Wafer test data outlier screening method

Abstract

The application discloses a kind of abnormal point screening methods of wafer test data, comprising the following steps: obtaining the original data set of wafer test, original data set includes multiple data items, and each data item at least includes first parameter and second parameter;Multiple data items are divided to obtain several parameter interval groups;Determine the dynamic threshold of the second parameter corresponding to each parameter interval group;For each data item, the determination threshold of the second parameter is calculated;The second parameter of each data item is compared with the determination threshold of the second parameter, if the second parameter is less than the determination threshold of the second parameter, then the data item is marked as abnormal point;Eliminate all data items marked as abnormal point, output remaining data item as clean data set.The determination threshold of data item is calculated based on the dynamic threshold of local statistical characteristics, to realize the efficient screening of abnormal point.

Description

Technical Field

[0001] This invention relates to the field of semiconductor chip testing technology, and more specifically, to a method for screening outliers in wafer test data. Background Technology

[0002] Semiconductor wafers are the fundamental carriers for manufacturing integrated circuits or optoelectronic devices. A single wafer typically integrates tens of thousands to hundreds of thousands of individual chip units. To ensure product yield and reliability, each chip unit must be tested for electrical performance or optical parameters before packaging. This process, known as wafer testing, is an indispensable part of quality control and yield analysis in semiconductor manufacturing.

[0003] In actual testing, probe stations collect multiple electrical and optical parameters, including voltage, current, emission wavelength, and light intensity, by contacting the chip electrodes with probes, forming large-scale wafer test data. This data is directly used for chip performance grading, yield screening, and process feedback. However, due to non-process factors such as poor probe contact, environmental noise, equipment transient fluctuations, and chip defects themselves, the raw test data often contains discrete outliers that do not conform to statistical laws. These outliers not only deviate from the normal performance distribution range of chips but also lack an explicit correlation with their spatial location on the wafer, causing conventional statistical analysis models to fail.

[0004] If data containing such anomalies is used directly for subsequent analysis, it is very easy for chips to be incorrectly classified into non-corresponding performance levels. This will not only affect the quality of product delivery, but also distort yield statistics, mislead the direction of process optimization, and reduce the credibility of quality control and production data reporting.

[0005] Existing technologies have addressed the problem of anomaly identification in wafer test data. For example, Chinese invention patent application CN120277314A proposes a discrete risk die detection method based on the slope difference of the cumulative probability distribution. This method identifies potentially high-risk dies (i.e., chips with performance discreteness caused by process defects) by recognizing features such as abrupt changes at the tail of the distribution and "no data in the next interval." This method is effective in handling systematic discreteness caused by process fluctuations. However, the core premise of this approach is that the anomalies originate from actual process defects and exhibit identifiable morphological changes in statistical distribution (such as thick tails and empty intervals). In actual production lines, many anomalies are generated by complex non-process factors, which neither meet the conditions for abrupt distribution changes nor possess interval empty characteristics, and do not conform to a normal distribution, making them difficult for this method to effectively capture.

[0006] Therefore, there is an urgent need for a screening method that focuses on the wafer test data itself and can accurately identify and eliminate outliers to ensure the accuracy of chip grading, the authenticity of yield analysis, and the effectiveness of process feedback. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method for screening outliers in wafer test data, aiming to solve the problem of decreased grading accuracy and insufficient data reliability caused by the presence of abnormal discrete points in semiconductor wafer test data.

