Semiconductor wafer defect detection method and system based on deep learning

Abstract

The invention belongs to the technical field of semiconductor device manufacturing, and particularly relates to a semiconductor wafer defect detection method and system based on deep learning, and the method comprises the steps: obtaining a radio frequency parameter observation value sequence, and carrying out the standardization and denoising processing; weighting the neural network feature extraction layer based on the change rate of the radio frequency parameter observation value to obtain a transient weighted feature value; obtaining a dynamic judgment threshold based on generalized Pareto distribution and a baseline drift compensation item; and comparing the transient weighted characteristic value with a dynamic judgment threshold value to obtain a defect detection result. According to the method, capturing of weak arc signals is enhanced through nonlinear amplification of transient fluctuation, the threshold value is dynamically adjusted according to the number of semiconductor wafers machined in an accumulated mode, the baseline drift influence caused by deposition of the inner wall of the reaction chamber is eliminated, and the defect detection rate is increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor device manufacturing. More particularly, the present application relates to a deep learning-based semiconductor wafer defect detection method and system. BACKGROUND

[0002] In the field of semiconductor device manufacturing, plasma etching process is the core step of defining circuit patterns. With the continuous miniaturization of process nodes, the stability of the internal electromagnetic environment of the reaction chamber directly determines the final processing yield of the semiconductor wafer. In order to monitor the stability of the etching process, the industry generally uses a fault detection and classification (FDC) system. The current mainstream monitoring strategy relies on statistical process control (SPC), that is, by collecting sensor data such as radio frequency voltage, current, and phase at the device end, a fixed upper and lower control limit is set based on historical normal data to determine whether the current process is abnormal.

[0003] However, the existing statistical control method faces severe challenges in advanced processes. The main reason is that plasma etching is not a strictly steady-state process, but includes complex transient stages such as plasma ignition and process gas switching. These transient stages with dramatic electromagnetic parameter fluctuations are often high-risk windows for defects such as micro-arc. Traditional statistical methods or standard deep learning models, such as recurrent neural networks (RNN), usually assign equal attention weights to data at different time steps when extracting features, making it difficult to effectively distinguish between normal fluctuations and actual defect-induced transient signal mutations in process transient stages, and thus causing the dilemma of missing weak defect signals or frequent false alarms for normal transient fluctuations.

[0004] In addition, another key technical bottleneck is the gradual drift of device state. During the continuous operation of the etching device, polymer by-products inevitably deposit on the inner wall of the reaction chamber, causing slow and nonlinear changes in the electrical impedance characteristics of the reaction chamber, which is known as process baseline drift. Traditional detection models usually assume that the data distribution of normal processes is static and unchanging, so the fixed dynamic decision threshold cannot adapt to the physical drift of the device baseline. As the number of processed semiconductor wafers increases, the baseline of the normal radio frequency parameter observation sequence may gradually approach or even exceed the preset dynamic decision threshold, causing the fault detection system to issue a large number of false alarms in the later maintenance period, forcing engineers to frequently adjust model parameters or perform unnecessary downtime cleaning. SUMMARY

[0005] To solve the technical problems that the conventional statistical process control method cannot distinguish between normal fluctuations and abnormal defects of process transients, and the false alarm rate is high due to equipment aging, the present application provides solutions in the following aspects.

[0006] In a first aspect, the present application provides a deep learning-based semiconductor wafer defect detection method, comprising: acquiring a sequence of radio frequency parameter observations through a sensor installed at the output end of a radio frequency matching network; performing standardization and denoising processing on the sequence of radio frequency parameter observations to obtain standardized radio frequency parameter observations; weighting a neural network feature extraction layer based on the rate of change of the standardized radio frequency parameter observations to obtain transient weighted feature values; obtaining a dynamic decision threshold based on a generalized Pareto distribution and a baseline drift compensation term, comparing the transient weighted feature values with the dynamic decision threshold to obtain a defect detection result of the semiconductor wafer.

[0007] The present application weights the neural network feature extraction layer based on the rate of change of the radio frequency parameter observations, and obtains the dynamic decision threshold based on the generalized Pareto distribution containing the baseline drift compensation term, which can dynamically adjust the attention weight according to the severity of signal fluctuations, and adjust the decision standard according to the number of accumulated processed semiconductor wafers, thereby enhancing the capture ability of weak arc signals in the transient process stage in semiconductor wafer defect detection, and reducing false alarms caused by impedance baseline drift due to the accumulation of deposits on the inner wall of the reaction chamber.

