A non-contact method and apparatus for measuring the flatness of an epitaxial wafer

Abstract

This invention discloses a non-contact method and apparatus for measuring the flatness of epitaxial wafers, belonging to the field of wafer processing technology. The method includes: acquiring raw and finished wafer morphology data based on a first measurement method; calculating the transfer coefficient and equipment fingerprint spectrum of the wafer morphology under the current production process; in actual production, acquiring first morphology data of the wafer before growth based on the first measurement method, and acquiring second morphology data of the wafer after growth based on the first morphology data and the transfer coefficient; selecting an equipment fingerprint spectrum according to the current production stage, and superimposing the second morphology data with the equipment fingerprint spectrum to obtain predicted morphology data; identifying defects in the predicted morphology data to generate a risk navigation map; correcting the first measurement method based on the risk navigation map to obtain a second measurement method; and scanning the wafer after growth is completed using the second measurement method. This invention improves the efficiency of detecting the flatness of large-diameter epitaxial wafers.

Description

Technical Field

[0001] This invention belongs to the field of wafer processing technology, specifically relating to a non-contact method and apparatus for measuring the flatness of epitaxial wafers. Background Technology

[0002] An epitaxial wafer refers to a wafer containing an epitaxial layer. The surface flatness of the epitaxial wafer directly affects the focusing accuracy of subsequent photolithography processes and the yield of the final device. Among existing flatness inspection technologies, non-contact optical interferometry is widely used for measuring wafer surface morphology due to its high precision and non-destructive characteristics. However, for large-diameter wafers, the limited field of view of a single interferometer makes it impossible to cover the entire wafer surface at once. Therefore, sub-aperture stitching interferometry is usually used for measurement. This technique controls the interferometer or stage to move along a predetermined path, acquiring a series of sub-aperture interferograms with overlapping regions. Utilizing the phase data consistency of the overlapping regions, mathematical algorithms are used to stitch these sub-aperture data into a complete wafer surface morphology.

[0003] However, existing sub-aperture stitching interferometry techniques typically employ fixed scanning paths and fixed sub-aperture overlap rates in practical applications. To prevent stitching misalignment or data loss, a relatively conservative overlap rate is often required, which leads to a significant increase in the number of sub-apertures acquired, reducing detection efficiency. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a non-contact method and apparatus for measuring the flatness of epitaxial wafers, thereby resolving the issues present in the background art.

[0005] To achieve the aforementioned objective, this invention proposes a non-contact method for measuring the flatness of epitaxial wafers, comprising: Training data is acquired based on the first measurement method, and the training data includes the original morphology data of the wafer before growth and the finished morphology data after growth. Based on the training data, the transfer coefficient of wafer morphology under the current production process is calculated, and the equipment fingerprint spectrum characterizing the influence of the growth equipment is extracted. In actual production, the first morphology data of the wafer under test before growth is obtained based on the first measurement method, and the second morphology data of the wafer after growth is obtained based on the first morphology data and the transfer coefficient. Select the equipment fingerprint spectrum according to the current production stage, and overlay the second morphology data with the equipment fingerprint spectrum to obtain the predicted morphology data; Defect identification is performed on the predicted topography data to generate a risk navigation map, which includes the distribution of defect risks on the wafer surface; Based on the risk navigation map, the first measurement method is modified to obtain a second measurement method. After the wafer growth is completed, the wafer is scanned based on the second measurement method to obtain the final product morphology data.

[0006] Furthermore, calculating the transfer coefficient of wafer morphology under the current manufacturing process based on the training data includes the following steps: Perform a fast Fourier transform on the morphology data in the training data to obtain the substrate power spectral density map of the corresponding original morphology data and the finished product power spectral density map of the corresponding finished product morphology data. The first feature matrix of the substrate power spectral density map, the second feature matrix of the process parameters, and the output matrix of the finished product power spectral density map are constructed. The first and second feature matrices are combined into the input matrix. A regression model between the input matrix and the output matrix is ​​established by the partial least squares algorithm. When calculating the weight vector in each iteration of the partial least squares algorithm, an objective function including smoothness regularization penalty is constructed. After the partial least squares algorithm converges, the regression coefficients corresponding to the input matrix are extracted from the regression model as the transfer coefficients.

[0007] Furthermore, extracting the device fingerprint spectrum characterizing the influence of the growth equipment includes the following steps: The finished product morphology data is constructed into an observation data matrix. The observation data matrix is ​​decomposed using a blind source separation algorithm to obtain component vectors. The energy proportion and spatial structure characteristics of each component vector are analyzed. Components with repeatability and specific geometric symmetry are identified as inherent components of the equipment. The identified inherent components of the equipment are then reconstructed into an equipment fingerprint map.

[0008] Further, defect identification is performed on the predicted morphology data to generate a risk navigation map, including the following steps: The predicted topography data is processed based on the mean filter kernel to obtain a local mean map. By calculating the difference between the predicted topography data and the local mean map, a local comparison map is generated. The absolute value of the local comparison map is used as the first risk map. Calculate the horizontal and vertical gradient maps of the predicted topography data, merge the horizontal and vertical gradient maps into a gradient magnitude map, calculate the mean and standard deviation of the gradient magnitude map, determine the screening threshold based on the mean and standard deviation, traverse the gradient magnitude map, select points greater than the screening threshold as candidate points, generate a circular path with each candidate point as the center, define each point within the circular path as a control point, connect the control points and candidate points to obtain a connecting line, calculate the angle between the connecting line and the horizontal direction, and obtain the height value of the control point, define the angle and height value as a data pair, calculate the Pearson correlation coefficient of the circular path data pair for each candidate point, and construct a second risk map based on the Pearson correlation coefficient; The first and second risk maps are merged into a risk navigation map.

