Semiconductor wafer etching method with detection function

Abstract

The invention relates to the technical field of semiconductor manufacturing, and discloses a semiconductor wafer etching method with a detection function. The method comprises the following steps: acquiring initial image data of a wafer surface through an optical detection device, and extracting geometric features and defect distribution information; processing through a multi-scale feature fusion algorithm, identifying an abnormal region and classifying defect types; dynamically adjusting process parameters of etching equipment according to defect conditions, and planning a self-adaptive etching path; during etching, a real-time monitoring module is used for collecting data, and etching parameters are corrected through a feedback control mechanism; and after etching, the etching precision is secondarily detected and verified by means of a multispectral imaging technology. According to the method, defects can be accurately positioned, the etching process is optimized, the etching precision and quality are effectively improved, the rejection rate is reduced, and the efficiency and reliability of semiconductor wafer manufacturing are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of production scheduling, and particularly to a semiconductor wafer etching method with a detection function. Background Art

[0002] In the field of semiconductor manufacturing, the wafer etching process is a key link determining the performance and quality of chips. With the continuous improvement of semiconductor integration and the continuous reduction of chip feature sizes, the requirements for wafer etching accuracy and uniformity have reached an unprecedented level. Traditional semiconductor wafer etching methods have many limitations and are difficult to meet the stringent requirements of modern semiconductor manufacturing.

[0003] Regarding the detection link before etching, the early detection technical means were relatively single, mainly relying on simple optical microscopes or electron microscopes for manual detection. This detection method is inefficient and easily interfered by human factors, resulting in poor accuracy and reliability of detection results. For tiny defects, especially sub-micron level defects, it is very difficult to comprehensively and accurately identify them manually. This makes some potential defects be ignored during the etching process, thus affecting the performance and yield rate of chips.

[0004] In terms of the control of etching process parameters, most traditional methods use fixed process parameters for etching. However, during the production process of semiconductor wafers, the types and distribution of defects on their surfaces are different, and unified process parameters cannot effectively handle different defects. For example, for defects with different depths and sizes, if the same etching rate and etching time are used, it may result in under-etching in some areas and over-etching in some other areas, seriously affecting the etching uniformity and accuracy. This will not only reduce the performance of chips but also may cause chips to malfunction during subsequent use, increasing production costs and product defect rates.

[0005] In the real-time monitoring link during the etching process, traditional methods also have obvious deficiencies. Due to the lack of effective real-time monitoring means, it is impossible to timely obtain the dynamic change information of the etching area, and thus it is difficult to timely and accurately adjust the etching process. Once abnormal situations occur during the etching process, such as etching rate fluctuations and local over-etching, it is very difficult to discover and take corresponding corrective measures in the first time. One can only detect after the etching is completed. If problems are found, rework is required, which will undoubtedly increase the production cycle and cost and reduce production efficiency.

[0006] In addition, the inspection after etching is also crucial. Traditional post-etch inspection methods can often only detect some obvious defects, and it is difficult to effectively detect potential defects such as tiny residues and microcracks. These potential defects may gradually expand during the subsequent use of the chip, affecting the long-term stability and reliability of the chip. Moreover, even if a defect is detected, it is difficult for traditional methods to quickly and accurately determine the type and location of the defect, and it cannot provide effective guidance for subsequent rework and correction. Summary of the Invention

[0007] The purpose of the present invention is to provide a semiconductor wafer etching method with a detection function to solve the problems presented in the above-mentioned background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A semiconductor wafer etching method with a detection function, the method includes: Step S1: Obtain the initial image data of the semiconductor wafer surface through an optical detection device, and extract the geometric features and defect distribution information on the wafer surface; Step S2: Based on the multi-scale feature fusion algorithm, process the initial image data, identify the abnormal areas on the wafer surface, and generate a defect type classification result; Step S3: Dynamically adjust the process parameters of the etching equipment according to the defect type and distribution density, and generate an adaptive etching path planning instruction; Step S4: During the etching process, use the real-time monitoring module to collect the dynamic change data of the etching area, and correct the etching parameters through the feedback control mechanism.

[0009] Preferably, the step S2 further includes: Step S21: Preprocess the initial image data, including noise suppression, contrast enhancement, and edge sharpening operations; Step S22: Decompose the image data using wavelet transform to extract feature maps at different resolutions; Step S23: Fuse the multi-scale feature maps through a convolutional neural network model to generate a high-confidence defect localization result; Step S24: Based on the support vector machine classifier, classify according to the defect morphological features to determine the defect type.