[0008] Preferably, the sequence of radio frequency parameter observations is acquired through a sensor installed at the output end of a radio frequency matching network, comprising: synchronously collecting radio frequency voltage, radio frequency current, phase angle, forward power and reflected power at a microsecond-level sampling rate using a radio frequency impedance sensor; synchronously aligning the collected data to form the sequence of radio frequency parameter observations.

[0009] The present application synchronously collects multi-dimensional data at a microsecond-level sampling rate using a radio frequency impedance sensor installed at the output end of a radio frequency matching network, and synchronously aligns the collected data to form a sequence of radio frequency parameter observations, which can obtain high-frequency physical signals containing complete process recipe cycles, thereby providing a digital basis reflecting millisecond-level microscopic state fluctuations inside the plasma for semiconductor wafer defect detection.

[0010] Preferably, the sequence of radio frequency parameter observations is standardized and denoised, comprising: mapping the sequence of radio frequency parameter observations to a zero-mean unit variance space using the Z-Score standardization method; filtering high-frequency random noise using a moving average filter.

[0011] This invention uses the Z-score normalization method to map the radio frequency parameter observation sequence to a zero-mean unit variance space and uses a moving average filter to filter out high-frequency random noise. This can eliminate the numerical differences between different physical dimensions such as power and phase angle, while retaining the principal component signal that reflects the characteristics of plasma dynamics, thus providing a high signal-to-noise ratio standard input for deep learning models.

[0012] Preferably, the transient weighted eigenvalues ​​satisfy the expression: In the formula, This represents the transient weighted feature value extracted at time t; The standard weight matrix representing a neural network; Indicates the bias term; Represents the observed radio frequency parameters at time t; This represents the observed radio frequency parameters at the previous moment; Represents the transient sensitivity coefficient; Indicates the nonlinear amplification index; Represents the hyperbolic tangent activation function; This represents absolute value operations.

[0013] This invention introduces a transient weighted feature value acquisition mechanism in the feature extraction layer. By nonlinearly amplifying the absolute difference value of the observed radio frequency parameters, the gain of feature extraction is dynamically adjusted, forcing the model to allocate higher weights to the moment when the signal undergoes a sharp change, thereby accurately capturing the weak arc signal hidden in rapid transient processes such as ignition or gas switching.

[0014] Preferably, the nonlinear amplification index is 2.0.

[0015] Preferably, the transient sensitivity coefficient is obtained based on historical fault samples through gradient-weighted class activation mapping analysis.

[0016] Preferably, the dynamic determination threshold satisfies the expression: In the formula, express The dynamic judgment threshold calculated at any given time; Indicates the index of discrete-time sampling points; express The location parameters of the generalized Pareto distribution are updated in real time; express The scale parameters of the generalized Pareto distribution are updated in real time; express The shape parameters of the generalized Pareto distribution are updated in real time; Indicates the preset confidence level; This indicates the current cumulative number of semiconductor wafers processed; This represents the baseline drift coefficient.

[0017] This invention obtains a dynamic judgment threshold based on a generalized Pareto distribution and a baseline drift compensation term. By introducing an additive compensation term related to the current cumulative number of semiconductor wafers processed, the judgment baseline can rise synchronously with the impedance background noise statistical baseline caused by the thickening of deposits on the inner wall of the reaction chamber, thereby maintaining a safe distance between the dynamic judgment threshold and the background noise baseline and reducing false alarms during equipment aging.

[0018] Preferably, the baseline drift coefficient is obtained by performing linear regression fitting on the historical mean data of the previous maintenance cycle.

[0019] Preferably, the confidence level is 99.9%.

[0020] Secondly, the present invention provides a semiconductor wafer defect detection system based on deep learning, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned semiconductor wafer defect detection method based on deep learning is implemented.

[0021] By adopting the above technical solution, a computer program for the semiconductor wafer defect detection method based on deep learning is generated and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0022] The beneficial effects of this invention are as follows: This invention, by embedding a dynamic attention weighting mechanism based on the rate of signal change into a recurrent neural network and using a nonlinear amplification exponent to process the transient fluctuation difference value, can enhance the model's sensitivity to weak defect signals in key process stages such as ignition and gas switching. This solves the problem that traditional methods have difficulty distinguishing between normal process transient fluctuations and abnormal arc defects, and improves the defect detection capability in semiconductor wafer processing.