[0009] Furthermore, merging the first and second risk maps into a risk navigation map includes the following steps: The first risk map is normalized to obtain the first standard map, and the second risk map is selectively normalized to obtain the second standard map. The signal contrast of the first standard map and the second standard map are calculated respectively. The signal contrast is obtained by calculating the ratio of peak value to average value in the image. The signal contrast index is normalized to obtain the corresponding dynamic weight. The first standard map and the second standard map are weighted and summed based on the dynamic weight to obtain the risk navigation map.

[0010] Furthermore, scanning the wafer based on the second measurement method includes the following steps: The first step length in the first measurement method is adjusted based on the risk navigation map, and a dynamic backtracking mechanism is added to obtain the second measurement method. The dynamic backtracking mechanism includes calculating the consistency residual between the measured topography data and the predicted topography data of the interferometric images captured at the current two first measurement points before executing the backtracking mechanism. If both the first stitching residual and the second stitching residual are higher than the quality threshold, but the consistency residual is lower than the preset critical threshold, then backtracking and rescanning will not be triggered.

[0011] The present invention also provides a non-contact epitaxial wafer flatness measuring device, which is used to implement the above-described method, and the device includes: The front-end module acquires training data based on the first measurement method. The training data includes the original morphology data of the wafer before growth and the finished morphology data after growth. Based on the training data, the transfer coefficient of the wafer morphology under the current production process is calculated, and the equipment fingerprint spectrum characterizing the influence of the growth equipment is extracted. In actual production, the prediction module acquires the first morphology data of the wafer before growth based on the first measurement method, acquires the second morphology data of the wafer after growth based on the first morphology data and the transfer coefficient, selects the equipment fingerprint spectrum according to the current production stage, and superimposes the second morphology data with the equipment fingerprint spectrum to obtain the predicted morphology data. The identification module performs defect identification on the predicted morphology data and generates a risk navigation map, which includes the defect risk distribution on the wafer surface. The adjustment module corrects the first measurement method based on the risk navigation map to obtain a second measurement method. After the wafer growth is completed, the wafer is scanned based on the second measurement method to obtain the final product morphology data.

[0012] The beneficial effects of this invention are as follows: This invention first acquires morphology data before and after wafer growth using a first measurement method, and uses this data as training data. Based on the training data, the transfer coefficient of the wafer morphology is calculated, and the equipment fingerprint spectrum characterizing the influence of the equipment is extracted. This completes the model building process of learning growth patterns and equipment characteristics from historical data. In actual production, the first morphology data of the wafer before growth is acquired, and combined with the calculated transfer coefficient and the equipment fingerprint spectrum, to accurately predict the morphology of the finished product after wafer growth. Subsequently, defect identification is performed on the predicted morphology data to generate a risk navigation map containing the distribution of potential defect risks. Finally, the first measurement method is modified based on the risk navigation map to perform intelligent scanning of the grown wafer. This solves the problem of low efficiency and difficulty in achieving defect detection rate in traditional fixed-path scanning, and improves the detection efficiency of flatness of large-diameter epitaxial wafers while ensuring measurement accuracy. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of a non-contact epitaxial wafer flatness measurement method according to the present invention. Figure 2 This is a schematic diagram illustrating the principle of the measurement path of this invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0015] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this application, a first script may be referred to as a second script, and similarly, a second script may be referred to as a first script.

[0016] like Figure 1 As shown, a non-contact epitaxial wafer flatness measurement method includes: Training data is obtained based on the first measurement method. The training data includes the original morphology data of the wafer before growth and the finished morphology data after growth.

[0017] In past production processes, measurements were taken using a primary measurement method to continuously accumulate historical production data, which was then used as training data after reaching a certain quantity. This primary measurement method is based on white light interferometry. Since the field of view of a single measurement using white light interferometry is smaller than the overall size of the wafer, a full wafer scan cannot be completed in one go. Therefore, during the measurement process, a scanning path covering the entire wafer surface is typically preset based on the size of the wafer being measured, such as a serpentine or spiral path. A mechanical scanning stage carries the wafer and moves it stepwise along this path. At each position, the measuring head performs a three-dimensional topographic measurement of the current field of view, acquiring sub-aperture three-dimensional topographic data. Furthermore, the path planning must ensure that there is an overlapping area between any two adjacent measurement fields of view. After the measurement is completed, the data within the overlapping area of ​​the two images is extracted. By calculating a two-dimensional cross-correlation function, the position with the highest correlation peak is found, and the two images are aligned and stitched together. This process is repeated to stitch together all sub-aperture three-dimensional topographic data, thereby obtaining an overall topographic image including the entire wafer surface.

[0018] In this embodiment, for the wafer before epitaxial growth, i.e. the substrate, a white light interferometer is used to perform a full-wafer scan to obtain an interference pattern. The phase of the interference is reconstructed to obtain the original morphology data. After the growth process is completed, i.e. after the epitaxial layer is grown on the wafer surface, the white light interferometer is used again to obtain the corresponding finished morphology data. The original morphology data and the finished morphology data of the same wafer are matched one-to-one to obtain training data.

