Prediction methods for semiconductor process measurement data

Abstract

This invention provides a method for predicting semiconductor process measurement data, comprising: providing historical measurement data of a wafer and historical raw process data; using an autoencoder to shorten the time step of the first time series data to obtain second time series data; using the second time series data and historical measurement data to fit and train a prediction model, the prediction model including a transformer encoder and a multilayer perceptron; inputting the raw process data to be predicted into the time step shortening model and the prediction model to obtain predicted measurement data. In this invention, shortening the time step of the time series data using an autoencoder improves overall computational efficiency without losing feature information; parallel computing using a transformer encoder improves computational speed; and a self-attention mechanism is used to acquire both the size and time step features of the data, thereby improving prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, and in particular to a method for predicting semiconductor process measurement data. Background Technology

[0002] In semiconductor manufacturing, inline measurement is typically performed on wafers after the semiconductor process is completed. By monitoring the inline measurement data, the semiconductor process can be monitored.

[0003] Due to the long time required for online measurement, in order to reduce the impact on process capacity and efficiency, current methods generally involve sampling and measuring a portion of wafers after the semiconductor process is completed. This cannot measure all wafers after the semiconductor process, and the measurement data also lags behind the process significantly, resulting in a higher risk of missed measurements and false alarms. On the other hand, increasing the frequency of wafer measurement will also bring huge cost pressures in terms of time, manpower, and resources.

[0004] At the same time, there are also methods that use data features of a few process parameters in semiconductor manufacturing to predict wafer measurement data. These data features include, for example, mean, extreme values, range, or slope. However, the accuracy of these methods in predicting measurement data in actual semiconductor manufacturing processes is low, and their practical application is not effective. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting semiconductor process measurement data, which can be used to obtain wafer measurement data accurately, quickly and at low cost in semiconductor processes.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting semiconductor process measurement data, comprising: providing historical measurement data of a wafer and historical raw process data of the semiconductor process corresponding to the historical measurement data, wherein the historical raw process data includes first time series data; using an autoencoder to shorten the time steps of the first time series data to obtain second time series data; using the second time series data and the historical measurement data to fit and train a prediction model, wherein the prediction model includes a transformer encoder and a multilayer perceptron, the transformer encoder is used to extract time series features, and the multilayer perceptron uses the time features to make predictions to obtain prediction results; and inputting the raw process data of the semiconductor process to be predicted into the autoencoder and the prediction model to obtain predicted measurement data for the corresponding wafer.

[0007] Optionally, both the historical raw process data and the raw process data to be predicted are FDC data of the semiconductor manufacturing process.

[0008] Optionally, the FDC data can be preprocessed to obtain the first time series data.

[0009] Optionally, the autoencoder includes an encoding module and a decoding module; the encoding module encodes the first time series data to obtain the second time series data; the decoding module has a structure symmetrical to the encoding module and is used to expand and reconstruct the second time series data to restore the first time series data.

[0010] Optionally, the encoding module includes a two-dimensional convolutional network, wherein the height of the convolutional kernel and the pooling kernel of the two-dimensional convolutional network are both 1, or the width of the convolutional kernel and the pooling kernel of the two-dimensional convolutional network are both 1.

[0011] Optionally, the transformer encoder includes two encoder layers, each of which includes a self-attention layer and a feedforward neural network layer.

[0012] Optionally, the self-attention layer includes a dropout layer, an attention layer, and a residual connection normalization layer. The dropout layer is used to prevent overfitting, the attention layer is used to focus on the correlation of sequence data at different time steps, and the residual connection normalization layer is used to prevent gradient explosion or gradient vanishing.

[0013] Optionally, the transformer encoder further includes an input module, which includes a mask layer and a position encoding layer.

[0014] Optionally, the multilayer perceptron includes a lambda layer, a mask pooling layer, a concatenate layer, and a dense layer.

[0015] Optionally, the multilayer sensor further includes a front-end input layer, which is connected in parallel to the concatenate layer. The front-end input layer contains measurement data of the wafer in the previous process of the semiconductor manufacturing process.

[0016] In summary, the semiconductor process measurement data prediction method provided by this invention involves: compressing the first time series data using an autoencoder to obtain a second time series data with a shortened time step, which can improve the computational efficiency of the subsequent prediction model without losing feature information; then processing the second time series data through the prediction model's transformer encoder to extract its time series features; and utilizing the self-attention mechanism and high parallel computing characteristics of the transformer encoder to improve the effectiveness, accuracy, and efficiency of obtaining time series features. Finally, predictive measurement data is obtained from the time series features using a multilayer perceptron, thereby achieving accurate, fast, and low-cost acquisition of wafer measurement data in semiconductor processes. Attached Figure Description

[0017] Those skilled in the art will understand that the accompanying drawings are provided to better understand the invention and do not constitute any limitation on the scope of the invention.