[0008] In a first aspect, the present invention provides a method for screening outliers in wafer test data, comprising the following steps: S1. Obtain the raw dataset for wafer testing. The raw dataset contains multiple data items, each of which includes at least a first parameter and a second parameter, wherein the first parameter and the second parameter are electrical parameters or optical parameters. S2. Based on the first parameter of multiple data items, divide the multiple data items into multiple equal-width parameter intervals to obtain several parameter interval groups; S3. Based on the number of data items in each parameter interval group and the value of the second parameter, determine the dynamic threshold of the second parameter corresponding to each parameter interval group; S4. For each data item, determine the parameter interval group it falls into, and select a parameter interval group adjacent to the parameter interval group; based on the first parameter, perform linear interpolation to calculate the interpolation threshold for the dynamic thresholds corresponding to the two parameter interval groups, and combine the preset offset of the second parameter to calculate the judgment threshold of the second parameter. S5. Compare the second parameter of each data item with the judgment threshold of the second parameter. If the second parameter is less than the judgment threshold of the second parameter, mark the data item as an outlier. S6. Remove all data items marked as outliers and output the remaining data items as a clean dataset.

[0009] The outlier screening method for wafer test data provided by this invention calculates the judgment threshold of data items based on dynamic thresholds of local statistical features, achieving efficient screening of outliers. Specifically, data items are first divided into multiple parameter interval groups according to a first parameter. Within each parameter interval group, the dynamic threshold corresponding to that parameter interval group is determined according to the number of data items and the value of a second parameter. This local design can determine the dynamic threshold based on the differences within different parameter interval groups, thereby avoiding misjudgments caused by a globally uniform threshold. Furthermore, when calculating the judgment threshold for each data item, not only is the dynamic threshold of the current parameter interval group used, but the dynamic thresholds of adjacent parameter interval groups are also referenced through a preset weighted mapping relationship, making the judgment threshold continuously change in the parameter space. This mechanism effectively reduces the deviation caused by noise or fluctuations in a single interval, suppresses threshold abrupt changes at interval boundaries, and improves the accuracy of the judgment threshold, thereby simultaneously improving the accuracy and recall of outlier identification. Moreover, this calculation process only relies on the statistically defined parameter interval group thresholds, maintaining high computational efficiency.

[0010] As an optional solution to the anomaly point screening method of the present invention, the specific steps of step S3 are as follows: Obtain the number of data items contained within the interval group of the parameter to be determined; Determine if the quantity is zero; if so, set the dynamic threshold of the second parameter corresponding to the parameter interval group to the preset default value. If the quantity is greater than 0 and less than or equal to the preset quantity threshold, then the dynamic threshold of the second parameter corresponding to the parameter interval group is set to the maximum value of the second parameter in the parameter interval group. If the quantity is greater than the preset quantity threshold, the second parameter in the parameter interval group is sorted, and the value of the second parameter at the preset quantile position is extracted as the dynamic threshold of the second parameter corresponding to the parameter interval group.

[0011] As an optional solution to the anomaly filtering method of the present invention, the data items also include auxiliary parameters, and the anomaly filtering method further includes auxiliary condition filtering: obtaining at least one auxiliary parameter and its corresponding preset condition range, and determining whether the value of the auxiliary parameter of each data item falls within the condition range; if it does not fall within the condition range, the corresponding data item is marked as an anomaly.

[0012] As an optional scheme of the anomaly point screening method of the present invention, the first parameter is an optical parameter and the second parameter is an electrical parameter.

[0013] As an optional scheme of the abnormal point screening method of the present invention, the electrical parameters include voltage parameters and current parameters; the optical parameters include wavelength parameters, light intensity parameters, luminous intensity parameters and light spot quality parameters.

[0014] As an optional scheme of the anomaly point screening method of the present invention, the first parameter is a wavelength parameter and the second parameter is a voltage parameter; in step S2, the wavelength parameter of each data item is rounded down, and based on the preset wavelength fluctuation range width, the rounded wavelength value is divided into multiple parameter ranges of equal width, thereby forming several parameter range groups.

[0015] As an optional solution to the anomaly point screening method of the present invention, in step S4, the determination threshold T of the second parameter... dynamic The calculation formula is: T dynamic =T int +X frac ×(T next -T int -offset Among them, T int T is the dynamic threshold corresponding to the parameter interval group into which the data item falls. next X is the dynamic threshold corresponding to the parameter interval group to the right of the parameter interval group into which the data item falls.frac This represents the decimal part of the wavelength parameter in the data item, and offset is the preset offset of the voltage parameter.