[0019] The transfer coefficient of wafer morphology under the current production process is calculated based on the training data, and the equipment fingerprint spectrum characterizing the influence of the growth equipment is extracted.

[0020] The transfer factor represents the pattern of how the original geometric features of the substrate are transferred, amplified, or suppressed after the growth process. For example, a point on the wafer surface may have a higher height than the reference plane before growth; after the growth process, the height of this point relative to the reference plane may be reduced or flattened. The device fingerprint refers to the fixed influence of the growth equipment's hardware characteristics or aging on the wafer morphology. By analyzing training data, features with consistent residuals are extracted as the device fingerprint. The specific calculation methods for the transfer factor and the extraction scheme for the fingerprint will be described in detail later.

[0021] In actual production, the first morphology data of the wafer under test before growth is obtained based on the first measurement method, and the second morphology data of the wafer after growth is obtained based on the first morphology data and the transfer coefficient.

[0022] After completing the above processing, in order to predict how the quality of the wafer itself will affect the final result in the actual production process, the first morphology data of the wafer before epitaxial production is still obtained by the first measurement method. Then, the first morphology data is linearly transformed by the transfer coefficient to obtain the second morphology data of the surface morphology of the wafer after growth.

[0023] Select the equipment fingerprint spectrum based on the current production stage, and overlay the second morphology data with the equipment fingerprint spectrum to obtain the predicted morphology data.

[0024] The epitaxial growth process of wafers is complex and influenced by numerous factors. Therefore, after obtaining the initial morphology data, it is necessary to compensate for the equipment's influence to improve the accuracy and reliability of the prediction. Since different growth machines or even the same machine at different maintenance cycles have varying environmental characteristics, it is essential to dynamically select the matching equipment fingerprint. For example, if the system identifies that the wafer under test is assigned to reaction chamber 2 for growth, it retrieves the corresponding equipment fingerprint from the database for reaction chamber 2, and numerically superimposes it with the second morphology data to obtain the predicted morphology data for the grown wafer.

[0025] Defects are identified from the predicted topography data, and a risk navigation map is generated, which includes the distribution of defect risks on the wafer surface.

[0026] The first measurement method is modified based on the risk navigation map to obtain the second measurement method. After the wafer growth is completed, the wafer is scanned based on the second measurement method to obtain the final product morphology data.

[0027] By analyzing anomalous features in the predicted topography data, such as local high points, pits, or abnormal gradients, regions that may form epitaxial defects are identified. In this embodiment, image processing algorithms are applied to scan the predicted topography data, and the data is divided into regions with various risk levels based on the scanning results. The specific algorithms for defect identification and the risk level classification strategies will be described in detail later.

[0028] The first measurement method is a standard scanning mode used to obtain a complete picture without prior knowledge. The second measurement method, however, is an intelligent scanning strategy derived from targeted modifications of the first method, guided by a risk navigation map. After generating the risk navigation map, the sampling step size and movement mechanism in the first measurement method are dynamically adjusted according to the distribution of risk levels, thus constructing the second measurement method. The second measurement method can focus on capturing high-risk areas and quickly pass through low-risk areas. Through these steps, the obtained finished product morphology data ensures a high detection rate of potential defects while significantly reducing the inspection time per wafer.

[0029] In this embodiment, obtaining training data based on the first measurement method includes the following steps: Measurement paths are defined on the wafer surface. The measurement paths consist of first measurement points and second measurement points. The first measurement points are spaced apart by a first step length, and the second measurement points are spaced apart by a second step length, which is smaller than the first step length. The second measurement points are located between the first measurement points.

[0030] In this step, two interferometers are used for measurement. One interferometer measures the first measurement point, and its large step size ensures rapid scanning. The second measurement point serves as an inserted verification point, providing additional information. If stitching problems occur after scanning with a large step size, the interferometric data obtained from the second measurement point can be used for auxiliary stitching. Figure 2 As shown, point A is the first measurement point, and point B is the second measurement point. There is a small first intersection area between adjacent points A, and a larger intersection area between points A and B. In this embodiment, the standard first step length is set to 8mm, the reduced first step length is 6mm, and the second step length is 4mm. The measurement path is defined as a series of staggered coordinate points; for example, the first measurement point sequence is [P1, P3, P5, ...], and the second measurement point sequence is [P2, P4, P6, ...]. Through this design, the system can cover both the first and second measurement points simultaneously with a single large step movement, achieving parallel data acquisition.

[0031] A first interferometer and a second interferometer are set up. The first interferometer and the second interferometer move simultaneously and measure the first measurement point and the second measurement point respectively. The shooting range of the interferometer is set so that in the captured interference image, there is a first intersection area between adjacent first measurement points and a second intersection area between the second measurement point and its adjacent first measurement point.

[0032] The first and second interferometers are mounted on the same precision displacement stage, with a fixed physical distance of the second step. As the stage moves, both interferometers simultaneously acquire data. To ensure data stitching, the field of view of the interferometers is set to be greater than the first step. In this embodiment, the imaging radius is set to 6 mm. After the first interferometer acquires data from P1 and P3 successively, a first intersection region with a width of 2 mm is formed between them because the field of view is greater than half the first step. Similarly, a second intersection region also exists between the data of P2 and its adjacent data from P1 and P3.