[0018] Figure 1 A flowchart of a method for predicting semiconductor process measurement data provided in an embodiment of this application;

[0019] Figure 2 A flowchart illustrating the prediction process of the autoencoder and prediction model provided in the embodiments of this application;

[0020] Figure 3 An architecture diagram of the prediction model provided in the embodiments of this application;

[0021] Figure 4 This is a schematic diagram showing a partial configuration of the transformer encoder provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, advantages, and features of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and are not drawn to scale, and are only used to facilitate and clarify the explanation of the embodiments of the invention. Furthermore, the structures shown in the drawings are often part of the actual structures. In particular, different figures may emphasize different aspects and may sometimes use different scales.

[0023] As used in this invention, the singular forms “a,” “an,” and “the” include plural objects; the term “or” is generally used to mean “and / or”; the term “a number” is generally used to mean “at least one”; and the term “at least two” is generally used to mean “two or more”. Furthermore, the terms “first,” “second,” and “third” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as “first,” “second,” or “third” may explicitly or implicitly include one or at least two of that feature, unless otherwise expressly indicated.

[0024] Figure 1 A flowchart illustrating a method for predicting semiconductor process measurement data provided in an embodiment of this application.

[0025] like Figure 1 As shown, the method for predicting semiconductor process measurement data provided in this embodiment includes:

[0026] S01: Provide historical measurement data of a wafer and historical raw process data of semiconductor manufacturing process corresponding to the historical measurement data, wherein the historical raw process data includes first time series data;

[0027] S02: Use an autoencoder to shorten the time step of the first time series data to obtain the second time series data;

[0028] S03: The prediction model is fitted and trained using the second time series data and the historical measurement data. The prediction model includes a transformer encoder and a multilayer perceptron. The transformer encoder is used to extract time series features, and the multilayer perceptron uses the time features to make predictions to obtain prediction results.

[0029] S04: Input the raw process data to be predicted for the semiconductor manufacturing process into the autoencoder and the prediction model to obtain the predicted measurement data of the corresponding wafer.

[0030] Figure 2 This is a schematic diagram of the prediction of the autoencoder and prediction model provided in this embodiment. The following will combine... Figure 2 This paper provides a detailed introduction to the prediction methods for semiconductor process measurement data.

[0031] First, step S01 is executed, providing historical measurement data of a wafer and historical raw process data of the semiconductor process corresponding to the historical measurement data. The historical raw process data includes first time series data.

[0032] Semiconductor manufacturing processes for wafers can include all processes such as photolithography, diffusion, thin film processing, polishing and cleaning, and etching. Specifically, the equipment used can be the main equipment (critical equipment) for the corresponding semiconductor process. Its raw process data can be the FDC data of the main equipment for that process. Wafer measurement data can include key parameters (parameters that affect downstream processes or customers) to alleviate some computational burden. Of course, the number of key parameters is not limited to one; it can be greater than or equal to two. Historical wafer measurement data is generally obtained by sampling a subset of wafers. The corresponding historical raw process data can be obtained from the FDC system based on the ID of the sampled wafer.

[0033] In this embodiment, all FDC data from the wafer during the semiconductor manufacturing process can be utilized to obtain more (all) dimensions of process data. No subjective filtering of the feature information in the FDC data is applied, thereby achieving higher accuracy in predicting measurement data. It should be understood that in a semiconductor manufacturing process (machine), up to dozens of sensors can simultaneously (or nearly simultaneously) acquire different FDC data (sensor data) to monitor fluctuations in the semiconductor process. If some sensor data that is subjectively considered potentially important (highly correlated) is selected and then processed using limited data such as mean, extreme values, and slope before being used for predicting wafer measurement data, the calculation process is not only complex and the generalization of the calculation method is poor, but a considerable amount of information is also filtered out during the calculation process, making it difficult to guarantee the accuracy of the calculation.

[0034] After obtaining the historical measurement data of the wafer and the corresponding semiconductor process FDC data for that wafer, the FDC data is preprocessed to ensure that the sensor data of all dimensions have the same form (sampling interval or time step, etc.) to obtain the first time series data. In the semiconductor process, each sensor data point of the wafer's FDC data is a time series data of a certain length step (lasting a certain time). However, the original sampling frequency of different sensor data may vary, and the FDC data fluctuates significantly in the early and late stages of the wafer semiconductor process (e.g., when the wafer is moved in or out). Therefore, all sensor data can be resampled or interpolated to have the same sampling rate, and a small amount of redundant data at the beginning and end of the time step can be removed to ensure the same time step (length). Of course, if the time step of some sensor data is still insufficient, the mean can be used to make up the difference. The first time series data X1 after the above processing can be, for example, as follows:

[0035]

[0036] Where n represents the number of sensors in the FDC data, m represents the time step (first time step), and x nmDenote the sensing data of the nth sensor at the mth time step.