[0016] As an optional solution to the anomaly point screening method of the present invention, step S1 further includes preset wavelength and electrical parameter range, and pre-filtering the data items by lower limit value.

[0017] Secondly, the present invention also provides an anomaly screening system for wafer test data, comprising: The data acquisition module is used to acquire the raw dataset for wafer testing; The parameter interval group construction module is used to divide multiple data items into multiple equal-width parameter intervals to obtain several parameter interval groups. The dynamic threshold calculation module is used to calculate the dynamic threshold of the second parameter corresponding to each parameter interval group; The linear interpolation judgment module is used to calculate the judgment threshold of the test element through linear interpolation and compare it with the second parameter; The anomaly marker module is used to mark data items that do not meet the conditions. The data output module is used to output the clean dataset after removing outliers.

[0018] The outlier screening system for wafer test data provided by this invention first divides data items into multiple parameter interval groups through a parameter interval group construction module, and then calculates the dynamic threshold corresponding to each parameter interval group through a dynamic threshold calculation module to avoid global misjudgment. Simultaneously, a linear interpolation judgment module calculates the judgment threshold of the test element through linear interpolation and compares it with a second parameter, thereby improving the accuracy of outlier identification.

[0019] Additional aspects and advantages of the invention will be set forth in part in the description which follows, some of which will become clear as the description proceeds, and others will be learned by practicing the invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the outlier screening method for wafer test data provided in an embodiment of this application.

[0022] Figure 2 This is a flowchart illustrating the dynamic threshold calculation method in Embodiment 1 of this application.

[0023] Figure 3 This is a schematic diagram of the abnormal point determination process in Embodiment 1 of this application.

[0024] Figure 4 This is a scatter plot comparing the data before and after outlier removal in Embodiment 1 of this application.

[0025] Figure 5 This is a schematic diagram of the auxiliary condition filtering and algorithm parameter configuration window in Embodiment 1 of this application.

[0026] Figure 6 This is a schematic diagram of the multidimensional variable selection and configuration interface in Embodiment 1 of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0028] This application provides a method for screening outliers in wafer test data, including the following steps: S1. Obtain the raw dataset for wafer testing. The raw dataset contains multiple data items, each of which includes at least a first parameter and a second parameter, where the first parameter and the second parameter are optical parameters and electrical parameters, respectively. S2. Based on the first parameter of multiple data items, divide the multiple data items into multiple equal-width parameter intervals to obtain several parameter interval groups; S3. Based on the number of data items in each parameter interval group and the value of the second parameter, determine the dynamic threshold of the second parameter corresponding to each parameter interval group. It should be noted that, for the differences in different parameter interval groups, the dynamic threshold corresponding to each parameter interval group is calculated separately to avoid misjudgment caused by a globally uniform threshold. S4. For each data item, determine the parameter interval group it falls into, and select a parameter interval group adjacent to the parameter interval group; based on the first parameter, perform linear interpolation on the dynamic thresholds corresponding to the two parameter interval groups, and calculate the judgment threshold of the second parameter by combining the preset offset of the second parameter; it should be noted that the calculation of the judgment threshold also uses the dynamic thresholds of adjacent parameter interval groups, which reduces the deviation caused by noise or fluctuation in a single interval, effectively suppresses the threshold mutation at the interval boundary, and improves the accuracy of the judgment threshold; S5. Compare the second parameter of each data item with the judgment threshold of the second parameter. If the second parameter is less than the judgment threshold of the second parameter, mark the data item as an outlier. S6. Remove all data items marked as outliers and output the remaining data items as a clean dataset.