[0033] After the interferometer acquires measurements of two first measurement points and the second measurement point between them, during the travel time to the next first measurement point, the first stitching residual between the first intersection areas of the two adjacent first measurement points is calculated. The first stitching residual is compared with a preset quality threshold. If the first stitching residual is not greater than the quality threshold, the interferometric image of the second measurement point is discarded. If the first stitching residual is greater than the quality threshold, the second stitching residual between the second measurement point and the adjacent first measurement point is calculated. If the second stitching residual is less than the quality threshold, the first step length between subsequent first measurement points is reduced. If the second stitching residual is greater than the quality threshold, the current travel is interrupted, the interferometer is controlled to retreat to the previous first measurement point, and a rescan is performed with a step size smaller than the second step length.

[0034] The first stitching residual reflects the consistency of the interferometric images of adjacent first measurement points in the overlapping region. Specifically, the interferometric image referred to in this embodiment is the image after phase reconstruction. This step calculates the stitching residual based on the following method: assuming that region 1 of the previous image and region 2 of the next image are stitched together, the difference between the corresponding point data of regions 1 and 2 is calculated, and the difference is summed to obtain the stitching residual, which is image differencing.

[0035] The intersection area of ​​the interferometric image corresponding to the first measurement point is small, which may result in poor stitching and an increase in stitching residual. When the first stitching residual is small, i.e., not greater than a preset quality threshold, the stitching effect is considered good. In this case, the data of the second measurement point acquired by the second interferometer is considered redundant information and is discarded. In this embodiment, the quality threshold is set to 5. If the first stitching residual between P1 and P3 is calculated to be 3, which is not greater than 5, the stitching quality is considered excellent. Then, the interferometric image of P2 is cleared from memory, and the next measurement is performed at the current step size. This strategy ensures that scanning measurements can be performed with a larger step size, thereby significantly improving the overall scanning efficiency.

[0036] When the first stitching residual is greater than the quality threshold, it indicates that the stitching effect between the first measurement points is poor, and direct stitching is unreliable. In this case, the interferometric image captured by the second measurement point is introduced for repair stitching. If the second stitching residual is less than the quality threshold, it indicates that although the interferometric image captured by the first measurement point with a large span has a poor stitching effect, the interferometric images of the two first measurement points can be well connected by the second measurement point inserted in the middle. This proves that although the region has large fluctuations, it can still be accurately restored by increasing the sampling point density. In this case, the data from the second measurement point is retained and utilized to complete the stitching of the current region. To prevent similar problems in subsequent measurements and reduce reliance on auxiliary repair, the subsequent first step length is reduced, that is, the spacing between the first measurement points is decreased. For example, if the stitching difference between P1 and P3 is 7, but the stitching difference between P2 and the two sides is only 2, the system uses P2 to complete the stitching and reduces the subsequent first step length from 8mm to 6mm.

[0037] When the second stitching residual is also not less than the quality threshold, it means that even with the introduction of a second measurement point for assistance, a good stitching cannot be achieved. This indicates that the measurement process in this area may have been severely disturbed. In this case, the current scan stroke is interrupted, and the stage is controlled to retract to the position of the last reliable first measurement point. Then, the first and second step lengths are simultaneously reduced to significantly increase the sampling density. Finally, the area is rescanned with a finer step size.

[0038] Specifically, this embodiment sets an upper limit on the number of rescans. If stitching still fails after reaching the upper limit in the same area, the area is recorded as a serious defect or measurement failure area, and the area is skipped to continue subsequent scans. When the first step length is reduced, if the first stitching residual is determined to be no greater than the quality threshold multiple times in a row, the first step length and the second step length are gradually restored to the initial set values ​​to improve scanning efficiency again.

[0039] In this embodiment, calculating the transfer coefficient of wafer morphology under the current manufacturing process based on training data includes the following steps: A fast Fourier transform is performed on the morphology data in the training data to obtain the substrate power spectral density map of the corresponding original morphology data and the finished product power spectral density map of the corresponding finished product morphology data.

[0040] The morphological features of a wafer surface include macroscopic warpage, mid-frequency ripples, and microscopic roughness, which are low-frequency, mid-frequency, and high-frequency information, respectively. In the spatial domain, these information are superimposed and greatly affected by position and phase, making it difficult to directly establish a stable transmission law. However, by using Fourier transform to enter the frequency domain, the height fluctuation information in the spatial domain is converted into the power spectral density distribution in the frequency domain, thereby obtaining the power spectral density map of the substrate and the power spectral density map of the finished product. This not only eliminates the registration error caused by phase difference, but also orthogonally decouples the morphological features of different spatial wavelengths.

[0041] The first feature matrix of the substrate power spectral density map, the second feature matrix of the process parameters, and the output matrix of the finished product power spectral density map are constructed. The first and second feature matrices are combined into the input matrix. A regression model between the input matrix and the output matrix is ​​established by the partial least squares algorithm. When calculating the weight vector in each iteration of the partial least squares algorithm, an objective function including smoothness regularization penalty is constructed. After the partial least squares algorithm converges, the regression coefficients corresponding to the input matrix are extracted from the regression model as the transfer coefficients.

