A method for detecting thickness uniformity of a VCSEL epitaxial wafer

Abstract

The present application belongs to the technical field of image processing, and particularly relates to a VCSEL epitaxial wafer thickness uniformity detection method, which comprises the following steps: obtaining a thickness atlas and a yield atlas of a VCSEL epitaxial wafer after spatial alignment; obtaining a global background thickness of each position in the thickness atlas; calculating the fluctuation standard deviation of the thickness deviation in the neighborhood of each position; obtaining a neighborhood anomaly score, a neighborhood failure rate, and a neighborhood reference parameter; calculating a problem attribution risk index according to the relative deviation and correlation weighted relationship among the product of the neighborhood anomaly score and the neighborhood failure rate of each position, the neighborhood reference parameter, and the Pearson correlation coefficient; and inferring an abnormal mode type and generating an attribution report according to the problem attribution risk index. The present application solves the problem of the lack of VCSEL epitaxial wafer thickness and yield correlation attribution capability in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to a VCSEL epitaxial wafer thickness uniformity detection method. BACKGROUND

[0002] As the core light source device of 3D sensing, optical communication and laser radar and other frontier technologies, the performance of VCSEL is highly dependent on the quality of epitaxial wafer growth, among which, the epitaxial wafer thickness uniformity is a fundamental physical parameter that determines the key performance indicators such as chip wavelength consistency, threshold current and photoelectric conversion efficiency. Therefore, accurate and effective detection of epitaxial wafer thickness uniformity is the prerequisite for ensuring the large-scale production yield and performance stability of VCSEL chips.

[0003] Currently, the detection method commonly used in the industry is to scan the entire epitaxial wafer after epitaxial growth using high-precision optical measurement equipment such as ellipsometer, white light interferometer, etc. to generate a thickness distribution map. Engineers find thickness abnormal areas by manually interpreting the map and combining the sampling test results of terminal chips to deduce process problems. This method can intuitively show the physical distribution of thickness and to a certain extent, it realizes the preliminary monitoring of thickness uniformity.

[0004] However, this method has significant limitations in depth analysis and cause positioning. First, manual interpretation of the map relies on the personal experience of engineers, and for complex or insignificant thickness abnormalities, the judgment standard is not the same, and the subjectivity is strong, which can easily lead to missed or misjudgment. More importantly, this method separates thickness abnormalities from yield loss, and cannot answer the question of whether the thickness deviation in a certain area is the real cause of the yield decline in that area, because yield loss can also be caused by substrate defects, process contamination and other factors. The existing method lacks an effective algorithm to evaluate the causal relationship between thickness abnormalities and yield loss, so it cannot free the analysis personnel from the massive thickness fluctuations to accurately lock the real cause of the problem, resulting in low problem attribution efficiency and long process optimization cycle. SUMMARY

[0005] To solve the technical problem of the lack of correlation and attribution of VCSEL epitaxial wafer thickness and yield in the prior art, the present application provides a VCSEL epitaxial wafer thickness uniformity detection method, comprising:

[0006] The thickness atlas and yield atlas of the VCSEL epitaxial wafer after spatial alignment are obtained; for each position in the thickness atlas, a weighted average of the original thickness values in the neighborhood of the position is performed based on the Euclidean distance of each point in the neighborhood to the position to obtain a global background thickness of the position; a difference between the original thickness value of each position in the thickness atlas and the global background thickness is calculated to obtain a thickness deviation; a fluctuation standard deviation of the thickness deviation in the neighborhood of each position is calculated; a ratio of the absolute value of the thickness deviation of each position to the fluctuation standard deviation is taken as a neighborhood anomaly score; an average yield value of each point in the neighborhood of each position is calculated, and a difference between 1 and the average yield value is taken as a neighborhood failure rate; an average value of the product of the neighborhood anomaly score and the neighborhood failure rate in the neighborhood of each position is calculated to obtain a neighborhood reference parameter; a problem attribution risk index is calculated according to the relative deviation and correlation weighted relationship among the product of the neighborhood anomaly score and the neighborhood failure rate of each position, the neighborhood reference parameter, and the Pearson correlation coefficient; an abnormal mode type is inferred according to the problem attribution risk index; and an attribution report is generated according to the abnormal mode type.