[0029] In some implementations, step S3 specifically involves the following steps: Obtain the number of data items contained within the interval group of the parameter to be determined; If the quantity is zero, the dynamic threshold of the second parameter corresponding to the parameter interval group is set to the preset default value. It should be noted that when the quantity is zero, it means there is no valid data. At this time, the preset default value is the minimum valid threshold of the second parameter. If the quantity is greater than 0 and less than or equal to the preset quantity threshold, then the dynamic threshold of the second parameter corresponding to the parameter interval group is set to the maximum value of the second parameter in the parameter interval group. It should be noted that when the amount of data is small, the dynamic threshold is set to the maximum value to remove abnormal discrete points in this part as much as possible. If the number exceeds a preset threshold, the second parameter in the parameter interval group is sorted, and the value of the second parameter at the preset quantile position is extracted as the dynamic threshold of the second parameter corresponding to the parameter interval group. It should be noted that when the data volume is large, determining the threshold through preset quantiles can effectively filter out interference from extreme outliers, ensuring that the judgment threshold can truly reflect the boundaries of the main data distribution. This makes the judgment threshold more representative and robust in a statistical sense, preserving the integrity and original characteristics of normal data distribution, and avoiding overly broad or strict judgment criteria due to individual noise points.

[0030] In some implementations, the data items also include auxiliary parameters, and the outlier filtering method further includes auxiliary condition filtering: obtaining at least one auxiliary parameter and its corresponding preset condition range, and determining whether the value of the auxiliary parameter of each data item falls within the condition range; if it does not fall within the range, the corresponding data item is marked as an outlier. It should be noted that combining auxiliary parameter condition pre-filtering provides high configuration flexibility and production adaptability, avoids the one-sidedness of setting a single parameter threshold, significantly improves the detection rate of complex failure modes (such as some parameters being normal but specific parameters being abnormal), improves the accuracy and robustness of outlier identification, and ensures the statistical consistency and authenticity of the output clean dataset in various electrical indicators.

[0031] In some implementations, the first parameter is an optical parameter, and the second parameter is an electrical parameter. It should be noted that by dividing parameter interval groups according to optical parameters and using electrical parameters as the criterion, it is possible to accurately identify anomalies where electrical performance deviates significantly from expectations under the same optical background.

[0032] In some implementations, electrical parameters include voltage and current parameters; optical parameters include wavelength, light intensity, luminous intensity, and spot quality parameters. It should be noted that electrical parameters (such as leakage current) are extremely sensitive to localized minute defects and exhibit significant local statistical characteristics. This fully leverages the invention's ability to suppress localized random noise, and the combination of linear interpolation to calculate thresholds effectively identifies anomalies.

[0033] In some implementations, the first parameter is a wavelength parameter and the second parameter is a voltage parameter. In step S2, the wavelength parameter of each data item is rounded down, and based on a preset wavelength fluctuation range width, the rounded wavelength value is divided into multiple parameter ranges of equal width, thereby forming several parameter range groups. It should be noted that by using rounding down and combining it with a preset wavelength fluctuation range width (e.g., 0.5nm or 1nm) for equal-width division, grains with physically similar wavelengths that belong to the same "equal wavelength band" or "same temperature field region" can be grouped into the same parameter range group.

[0034] In some implementations, in step S4, the determination threshold T of the second parameter dynamic The calculation formula is: T dynamic =T int +X frac ×(T next -T int -offset Among them, T int T is the dynamic threshold corresponding to the parameter interval group into which the data item falls. next X is the dynamic threshold corresponding to the parameter interval group to the right of the parameter interval group into which the data item falls. frac The value is the decimal part of the first parameter in the data item, and offset is the preset offset of the voltage parameter.

[0035] It should be noted that regarding the interpolation term T... int +X frac ×(T next -T int The present invention generates the fractional part X after rounding the wavelength. frac As an interpolation weighting coefficient, this decimal part naturally represents the relative position of the current grain within its wavelength range. Using this position as a weighting coefficient, a dynamic threshold corresponding to the parameter interval group into which the data item falls is linearly weighted, generating a judgment threshold that continuously and gradually changes with wavelength. Regarding the offset term, subtracting offset from the formula aims to identify abnormal points where the forward voltage is too low (low voltage usually indicates leakage or short circuit defects). Offset serves as a preset safety margin to control the stringency of the judgment boundary, avoiding the misjudgment of good products as abnormal points due to normal process fluctuations or test noise, thereby achieving a balance between the detection rate and the false alarm rate.

