Semiconductor etching uniformity control system based on remote plasma source

Abstract

The invention relates to the technical field of semiconductor etching, and discloses a semiconductor etching uniformity control system based on a remote plasma source. The system comprises a multi-source data acquisition and processing module which is used for collecting and preprocessing multi-source sensor data in a remote plasma source etching process; the historical etching data indexing module is used for indexing historical etching data according to the current etching target parameters; the etching condition classification module is used for classifying the etching condition information in the historical etching data to obtain a plurality of classified etching conditions; the data optimization and dimension reduction module is used for carrying out optimization and dimension reduction on the plurality of classified data sets to obtain a dimension reduction data sequence; the etching trend analysis module is used for carrying out etching uniformity trend analysis according to the dimension reduction data sequence to obtain a change rate and a degradation rate; and the uniformity control decision module is used for performing control decision according to the change rate and the degradation rate so as to obtain an etching uniformity control scheme. The system can improve the accuracy and adaptability of etching uniformity control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor etching, in particular to a semiconductor etching uniformity control system based on a remote plasma source. BACKGROUND

[0002] In the field of semiconductor manufacturing, etching process as a key link of pattern transfer, its uniformity directly affects the performance and yield of chips. With the continuous reduction of integrated circuit feature size, the requirement for etching precision is increasingly stringent, especially in advanced processes such as 3D NAND and FinFET, small deviation of etching uniformity may lead to device failure. Remote plasma source has become one of the mainstream technologies for advanced semiconductor etching due to its high plasma density and low damage, but in practical application, etching uniformity still faces many challenges.

[0003] The etching process involves plasma parameters, gas flow, RF power, wafer temperature and other multi-dimensional variables. The dynamic fluctuations of these parameters will directly affect the etching rate and profile consistency. For example, uneven spatial distribution of plasma density will cause differences in etching rate in different areas of the wafer, and instantaneous changes in gas flow may cause local imbalance in etching chemical reaction. In addition, the aging of equipment components after long-term operation, such as the accumulation of deposition layer on the reaction chamber wall and the decline of RF power output stability, will gradually exacerbate the drift of etching uniformity.

[0004] Traditional etching uniformity control methods are mostly based on pre-set fixed parameter recipes, maintaining process stability through regular calibration, but this approach is difficult to cope with real-time disturbances in complex dynamic environments. Some improved solutions introduce feedback adjustment of a single sensor, such as monitoring plasma state through optical emission spectroscopy or measuring wafer temperature through thermocouple, but the amount of information from a single data source is limited, making it difficult to fully reflect the overall state of the etching system, resulting in insufficient control accuracy. At the same time, the process rules contained in historical etching data have not been fully explored, and cannot provide effective reference for real-time control, so that the system needs to be repeatedly adjusted to achieve the desired effect when facing new etching targets, seriously affecting production efficiency.

[0005] With the development of semiconductor devices towards higher integration and more complex structure, higher requirements are put forward for the control of etching uniformity, and the limitations of existing control methods in multi-parameter coupling analysis, dynamic trend prediction and adaptive decision-making are increasingly prominent, requiring a systematic solution that can integrate multi-source information, exploit the value of historical data and achieve precise control. SUMMARY

[0006] The purpose of the present application is to provide a semiconductor etching uniformity control system based on a remote plasma source to solve the problems raised in the background.

[0007] To achieve the above object, the application provides a semiconductor etching uniformity control system based on a remote plasma source, which comprises:

[0008] The core functional modules of the system work together to achieve etching uniformity control. The multi-source data acquisition and processing module receives various sensor data from the interior of the etching chamber and the remote plasma source, including but not limited to plasma emission spectrum (OES), radio frequency power, gas flow, chamber pressure, substrate temperature, etc. This module groups the received raw sensor data by data type, applies time window resampling technology to each group of data, and ensures that data streams of different sampling rates are aligned on the time stamp. After alignment, the data is processed by unit standardization and normalization to eliminate dimensional differences and form a preprocessed data stream.

[0009] The historical etching data index module receives the target parameter setting values of the current etching process. Based on these target parameters, the module performs a query operation in the historical process database to retrieve a set of historical etching process records similar or matching the current target parameters.

[0010] The etching condition classification module analyzes the retrieved set of historical etching data in depth. This module extracts etching condition information from each historical record, including process parameters, equipment status, environmental parameters, etc. Using clustering algorithms or rule-based methods, these historical etching condition information is classified into multiple groups of classified etching conditions with similar characteristics, each group corresponding to a classified data set.

[0011] The data optimization and dimension reduction module processes each classified data set. This module applies dimension reduction techniques to extract key features from each classified data set, removes redundant information, and compresses high-dimensional historical data into low-dimensional data sequences that better reflect the essential characteristics of the data, i.e. dimension-reduced data sequences. These sequences usually retain the key patterns in the original data related to etching uniformity changes.