[0007] In the prior art, the standards for manually judging the thickness abnormality of the epitaxial wafer are not unified, and the judgment is prone to be missed or mistaken, and the effective correlation between the thickness and the yield cannot be established, so it is difficult to determine whether the thickness abnormality really leads to yield loss, and other causes such as test failure and impurities may also be attributed to the thickness problem, resulting in low attribution efficiency and blind process optimization. The present application solves this problem through a multi-step design, that is, first, the thickness and yield atlases are accurately aligned to ensure data matching; then, a reasonable thickness reference and fluctuation range are calculated to distinguish normal fluctuation from abnormality; then, the regional failure condition is counted, and the regional reference is set to exclude accidental interference and inherent fluctuation; finally, the causal relationship between the thickness and the failure is accurately judged by the risk index, the abnormal mode is identified, and the attribution is matched to give suggestions. The whole process does not require manual subjective judgment, can accurately lock the real thickness problem, avoids blind adjustment, shortens the process improvement time, greatly improves the accuracy and efficiency of thickness detection and problem attribution, reduces the production of unqualified products, and solves the core pain points of the prior art in the correlation and attribution of thickness and yield.

[0008] Preferably, the thickness atlas and yield atlas of the VCSEL epitaxial wafer after spatial alignment are obtained, comprising:

[0009] For a single VCSEL epitaxial wafer, the thickness distribution data after growth is obtained, denoted as a thickness atlas; at the same time, the yield distribution data of the same epitaxial wafer after the chip manufacturing and testing process is obtained, denoted as a yield atlas; and an image registration algorithm based on feature points is used to accurately spatially align the thickness atlas and the yield atlas.

[0010] Preferably, the global background thickness satisfies the following expression:

[0011] ;

[0012] In the formula, represents the global background thickness of the point in the thickness map; represents the set of all neighboring points of the point in the thickness map, denoted as the neighborhood set, with a neighborhood size of a circular area with a radius of 3 λ ; represents the original thickness value of any point in the neighborhood ; represents the neighborhood distance between the point and the point in the thickness map; represents the smoothing bandwidth parameter.

[0013] The present application can clearly distinguish the thickness variation of the whole epitaxial wafer and the local abnormal condition by calculating the normal thickness reference of each position in a specific way, and will not be disturbed by the natural thickness variation from the edge to the center, avoiding regarding the normal overall variation as a local abnormality. At the same time, the set neighborhood range can reduce the number of points to be calculated under the premise of ensuring the accuracy of calculation, speed up the calculation speed, and without spending too much time on complex calculation, so that the subsequent identification of local thickness abnormality can have an accurate normal reference as a reference, and improve the reliability of abnormal identification.

[0014] Preferably, the fluctuation standard deviation satisfies the following expression:

[0015] ;

[0016] In the formula, represents the fluctuation standard deviation of the point in the thickness map; represents the thickness deviation of the point j in the neighborhood set of the point in the thickness map; represents the neighborhood set of the point in the thickness map; is the number of points in the neighborhood set of the point in the thickness map, > 1; is the neighborhood average deviation of the point in the thickness map.

[0017] The present application can avoid mistaking normal thickness deviation as abnormal or missing the real abnormal, and make the judgment of thickness deviation more in line with the actual situation of different regions, and provide a reasonable basis for subsequent measurement of abnormal degree.

[0018] Preferably, as the neighborhood abnormal score, it includes:

[0019] The absolute difference between the original thickness value of the point in the thickness map and its global background thickness is calculated, and the absolute difference between the original thickness value of the point in the thickness map and its global background thickness is divided by the standard deviation of the neighborhood within which the point is located, and a very small positive number is added to the denominator of the ratio , and the result obtained is taken as the neighborhood abnormal score.

[0020] Preferably, the neighborhood failure rate satisfies the following expression:

[0021] ;

[0022] In the formula, represents the neighborhood failure rate of the point in the thickness map; represents the number of points in the neighborhood set; represents the yield map and locates the yield value of the point in the neighborhood set of the point in the thickness map; represents the neighborhood set of the position in the thickness map.

[0023] The present application can avoid being misled by accidental failure of a single point, such as occasional unqualified points, and will not judge that the entire region has a problem. By statistically analyzing the overall situation in the region, the real performance level of the region can be reflected, the relationship between thickness and region failure can be established, and the judgment will not be affected by the special situation of a single point. The evaluation of the performance of the region is more objective.