[0037] A first time series data corresponds to the historical measurement data of a wafer (or the same batch of wafers with the same original process data). The sample data set includes multiple first time series data and corresponding historical measurement data. The more data copies in the sample data set, the more beneficial it is for model training.

[0038] Next, execute step S02, and use an autoencoder to perform time step shortening processing on the first time series data to obtain second time series data.

[0039] The autoencoder can perform feature extraction on the first sequence data X1 in terms of time steps to obtain compressed time steps under the condition of lossless feature information, and can also perform noise reduction processing on the first sequence data X1 during this process. The autoencoder includes an encoding module and a decoding module. The encoding module performs feature extraction on the first time series data to obtain second time series data. The decoding module has a symmetric structure relative to the encoding module and is used to unfold and reconstruct the second time series data to restore the first time series data to ensure no loss of feature information.

[0040] Specifically, the encoding module of the autoencoder can be a convolutional neural network (CNN). This convolutional neural network includes at least one hidden layer, and this hidden layer includes a two-dimensional convolutional layer, an activation function layer, and a max pooling layer. Preferably, the shape (size) of the convolutional kernel of the two-dimensional convolutional layer and the pooling kernel of the max pooling layer is rectangular, and the width or height of this rectangle is 1 at the same time, so as to encode the sensing data of each dimension along the time steps respectively to compress the time steps without affecting the data features between different sensors. In this embodiment, the encoding module of the autoencoder includes one hidden layer. The convolutional kernel size of the two-dimensional convolutional network is (3,1), the activation function is LeakyReLu, the size of the max pooling layer is (4,1), and the padding type is'same' (keeping the size unchanged). After the autoencoder is unsupervised trained using the encoding module and the decoding module, the output of the encoding module is used as the second time series data X2. The second time series data X2 can be, for example

[0041]

[0042] where n represents the number of sensors in the FDC data, h represents the time step after time step shortening (i.e., the second time step), and h < m, that is, the second time step is less than the first time step.

[0043] All first time series data are compressed by time step to obtain second time series data. The second time series data is then used to replace the corresponding first time series data in the sample dataset to form an updated sample dataset.

[0044] Next, step S03 is executed, which uses the second time series data and historical measurement data to fit and train the prediction model. The prediction model includes a transformer encoder and a multilayer perceptron. The transformer encoder is used to extract time series features, and the multilayer perceptron uses the time features to make predictions to obtain prediction results.

[0045] The specific process of building and training a prediction model may include, for example, the following steps:

[0046] S031: Divide the updated sample dataset into a training set and a validation set (or test set);

[0047] S032: Construct a prediction model, which includes a transformer encoder and a multilayer perceptron;

[0048] S033: Use the training set to train the hyperparameters of the prediction model, and use the validation set to validate (or test) the model.

[0049] In step S031, the sample dataset can be randomly divided into a training set and a validation set in a ratio of 8:2 or 7:3. The second time series data is used as input and the corresponding historical measurement data is used as output (label) in the training set and validation set.

[0050] In step S032, as Figure 3 As shown, the prediction model includes an input module, a transformer encoder, and a multilayer perceptron (MLP). The input module includes a masking layer and a positional encoding layer. The masking layer uses a mask tensor to pad the time series data and mask invalid regions, while the positional encoding layer encodes the time series data at time steps. The transformer encoder includes one or more encoder layers with similar structures, which are used to extract time series features. The multilayer perceptron includes lambda layers, max pooling layers, pooling layers, concatenation layers, and dense layers to regress the time series features to obtain the prediction structure.

[0051] Specifically, in this embodiment, the encoder layer has two layers, such as... Figure 4As shown, each encoder layer includes a self-attention layer and a feedforward neural network layer. The self-attention layer includes a first dropout layer, an attention layer, and a residual connection normalization layer. The feedforward neural network layer includes a convolutional layer, a second dropout layer, and a residual connection normalization layer. The attention layer focuses on the influence of different time steps and includes a third dropout layer to prevent overfitting. The convolutional layers are two-dimensional, with the kernel size being 1 (rectangular) in width or height. The random probabilities of the first, second, and third dropout layers can be, for example, 0.5.