[0012] The etching trend analysis module receives the reduced dimension data sequence. This module applies time series analysis techniques to model and analyze the reduced dimension data sequence. The core objective of the analysis is to calculate the rate of change of the etching uniformity index over time, characterizing how fast the uniformity is changing, and the degradation rate, characterizing the degree of deviation from the ideal state or the deterioration trend. These quantitative indicators reflect the current or predicted etching uniformity dynamics. The uniformity control decision module makes decisions based on the rate of change and degradation rate information output by the etching trend analysis module. This module usually contains a pre-set control rule library or more advanced decision algorithms. The decision-making process involves evaluating the current uniformity state and its development trend, predicting the possible consequences if no intervention is made, and then generating a corresponding control scheme. The specific form of the control scheme includes adjusting the operating parameters of the remote plasma source, adjusting the chamber pressure, optimizing the substrate temperature distribution, changing the etching time, or triggering the equipment maintenance program, etc., aiming to correct or optimize the etching uniformity in real time. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A timing diagram for the remote plasma source-based semiconductor etching uniformity control system described in the present invention;

[0014] Figure 2 A flowchart for conditional matching correction and decision control;

[0015] Figure 3 A flowchart for plasma state reward-driven decision-making;

[0016] Figure 4 A flowchart for dynamic device association and trend analysis;

[0017] Figure 5 A flowchart for benchmark score-driven probabilistic dimension reduction.

[0018] Example 1: Refer to Figure 2 The conditional matching correction module receives the real-time preprocessed data stream output by the multi-source data acquisition and processing module, which contains the actual parameter state at the current etching process running time, such as radio frequency power instantaneous value, chamber pressure reading, gas flow measured value and substrate temperature distribution, etc. This module quantifies the current etching conditions into a feature vector, whose dimension is consistent with the classification feature dimension defined by the historical etching condition classification module. The matching process uses the weighted Euclidean distance algorithm: calculate the distance value of the current feature vector and each classified etching condition standard feature vector (representing the typical parameter combination of historical statistics under this classification). Different parameters are assigned different weight coefficients in the distance calculation, and the weight is dynamically set according to the sensitivity of the parameter to uniformity. The matching result is output as the first K classified conditions with the smallest distance, i.e. the matching etching condition set. For each matching condition, the module calculates the deviation vector of the current actual parameter from its standard parameter, such as radio frequency power deviation , pressure deviation , main gas flow bias . The bias vector is input to a correction function, which is a multi-layer perceptron model trained based on historical data, with its hidden layers containing non-linear activation functions. The model takes the aforementioned bias as input and outputs a correction factor matrix. This matrix is subjected to Hadamard product operation with the original change rate and degradation rate provided by the etching trend analysis module, generating the corrected change rate and corrected degradation rate. The correction process introduces the sign directionality of the bias parameter: if the current power is higher than the standard value and the history shows that high power accelerates uniformity degradation, then the degradation rate correction factor is greater than 1; otherwise it is less than 1. The uniformity control decision module receives the corrected trend indicators and generates control instructions in combination with the pre-set rule library, finally outputting an etching uniformity control scheme containing a parameter adjustment sequence.

[0019] Example 2: see Figure 3 , the plasma state generation module receives initial plasma parameters, including the average intensity of specific wavelengths extracted by the optical emission spectrum sensor, the electron density estimate value, and the initial reading of the electron temperature. The initial latent space encoding state is generated by a pre-trained variational autoencoder: the encoder network contains a multi-layer convolutional structure and a fully connected layer, which compresses the high-dimensional input into a fixed-dimensional latent variable. The sequence prediction model adopts a gated recurrent unit architecture, taking the aforementioned latent variable as the initial state and combining time step encoding to output a predicted sequence of plasma states for multiple future time steps. The sequence includes the predicted electron density values, characteristic spectral line intensity values, and spatial uniformity indices at each time point.

[0020] The reward calculation module obtains real-time actual plasma state data: the actual spectral line intensity at a specific time is collected by the spectrum sensor, and the actual electron density is measured by the probe. The progress reward value is calculated based on the offset between the actual etching progress and the predicted time axis: when the offset is less than a set threshold, a positive reward is given, and when the offset exceeds the threshold, a negative reward is given in proportion to the offset. The exploration reward value gives a fixed reward value to operations that deviate from the historical optimal parameters but are still within the safe range. The total reward value is a weighted combination of the basic reward, the progress reward, and the exploration reward, and the weight coefficients are dynamically adjusted through reinforcement learning algorithms.

[0021] The uniformity control decision module adopts a deep deterministic policy gradient framework: the execution network outputs a continuous action vector, and the evaluation network analyzes the state-action value function. The decision-making process aims to maximize the cumulative discounted reward, generating an etching uniformity optimization scheme containing multiple control actions.