[0042] The two-dimensional substrate power spectral density map is flattened into a one-dimensional vector, which serves as the row vector of the first feature matrix. This matrix represents the initial morphological spectral characteristics of the substrate. At the same time, the corresponding epitaxial growth process parameters, such as temperature, pressure, gas flow rate, and rotation speed, are used as the second feature matrix. The first feature matrix and the second feature matrix are horizontally concatenated to form a joint input matrix. The flattened finished product power spectral density map is used as the output matrix.

[0043] This embodiment uses partial least squares (PLS) to establish a regression model. Traditional PLS only pursues maximizing the covariance between input and output, which easily leads to drastic fluctuations in the calculated regression coefficient vector (i.e., weight vector) at adjacent frequency points and contains a large number of non-zero values ​​caused by measurement noise. This contradicts the actual physical transfer law, because the transfer of morphology during epitaxial growth is usually smooth. Therefore, this embodiment constructs an improved objective function that includes a smoothness regularization penalty in the iterative calculation of the weight vector using PLS.

[0044] The objective function is set as follows: on the basis of maximizing the input and output covariance, a smoothing penalty term based on the L2 norm is subtracted. The smoothing penalty term can be in the form of the second difference of the weight vector or the Laplace operator, which constrains the regression coefficients of adjacent frequency points to keep the changes continuous and smooth, and prevents the model from overfitting noise. With the function of the above objective function described, those skilled in the art can easily think of the specific formula construction method, which will not be described in detail here.

[0045] Once the partial least squares method converges, a regression coefficient vector corresponding specifically to the input matrix is ​​extracted from the regression model. This vector represents the proportion by which the substrate components at various spatial frequencies are linearly transmitted to the epitaxial layer surface. In subsequent calculations, the newly measured first morphology data undergoes a Fourier transform and is used to construct an input matrix with process parameters. This matrix is ​​then multiplied by the regression coefficient vector to predict the ideal predicted morphology spectrum that the wafer should theoretically generate under standard processes. Finally, an inverse Fourier transform is used to restore the second morphology data in the spatial domain.

[0046] In this embodiment, extracting the device fingerprint spectrum characterizing the influence of the growth equipment includes the following steps: The finished product morphology data is constructed into an observation data matrix. The blind source separation algorithm is used to decompose the observation data matrix to obtain component vectors. The energy ratio and spatial structure characteristics of each component vector are analyzed. Components with repeatability and specific geometric symmetry are identified as inherent components of the equipment. The identified inherent components of the equipment are then reshaped into an equipment fingerprint map.

[0047] In this embodiment, the device fingerprint spectrum represents the repetitive systematic morphological deviation introduced by the hardware characteristics of a specific production device. For example, due to the characteristics of the device, after the growth is completed, a certain area of ​​the wafer will always have an abnormal thickness of 1 mm. This deviation is superimposed on the second morphological data to obtain the predicted morphological data.

[0048] When acquiring equipment fingerprint data, the finished morphology data of N wafers from the same process batch are first selected. The morphology data of each wafer is a two-dimensional image of M×M pixels. Each two-dimensional image is flattened to obtain N row vectors. Then, the N row vectors are stacked vertically to form an observation data matrix. Each row in the matrix represents the morphology of a wafer, and each column corresponds to a specific spatial location point on the wafer.

[0049] The observation data matrix is ​​then decomposed using a blind source separation algorithm. This embodiment employs Independent Component Analysis (ICA). The specific process of ICA is existing technology and will not be described here. Each component vector decomposed by ICA is evaluated from two dimensions: energy proportion and spatial structure. In terms of energy proportion, the contribution of each component vector to the total variance of the original observation data is calculated. Equipment-inherent components, as recurring deviations, typically have a high and stable energy proportion, while non-systematic components such as random noise have lower energy. For spatial structure, each one-dimensional component vector is reconstructed into a two-dimensional image for visualization analysis. For example, one might observe a grid pattern corresponding to the heater array, concentric circles or spiral patterns related to wafer tray rotation, radial distribution patterns caused by gas inlet spray heads, or overall tilting or warping caused by cavity asymmetry. In contrast, other source components often exhibit randomly located, geometrically irregular speckles or high-frequency noise patterns.

[0050] Finally, the component vectors that simultaneously satisfy the requirements of high energy percentage and possess a clear geometric structure and repeatability are identified as intrinsic components of the device. The identified intrinsic component vectors are then reconstructed into a two-dimensional image, namely the device fingerprint, which includes the biases introduced by the device.

[0051] In this embodiment, defect identification is performed on the predicted morphology data to generate a risk navigation map, including the following steps: The predicted topography data is processed by a mean filter kernel to obtain a local mean map. By calculating the difference between the predicted topography data and the local mean map, a local comparison map is generated. The absolute value of the local comparison map is used as the first risk map.

[0052] Calculate the horizontal and vertical gradient maps of the predicted topographic data, merge them into a gradient magnitude map, calculate the mean and standard deviation of the gradient magnitude map, determine the screening threshold based on the mean and standard deviation, traverse the gradient magnitude map, and select points greater than the screening threshold as candidate points. Generate a circular path centered on each candidate point, define each point within the circular path as a control point, connect the control points and candidate points to obtain a connecting line, calculate the angle between the connecting line and the horizontal direction, and simultaneously obtain the height value of the control point. Define the angle and height value as a data pair, calculate the Pearson correlation coefficient of the circular path data pair for each candidate point, and construct a second risk map based on the Pearson correlation coefficient.

[0053] The first and second risk maps are merged into a risk navigation map.