[0024] Preferably, the neighborhood reference parameter is obtained, including:

[0025] In the neighborhood set of the point in the thickness map, the average value of the product of the neighborhood failure rate and the neighborhood abnormal score of all neighborhood points is calculated, and this average value is taken as the neighborhood reference parameter.

[0026] Preferably, the problem attribution risk index satisfies the following expression:

[0027] ;

[0028] wherein, represents the problem attribution risk index of the point k in the thickness map; 、 represents the neighborhood anomaly score and the neighborhood failure rate of the point in the thickness map; represents the Pearson correlation coefficient of the data sequence composed of the neighborhood anomaly scores of all neighborhood points in the neighborhood set of the point k in the thickness map and the data sequence composed of the neighborhood failure rates of all neighborhood points in the same position; represents the neighborhood reference parameter; represents the maximum value function; represents the exponential function with the natural constant as the base number; represents a very small positive number, which ensures that the denominator is not 0.

[0029] The present application can accurately distinguish whether the thickness problem really leads to failure by calculating the risk index through this method, and will not attribute the failure caused by other reasons such as test failure and impurities to the thickness problem, thereby excluding irrelevant interference. At the same time, the risk difference of different regions can be amplified, so as to facilitate and quickly find out the key problem area. The calculation will not be wrong due to extreme conditions. Engineers do not need to blindly search in a large amount of data, and can quickly lock the problem caused by thickness, thereby improving the accuracy and efficiency of problem positioning.

[0030] Preferably, the abnormal mode type is inferred according to the problem attribution risk index, including:

[0031] The problem attribution risk indexes of all points in the thickness map are rendered into a pseudo-color map to form a problem attribution map; the regions highlighted in the problem attribution map are subjected to image segmentation processing to extract independent abnormal regions; morphological feature extraction is performed on each independent abnormal region to form an abnormal feature parameter set; and the abnormal feature parameter set is input into a pre-trained pattern recognition model to identify the abnormal mode type of the abnormal region.

[0032] Preferably, the attribution report is generated according to the abnormal mode type, including:

[0033] The identified abnormal mode type is matched with a pre-established process knowledge base, so as to infer the potential process attribution leading to yield loss, and generate an attribution report; and the system automatically generates a targeted process optimization suggestion according to the attribution report.

[0034] The beneficial effects of the present application are that: as the core device of the frontiers such as 3D sensing, optical communication, laser radar, etc., the quality stability of the VCSEL product directly affects the landing and development of the downstream technology. The traditional detection method relying on manual experience not only has low efficiency and large error, but also delays the opportunity for process improvement, causing a lot of production waste. The present application changes the traditional extensive detection mode through the systematic and intelligent thickness detection and problem attribution scheme, which can efficiently and accurately find the thickness problems in production, accurately locate the process reasons behind, and give specific improvement suggestions, so that the production adjustment is more targeted. This not only can stabilize the production quality of the VCSEL epitaxial wafer, improve the proportion of qualified products, and reduce the material and time cost, but also can help enterprises improve the production control ability and reduce the market risk caused by quality problems. From the industry level, the application of the present application can promote the transformation of VCSEL production from experience-driven to data-driven, improve the production efficiency and product competitiveness of the whole industry, provide reliable device guarantee for the technical breakthrough of related frontiers, and help the high-quality development of the whole industry chain. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a flow chart schematically showing a VCSEL epitaxial wafer thickness uniformity detection method in the present application. DETAILED DESCRIPTION

[0036] The embodiment of the present application discloses a VCSEL epitaxial wafer thickness uniformity detection method, referring to Figure 1 , including steps S1-S4:

[0037] S1: Obtain the thickness map and yield map of the VCSEL epitaxial wafer after spatial alignment; for each position in the thickness map, perform weighted average on the original thickness values in the neighborhood based on the Euclidean distance from each point in the neighborhood to the position, to obtain the global background thickness of the position.

[0038] It should be noted that the production of VCSEL epitaxial wafer is a complex process involving multiple process steps from epitaxial growth to chip manufacturing and testing. In order to monitor and evaluate the production process and product quality from different dimensions, it is necessary to obtain the thickness map and yield map of the VCSEL epitaxial wafer. The thickness map is the thickness distribution data obtained by a specific measurement method after the growth of a single VCSEL epitaxial wafer is completed. It reflects the thickness information of each position on the epitaxial wafer and presents the spatial distribution of the thickness in a graphical manner, like a panoramic photo of the thickness condition of the epitaxial wafer. The yield map is the yield distribution data obtained after the same epitaxial wafer goes through the chip manufacturing and testing process. It clearly shows the qualification and performance distribution of the chip at different positions, indicating which areas have high yield and which areas may have problems. The qualified area is marked as 1 and the unqualified area is marked as 0, reflecting the performance results of different areas on the epitaxial wafer at the testing moment. In the correlation analysis of VCSEL epitaxial wafer thickness and yield, in order to make the data correspond one by one, ensure the effectiveness of statistical analysis, and exclude the interference of invalid areas, it is necessary to ensure that the epitaxial wafer area involved in the two is completely consistent.