[0023] This invention obtains a generalized Pareto distribution model that includes a baseline drift compensation term, and dynamically adjusts the dynamic judgment threshold based on the cumulative number of semiconductor wafers processed. This automatically offsets the influence of RF parameter impedance baseline drift caused by polymer deposition on the inner wall of the reaction chamber, thereby preventing false alarms caused by normal signals touching fixed thresholds in the later stages of the process and ensuring the long-term operational stability of the semiconductor wafer etching equipment.

[0024] This invention combines microsecond-level multidimensional radio frequency parameter acquisition with deep learning feature analysis. Through the collaborative work of data standardization preprocessing and adaptive threshold determination, it can extract pure process features and make accurate judgments in complex electromagnetic interference environments. This provides a highly adaptable and robust monitoring solution for plasma etching processes in advanced manufacturing processes, reducing the risk of semiconductor wafer scrap. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a semiconductor wafer defect detection method based on deep learning according to the present invention; Figure 2 This is a schematic diagram illustrating the transient weighted features; Figure 3 This is a schematic diagram illustrating the dynamic threshold determination. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] This invention discloses a semiconductor wafer defect detection method based on deep learning, referring to... Figure 1 This includes steps S1-S4: S1. Obtain the sequence of radio frequency parameter observations through a sensor installed at the output of the radio frequency matching network.

[0029] It should be noted that in semiconductor plasma etching processes, the stability of the electromagnetic environment inside the reaction chamber directly determines the processing yield of semiconductor wafers. Traditional monitoring methods often have insufficient sampling frequencies, making it difficult to capture millisecond-level microscopic state fluctuations, which are often early signals of defect generation. To comprehensively capture the state changes inside the plasma, this invention establishes a high-frequency data acquisition channel covering multiple physical dimensions, converting simulated physical signals into digital time-series data, laying the physical foundation for subsequent in-depth analysis.

[0030] Specifically, this invention uses an RF impedance sensor installed at the output of the RF matching network to synchronously acquire RF voltage, RF current, phase angle, forward power, and reflected power at a microsecond-level sampling rate. The system synchronizes and aligns the timing data of these different physical dimensions to obtain a sequence of RF parameter observations containing the complete process formulation cycle, thereby completing the digital mapping of the underlying physical signals.

[0031] S2. The radio frequency parameter observation sequence is standardized and denoised to obtain standardized radio frequency parameter observations.

[0032] It should be noted that the collected RF parameter observations exhibit significant differences in units. For example, power data may reach kilowatt levels, while phase angle data may only be at the degree level. This imbalance in numerical scale can lead to difficulties in subsequent model training convergence. Furthermore, the complex electromagnetic environment of industrial sites introduces a large amount of high-frequency random noise, which can mask the true process characteristics. Therefore, before feature extraction, the raw data must be preprocessed to eliminate the influence of units and improve the signal-to-noise ratio.

[0033] Specifically, this invention employs the Z-Score normalization method to map the RF parameter observations of each channel to a distribution space with zero mean and unit variance, thereby eliminating the influence of different physical dimensions. Furthermore, a moving average filter is used to smooth the non-stationary RF parameter observation sequence, filtering out high-frequency random noise components that lack process meaning and retaining the principal component signals reflecting plasma dynamics characteristics, thus reconstructing a pure process curve.

[0034] S3. The feature extraction layer of the neural network is weighted based on the rate of change of standardized radio frequency parameter observations to obtain transient weighted feature values.

[0035] It should be noted that standard recurrent neural network models typically assign the same level of attention to all inputs at all time steps when processing time-series data. However, in semiconductor wafer processing, defects often occur during transient phases where plasma states change drastically, such as the moment of ignition or gas switching. Standard models tend to overlook these brief but crucial fluctuations, leading to missed detections. Therefore, this invention introduces a mechanism in the feature extraction layer that dynamically adjusts the attention weights based on the severity of signal changes, forcing the model to focus on high-risk transient processes.

[0036] Specifically, this invention incorporates a rate-of-change-sensitive module into the input gating unit of a recurrent neural network, which dynamically adjusts the gain of feature extraction based on the intensity of fluctuations in the real-time signal.

[0037] The transient weighted eigenvalues ​​satisfy the expression:

[0038] In the formula, This represents the transient weighted feature value extracted at time t; The standard weight matrix representing a neural network; Indicates the bias term; Represents the observed radio frequency parameters at time t; This represents the observed radio frequency parameters at the previous moment; Represents the transient sensitivity coefficient; Indicates the nonlinear amplification index; Represents the hyperbolic tangent activation function; This represents absolute value operations.