[0022] Example 3: see Figure 4, the etching device correlation graph built by the dynamic device correlation updating module adopts an attribute graph model: the node types include a radio frequency subsystem, a gas path subsystem, a vacuum subsystem, and a sensing subsystem. Edge attributes include connection types and influence coefficients. Graph updating is realized through a time sequence graph convolution network (TGCN): when the multi-source data acquisition and processing module inputs new sensor data, the TGCN extracts the node feature change amount , through a graph convolution layer

[0023]

[0024] wherein: : a feature vector of a node v at an (l+1)th layer, : a ReLU activation function, : a neighbor node set of the node v, : a trainable weight matrix of an lth layer, : a feature vector of a neighbor node u at the lth layer, : a feature change amount of the node v, : a number of neighbor nodes.

[0025] In the updated graph, if the abnormality degree of a certain node exceeds a threshold value, the correlation path analysis is triggered. The fault diagnosis and prediction module adopts a graph attention mechanism to calculate the propagation influence weight of the abnormal node on the wafer uniformity node, and combines an LSTM time sequence model to identify a fault mode. The device life prediction module builds a physical degradation model for a radio frequency power supply:

[0026] wherein: : a remaining use life (hours) of the radio frequency power supply, : a rated life coefficient (1.2x10^7 is taken), : a historical average operating power (kW), : a cumulative working time (hours) When the fault prediction confidence is greater than 80% or the RUL is less than 200 hours, the device aging state is marked as "critical". The uniformity control decision module introduces a device constraint in the control scheme: if the device state is "critical", the radio frequency power fluctuation range is limited to ≤±5%, and a maintenance suggestion code is appended at the end of the scheme.

[0027] Embodiment 4: refer to Figure 5 After the etching condition classification module completes the historical data classification, 20 batches of historical etching data are extracted for the classified group A (condition: pressure = 50-55 mTorr, power = 1500-1600 W). The historical uniformity score calculation adopts a wafer in-plane uniformity formula: score = 100x(1-3σ / μ), wherein σ is an etching depth standard deviation, and μ is a mean value. An example of the extracted historical uniformity score set is shown in Table 1.

[0028] Table : historical etch uniformity score dataset.

[0029]

[0030] The data optimization dimension reduction module selects the score 92.3 of the batch BATCH_05 as the first reference uniformity score. The allocation probability is calculated using a Gaussian kernel function and then normalized. The weighted covariance matrix is constructed using the probability weighted principal component analysis (WPCA), and the eigenvector corresponding to the maximum eigenvalue is taken as the dimension reduction direction. The original 6-dimensional score set is projected into a 1-dimensional space to obtain the dimension reduction score set [-0.41, -0.18, 2.97, -1.12, 0, 0.74]. The sequence is generated in the order of batch time: [-1.12 (BATCH_04), -0.41 (BATCH_01), -0.18 (BATCH_02), 0 (BATCH_05), 0.74 (BATCH_06), 2.97 (BATCH_03)]. The etching trend analysis module performs first-order difference on the sequence to obtain the change rate sequence [0.71, 0.23, 0.18, 0.74, 2.23], and the degradation rate is calculated as the root mean square deviation of the current value relative to the target value 1.87.

[0031] In Example 5, the original sensor data received by the multi-source data acquisition and processing module is divided into seven categories according to physical characteristics: spectral data group, power data group, pressure data group, flow data group, temperature data group, position data group, and vacuum data group. The time window resampling sets the window length to 500 ms and the sliding step to 250 ms. The spectral data group has an original sampling rate of 1 kHz, and the root mean square value of the intensity of each channel is calculated within the window as the representative value. The temperature group has a sampling rate of 10 Hz, which is supplemented to 500 ms intervals through cubic spline interpolation. Unit normalization converts OES intensity to relative intensity units (RIU), pressure to mTorr, and power to kW. The normalization process uses a dynamic range:

[0032]

[0033] wherein: : normalized parameter value, : original parameter value, : minimum value of the parameter in the last 24 hours, : The maximum value of this parameter in the last 24 hours. Data quality detection adopts double mechanism: missing values are filled by linear interpolation of valid data points before and after; the sliding window Z-score method is used to detect abnormal values, which are considered abnormal when |Z|>3 and are replaced by the median value in the window. Data verification includes integrity check (continuous missing window number <3), range verification (normalized value ∈ [0,1]), and logical consistency. Data segments that fail the verification trigger resampling requests, and the preprocessed data stream is output to the downstream module, with a data structure of timestamp-aligned two-dimensional tensor.