[0010] Preferably, the dynamic adjustment of process parameters in the step S3 includes: Step S31: Construct a mapping relationship model between the etching rate and process parameters, and define the correlation function of etching depth, gas flow rate, and plasma intensity; Step S32: Calculate the initial etching parameter combination according to the size and depth requirements of the defect area; Step S33: Use the particle swarm optimization algorithm to iteratively adjust the parameter combination until the etching uniformity constraint condition is satisfied; Step S34: Input the optimized parameters into the etching equipment to generate a path instruction sequence including the priority etching area.

[0011] Preferably, the specific steps of the feedback control mechanism in step S4 include: Step S41: Measure the depth change of the etching area in real time through a laser interferometer to generate a dynamic thickness distribution map; Step S42: Compare the deviation between the actual thickness and the target thickness and calculate the local etching compensation coefficient; Step S43: Adjust the plasma energy output based on a fuzzy logic controller to suppress the etching rate fluctuation; if local over-etching is detected, pause the etching of the current area and mark it as a secondary correction area.

[0012] Preferably, the method further includes: Step S5: Perform a secondary inspection on the surface of the etched wafer through multispectral imaging technology to verify the etching accuracy; the specific steps include: Step S51: Collect the reflection spectral data of the etching area and extract the intensity distribution at the characteristic wavelength; Step S52: Construct an association database between spectral features and etching defects to match abnormal spectral patterns; Step S53: If residues or microcracks are detected, generate a rework instruction and update the etching path planning.

[0013] Preferably, the fitness function of the particle swarm optimization algorithm in step S33 is defined as:

[0014] where, is the etching depth deviation, is the standard deviation of etching uniformity, is the process time, 、 、 are dynamic weight factors.

[0015] Preferably, the input variables of the fuzzy logic controller in step S43 include the absolute value of the deviation, the deviation change rate, and the historical compensation amount, and the output variable is the correction percentage of the plasma energy; The fuzzy rule base of the fuzzy logic controller includes: if the deviation is large and the change rate is positive, then significantly increase the energy output; if the deviation is small and the change rate is negative, then slightly adjust the energy output.

[0016] Preferably, the spectral feature matching method in step S52 includes: Step S521: Normalize the reflection spectral data to eliminate ambient light interference; Step S522: Reduce the dimensionality through principal component analysis and extract the first three principal components as feature vectors; Step S523: Calculate the cosine similarity between the feature vectors and the database samples to determine whether there is a defect pattern.

[0017] Preferably, the adjustment strategy of the dynamic weight factor includes: Set the initial weight according to the etching stage, and increase it in the depth accuracy stage and increase it in the uniformity stage Real-time monitor the stability of the etching rate. If the fluctuation exceeds the threshold, then decrease it and increase ; Optimize the weight combination through the genetic algorithm to minimize the global etching error of the fitness function.

[0018] Preferably, the optimization method of the fuzzy rule base includes: training a neural network model based on historical etching data to generate initial fuzzy rules, dynamically updating the rule weights through reinforcement learning, preferentially retaining the rules with significant compensation effects. If the compensation is ineffective for multiple consecutive times, then reset the rule base and trigger an artificial intervention signal.

[0019] Compared with the prior art, the beneficial effects of the present invention are: In the detection stage before etching, use an optical detection device to obtain the initial image data of the wafer surface and process it through a multi-scale feature fusion algorithm. It can not only accurately extract the geometric features and defect distribution information on the wafer surface, but also identify abnormal areas and accurately classify the defect types. Compared with traditional detection methods, this greatly improves the detection accuracy and efficiency, effectively avoids subsequent problems caused by detection omissions, and provides a reliable basis for the accurate implementation of the subsequent etching process.

[0020] In the link of adjusting the etching process parameters, dynamically adjust the process parameters of the etching equipment according to the defect type and distribution density. By constructing a mapping relationship model between the etching rate and the process parameters and using the particle swarm optimization algorithm to iteratively optimize the parameter combination, the uniformity and accuracy of the etching are ensured. Compared with traditional fixed-parameter etching, this adaptive adjustment method can formulate an optimal etching plan for different defect situations, effectively reduce the occurrence of under-etching or over-etching phenomena, and improve the performance and yield of the chip. For example, when processing wafers with complex defect distributions, parameters such as the etching depth, gas flow rate, and plasma intensity can be accurately adjusted according to the actual situation, making the etching effect more ideal and greatly improving the product quality.