[0039] In the formula, the left side of the expression This is the standard hyperbolic tangent activation function part, used to extract basic features; the parenthetical terms on the right side of the expression. Since the input data has been standardized, the absolute difference of the signal This directly reflects the fluctuation factor relative to the background noise. As the process enters a rapidly changing transition region, the value of this difference increases, leading to a larger value in the entire weighting term, which in turn drives the transient weighted feature value extracted at that moment. This mechanism forces the model to allocate higher weights to moments of dramatic change during gradient descent, thereby accurately capturing the weak arc signals hidden in rapid transients and effectively avoiding numerical instability that may result from using the mean as the denominator.

[0040] It should be further noted that the nonlinear amplification index in this invention... The value range is set from 1.5 to 3.0. If the value is set too small, the amplification effect on transient features is not obvious, making it difficult to distinguish between normal fluctuations and abnormal jumps; if the value is set too large, it will cause feature values ​​to diverge, affecting model stability. This invention selects 2.0 as an empirical value, that is, adopts quadratic amplification, which can significantly improve the detection rate of short-time arc signals while ensuring numerical stability.

[0041] It should be noted that the transient sensitivity coefficient was determined by performing gradient-weighted class activation mapping (Grad-CAM) analysis on historical failure samples; the standard weight matrix and bias terms of the neural network were obtained during the offline training phase by iteratively updating and converging based on a large amount of historical data from normal processes using the backpropagation algorithm.

[0042] For example, Figure 2This is a schematic diagram of transient weighted features. The diagram shows a comparison of the changes in standard eigenvalues ​​and transient weighted eigenvalues ​​when weak defects occur during transient phases of the process, such as ignition or gas switching. Due to the introduction of a nonlinear amplification mechanism based on the signal rate of change, the transient weighted eigenvalues ​​are significantly amplified when the signal fluctuates violently, thus highlighting the weak defect signals hidden in the normal transient background, while the standard eigenvalues ​​are unable to distinguish between normal fluctuations and defect signals.

[0043] S4. Based on the generalized Pareto distribution and baseline drift compensation term, obtain the dynamic judgment threshold, compare the transient weighted feature value with the dynamic judgment threshold, and obtain the defect detection result of the semiconductor wafer.

[0044] It should be noted that while the generalized Pareto distribution can effectively describe the tail characteristics of data, polymers gradually deposit on the inner walls of the reaction chamber during the continuous processing of semiconductor wafers as the equipment runs. This deposition causes a nonlinear physical drift in the system's impedance baseline, manifested as a monotonically shifting statistical center of the data over time. If only the standard generalized Pareto distribution is used to calculate a fixed threshold, the dynamic threshold cannot shift synchronously with the baseline, leading to false alarms caused by normal signals touching the threshold in later stages of the process. Therefore, a compensation mechanism capable of sensing equipment aging must be introduced to construct a dynamic threshold.

[0045] Specifically, the present invention updates the distribution parameters in real time based on the historical abnormal score sequence and calculates a dynamic judgment threshold including aging compensation.

[0046] The dynamic threshold satisfies the expression:

[0047] In the formula, express The dynamic judgment threshold calculated at any given time; Indicates the index of discrete-time sampling points; express The location parameters of the generalized Pareto distribution are updated in real time; express The scale parameters of the generalized Pareto distribution are updated in real time; express The shape parameters of the generalized Pareto distribution are updated in real time; Indicates the preset confidence level; This indicates the current cumulative number of semiconductor wafers processed; This represents the baseline drift coefficient.

[0048] In the formula, the first part of the expression This forms a dynamic baseline, and the second part is the safety margin term of the standard generalized Pareto distribution. This varies with the cumulative number of semiconductor wafers processed. As the volume increases, the thickening of deposits on the inner wall of the reaction chamber causes a positive shift in the statistical baseline of the background noise, i.e., the location parameters... The physical background value increases, at which point the additive compensation term... This increases the baseline value, driving the final dynamic judgment threshold. Synchronously shift upwards. This additive compensation mechanism ensures that the dynamic judgment threshold and the background noise baseline always maintain a constant safe distance, avoiding false alarms caused by baseline drift due to equipment aging.

[0049] It should be noted that the baseline drift coefficient is obtained by linear regression fitting to the historical mean data of the previous maintenance cycle. This value reflects the impedance increment caused by semiconductor wafer processing. , , It is obtained by real-time fitting of historical outlier scores within the sliding window using the maximum likelihood estimation method.