[0036] In some embodiments, step S1 further includes preset wavelength and electrical parameter ranges, and pre-filtering the data items using lower limit values. It should be noted that this invention is aimed at silicon-based LED product process platforms, where the emission wavelength of the normal epitaxial structure must fall within a certain range, such as the 500-600nm range. By combining the preset upper and lower limits of voltage or current based on production experience, the effective range of the algorithm can be defined.

[0037] This application also provides an outlier screening system for wafer test data, including: The data acquisition module is used to acquire the raw dataset for wafer testing; The parameter interval group construction module is used to divide multiple data items into multiple equal-width parameter intervals to obtain several parameter interval groups. The dynamic threshold calculation module is used to calculate the dynamic threshold of the second parameter corresponding to each parameter interval group; The linear interpolation judgment module is used to calculate the judgment threshold of the test element through linear interpolation and compare it with the second parameter; The anomaly marker module is used to mark data items that do not meet the conditions. The data output module is used to output the clean dataset after removing outliers.

[0038] The following are some embodiments of this application, and the embodiments of the present invention will further describe the technical solution of the present invention in detail. Example 1

[0039] Please see Figure 1 This embodiment provides a method for screening outliers in wafer test data. The wafer in this embodiment is a semiconductor light-emitting chip wafer, and the method specifically includes the following steps: Step S1: Obtain the raw dataset for wafer testing. The raw dataset contains data items for multiple chip units. Each data item includes a first parameter, a second parameter, and auxiliary parameters.

[0040] Specifically, in this embodiment, the first parameter is the wavelength parameter (ValueX), with the unit being nanometers (nm) and a value range of 500-600nm; the second parameter is the voltage parameter (ValueY), with the unit being volts (V) and a value range of 0-10V. Auxiliary parameters include the current parameter, light intensity parameter, luminous intensity parameter, and light spot quality parameter.

[0041] The preset lower limit of the voltage parameter is set to 1.4V, and data items with voltage parameters lower than 1.4V are pre-selected to eliminate equipment noise interference.

[0042] Step S2: Round down the wavelength parameter of each data item to obtain the integer wavelength value: int WaveLength=Floor(valueX) Where valueX represents the wavelength parameter for each data item, int WaveLength This represents the integer wavelength value obtained by rounding down the wavelength. Floor indicates the rounding down operation.

[0043] The integer wavelength values ​​of all data items are stored in a hash set to obtain a candidate wavelength segment set. For example, in this embodiment, the wavelength values ​​of the data items cover 500nm, 501nm, 502nm...600nm, so the candidate wavelength segment set contains 101 integer wavelength points. If an integer wavelength point has no corresponding data item, it will not participate in subsequent calculations, thus achieving computational pruning.

[0044] Then, iterate through each integer wavelength point in the candidate wavelength range set. Taking the current integer wavelength point as the center, obtain all data items whose wavelength falls within the interval [waveLength-f, waveLength+f], where waveLength is the integer wavelength point and f is a preset fluctuation range parameter. In this embodiment, f=0.2nm, that is, take the data items within a 0.2nm range before and after it. Data items within the same wavelength interval constitute a wavelength interval group. Thus, divide all data items according to wavelength intervals to obtain several wavelength interval groups.

[0045] S3. Based on the number of data items and the values ​​of voltage parameters within each wavelength interval group, determine the dynamic threshold of the voltage parameter corresponding to each wavelength interval group, such as... Figure 2 As shown.