[0021] The real-time monitoring and feedback control mechanism during the etching process is a major highlight of the present invention. By using the real-time monitoring module to collect dynamic change data of the etching area, the etching depth change is measured in real time through a laser interferometer, the deviation between the actual thickness and the target thickness is compared in a timely manner, and the plasma energy output is adjusted based on a fuzzy logic controller, effectively suppressing the etching rate fluctuation. Once local over-etching is detected, the etching of the current area can be quickly paused and marked as a secondary correction area, achieving precise control of the etching process. This real-time monitoring and feedback adjustment mechanism can promptly discover and solve problems occurring during the etching process, avoid the accumulation and expansion of problems, further improve the etching accuracy and product reliability, while reducing rework caused by etching mistakes, lowering production costs, and increasing production efficiency.

[0022] The secondary detection link after etching is also of great significance. By collecting the reflection spectrum data of the etching area through multi-spectral imaging technology and constructing a correlation database between spectral features and etching defects, tiny defects such as residues or micro-cracks on the surface of the wafer after etching can be accurately detected. Once a defect is detected, a rework instruction can be generated in a timely manner and the etching path planning can be updated, ensuring the quality of the final product. This comprehensive secondary detection method makes up for the deficiencies of traditional detection methods, effectively improves the quality and stability of the product, and enhances the competitiveness of the product in the market.

[0023] From the perspective of the overall production process, the etching method of the present invention realizes the full-process intelligent and precise management from pre-etching detection, etching process control to post-etching detection through the optimization and coordinated work of each link. It not only improves the etching quality and production efficiency of semiconductor wafers, but also reduces production costs and the scrap rate, provides strong technical support for the development of the semiconductor manufacturing industry, helps to promote the semiconductor industry to move towards higher precision and higher integration, and has broad application prospects and great economic value. Brief Description of the Drawings

[0024] Figure 1 It is the working principle diagram of the semiconductor wafer etching method with detection function described in the present invention; Figure 2 It is the flow chart of dynamically adjusting process parameters; Figure 3 It is the flow chart related to the fuzzy logic controller; Figure 4 It is the flow chart of the fuzzy rule base optimization method. Detailed Embodiment

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0026] Please refer to Figures 1 to 4 , the present invention provides a semiconductor wafer etching method with a detection function, aiming to improve the accuracy and efficiency of the semiconductor wafer etching process, effectively handle the defects on the wafer surface, and reduce the production cost. The following details the specific implementation manners of this method.

[0027] Step S1: Use an optical detection device to detect the surface of the semiconductor wafer to obtain initial image data. In actual operation, a high-resolution optical microscope or electron microscope can be selected as the optical detection device. These devices can clearly capture the microscopic features of the wafer surface, and the obtained initial image data contains rich information. Through professional image analysis software, geometric features of the wafer surface, such as the flatness of the wafer and the edge shape, are extracted from it, and at the same time, the defect distribution information on the wafer surface is identified, such as the positions and approximate ranges of defects such as scratches and holes.

[0028] Step S2: Based on the multi-scale feature fusion algorithm, process the initial image data obtained in Step S1. First, preprocess the image. Through specific image processing algorithms, the noise in the image is suppressed, the contrast of the image is enhanced, so that the defect features are more obvious, and at the same time, the image edges are sharpened for more accurate feature extraction in the subsequent steps. Then, use wavelet transform technology to decompose the image data to obtain feature maps at different resolutions. These feature maps show the information of the wafer surface at different scales. Then, with the help of a convolutional neural network model, fuse the multi-scale feature maps. Through the training and learning of the network, a high-confidence defect localization result is generated to clarify the exact position of the defect on the wafer surface. Finally, based on a support vector machine classifier, classify according to the morphological features of the defects to accurately determine the defect types, such as physical damage type defects, material impurity type defects, etc.

[0029] Step S3: According to the defect type and distribution density determined in Step S2, dynamically adjust the process parameters of the etching equipment and generate an adaptive etching path planning instruction. First, construct a mapping relationship model between the etching rate and process parameters, and determine the correlation function between the etching depth, gas flow rate, and plasma intensity. Calculate the initial combination of etching parameters based on the size and depth requirements of the defect area. In order to make the etching process meet the uniformity constraint conditions, use the particle swarm optimization algorithm to iteratively adjust the parameter combination. After multiple iterations and optimizations, input the obtained optimized parameters into the etching equipment to generate a path instruction sequence including the priority etching area, ensuring that the etching equipment can etch the wafer efficiently and precisely according to the planned path.