[0039] Specifically, for a single VCSEL epitaxial wafer, the thickness distribution data after growth is obtained, denoted as a thickness map; at the same time, the yield distribution data after the chip manufacturing and testing process is obtained, denoted as a yield map; and a feature point-based image registration algorithm is used to accurately align the thickness map and the yield map in space.

[0040] It should be noted that to accurately identify the local thickness anomaly affecting the yield in the thickness map, the normal thickness reference, i.e. the global background thickness of the epitaxial wafer, needs to be determined first. If the original thickness value is directly analyzed, it is easy to be disturbed by the overall thickness trend of the wafer, such as the natural gradual change from the edge to the center, which cannot distinguish between the overall trend and the local anomaly, resulting in loss of accuracy in subsequent anomaly identification. Therefore, the global background thickness at point k needs to be calculated.

[0041] Preferably, the original thickness value of each position in the thickness map and the Euclidean distance from each position to its neighborhood points are obtained, denoted as neighborhood distance; the smoothing bandwidth parameter is automatically determined based on the physical scale experience rule according to the wafer physical size and data resolution; and the global background thickness is calculated according to the neighborhood distance and the smoothing bandwidth parameter in the thickness map, including:

[0042] The global background thickness satisfies the following expression:

[0043] ;

[0044] In the formula, Indicates the midpoint of the thickness map Global background thickness; Indicates position in thickness map The set of all neighboring points is denoted as the neighborhood set, and the size of the neighborhood is a radius of 3. λ A circular area; Representing the neighborhood One point of responsibility The original thickness value; In the thickness map, the neighboring points With position The neighborhood distance; This represents the smooth bandwidth parameter.

[0045] In the formula, This represents a Gaussian kernel function used as a weighting factor, whose value varies with the neighborhood points. To the center position distance It increases and then decreases, and the rate of decrease is determined by the smoothing bandwidth parameter; Indicates the position in the thickness map The thickness value of each point in the neighborhood Assign appropriate weights; Indicates the position in the thickness map The sum of the original thickness values ​​of all neighboring points in the neighborhood set multiplied by their corresponding weights reflects the weighted thickness accumulation. Indicates position in thickness map The sum of the weights of all points within the neighborhood set is used to normalize the weighted sum of the numerator, making the final result... Within a reasonable numerical range, it can accurately reflect the position in the thickness map. And the background thickness of the neighborhood; Indicates the position in the thickness map The original thickness values ​​of all points in the neighborhood set are weighted according to the rule that the closer the neighborhood point is to position k, the greater the weight, and the farther the distance, the smaller the weight. Then, the weighted sum is divided by the total weight to obtain the global background thickness at position k, which reflects the background thickness of the position and the surrounding area.

[0046] It should be noted that setting the neighborhood size to a circular region with a radius of 3λ is based on the rapid decay characteristic of the Gaussian kernel function, achieving an optimal balance between computational efficiency and numerical accuracy. Specifically, when the distance between a neighboring point and the center point reaches 3λ, the thickness value of points further away contributes very little to the final calculation result of the background thickness and can be ignored. Therefore, by setting this, the theoretically infinite calculation range can be reduced to a finite local region without sacrificing computational accuracy, thereby greatly improving the algorithm's running efficiency.

[0047] Thus, the global background thickness of any point in the thickness map is obtained.

[0048] S2: Calculate the difference between the original thickness value of each position in the thickness map and the global background thickness to obtain the thickness deviation; calculate the fluctuation standard deviation of the thickness deviation in the neighborhood of each position; take the ratio of the absolute value of the thickness deviation of each position to the fluctuation standard deviation as the neighborhood anomaly score.

[0049] It should be noted that after obtaining the global background thickness of any point in the thickness map, the degree of deviation of each point from the background can be directly calculated, but the absolute size of this deviation value is not sufficient to explain its abnormality. A deviation value may be significant in a flat area, but it may be normal in a region with severe fluctuations, so it is necessary to measure the fluctuation level of the region first.