[0046] Specifically, let n be the number of data items in the wavelength interval group, V be the set of voltage parameter values ​​in the wavelength interval group, and T be the dynamic threshold of the voltage parameter corresponding to the wavelength interval group.

[0047] If n=0, it indicates that there are no valid data items in this wavelength interval group, and the dynamic threshold T of the voltage parameter is set to the preset default value. In this embodiment, the preset default value is 1.4V (T=1.4V), which is consistent with the preset lower limit value in step S1.

[0048] If 1≤n≤5, it indicates that there are fewer data items in the wavelength interval group, which usually means that the interval contains discrete points. Therefore, the dynamic threshold T of the voltage parameter is set as the maximum value of the voltage parameter in the wavelength interval group: T=Max(V).

[0049] If n > 5, it indicates that there are sufficient data items within the wavelength interval group. Therefore, the dynamic threshold T for the voltage parameter is calculated using the quantile method. The voltage values ​​are sorted in ascending order, and the value at the p-th quantile position is taken as the threshold T. sortedV = sort(V) index = Floor(n*p) T = sortedV[index] Where sort represents sorting operation, sortedV represents a one-dimensional array obtained by sorting set V in ascending order, p represents the preset quantile ratio, index represents the subscript position of the target quantile in the sortedV array, and floor represents the floor operation.

[0050] In this embodiment, p=0.09, i.e., the 9th percentile. Taking the wavelength interval group [500nm-0.2nm, 500nm+0.2nm] as an example, there are 12 data items in this wavelength interval group, and the voltage parameter values ​​are [1.486, 1.513, 1.532, 1.511, 1.487, 1.527, 1.518, 1.514, 1.479, 1.509, 1.51, 1.521]. The value with index 1 after sorting (12×0.09≈1) is set as the dynamic threshold T of the voltage parameter of this wavelength interval group, i.e., T=1.486V.

[0051] Step S4: For each data item, determine the wavelength interval group it falls into, and select a parameter interval group adjacent to the wavelength interval group; based on the wavelength parameter, perform linear interpolation on the dynamic thresholds corresponding to the two wavelength interval groups, and calculate the judgment threshold of the voltage parameter in combination with the preset offset of the voltage parameter.

[0052] Specifically, let the wavelength parameter of a certain data item take the value X, where the integer part of the wavelength is X. int =Floor(X), where the decimal part of the wavelength is X. frac =XX int Get X int and X int The dynamic threshold T of the voltage parameters for the two wavelength intervals corresponding to +1 int and T next .

[0053] Calculate the threshold for this data item: T dynamic =T int +X frac ×(T next -T int -offset Among them, T int T is the dynamic threshold corresponding to the wavelength interval group into which the data item falls. next X is the dynamic threshold corresponding to the wavelength interval group to the right of the wavelength interval group into which the data item falls. frac The value is the fractional part of the wavelength parameter in the data item, and offset is the preset offset of the voltage parameter. In this embodiment, offset = 0.02V.

[0054] Step S5: Set the voltage parameter value V for each data item. t The threshold T for determining voltage parameters dynamic If the voltage parameter is less than the judgment threshold T, the comparison is made. dynamic If the condition is met, then mark the data item as an outlier; otherwise, retain it.

[0055] Specifically, for example, in a certain data item, the wavelength parameter value is X = 500.3 nm, and the voltage parameter value is V. t =1.441V. The judgment threshold T obtained through step S4. dynamic =1.486 + 0.3 × (1.553 - 1.486) - 0.02 = 1.4861V. Since V t Less than T dynamic If so, then mark the data item as an outlier.

[0056] For each data item with wavelength parameters within the range of 500nm-600nm, the execution of the judgment is illustrated as follows: Figure 3 As shown, Figure 3 This is merely a schematic diagram illustrating the principle of the dynamic threshold linear interpolation process. The entire wavelength band in the diagram is drawn directly using linear interpolation, without considering the influence of the number of neighboring elements on the threshold value. In actual operation, it is necessary to strictly follow the method described in Example 1 to determine the dynamic threshold of the corresponding wavelength interval group based on the number of data items within the wavelength interval group, and then perform the subsequent interpolation judgment operation.