[0030] Step S4: During the etching process, use the real-time monitoring module to collect the dynamic change data of the etching area, and correct the etching parameters through the feedback control mechanism. Use a laser interferometer to measure the depth change of the etching area in real time, generate a dynamic thickness distribution map, compare the thickness measured in real time with the preset target thickness, calculate the deviation between the two, and then obtain the local etching compensation coefficient. Based on the deviation situation, adjust the plasma energy output based on a fuzzy logic controller to suppress the etching rate fluctuation. If local over-etching is detected, pause the etching of the current area and mark this area as a secondary correction area for subsequent targeted processing.

[0031] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0032] Embodiment 1: In this embodiment, the specific process of processing image data by the multi-scale feature fusion algorithm in Step S2 is emphasized. After obtaining the initial image data, preprocessing is first performed. For noise suppression, the Gaussian filtering algorithm is used. By setting appropriate Gaussian kernel size and standard deviation, smooth the noise in the image, remove random noise points generated by factors such as equipment noise and environmental interference, and make the image clearer. In terms of contrast enhancement, use the histogram equalization technique to adjust the gray level distribution of the image, evenly distribute the gray level values of the image within the entire gray level range, thereby enhancing the contrast of the image and making the originally unobvious defect features more prominent. For edge sharpening, use the Laplace operator to perform convolution operations on the image to highlight the edge information in the image and make the edges of the defects clearer and distinguishable.

[0033] When decomposing the image data by wavelet transform, select an appropriate wavelet basis function, such as the Daubechies wavelet. Decompose the image into sub-bands with different resolutions, including a low-frequency sub-band and multiple high-frequency sub-bands. The low-frequency sub-band contains the main contour information of the image, and the high-frequency sub-bands contain the detail information of the image, such as the edges and textures of the defects. Through this multi-resolution decomposition, the image features can be observed from different scales.

[0034] In the construction of the convolutional neural network model, a classic convolutional neural network architecture such as VGGNet is selected. During network training, the multi-scale feature maps obtained through wavelet transform are used as inputs, and through multiple convolutional, pooling, and fully connected layer operations, automatic feature extraction and fusion are achieved. After the network is trained with a large amount of labeled wafer defect image data, it can accurately locate the defects and generate defect localization results with high confidence.

[0035] When classifying defects based on a support vector machine classifier, first extract the morphological features of the defects, such as the area, perimeter, shape complexity, etc. of the defects. These features are used as inputs to the support vector machine, and a classification model is obtained through training. During the training process of the support vector machine, an optimal classification hyperplane is searched for, so that different types of defects can be maximally separated in the feature space, thereby achieving accurate defect type classification.

[0036] Example 2: During the process of dynamically adjusting process parameters in step S3, the specific implementation methods for constructing the mapping relationship model between the etching rate and process parameters and determining the correlation function coefficients are as follows: I. Principle of mapping relationship model construction and data fitting The mapping relationship model between the etching rate and process parameters is constructed based on the physical mechanism of plasma etching. Taking the etching depth as the target variable and the gas flow rate and plasma intensity as input variables, a multi-variable linear regression model is established. This model assumes a linear correlation between the etching depth and the gas flow rate and plasma intensity, and also includes a constant term to represent the comprehensive influence of background factors such as substrate material characteristics and equipment inherent parameters. The specific modeling process is as follows: The orthogonal experiment method is used to design multiple groups of etching experiments, covering the typical working ranges of the gas flow rate (Q) and plasma intensity (I). For example, the gas flow rate range is set to 50 sccm to 200 sccm, divided into 4 levels at 50 sccm intervals; the plasma intensity range is 100 W to 400 W, divided into 4 levels at 100 W intervals, forming a total of 16 experimental combinations of 4×4. Each group of experiments uses the same type of semiconductor wafer, and etching is carried out at a fixed etching time (such as t = 300 s), and the etching depth (D) is measured by a step profiler. Each experimental combination is repeated 3 times to reduce random errors.