[0050] Specifically, according to the difference between the global background thickness and the original thickness value of all points in the thickness map, the thickness deviation of each point is calculated, denoted as thickness deviation; the number of pixel points in the neighborhood set of all points in the thickness map is counted, and the average value of the thickness deviation of the pixel points in the neighborhood set is calculated, denoted as neighborhood average deviation; according to the thickness deviation, the number of pixel points and the neighborhood average deviation, the fluctuation standard deviation of the thickness deviation in the neighborhood of all points in the thickness map and the neighborhood average deviation is calculated, denoted as fluctuation standard deviation, including:

[0051] The fluctuation standard deviation satisfies the following expression:

[0052] ;

[0053] In the formula, denotes the fluctuation standard deviation of point in the thickness map; denotes the thickness deviation of point j in the neighborhood set of point in the thickness map; denotes the neighborhood set of position in the thickness map; is the number of points in the neighborhood set of point in the thickness map, > 1; is the neighborhood average deviation of position in the thickness map.

[0054] In the formula, denotes the square of the difference between each thickness deviation and its neighborhood average deviation in the neighborhood of point in the thickness map, which is used to evaluate the degree of deviation of the thickness deviation of point from the neighborhood average level, denoted as deviation degree; denotes the square of the difference between each thickness deviation and its neighborhood average deviation in the neighborhood of point the sum of the deviation degrees of all points in the neighborhood set of the point the variance of the deviation degrees of the thickness deviations in the neighborhood of the point from the average level of the neighborhood, denoted as neighborhood variance; denotes the square root of the neighborhood variance, thereby obtaining the neighborhood standard deviation, which evaluates the fluctuation degree of the deviation degrees in the neighborhood of the point .

[0055] It should be noted that in the VCSEL epitaxial wafer thickness anomaly detection, only the thickness deviation, that is, the difference between the original thickness and the global background thickness, cannot accurately determine the anomaly. The thickness fluctuation characteristics of different regions of the epitaxial wafer are different. For example, the thickness fluctuation in the central region is gentle, and a small deviation may be an anomaly; the edge region fluctuates greatly due to the growth process, and a large deviation may be within the normal range. If the regional fluctuation level is not combined for judgment, it is easy to misjudge normal as abnormal or miss the judgment of abnormal as normal. At the same time, when calculating the fluctuation standard deviation, there may be a situation that the fluctuation of part of the region is extremely small, that is, the standard deviation tends to 0, and directly using it as the denominator will cause the numerical calculation to fail, so a small constant needs to be introduced, and the neighborhood anomaly score needs to be standardized by the ratio of the deviation to the fluctuation level, so that the anomaly degree of different regions of the whole wafer has a unified measurement standard, and the thickness anomaly position that has a practical impact on the subsequent steps can be accurately screened out.

[0056] Preferably, the original thickness value and the global background thickness of the point in the thickness atlas, and the fluctuation standard deviation are obtained; a small positive number is set; the neighborhood anomaly score of the point in the thickness atlas is calculated, denoted as neighborhood anomaly score, including:

[0057] The neighborhood anomaly score satisfies the following expression:

[0058] ;

[0059] In the formula, denotes the neighborhood anomaly score of the point in the thickness atlas; denotes the original thickness value of the point in the thickness atlas; denotes the global background thickness of the point in the thickness atlas; denotes the fluctuation standard deviation of the point in the thickness atlas; denotes a small positive number, which ensures that the denominator is not 0.

[0060] In the formula, denotes the neighborhood anomaly score of the point the absolute difference between the original thickness value of the point and its global background thickness, reflecting the degree of thickness deviation from the neighborhood background; represents the standard deviation of the thickness of the point in the thickness map plus a very small positive number , used as the denominator of the normalization; represents the deviation of the original thickness value of the point from the global background thickness, compared with the overall fluctuation level of such deviation in the neighborhood of the point in the thickness map, such normalization processing realizes the standardized analysis of the anomaly, the greater, indicating that the thickness deviation of the point in the thickness map relative to its neighborhood background is more significant in its neighborhood background, so the neighborhood anomaly score of the point in the thickness map is greater.