[0052] In a multilayer perceptron, the concatenate layer fuses (concatenates) the features extracted by the lambda layer, max pooling layer, and mean pooling layer before outputting them to the dense layer. The activation function of the dense layer can be sigmoid. Preferably, the multilayer perceptron may also include a pre-step input layer connected in parallel to the concatenate layer, which contains measurement data from the previous process (step) of the wafer in the semiconductor process, so that the corresponding measurement data can be obtained directly using the pre-step measurement data without the need for other transformations.

[0053] In step S033, the loss function, network optimization algorithm and all hyperparameters required for training are constructed. The prediction model is trained using the training data to obtain the weights and parameters of the prediction model. The prediction model is then validated using the validation set.

[0054] Next, step S04 is executed, in which the original process data to be predicted for the semiconductor manufacturing process is input into the time step shortening model and the prediction model to obtain the predicted measurement data of the corresponding wafer.

[0055] By utilizing the aforementioned time-step reduction model (including preprocessing) and prediction model, not only can predicted measurement data be obtained for each wafer (or batch of wafers) within approximately 2 minutes of acquiring the original process data under relatively limited hardware and software conditions, but this predicted measurement data also boasts high accuracy. This enables full measurement of wafers in semiconductor manufacturing processes, facilitating monitoring of these processes. It should be understood that the aforementioned time-step reduction model shortens the time step by removing redundant and noisy data from the time-series data to improve subsequent computational efficiency. The prediction model uses all the numerical and temporal characteristics of the sensor data from each dimension (sensor) of the input time-series data to predict the measurement data, thus achieving superior accuracy.

[0056] In summary, the semiconductor process measurement data prediction method provided by this invention involves: compressing the first time series data using an autoencoder to obtain a second time series data with a shortened time step, which can improve the computational efficiency of the subsequent prediction model without losing feature information; then processing the second time series data through the prediction model's transformer encoder to extract its time series features; and utilizing the self-attention mechanism and high parallel computing characteristics of the transformer encoder to improve the effectiveness, accuracy, and efficiency of obtaining time series features. Finally, predictive measurement data is obtained from the time series features using a multilayer perceptron, thereby achieving accurate, fast, and low-cost acquisition of wafer measurement data in semiconductor processes.

[0057] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A method for predicting semiconductor process measurement data, characterized in that, include: Provide historical measurement data of a wafer and historical raw process data of semiconductor manufacturing process corresponding to the historical measurement data, wherein the historical raw process data includes first time series data; An autoencoder is used to shorten the time step of the first time series data to obtain the second time series data. The autoencoder includes an encoding module and a decoding module. The encoding module encodes the first time series data to obtain the second time series data. The encoding module includes a two-dimensional convolutional network. The height of the convolution kernel and the pooling kernel of the two-dimensional convolutional network are both 1, or the width of the convolution kernel and the pooling kernel of the two-dimensional convolutional network are both 1. The prediction model is fitted and trained using the second time series data and the historical measurement data. The prediction model includes a transformer encoder and a multilayer perceptron. The transformer encoder is used to extract time series features, and the multilayer perceptron uses the time series features to make predictions to obtain prediction results. The multilayer perceptron includes a lambda layer, a max pooling layer, an average pooling layer, a preprocessing input layer, a concatenate layer, and a dense layer. The concatenate layer fuses the features extracted by the lambda layer, the max pooling layer, the average pooling layer, and the preprocessing input layer, and then outputs them to the dense layer. The preprocessing input layer is the measurement data of the wafer in the previous process of the semiconductor manufacturing process. The raw process data to be predicted for the semiconductor manufacturing process is input into the autoencoder and the prediction model to obtain the predicted measurement data of the corresponding wafer. The historical raw process data and the raw process data to be predicted are both FDC data of the semiconductor manufacturing process.

2. The method for predicting semiconductor process measurement data according to claim 1, characterized in that, The FDC data is preprocessed to obtain the first time series data.

3. The method for predicting semiconductor process measurement data according to claim 1, characterized in that, The decoding module has a structure symmetrical to the encoding module and is used to expand and reconstruct the second time series data to restore the first time series data.

4. The method for predicting semiconductor process measurement data according to claim 1, characterized in that, The transformer encoder includes two encoder layers, each of which includes a self-attention layer and a feedforward neural network layer.

5. The method for predicting semiconductor process measurement data according to claim 4, characterized in that, The self-attention layer includes a dropout layer, an attention layer, and a residual connection normalization layer. The dropout layer is used to prevent overfitting, the attention layer is used to focus on the correlation of sequence data at different time steps, and the residual connection normalization layer is used to prevent gradient explosion or gradient vanishing.

6. The method for predicting semiconductor process measurement data according to claim 4, characterized in that, The transformer encoder also includes an input module, which includes a mask layer and a position encoding layer.

Images