[0054] The predicted topography data map is essentially a two-dimensional matrix, where each element represents the height value of its corresponding coordinate point. To generate a first risk map characterizing local height anomalies, the predicted topography data map undergoes mean filtering. For example, using a mean filter kernel of a preset size (5x5 or 7x7 square kernel), a two-dimensional convolution operation is performed to traverse the predicted topography data map, calculating the average height within the neighborhood of each pixel, thus generating a local mean map. Next, a local comparison map is obtained by performing pixel-by-pixel matrix subtraction between the predicted topography data map and the local mean map. The pixel values ​​in the local comparison map represent the relative height difference of each point relative to its surrounding background. Finally, the absolute values ​​of all pixel values ​​in the local comparison map are taken to generate the first risk map. In the first risk map, the value of each pixel represents the absolute deviation of that point's height from the average height of its neighborhood. A larger value indicates a greater risk of a protrusion or depression at that location.

[0055] In wafer epitaxial growth, while the aforementioned defect identification methods can effectively detect problems such as bumps and pits, they cannot identify crystal defects like spiral dislocations. Spiral dislocations are a special topological structure formed during wafer growth due to substrate lattice dislocations. They do not appear as isolated bumps or pits, but rather as a series of continuously rising or falling atomic steps centered on the dislocation core. Because the height difference between these steps is extremely small, their contrast in the morphology image is not significant, and they are easily ignored as background noise. However, the presence of spiral dislocations can severely affect the electrical performance and reliability of devices. Therefore, this embodiment also proposes the following method.

[0056] Horizontal and vertical gradient maps are obtained by convolving the predicted topography data with the Sobel or Prewitt gradient operator. These maps are quantized maps of the rate of height change in the x and y directions. Next, the horizontal and vertical gradient maps are merged into a single gradient magnitude map by calculating the Euclidean norm pixel-by-pixel. Each point in the gradient magnitude map represents the degree of height change at the corresponding location in the original topography map. To filter out spiral dislocation nuclei from these regions, the mean and standard deviation of all pixel values ​​in the entire gradient magnitude map are calculated, and a screening threshold is determined based on this, for example, using a three-standard-deviation rule. The gradient magnitude map is then traversed, and all pixel values ​​greater than the screening threshold are marked as candidate points.

[0057] On the gradient magnitude map, for each candidate point, a virtual circular path is generated with that point as the center and within a preset radius R (e.g., 5 pixels). Several control points are uniformly sampled along this path. For each control point, the system calculates the angle between the line connecting it to the central candidate point and the horizontal direction, and simultaneously obtains the height value of the control point itself. Thus, for a candidate point, a set of data pairs consisting of the angle and height is obtained. Finally, using the Pearson correlation coefficient formula, the linear correlation between the angle and height in all data pairs corresponding to a candidate point is calculated. For screw dislocations, their height monotonically increases or decreases approximately linearly with the angle of rotation around the core, so the absolute value of their Pearson correlation coefficient is very close to 1. The absolute value of the Pearson correlation coefficient calculated for each candidate point is used as the risk value at that point, and the positions of other non-candidate points are assigned a value of 0, thereby constructing a second risk map. In the second risk map, the pixel value is a dimensionless correlation measure between 0 and 1; the closer the value is to 1, the higher the probability that the point is the core of a screw dislocation.

[0058] In this embodiment, merging the first risk map and the second risk map into a risk navigation map includes the following steps: The first risk map is normalized to obtain the first standard map. The second risk map is selectively normalized to obtain the second standard map. The signal contrast of the first standard map and the second standard map are calculated respectively. The signal contrast is obtained by calculating the ratio of peak value to average value in the image. The signal contrast index is normalized to obtain the corresponding dynamic weight. The first standard map and the second standard map are weighted and summed based on the dynamic weight to obtain the risk navigation map.

[0059] To obtain the final risk navigation map, the first and second risk maps generated above are merged. Before merging, to address the issue of inconsistent dimensions between the two maps, the first and second risk maps are first normalized, linearly mapping their numerical ranges to the range of 0-1.

[0060] However, if the second risk map is directly normalized, a point with a very low value (e.g., a Pearson correlation coefficient of only 0.3) may be mistakenly magnified to 1, thus losing the measurement of the true severity of the defect; while if no normalization is performed at all, wafers with indistinct spiral features may be covered by the first risk map during the fusion process, resulting in missed detections.

[0061] Therefore, this implementation sets a judgment threshold, such as 0.8, to obtain the actual maximum value of the second risk map within the global range for the theoretical value already in the 0-1 interval. If the actual maximum value is less than the judgment threshold, the spiral feature signal on the current wafer is determined to be weak. In this case, to preserve the absolute physical meaning of the defect intensity, the original second risk map is directly used as the normalized second risk map. Conversely, if the actual maximum value is greater than or equal to the judgment threshold, the current wafer is determined to have at least one spiral defect, and the second risk map is normalized to appropriately enhance its feature intensity.

[0062] After normalization, a set of weighting coefficients w1 and w2, adjustable according to current process concerns, are used to weight and sum the two normalized risk maps. For example, the risk navigation map = w1 * first standard map + w2 * second standard map. The salience of defect features can vary significantly between different wafers or different regions of the same wafer. For instance, in one map, a local height anomaly signal might be very prominent while a spiral feature signal might be unclear; in this case, the system should place more trust in the indication of the first standard map, and vice versa. Fixed weights cannot adapt to such variations, potentially amplifying noise or suppressing critical signals, while a dynamic weighting mechanism ensures that the fusion result is always dominated by the most definitive and reliable information at the moment.