[0061] S3: Calculate the average yield value of each point in the neighborhood of each position, and take the difference between 1 and the average yield value as the neighborhood failure rate; in the neighborhood of each position, calculate the average value of the product of the neighborhood anomaly score and the neighborhood failure rate to obtain the neighborhood reference parameter; according to the relative deviation and correlation weighted relationship among the product of the neighborhood anomaly score and the neighborhood failure rate of each position, the neighborhood reference parameter and the Pearson correlation coefficient, the problem attribution risk index is calculated.

[0062] It should be noted that the local anomaly score calculated only by the thickness map can only reflect the degree of thickness deviation of the position in the physical layer, and to determine whether such thickness anomaly really affects the yield value of the corresponding point in the neighborhood, it is necessary to investigate the failure condition of the corresponding area in combination with the yield map. Because the yield value of a single area may be affected by random factors such as test noise and local small defects, there may be accidental failure, and by calculating the neighborhood failure rate, the collective failure condition of all points in the neighborhood of the point k in the thickness map can be effectively smoothed, and whether the performance problem of the collective failure of all points in the neighborhood occurs in the area can be accurately reflected, and whether there is a significant performance failure problem in the area in the neighborhood.

[0063] Specifically, the yield value of each point in the neighborhood set of the point in the thickness map is obtained, and the number of points in the neighborhood set is obtained, and the neighborhood failure rate of the point in the thickness map is calculated, denoted as neighborhood failure rate, including:

[0064] The neighborhood failure rate satisfies the following expression:

[0065] ;

[0066] In the formula, neighborhood failure rate of a point in the thickness map; number of points in the neighborhood set of a point yield map and locating the yield value of a point in the neighborhood set of a point in the thickness map in the neighborhood set of a point in the thickness map neighborhood set of a point

[0067] wherein, sum of yield values of all points in the neighborhood of a point in the thickness map; proportion of qualified chips in the neighborhood area of a point in the thickness map, i.e., neighborhood average yield; neighborhood average yield minus 1 collective failure of all points in the neighborhood of a point in the thickness map, effectively smoothing the randomness of individual point failure.

[0068] It should be noted that different regions of the VCSEL epitaxial wafer have significant inherent fluctuation differences due to process influences such as airflow and temperature, and traditional global reference is prone to misjudgment of normal fluctuation as a problem. The present application uses the average value of the product of the neighborhood failure rate and the neighborhood anomaly score of point k as the neighborhood reference parameter to exclude the fluctuation interference in the neighborhood.

[0069] Preferably, in the neighborhood set of a point in the thickness map, the average value of the product of the neighborhood failure rate and the neighborhood anomaly score of all neighborhood points is calculated and used as the neighborhood reference parameter.

[0070] It should be noted that VCSEL epitaxial wafer yield loss can be caused by various factors such as probe test failure, substrate impurities, and electrode preparation defects, and these factors are not all related to thickness. If the positive correlation between thickness anomalies and yield loss is not screened, yield loss caused by non-thickness reasons will be mistakenly attributed to thickness problems. Meanwhile, some regions have joint deviations of thickness anomalies and yield loss, but the positive correlation between them is very weak, which indicates that yield loss is more likely to be caused by other factors. If the degree of deviation is not positively correlated and weighted, the thickness-related risk of the region will be exaggerated.

[0071] Preferably, according to the neighborhood reference parameter, the fluctuation standard deviation, the neighborhood failure rate, and the neighborhood anomaly score, the problem attribution risk index of a point in the thickness map is calculated, denoted as problem attribution risk index, including:

[0072] The problem attribution risk index satisfies the following expression:

[0073] ;

[0074] In the formula, This represents the attribution risk index of point k in the thickness map. , Indicates the midpoint of the thickness map Neighborhood anomaly score, neighborhood failure rate; The Pearson correlation coefficient represents the data sequence of neighborhood anomaly scores of all neighboring points in the neighborhood set of point k in the thickness map, arranged by position, and the data sequence of neighborhood failure rates of all neighboring points arranged by the same position. Represents the neighborhood reference parameters; Represents the maximum value function; Represents an exponential function with the natural constant as its base; It represents a very small positive number, and guarantees that the denominator is not 0.