[0050] It should be further noted that the confidence level in this invention... The value is set to a range of 99.5% to 99.99%. If this value is set too low, a large number of normal process fluctuations will be misjudged as defects, increasing the false alarm rate; if this value is set too high, weak early arc signals may be missed, increasing the risk of missed alarms. This invention selects 99.9% as an empirical value, which can effectively capture destructive defect signals while ensuring an extremely low false alarm rate.

[0051] For example, Figure 3 This is a schematic diagram of a dynamic judgment threshold. The diagram shows that as the cumulative number of semiconductor wafers processed increases, the RF characteristic baseline gradually rises due to the accumulation of deposits on the inner wall of the reaction chamber. A fixed dynamic judgment threshold cannot adapt to this physical drift, causing normal signals to touch the threshold in the later stages of the process, thus triggering false alarms. In contrast, the dynamic judgment threshold, by introducing a baseline drift compensation term, can rise synchronously with the background noise baseline, always maintaining a reasonable safety margin and effectively avoiding false alarms caused by equipment aging.

[0052] This invention also discloses a semiconductor wafer defect detection system based on deep learning, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a semiconductor wafer defect detection method based on deep learning according to the present invention.

[0053] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A deep learning-based semiconductor wafer defect detection method, characterized by, The method comprises the following steps: Obtaining a sequence of radio frequency parameter observation values through a sensor installed at the output end of a radio frequency matching network; Standardizing and denoising the sequence of radio frequency parameter observation values to obtain standardized radio frequency parameter observation values; Weighting a neural network feature extraction layer based on the rate of change of the standardized radio frequency parameter observation values to obtain transient weighted feature values; Obtaining a dynamic decision threshold value based on a generalized Pareto distribution and a baseline drift compensation term, comparing the transient weighted feature values with the dynamic decision threshold value, and obtaining a defect detection result of a semiconductor wafer. 2.The deep learning-based semiconductor wafer defect detection method of claim 1, wherein, The method of obtaining a sequence of radio frequency parameter observation values through a sensor installed at the output end of a radio frequency matching network comprises the following steps: Synchronously collecting radio frequency voltage, radio frequency current, phase angle, forward power and reflected power at a microsecond level sampling rate using a radio frequency impedance sensor; and synchronously aligning the collected data to form the sequence of radio frequency parameter observation values. 3.The deep learning-based semiconductor wafer defect detection method of claim 1, wherein, The method of standardizing and denoising the sequence of radio frequency parameter observation values comprises the following steps: Mapping the sequence of radio frequency parameter observation values to a zero-mean unit-variance space using a Z-Score standardization method; and filtering high-frequency random noise using a sliding average filter. 4.The deep learning-based semiconductor wafer defect detection method of claim 1, wherein, The transient weighted feature values satisfy the expression: ； In the formula, represents a transient weighting characteristic value extracted at time t; represents a standard weight matrix of a neural network; represents a bias term; represents a radio frequency parameter observation value at time t; represents a radio frequency parameter observation value at the previous time; represents a transient sensitivity coefficient; represents a nonlinear amplification index; represents a hyperbolic tangent activation function; represents an absolute value operation. 5.The deep learning-based semiconductor wafer defect detection method of claim 4, wherein, The value of the non-linear amplification exponent is 2.

0. 6.The deep learning based semiconductor wafer defect detection method of claim 4, wherein, The transient sensitivity coefficient is obtained based on historical fault samples through gradient-weighted class activation mapping analysis.

7. The deep learning-based semiconductor wafer defect detection method of claim 1, wherein, The dynamic decision threshold value satisfies the expression: ； In the formula, represents a dynamic determination threshold calculated at the moment; represents a discrete time sampling point index; represents a generalized Pareto distribution position parameter updated at the moment; represents a generalized Pareto distribution scale parameter updated at the moment; represents a generalized Pareto distribution shape parameter updated at the moment; represents a preset confidence level; represents a current cumulative number of processed semiconductor wafers; represents a baseline drift coefficient. 8.The deep learning-based semiconductor wafer defect detection method of claim 7, wherein, The baseline drift coefficient is obtained by linear regression fitting of the historical mean data of the previous maintenance period. 9.The deep learning based semiconductor wafer defect detection method of claim 7, wherein, The value of the confidence level is 99.9%.

10. A deep learning-based semiconductor wafer defect detection system, characterized by, The method comprises the following steps: A processor and a memory, the memory storing computer program instructions, when the computer program instructions are executed by the processor, a semiconductor wafer defect detection method based on deep learning according to any one of claims 1-9 is realized.

Images