[0063] Specifically, when determining the dynamic weights, the signal contrast index is calculated for the first and second standard images. Signal contrast is the ratio of the image's peak value to its average value. For the first standard image, its signal contrast is calculated. The physical meaning of signal contrast is: if most areas of the image are close to 0, and only a few points are close to 1, then its average value will be very small, while the peak value will be large, resulting in a very high signal contrast. This indicates that the signal in the image is very clear and focused. Conversely, if an image is filled with uniform noise, the difference between its peak value and average value will be small, resulting in a very low signal contrast value. This indicates that the information provided by the image is chaotic and unreliable. After obtaining the signal contrast of each of the two standard images, the signal contrast is normalized to generate the dynamic weights.

[0064] In this embodiment, scanning the wafer based on the second measurement method includes the following steps: The first step length in the first measurement method is adjusted based on the risk navigation map, and a dynamic backtracking mechanism is added to obtain the second measurement method. The dynamic backtracking mechanism includes calculating the consistency residual between the measured topography data and the predicted topography data of the interferometric images captured at the current two first measurement points before executing the backtracking mechanism. If both the first stitching residual and the second stitching residual are higher than the quality threshold, but the consistency residual is lower than the preset critical threshold, then backtracking and rescanning will not be triggered.

[0065] After obtaining the risk navigation map, the first step length is adaptively set based on it. Specifically, firstly, the evaluation range is set. In the risk navigation map, for each preset first measurement point, the values ​​of all first measurement points within its subsequent evaluation range are obtained, and the largest value is extracted as the maximum risk value R_max. The next step length S_next is then calculated based on the maximum risk value. The specific calculation formula is S_next = ⌈S_1 - R_max * K⌉, where S_1 is the currently set first step length, K is a preset step length adjustment coefficient (its value is less than the currently set first step length), and ⌈⌉ is rounded up. Through this step, a large step length high-speed scan can be automatically used in areas predicted to be flat, while a small step length scan can be automatically switched in areas predicted to be complex.

[0066] After obtaining the first and second stitching residuals, if both are higher than the quality threshold, the consistency is calculated. This is done by registering and aligning the measured topographic data of the first and second measurement points with the corresponding measured topographic data portions in the predicted topographic data, and then performing difference calculations to obtain the consistency residual. If the consistency residual is less than the critical threshold, it indicates that the predicted result matches the actual detection result, and the predicted result can be used for auxiliary stitching to avoid the impact of backtracking and rescanning on efficiency. Otherwise, backtracking and rescanning are performed to ensure scanning accuracy.

[0067] The present invention also provides a non-contact epitaxial wafer flatness measuring device for implementing the above-described method, the device comprising: The front-end module acquires training data based on the first measurement method. The training data includes the original morphology data of the wafer before growth and the finished morphology data after growth. Based on the training data, the transfer coefficient of the wafer morphology under the current production process is calculated, and the equipment fingerprint spectrum characterizing the influence of the growth equipment is extracted.

[0068] In actual production, the prediction module acquires the first morphology data of the wafer before growth based on the first measurement method, acquires the second morphology data of the wafer after growth based on the first morphology data and the transfer coefficient, selects the equipment fingerprint spectrum according to the current production stage, and superimposes the second morphology data with the equipment fingerprint spectrum to obtain the predicted morphology data.

[0069] The identification module identifies defects in the predicted morphology data and generates a risk navigation map, which includes the distribution of defect risks on the wafer surface.

[0070] The adjustment module corrects the first measurement method based on the risk navigation map to obtain the second measurement method. After the wafer growth is completed, the wafer is scanned based on the second measurement method to obtain the final product morphology data.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

[0073] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-contact method for measuring the flatness of epitaxial wafers, characterized in that, include: Training data is acquired based on the first measurement method, and the training data includes the original morphology data of the wafer before growth and the finished morphology data after growth. Based on the training data, the transfer coefficient of wafer morphology under the current production process is calculated, and the equipment fingerprint spectrum characterizing the influence of the growth equipment is extracted. In actual production, the first morphology data of the wafer under test before growth is obtained based on the first measurement method, and the second morphology data of the wafer after growth is obtained based on the first morphology data and the transfer coefficient. Select the equipment fingerprint spectrum according to the current production stage, and overlay the second morphology data with the equipment fingerprint spectrum to obtain the predicted morphology data; Defect identification is performed on the predicted topography data to generate a risk navigation map, which includes the distribution of defect risks on the wafer surface; Based on the risk navigation map, the first measurement method is modified to obtain a second measurement method. After the wafer growth is completed, the wafer is scanned based on the second measurement method to obtain the final product morphology data.