[0075] In the formula, This means that only the positive correlation portion of the Pearson correlation coefficient is retained, i.e. When >0, take Otherwise, take 0 to ensure that subsequent calculations only focus on the positive correlation between thickness anomalies and yield loss, and exclude interference from the two being unrelated or negatively related. Represents the midpoint of the thickness map The combined influence strength of neighborhood anomaly score and neighborhood failure rate is such that the larger the neighborhood anomaly score and the larger the neighborhood failure rate, the larger the product and the stronger the correlation between the two problems. Indicates the midpoint of the thickness map The degree of deviation of the product of the neighborhood anomaly score and the neighborhood failure rate relative to the neighborhood baseline parameter; This indicates that the degree of deviation is weighted. In the middle, only the positive correlation is strong. >0, and deviates significantly. The larger the value, the greater the response of the problem attribution risk index. If there is no positive correlation, even Even if it is large, it will not lead to a large response in the problem attribution risk index, thus avoiding misjudgment; In the middle, when The larger, The closer it is to 0, the closer it is to 1; the exponential function amplifies the changes in risk-related assessment data non-linearly. In the middle, the smaller the denominator, the better. The larger the value, the stronger the positive correlation and the greater the deviation from the normal baseline. The larger the value, the higher the risk of attributing the problem to the cause, and vice versa. By whether there is a real causal relationship between the thickness anomaly and yield failure, and output a problem attribution risk index, to determine whether the thickness map point k is really caused by the thickness problem leading to yield decline.

[0076] S4: According to the problem attribution risk index, infer the abnormal mode type; according to the abnormal mode type, generate an attribution report.

[0077] Specifically, render the problem attribution risk index of all thickness map points into a pseudo-color map to form a problem attribution map; perform image segmentation processing on the highlighted area in the problem attribution map to extract independent abnormal areas; perform morphological feature extraction on each independent abnormal area to form an abnormal feature parameter set; input the abnormal feature parameter set into a pre-trained pattern recognition model to identify the abnormal mode type of the abnormal area.

[0078] It should be noted that the morphological feature extraction of each independent abnormal area is to calculate its area, perimeter, aspect ratio, circularity or principal axis direction and other geometric parameters; the input of the abnormal feature parameter set into the pre-trained pattern recognition model is a support vector machine; and the abnormal mode type includes ring-shaped abnormality, stripe abnormality, center point abnormality or edge collapse.

[0079] Preferably, the identified abnormal mode type is matched with a pre-established process knowledge base to infer the potential process attribution leading to yield loss and generate an attribution report; and the system automatically generates targeted process optimization suggestions according to the attribution report.

[0080] It should be noted that the process knowledge base stores different abnormal mode types and VCSEL epitaxial growth process parameters, such as MOCVD reaction temperature, gas flow, wafer rotation speed, etc. The system automatically generates targeted process optimization suggestions, such as adjusting the temperature field distribution of the MOCVD equipment, optimizing the gas flow uniformity or calibrating the wafer rotation mechanism, to eliminate or reduce the thickness anomaly leading to yield loss.

[0081] At this point, the uniformity detection of the VCSEL epitaxial wafer thickness is completed.

[0082] Although the present specification has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that many modifications, changes and substitutions can be made without departing from the idea and spirit of the present application.

Claims

1. A method for detecting the thickness uniformity of VCSEL epitaxial wafers, characterized in that, The method comprises the following steps: obtaining a thickness atlas and a yield atlas of a VCSEL epitaxial wafer after spatial alignment; for each position in the thickness atlas, performing a weighted average of the original thickness values in the neighborhood of the position based on the Euclidean distance of each point in the neighborhood to the position to obtain a global background thickness of the position; calculating the difference between the original thickness value of each position in the thickness atlas and the global background thickness to obtain a thickness deviation; calculating the fluctuation standard deviation of the thickness deviation in the neighborhood of each position; taking the ratio of the absolute value of the thickness deviation of each position to the fluctuation standard deviation as a neighborhood anomaly score; calculating the average yield value of each point in the neighborhood of each position, and taking the difference between 1 and the average yield value as a neighborhood failure rate; calculating the average value of the product of the neighborhood anomaly score and the neighborhood failure rate in the neighborhood of each position to obtain a neighborhood reference parameter; calculating a problem attribution risk index according to the relative deviation and correlation weighted relationship among the product of the neighborhood anomaly score and the neighborhood failure rate of each position, the neighborhood reference parameter, and the Pearson correlation coefficient; inferring an abnormal pattern type according to the problem attribution risk index; generating an attribution report according to the abnormal pattern type.