[0037] The mean value of the measurement data for each group of experiments is calculated to obtain the average etching depth corresponding to the gas flow rate and plasma intensity. At the same time, abnormal data points (such as data with a deviation from the mean value exceeding 20%) are excluded to avoid interference from equipment fluctuations or operation errors on the model.

[0038] Based on the least squares principle, a linear regression equation for the etching depth D with respect to the gas flow rate Q and the plasma intensity I is established:

[0039] where, is the influence coefficient of the gas flow rate on the etching depth (unit: nm / (sccm·s)), is the influence coefficient of the plasma intensity on the etching depth (unit: nm / (W·s)), is the constant term (unit: nm), reflecting comprehensive factors such as the etching time and the etching rate constant of the substrate material. By minimizing the mean square error (MSE) between the measured etching depth and the model predicted value, the coefficients , , are solved. The specific optimization objective is:

[0040] In the formula, is the number of effective experimental samples, , , are the measured values of the etching depth, gas flow rate, and plasma intensity of the th group of experiments, respectively.

[0041] The model is verified using the reserved experimental data not involved in model building (such as selecting 2 groups of experiments as the validation set), and the relative error between the predicted etching depth and the measured value is calculated. If the average relative error is less than 5%, the model accuracy is considered to meet the process requirements; if the error exceeds the limit, the experimental range is expanded or non-linear terms (such as quadratic terms, interaction terms) are introduced to correct the model until the expected accuracy is achieved.

[0042] II. Experimental determination method of the correlation function coefficients The coefficients , , are experimentally determined through the following steps: 1. Determination of the constant term : Under the theoretical condition that both the gas flow rate and the plasma intensity are zero, corresponds to the theoretical value of the etching depth. However, in actual operation, plasma etching requires maintaining the minimum gas flow rate and power. Therefore, it is indirectly obtained by the extrapolation method: Select the minimum non-zero experimental values of the gas flow rate and the plasma intensity (such as ), measure the etching depth , and establish a reference equation in combination with the model equation:

[0043] Subsequently, solve the simultaneous equations through multiple sets of experimental data.

[0044] 2. Partial regression coefficient , Separation calculation Fix the plasma intensity and measure the influence of gas flow rate: Keep the plasma intensity constant (such as ), change the gas flow rate to conduct multiple sets of etching experiments, and obtain the linear relationship between the etching depth and the gas flow rate. At this time, the model is simplified to , and determine by fitting the slope of this set of data linearly.

[0045] Fix the gas flow rate and measure the influence of plasma intensity: Keep the gas flow rate constant (such as ), change the plasma intensity to conduct multiple sets of etching experiments, and the model is simplified to , and determine by fitting the slope of this set of data linearly.

[0046] Substitute the , preliminary calculated values into the reference equation to solve and obtain . To improve the accuracy, use the least squares method to globally fit all experimental data, and solve the system of multiple linear equations through matrix operations or iterative algorithms (such as the Newton-Raphson method) to obtain the final coefficient values.

[0047] After obtaining the initial combination of etching parameters, use the Particle Swarm Optimization (PSO) algorithm for iterative optimization. The specific process is as follows: Take the gas flow rate Q and the plasma intensity I as the two-dimensional position vector of the particle , randomly generate N particles (such as N = 50) within the feasible region determined by the experiment, and randomly assign the initial velocity vector of each particle within the feasible region as well.

[0048] The fitness function comprehensively considers the etching depth deviation ( ), the standard deviation of etching uniformity ( ), and the process time (T), and its form is:

[0049] Among them, represents the fitness function, , , are dynamic weight factors (the value range is 0 to 1, and ), and they are dynamically adjusted according to the etching stage: At the initial stage of etching, give priority to ensuring the depth accuracy ( ), and in the middle stage, focus on uniformity ( ), and the later optimization efficiency ( ). This embodiment details the specific process of dynamically adjusting process parameters in step S3. When constructing the mapping relationship model between the etching rate and process parameters, a large amount of experimental data is used for fitting. The etching depth and the gas flow rate and the plasma intensity of the correlation function can be expressed as , where , , are coefficients determined by experiments, which reflect the influence degree of gas flow rate and plasma intensity on the etching depth.

[0050] Calculate the initial etching parameter combination according to the size and depth requirements of the defect area. Assume that the target etching depth of a certain defect area is known as , and the gas flow rate and the plasma intensity are preliminarily calculated through the correlation function, that is (here, it is first assumed that is an empirical initial value).

