convolution block, a third convolution block, a fully connected layer and an output layer cascaded in sequence; the first convolution block, the second convolution block, and the third convolution block all include a first convolution layer, a second convolution layer and a pooling layer cascaded in sequence. Among them, the number of convolution kernels of the convolution layer is 32, 32, 64, 64, 128, and 128 respectively, the convolution kernel size is 3, and the activation function uses the ReLU function; after a maximum pooling operation, the length and width of the calculation domain are reduced to 1 / 2 of the original respectively; the number of neurons in the fully connected layer is 1024, and the activation function uses the ReLU function; the regularization coefficient is set to 0.01 to prevent overfitting, and then the output is mapped to the specified output space dimension (49 is used in this embodiment); to avoid overfitting, a random disconnection layer (Drop-out) is set in this network, and the random disconnection rate is 0.5; the optimizer is the Adam optimizer, the loss function is the RMSE root mean square loss error, and the learning rate is 0.001; the number of training epochs=1000, and the batch size batch_size=16.

[0031] Use the trained model to predict the photoresist film thickness.

[0032] 4) Taking advantage of the selectivity of the number of features of the one-dimensional convolutional neural network, the number and selection of features of the trained and tested regression prediction model are optimized. Too few features may not be enough to capture the essence of the data, resulting in the model being unable to learn the key information in the data. The performance of the model may be limited and fail to achieve the expected accuracy or efficiency. Too many features will increase the complexity of the model, and may learn unnecessary noise as features, reducing the stability of the model.

[0033] The feature selection and quantity in the regression model will affect the output of the model. The feature selection and quantity optimization of the regression prediction model established by the one-dimensional convolutional neural network can be performed in the following ways, including: Increase or decrease the number of input features, and by comparing the combinations of several different sets of features, find a set of feature combinations that can make the regression prediction model perform best. The standard for evaluating the regression model is the root mean square error of the prediction results. The closer it is to 0, the higher the coefficient of determination The closer it is to 1, the better the model effect is. and the coefficient of determination The calculation formula is: In the formula, is the number of samples, , are the predicted value, true value and the average of the true value, respectively.

[0034] In this embodiment, it is generally believed that the rotation speed has the greatest impact on the film thickness. Therefore, the Spin_rotation_speed (rotation speed) with the standard deviation in the middle is used as the main feature. On this basis, 1-2 other features are added to find a set of feature combinations that can make the regression prediction model perform best. These are used as the feature variable input of the final regression prediction model, and the optimized final regression prediction model is used to predict the thickness of the photoresist film of the spin-spinning film. The results are shown in Figure 6 , Figure 7 , proving that the regression model obtained by selecting the combination of the two features Spin_rotation_speed (rotation speed) and RRC_Nozzle_flow (RRC (Reduced ResistConsumption) flow) as input has the best effect.

[0035] In order to verify the effectiveness of the method of the present invention, a new set of machine data was selected and input into the model for prediction. The results obtained are shown in Figure 8 , where the error of the point where the film thickness prediction differs most from the true value does not exceed 0.15%, proving that this method is effective.

Claims

1. A method for predicting photoresist film thickness using a convolutional neural network based on FDC data, characterized in that: The following steps are involved: Step S1, during the photoresist coating process on the wafer surface, collecting the machine data of the Track machine when rotating and coating and the photoresist film thickness, and pre-processing the machine data; Each set of preprocessed machine data contains m process parameter variables, where m is an integer greater than 1; Step S2, screening the process parameter variables of the preprocessed machine data, then constructing a data set and dividing it into a training set and a test set; the label of the data set adopts the photoresist film thickness collected in step S1; Step S3: construct a regression prediction model based on a one-dimensional convolutional neural network, and perform training and testing using a training set and a test set; Step S4, collecting the machine data recorded by the Track machine during the rotating coating, pre-processing the machine data and screening the process parameter variables, and then inputting the data into the trained and tested regression prediction model based on the one-dimensional convolutional neural network to obtain the photoresist film thickness prediction data.

2. The method for predicting photoresist film thickness using a convolutional neural network based on FDC data according to claim 1, characterized in that: The machine data include the photoresist temperature, tray temperature, cavity temperature, cavity humidity, rotation speed, exhaust pressure during rotation and photoresist flow rate of the wafer passing through the four cavities ADH, SCPL, COT and CGHP of the Track machine during the wafer rotation and coating process.

3. The method for predicting photoresist film thickness using a convolutional neural network based on FDC data according to claim 1, characterized in that: The preprocessing includes data cleaning and normalization.

4. The method for predicting photoresist film thickness using a convolutional neural network based on FDC data according to claim 1, characterized in that: The process parameter variable screening specifically includes deleting the variables with standard deviation less than the threshold range. process parameter variables.

5. The method for predicting photoresist film thickness using a convolutional neural network based on FDC data according to claim 1, characterized in that: The regression prediction model based on the one-dimensional convolutional neural network in step S3 includes an input layer, a first convolution block, a second convolution block, a third convolution block, a fully connected layer and an output layer cascaded in sequence; the first convolution block, the second convolution block, and the third convolution block have the same structure, and all include a first convolution layer, a second convolution layer and a pooling layer cascaded in sequence.

6. The method for predicting photoresist film thickness using a convolutional neural network based on FDC data according to claim 5, characterized in that: The number of convolution kernels in the convolution layer in the first convolution block is 32, and the convolution kernel size is 3; the number of convolution kernels in the convolution layer in the second convolution block is 64, and the convolution kernel size is 3; the number of convolution kernels in the convolution layer in the third convolution block is 128, and the convolution kernel size is 3.

7. A photoresist film thickness prediction system implementing the method according to any one of claims 1 to 6, characterized in that: Includes the following modules: Data acquisition module, used to obtain machine data during wafer manufacturing process; Data processing module, used for preprocessing machine data and screening process parameter variables; The film thickness prediction module inputs the processed machine data into the trained and tested regression prediction model based on the one-dimensional convolutional neural network to obtain the prediction result of the photoresist film thickness.

8. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the method according to any one of claims 1 to 6.

9. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the method described in any one of claims 1-6.

Images