2. The method for detecting the thickness uniformity of a VCSEL epitaxial wafer according to claim 1, characterized in that, The method comprises the following steps: for a single VCSEL epitaxial wafer, obtaining thickness distribution data after growth, denoted as a thickness atlas; at the same time, obtaining yield distribution data after chip manufacturing and testing processes, denoted as a yield atlas; and performing accurate spatial alignment of the thickness atlas and the yield atlas by using a feature point-based image registration algorithm.

3. The method of claim 1, wherein the method further comprises: The global background thickness satisfies the following expression: ； wherein, denotes the global background thickness of the point in the thickness map; denotes the set of all neighboring points of the point in the thickness map, denoted as the neighborhood set, with a neighborhood size of a circular region with a radius of 3 λ ; denotes the original thickness value of any point in the neighborhood ; denotes the neighborhood distance between the point and the point in the thickness map; denotes the smoothing bandwidth parameter.

4. The method of claim 1, wherein the method further comprises: The fluctuation standard deviation satisfies the following expression: ； wherein denotes the standard deviation of the fluctuations of the points of the thickness map; denotes the thickness deviation of the points j of the neighborhood set of the points of the thickness map; denotes the thickness deviation of the points j of the neighborhood set of the points of the thickness map; denotes the thickness deviation of the points j of the neighborhood set of the points of the thickness map; denotes the neighborhood set of the points of the thickness map at position denotes the neighborhood set of the points of the thickness map at position is the number of points within the neighborhood set of the points of the thickness map at position > 1; is the neighborhood average deviation of the points of the thickness map at position is the neighborhood average deviation of the points of the thickness map at position is the neighborhood average deviation of the points of the thickness map at position 5. The method of claim 1, wherein the method further comprises: The neighborhood anomaly score comprises: Calculate the midpoint of the thickness map The absolute difference between the original thickness value and its global background thickness, and the midpoint of the thickness map. The absolute difference between the original thickness value and its global background thickness and the midpoint of the thickness map. Compare the standard deviations of the fluctuations within the same neighborhood, and add a very small positive number to the denominator of the ratio. The results are used as neighborhood anomaly scores.

6. The method of claim 1, wherein the method further comprises: The neighborhood failure rate satisfies the following expression: ； wherein represents the neighborhood failure rate of a point in the thickness map; represents the neighborhood set of a point in the thickness map; represents the number of points in the neighborhood set of a point; represents the yield map and locates the yield value of a point in the neighborhood set of a point in the thickness map; represents the neighborhood set of a point in the thickness map; represents the yield map and locates the yield value of a point in the neighborhood set of a point in the thickness map; represents the neighborhood set of a point in the thickness map; represents the neighborhood set of a point in the thickness map.

7. The method of claim 1, wherein the method further comprises: The neighborhood reference parameter is obtained by: In the neighborhood set of the point of the thickness map, the average of the product of the neighborhood failure rate and the neighborhood anomaly score of all neighborhood points is calculated and used as the neighborhood reference parameter.

8. The method of claim 1, wherein the method further comprises: The problem attribution risk index satisfies the following expression: ； wherein, represents the point k in the thickness map's risk index of being attributed to a problem; , represents the point k in the thickness map's neighborhood anomaly score, neighborhood failure rate; represents the Pearson correlation coefficient of the data sequence of the neighborhood anomaly scores of all neighborhood points in the neighborhood set of the point k in the thickness map and the data sequence of the neighborhood failure rates of all neighborhood points in the same neighborhood set of the point k in the thickness map; represents the neighborhood reference parameter; represents the maximum function; represents the exponential function with the natural constant as the base number; represents the minimum positive number, which ensures the denominator not to be 0.​ 9. The method of claim 1, wherein the method further comprises: The abnormal pattern type is inferred according to the problem attribution risk index, which comprises: rendering the problem attribution risk index of all points in the thickness atlas into a pseudo-color atlas to form a problem attribution map; performing image segmentation processing on the highlighted area in the problem attribution map to extract independent abnormal areas; performing morphological feature extraction on each independent abnormal area to form an abnormal feature parameter set; inputting the abnormal feature parameter set into a pre-trained pattern recognition model to identify the abnormal pattern type of the abnormal area.

10. The method of claim 1, wherein the method further comprises: The attribution report is generated according to the abnormal pattern type, which comprises: matching the identified abnormal pattern type with a pre-established process knowledge base to infer potential process attributions causing yield loss, and generating an attribution report; the system automatically generates targeted process optimization suggestions according to the attribution report.

Images