[0051] Use the particle swarm optimization algorithm to iteratively adjust the parameter combination. In the particle swarm optimization algorithm, each particle represents a set of etching parameters (gas flow rate and plasma intensity). Define the fitness function , where is the etching depth deviation, that is, the difference between the actual etching depth and the target etching depth, reflecting the accuracy of the etching depth; is the standard deviation of etching uniformity, which is used to measure the uniformity of the etching process in different regions. The smaller the value, the better the etching uniformity; is the process time, representing the time spent in the etching process. The shorter the , the higher the production efficiency; , , are dynamic weight factors, which are used to balance the importance of different indicators in the optimization process.

[0052] In the algorithm iteration process, the particle continuously adjusts its speed and position according to its own historical optimal position and the global optimal position of the group, that is, adjusts the etching parameters. After multiple iterations, when the etching uniformity constraint condition is met (such as is less than a certain set threshold), the optimized parameter combination is obtained. Input the optimized parameters into the etching equipment to generate a path instruction sequence including the priority etching area. According to the severity and position distribution of the defects, determine the priority etching area, and give priority to processing the defect areas that have a greater impact on the wafer performance, so as to improve the etching efficiency and wafer quality.

[0053] Example 3: This example focuses on the specific implementation of the feedback control mechanism in step S4. The depth change of the etching area is measured in real time by a laser interferometer. The laser interferometer utilizes the principle of light interference. When the laser irradiates the etching area, the reflected light interferes with the reference light, and the depth change of the etching area is calculated by detecting the change of the interference fringes. According to the number of moving interference fringes and the wavelength of the laser, the change amount of the etching depth can be accurately calculated, thereby generating a dynamic thickness distribution map.

[0054] Compare the deviation between the actual thickness and the target thickness, and calculate the local etching compensation coefficient. Let the actual thickness be , and the target thickness be , then the deviation . The local etching compensation coefficient can be calculated by the formula . The value of reflects the relative magnitude of the thickness deviation during the etching process and is used to adjust the etching parameters subsequently.

[0055] Adjust the plasma energy output based on a fuzzy logic controller. The input variables of the fuzzy logic controller include the absolute value of the deviation , the rate of change of the deviation , and the historical compensation amount , and the output variable is the correction percentage of the plasma energy . The fuzzy rule base of the fuzzy logic controller includes: if is large and is positive, then significantly increase the energy output; if is small and is negative, then slightly adjust the energy output. For example, when is greater than the set larger threshold and is greater than 0, can be set to a large value, such as increasing the plasma energy output by 10% - 20%; when is less than the set smaller threshold and is less than 0, is set to a small value, such as increasing the plasma energy output by 1% - 3%. If local over-etching is detected, that is, and exceeds a certain allowable range, pause the etching of the current area and mark it as a secondary correction area for subsequent additional processing, such as re-etching after adjusting the etching parameters or using other repair methods.

[0056] Example 4: This embodiment details the process of secondary inspection of the etched wafer surface through hyperspectral imaging technology in step S5. Reflectance spectral data of the etched area is collected by using a hyperspectral camera to photograph the surface of the etched wafer. The hyperspectral camera can simultaneously obtain the reflected light information of multiple different wavelengths. During the photographing process, ensure that the shooting angle, lighting conditions, etc. of the camera remain consistent to guarantee the accuracy and comparability of the collected data. Extract the intensity distribution at the characteristic wavelengths. Based on the previous spectral studies of different etching defects, determine several key characteristic wavelengths, such as specific wavelengths related to common residues or microcracks. Analyze the intensity distribution of the reflected light at these characteristic wavelengths to observe whether there are abnormal intensity changes.

[0057] Construct a correlation database between spectral features and etching defects. Through a large number of experiments, collect the reflected spectral data of different types of etching defects under hyperspectral conditions, and establish the correlation between spectral features (such as intensity values at specific wavelengths, intensity ratios between different wavelengths, etc.) and the corresponding defect types. When matching abnormal spectral patterns, first preprocess the collected reflected spectral data.

[0058] Perform normalization processing on the reflected spectral data to eliminate ambient light interference. Use the normalization formula , where is the original spectral intensity value, and are the minimum intensity value and the maximum intensity value among all the collected data at this wavelength respectively, is the normalized intensity value. In this way, the data collected under different conditions is made comparable.