2. The method according to claim 1, characterized in that, Obtaining training data based on the first measurement method includes the following steps: A measurement path is defined on the wafer surface. The measurement path consists of a first measurement point and a second measurement point. The first measurement point is spaced apart by a first step length, and the second measurement point is spaced apart by a second step length. The second step length is smaller than the first step length. The second measurement point is between the first measurement point. Set up a first interferometer and a second interferometer. The first interferometer and the second interferometer move simultaneously and measure the first measurement point and the second measurement point respectively. Set the shooting range of the interferometer so that in the captured interference image, there is a first intersection area between adjacent first measurement points and a second intersection area between the second measurement point and its adjacent first measurement point. After the interferometer acquires measurements of two first measurement points and the second measurement point between them, during the travel time to the next first measurement point, the first stitching residual between the first intersection areas of the two adjacent first measurement points is calculated. The first stitching residual is compared with a preset quality threshold. If the first stitching residual is not greater than the quality threshold, the interferometric image of the second measurement point is discarded. If the first stitching residual is greater than the quality threshold, the second stitching residual between the second measurement point and the adjacent first measurement point is calculated. If the second stitching residual is less than the quality threshold, the first step length between subsequent first measurement points is reduced. If the second stitching residual is greater than the quality threshold, the current travel is interrupted, the interferometer is controlled to retreat to the previous first measurement point, and a rescan is performed with a step size smaller than the second step length.

3. The method according to claim 1, characterized in that, Calculating the transfer coefficient of wafer morphology under the current manufacturing process based on the training data includes the following steps: Perform a fast Fourier transform on the morphology data in the training data to obtain the substrate power spectral density map of the corresponding original morphology data and the finished product power spectral density map of the corresponding finished product morphology data. The first feature matrix of the substrate power spectral density map, the second feature matrix of the process parameters, and the output matrix of the finished product power spectral density map are constructed. The first and second feature matrices are combined into the input matrix. A regression model between the input matrix and the output matrix is ​​established by the partial least squares algorithm. When calculating the weight vector in each iteration of the partial least squares algorithm, an objective function including smoothness regularization penalty is constructed. After the partial least squares algorithm converges, the regression coefficients corresponding to the input matrix are extracted from the regression model as the transfer coefficients.

4. The method according to claim 1, characterized in that, Extracting the equipment fingerprint profile characterizing the influence of the growth equipment includes the following steps: The finished product morphology data is constructed into an observation data matrix. The observation data matrix is ​​decomposed using a blind source separation algorithm to obtain component vectors. The energy proportion and spatial structure characteristics of each component vector are analyzed. Components with repeatability and specific geometric symmetry are identified as inherent components of the equipment. The identified inherent components of the equipment are then reconstructed into an equipment fingerprint map.

5. The method according to claim 2, characterized in that, Defect identification is performed on the predicted morphology data to generate a risk navigation map, including the following steps: The predicted topography data is processed based on the mean filter kernel to obtain a local mean map. By calculating the difference between the predicted topography data and the local mean map, a local comparison map is generated. The absolute value of the local comparison map is used as the first risk map. Calculate the horizontal and vertical gradient maps of the predicted topography data, merge the horizontal and vertical gradient maps into a gradient magnitude map, calculate the mean and standard deviation of the gradient magnitude map, determine the screening threshold based on the mean and standard deviation, traverse the gradient magnitude map, select points greater than the screening threshold as candidate points, generate a circular path with each candidate point as the center, define each point within the circular path as a control point, connect the control points and candidate points to obtain a connecting line, calculate the angle between the connecting line and the horizontal direction, and obtain the height value of the control point, define the angle and height value as a data pair, calculate the Pearson correlation coefficient of the circular path data pair for each candidate point, and construct a second risk map based on the Pearson correlation coefficient; The first and second risk maps are merged into a risk navigation map.

6. The method according to claim 5, characterized in that, The steps involved in merging the first and second risk maps into a risk navigation map are as follows: The first risk map is normalized to obtain the first standard map, and the second risk map is selectively normalized to obtain the second standard map. The signal contrast of the first standard map and the second standard map are calculated respectively. The signal contrast is obtained by calculating the ratio of peak value to average value in the image. The signal contrast index is normalized. The first standard map and the second standard map are weighted and summed based on dynamic weights to obtain the risk navigation map.

7. The method according to claim 5, characterized in that, Scanning the wafer based on the second measurement method includes the following steps: The first step length in the first measurement method is adjusted based on the risk navigation map, and a dynamic backtracking mechanism is added to obtain the second measurement method. The dynamic backtracking mechanism includes calculating the consistency residual between the measured topography data and the predicted topography data of the interferometric images captured at the current two first measurement points before executing the backtracking mechanism. If both the first stitching residual and the second stitching residual are higher than the quality threshold, but the consistency residual is lower than the preset critical threshold, then backtracking and rescanning will not be triggered.

8. A non-contact epitaxial wafer flatness measuring device, used to implement the method as described in any one of claims 1-7, characterized in that, The device includes: The front-end module acquires training data based on the first measurement method. The training data includes the original morphology data of the wafer before growth and the finished morphology data after growth. Based on the training data, the transfer coefficient of the wafer morphology under the current production process is calculated, and the equipment fingerprint spectrum characterizing the influence of the growth equipment is extracted. In actual production, the prediction module acquires the first morphology data of the wafer before growth based on the first measurement method, acquires the second morphology data of the wafer after growth based on the first morphology data and the transfer coefficient, selects the equipment fingerprint spectrum according to the current production stage, and superimposes the second morphology data with the equipment fingerprint spectrum to obtain the predicted morphology data. The identification module performs defect identification on the predicted morphology data and generates a risk navigation map, which includes the defect risk distribution on the wafer surface. The adjustment module corrects the first measurement method based on the risk navigation map to obtain a second measurement method. After the wafer growth is completed, the wafer is scanned based on the second measurement method to obtain the final product morphology data.

Images