[0034] It should be noted that the relational terms herein, such as first and second, are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0035] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A remote plasma source based semiconductor etch uniformity control system, comprising: The system comprises: a multi-source data collection processing module, configured to collect and pre-process multi-source sensor data from a remote plasma source etching process; a historical etching data indexing module, configured to obtain a current etching target parameter, and index historical etching data according to the current etching target parameter; an etching condition classification module, configured to classify etching condition information in the historical etching data, and obtain a plurality of classified etching conditions; a data optimization and dimension reduction module, configured to optimize and reduce dimensions of a plurality of classified data sets, and obtain a reduced dimension data sequence; an etching trend analysis module, configured to perform etching uniformity trend analysis according to the reduced dimension data sequence, and obtain a change rate and a degradation rate; a uniformity control decision module, configured to perform control decision according to the change rate and the degradation rate, and obtain an etching uniformity control scheme.

2. The remote plasma source based semiconductor etch uniformity control system of claim 1, wherein, The system further comprises a condition matching correction module, configured to match a current etching condition with the plurality of classified etching conditions, obtain a matching etching condition, and correct the change rate and the degradation rate according to a deviation of a standard matching condition between the current etching condition and the matching etching condition, and obtain a corrected change rate and a corrected degradation rate; The uniformity control decision module is configured to perform control decision according to the corrected change rate and the corrected degradation rate.

3. The remote plasma source based semiconductor etch uniformity control system of claim 1, wherein, The system further comprises a plasma state generation module, configured to generate an expected plasma state sequence; and a reward calculation module, configured to calculate a reward value according to a deviation between an actual plasma state and the expected plasma state sequence; The uniformity control decision module is configured to perform control decision according to the reward value.

4. The remote plasma source based semiconductor etch uniformity control system of claim 1, wherein, The system further comprises a dynamic device association updating module, configured to establish an etching device association graph, and dynamically update the etching device association graph based on data pre-processed by the multi-source data collection processing module; The etching trend analysis module is configured to perform etching uniformity trend analysis based on the etching device association graph.

5. The remote plasma source based semiconductor etch uniformity control system of claim 2, wherein, The etching condition classification module is further configured to extract a historical uniformity information set and a plurality of historical etching data sets under the plurality of classified etching conditions, and process to obtain a plurality of historical uniformity score sets; The data optimization and dimension reduction module is configured to optimize and reduce dimensions of the plurality of historical uniformity score sets.

6. The remote plasma source-based semiconductor etch uniformity control system of claim 5, wherein, The data optimization dimension reduction module is configured to select a first historical uniformity score set from the plurality of historical uniformity score sets, obtain a first benchmark uniformity score, and assign a distribution probability according to a difference between other historical uniformity scores in the first historical uniformity score set and the first benchmark uniformity score, to obtain a first basic probability distribution, wherein the size of the difference and the size of the distribution probability are negatively correlated, and then optimize and reduce dimensions of the first historical uniformity score set according to the first basic probability distribution, to obtain a first reduced-dimension uniformity score set, and arrange the first historical uniformity score set in a time sequence to obtain a first historical uniformity score sequence; The etching trend analysis module is configured to analyze etching uniformity trends according to the first historical uniformity score sequence.

7. The remote plasma source based semiconductor etch uniformity control system of claim 3, wherein, The plasma state generation module is configured to generate an initial latent space encoding state based on initial plasma state parameters, and predict an etching process state sequence according to the initial latent space encoding state. The reward calculation module is configured to calculate a distance reward value, a progress reward value, and an exploration reward value according to a deviation between the actual plasma state and the predicted etching process state sequence, and synthesize a total reward value. The uniformity control decision module is configured to make a control decision according to the total reward value.

8. The remote plasma source-based semiconductor etch uniformity control system of claim 4, wherein, The system further comprises a fault diagnosis prediction module configured to predict potential faults based on the etching equipment association graph and analysis results of the etching trend analysis module. The equipment life prediction module is configured to construct an equipment virtual model, and analyze equipment aging conditions based on prediction results of the fault diagnosis prediction module. The uniformity control decision module is configured to make a control decision based on the equipment aging conditions.

9. The remote plasma source-based semiconductor etch uniformity control system of claim 1, wherein, The multi-source data acquisition and processing module is configured to receive multi-source sensor data, group the same type of data, and perform time window resampling synchronization, and then perform unit standardization and normalization processing to unify data ranges and units. The historical etching data indexing module is configured to index historical etching data based on the processed data of the multi-source data acquisition and processing module.

10. The remote plasma source-based semiconductor etch uniformity control system of claim 9, wherein, The multi-source data acquisition and processing module is further configured to detect and process missing values and abnormal values, fill in missing values by linear interpolation, replace abnormal values by mean filling, verify data quality, and output preprocessed data. The etching condition classification module is configured to classify based on the preprocessed data.

Images