[0059] Reduce the dimension through principal component analysis and extract the first three principal components as feature vectors. Principal component analysis can transform multiple related spectral feature variables into a few uncorrelated principal components, while retaining most of the original data information, reducing the data dimension, and improving the calculation efficiency. Calculate the cosine similarity between the feature vector and the database samples to determine whether there is a defect pattern. Let the feature vector be , and the database sample vector be . The cosine similarity calculation formula is . When the similarity is lower than a certain set threshold, it is determined that there is an abnormal spectral pattern, that is, there may be an etching defect. If residues or microcracks are detected, a rework instruction is generated and the etching path planning is updated to reprocess the defective area to ensure that the quality of the wafer meets the requirements.

[0060] Example 5: This embodiment deeply explores the adjustment strategy of the dynamic weight factor in the fitness function of the particle swarm optimization algorithm in step S33. Set the initial weight according to the etching stage. At the beginning stage of etching, more attention is paid to the accuracy of the etching depth. At this time, Set to a larger value, such as 0.5, Set to 0.3, Set to 0.2. As the etching process progresses and enters the uniformity stage, to ensure the uniformity of etching, increase the value of, for example, adjust to 0.5, and at the same time appropriately reduce the value of, such as adjusting to 0.3, Keep it unchanged or make fine adjustments.

[0061] Monitor the stability of the etching rate in real time. If the fluctuation exceeds the threshold, then reduce and increase . The stability of the etching rate can be measured by monitoring the change rate of the etching depth over a period of time. Let the change amount of the etching depth from time to be , then the etching rate . Set an etching rate fluctuation threshold , when (where is the average etching rate), it is considered that the etching rate fluctuation exceeds the threshold. At this time, to ensure the uniformity of etching, reduce the value of, such as reducing from 0.2 to 0.1, and at the same time increase the value of, such as increasing from 0.3 to 0.4.

[0062] Optimize the weight combination through the genetic algorithm to minimize the global etching error of the fitness function. The genetic algorithm simulates the biological evolution process and optimizes the weight combination ( , , ) through operations such as selection, crossover, and mutation. In each generation of evolution, calculate the fitness function value corresponding to each weight combination, select the weight combination with better fitness for crossover and mutation operations to generate new weight combinations. After multiple generations of evolution, gradually find the weight combination that minimizes the fitness function, thereby optimizing the etching process, reducing the global etching error, and improving the quality and efficiency of wafer etching.

[0063] Example 6: This example details the optimization method of the fuzzy rule base of the fuzzy logic controller in step S43. Train a neural network model based on historical etching data to generate initial fuzzy rules. Collect a large amount of historical etching data, including information such as the absolute value of the deviation, the deviation change rate, the historical compensation amount, and the corresponding plasma energy correction percentage during the etching process. Use these data as the training set to train the neural network model, such as using a multi-layer perceptron (MLP). Through the learning of the neural network, mine the potential laws in the data to generate initial fuzzy rules.

[0064] Dynamically update the rule weights through reinforcement learning, and preferentially retain the rules with significant compensation effects. Reinforcement learning adjusts its own behavior strategy continuously according to the reward mechanism through the interaction between the agent and the environment. In the optimization of the fuzzy rule base, regard the fuzzy rules as the behaviors of the agent. When a certain rule can effectively reduce the etching deviation during the actual etching process, that is, when the compensation effect is significant, increase the weight of this rule; on the contrary, if the compensation effect of a certain rule is not good, then reduce its weight.

[0065] If the compensation is ineffective for multiple consecutive times, reset the rule base and trigger an artificial intervention signal. Set a threshold for the number of consecutive ineffective compensations such that when, during the etching process of a certain area, after compensating according to the fuzzy rules for consecutive times and the etching deviation still cannot be effectively controlled, it is considered that there may be a problem with the rule base. At this time, reset the rule base to its initial state and trigger an artificial intervention signal. The manual operator can check and analyze the etching process, find the cause of the problem, such as whether there is equipment failure, unreasonable process parameter settings, etc., and then readjust the rule base or perform other necessary processing to ensure the smooth progress of the etching process and guarantee the etching quality of the wafer.

[0066] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0067] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A semiconductor wafer etching method with a detection function, characterized in that, It includes the following steps: Step S1: Obtain the initial image data of the semiconductor wafer surface through an optical detection device, and extract the geometric features and defect distribution information on the wafer surface; Step S2: Based on the multi-scale feature fusion algorithm, process the initial image data, identify the abnormal areas on the wafer surface, and generate the defect type classification result; Step S3: Dynamically adjust the process parameters of the etching equipment according to the defect type and distribution density, and generate the adaptive etching path planning instruction; Step S4: During the etching process, use the real-time monitoring module to collect the dynamic change data of the etching area, and correct the etching parameters through the feedback control mechanism.