[0057] Figure 3 It contains two dynamic threshold curves, one above the other. The upper blue solid line is the dynamic threshold curve obtained by linear interpolation of the dynamic thresholds of adjacent wavelength interval groups. The lower red dashed line is the judgment threshold baseline obtained by the dynamic threshold after a preset offset. When the data point falls above the judgment threshold baseline, it is marked as a normal green dot, which is a green circle in the figure; when the data point falls below the judgment threshold baseline, it is marked as an abnormal point, which is a red cross in the figure.

[0058] Step S6: Obtain the configuration conditions for one or more auxiliary parameters. For example, the condition range for configurable current parameters is (0mA, 0.01mA). If the value of the current parameter of a data item is not within this range, it will be marked as an anomaly regardless of whether the value of the voltage parameter meets the threshold condition.

[0059] Step S7: Remove all data items marked as outliers and output the remaining data items as a clean dataset. This dataset can be directly used for chip performance level classification.

[0060] In the method of this embodiment, the data can be visualized. A scatter plot of the data before and after outlier removal is generated in the same coordinate system using a graphical user interface to demonstrate the algorithm's effect, such as... Figure 4 As shown. Figure 4 The visual comparison intuitively demonstrates the technical effect of the present invention: before removal (e.g., represented by red dots in the figure), the original scatter plot has a large number of diffuse discrete points inside and around the main distribution band, and the data boundaries are blurred; after removal by the method of the present invention (e.g., represented by blue dots in the figure), the envelope boundary of the main distribution band becomes clear and smooth. This proves that the judgment threshold based on linear interpolation of adjacent intervals can accurately remove abnormal discrete points that deviate from local statistical characteristics without destroying the original distribution pattern of the data, making the outline of the main data more compact, and providing a high signal-to-noise ratio data foundation for subsequent chip classification, yield analysis and process feedback.

[0061] It also allows configuration of multidimensional variables, initial screening indicators, and algorithm parameters, such as... Figures 5 to 6 The diagram is provided for operators to verify the effectiveness and reliability of the algorithm under different parameters. Figure 5 This demonstrates the system's multi-dimensional variable configuration capabilities. Operators can flexibly specify a first parameter (such as the dominant wavelength WLD) from dozens of electrical and optical parameters as the interval division dimension, while simultaneously selecting one to four second parameters (such as VF1, VF4, VF5, VF6, etc.) to perform multi-round, multi-dimensional discrete elimination. This design possesses high configuration flexibility and production adaptability, avoiding the one-sidedness of setting a single parameter threshold, and ensuring the statistical consistency and authenticity of the output clean dataset across various electrical indicators. Figure 6 The interface for adjusting initial screening indicators and algorithm parameters is displayed. Operators can preset the range of electrical parameters based on prior experience with the process platform to define the effective range of the algorithm; simultaneously, they can fine-tune algorithm parameters such as unit wavelength statistical fluctuation (interval width), quantile coefficient (quantile position), adjustable offset, and baseline (preset default value). This configuration panel provides process engineers with an intuitive parameter adjustment entry point, enabling them to adjust parameters in different product batches or process change scenarios, combined with visual feedback from the graphical user interface (see...). Figure 4 This allows for rapid verification and locking of the optimal parameter combination, ensuring the robustness and reliability of the algorithm in diverse production environments.

[0062] 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 application.