2. The semiconductor wafer etching method with a detection function according to claim 1, wherein The step S2 further includes: Step S21: Preprocess the initial image data, including noise suppression, contrast enhancement, and edge sharpening operations; Step S22: Decompose the image data by wavelet transform to extract the feature maps at different resolutions; Step S23: Fuse the multi-scale feature maps through a convolutional neural network model to generate a high-confidence defect localization result; Step S24: Based on the support vector machine classifier, classify according to the defect morphological features to determine the defect type.

3. The semiconductor wafer etching method with a detection function according to claim 1, wherein The dynamic adjustment of process parameters in the step S3 includes: Step S31: Construct a mapping relationship model between the etching rate and process parameters, and define the correlation function of the etching depth, gas flow rate, and plasma intensity; Step S32: Calculate the initial etching parameter combination according to the size and depth requirements of the defect area; Step S33: Use the particle swarm optimization algorithm to iteratively adjust the parameter combination until the etching uniformity constraint condition is satisfied; Step S34: Input the optimized parameters into the etching equipment to generate a path instruction sequence including the priority etching area.

4. The semiconductor wafer etching method with a detection function according to claim 1, characterized in that, The specific steps of the feedback control mechanism in the step S4 include: Step S41: Measure the depth change of the etching area in real time through a laser interferometer to generate a dynamic thickness distribution map; Step S42: Compare the deviation between the actual thickness and the target thickness, and calculate the local etching compensation coefficient; Step S43: Adjust the plasma energy output based on the fuzzy logic controller to suppress the etching rate fluctuation; if local over-etching is detected, pause the etching of the current area and mark it as the secondary correction area.

5. The semiconductor wafer etching method with a detection function according to claim 1, characterized in that, It also includes: Step S5: Perform secondary detection on the etched wafer surface through multi-spectral imaging technology to verify the etching accuracy; The specific steps include: Step S51: Collect the reflection spectrum data of the etching area and extract the intensity distribution at the characteristic wavelengths; Step S52: Construct an association database between the spectral features and etching defects to match the abnormal spectral patterns; Step S53: If residues or micro-cracks are detected, generate a rework instruction and update the etching path planning.

6. The semiconductor wafer etching method with a detection function according to claim 3, characterized in that, The fitness function of the particle swarm optimization algorithm in the step S33 is defined as: Among them, is the etching depth deviation, is the standard deviation of etching uniformity, is the process time, , , are dynamic weight factors.

7. The semiconductor wafer etching method with a detection function according to claim 4, wherein The input variables of the fuzzy logic controller in the step S43 include the absolute value of the deviation, the deviation change rate, and the historical compensation amount, and the output variable is the correction percentage of the plasma energy; The fuzzy rule base of the fuzzy logic controller includes: if the deviation is large and the change rate is positive, then significantly increase the energy output; if the deviation is small and the change rate is negative, then slightly adjust the energy output.

8. The semiconductor wafer etching method with a detection function according to claim 5, characterized in that, The spectral feature matching method in step S52 includes: Step S521: Normalize the reflection spectral data to eliminate ambient light interference; Step S522: Reduce the dimension through principal component analysis and extract the first three principal components as feature vectors; Step S523: Calculate the cosine similarity between the feature vectors and the database samples to determine whether there is a defect pattern.

9. The semiconductor wafer etching method with a detection function according to claim 6, characterized in that, The adjustment strategy of the dynamic weight factor includes: Set the initial weight according to the etching stage, and improve in the depth accuracy stage , and improve in the uniformity stage , monitor the stability of the etching rate in real time. If the fluctuation exceeds the threshold, then reduce and increase ; Optimize the weight combination through a genetic algorithm to minimize the global etching error of the fitness function.

10. The semiconductor wafer etching method with a detection function according to claim 7, characterized in that, The optimization method of the fuzzy rule base includes: training a neural network model based on historical etching data to generate initial fuzzy rules, dynamically updating the rule weights through reinforcement learning, preferentially retaining the rules with significant compensation effects, and if the compensation is ineffective for multiple consecutive times, reset the rule base and trigger an artificial intervention signal.

Images