[0063] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for screening outliers in wafer test data, characterized in that, Includes the following steps: S1. Obtain the raw dataset for wafer testing. The raw dataset contains multiple data items, each of which includes at least a first parameter and a second parameter, wherein the first parameter and the second parameter are electrical parameters or optical parameters. S2. Based on the first parameter of multiple data items, divide the multiple data items into multiple equal-width parameter intervals to obtain several parameter interval groups; S3. Based on the number of data items in each parameter interval group and the value of the second parameter, determine the dynamic threshold of the second parameter corresponding to each parameter interval group; S4. For each data item, determine the parameter interval group it falls into, and select a parameter interval group adjacent to the parameter interval group; based on the first parameter, perform linear interpolation to calculate the interpolation threshold for the dynamic thresholds corresponding to the two parameter interval groups, and combine the preset offset of the second parameter to calculate the judgment threshold of the second parameter. S5. Compare the second parameter of each data item with the judgment threshold of the second parameter. If the second parameter is less than the judgment threshold of the second parameter, mark the data item as an outlier. S6. Remove all data items marked as outliers and output the remaining data items as a clean dataset.

2. The method for screening outliers in wafer test data according to claim 1, characterized in that, The specific steps of step S3 are as follows: Obtain the number of data items contained within the interval group of the parameter to be determined; Determine whether the quantity is zero; if so, set the dynamic threshold of the second parameter corresponding to the parameter interval group to a preset default value. If the quantity is greater than 0 and less than or equal to a preset quantity threshold, then the dynamic threshold of the second parameter corresponding to the parameter interval group is set to the maximum value of the second parameter in the parameter interval group; If the quantity is greater than the preset quantity threshold, the second parameter in the parameter interval group is sorted, and the value of the second parameter at the preset quantile position is extracted as the dynamic threshold of the second parameter corresponding to the parameter interval group.

3. The method for screening outliers in wafer test data according to claim 1, characterized in that, The data items also include auxiliary parameters, and the anomaly filtering method further includes auxiliary condition filtering: obtaining at least one auxiliary parameter and its corresponding preset condition range, and determining whether the value of the auxiliary parameter of each data item falls within the condition range. If the data item is not included, the corresponding data item will be marked as an outlier.

4. The method for screening outliers in wafer test data according to claim 1, characterized in that: The first parameter is an electrical parameter, and the second parameter is an optical parameter; or, the first parameter is an optical parameter, and the second parameter is an electrical parameter.

5. The method for screening outliers in wafer test data according to claim 1, characterized in that: The electrical parameters include voltage parameters and current parameters; the optical parameters include wavelength parameters, light intensity parameters, luminous intensity parameters, and light spot quality parameters.

6. The method for screening outliers in wafer test data according to claim 1, characterized in that: The first parameter is a wavelength parameter, and the second parameter is a voltage parameter. In step S2, the wavelength parameter of each data item is rounded down, and based on the preset wavelength fluctuation range width, the rounded wavelength value is divided into multiple parameter ranges of equal width, thereby forming several parameter range groups.

7. The method for screening outliers in wafer test data according to claim 6, characterized in that, In step S4, the determination threshold T of the second parameter dynamic The calculation formula is: T dynamic =T int +X frac ×(T next -T int )-offset Among them, T int T is the dynamic threshold corresponding to the parameter interval group into which the data item falls. next X is the dynamic threshold corresponding to the parameter interval group to the right of the parameter interval group into which the data item falls. frac The value is the decimal part of the first parameter in the data item, and offset is a fixed offset preset based on the statistical characteristics of the data item.

8. The method for screening outliers in wafer test data according to claim 6, characterized in that: Step S1 also includes setting a preset wavelength and electrical parameter range, and pre-filtering the data items by setting a lower limit value.

9. An outlier screening system for wafer test data, characterized in that, include: The data acquisition module is used to acquire the raw dataset for wafer testing; The parameter interval group construction module is used to divide multiple data items into multiple equal-width parameter intervals to obtain several parameter interval groups. The dynamic threshold calculation module is used to calculate the dynamic threshold of the second parameter corresponding to each parameter interval group; The linear interpolation judgment module is used to calculate the judgment threshold of the test element through linear interpolation and compare it with the second parameter; The anomaly marker module is used to mark data items that do not meet the conditions. The data output module is used to output the clean dataset after removing outliers.

Images