Defect evaluation method for semiconductor substrate

Abstract

The present invention is a method for evaluating defects of a semiconductor substrate, comprising: a first step for detecting defects on the surfaces of a plurality of semiconductor substrates; a second step for acquiring a microscopic image of the defect; a third step for performing component analysis on whether the defect is a Ni defect; a fourth step for classifying the type of the defect on the basis of the microscopic image and the result of the component analysis; a fifth step for performing machine learning on the microscopic images of the various defects classified in the fourth step by using an image classification means; a sixth step for applying the image classification means, which has been subjected to machine learning in the fifth step, to a microscopic image of an unknown defect to estimate the type of the unknown defect; and a seventh step for layering the type of the unknown defect estimated in the sixth step into the Ni defect and the non-Ni defect and integrating the Ni defect and the non-Ni defect. As a result, a method for easily evaluating Ni defects on the surface of a semiconductor substrate is provided.

Description

Technical Field

[0001] This invention relates to a method for evaluating defects present on the surface of a semiconductor substrate. Background Technology

[0002] Several methods for evaluating trace metal contamination in semiconductor crystal substrates are known. As electrical analysis methods, deep-level transient spectroscopy (DLS) or minority carrier lifetime measurement can be cited. For example, in p-type silicon crystals, Fe contamination can be evaluated using these methods because active Fe-B pairs are formed electrically. As a chemical analysis method, inductively coupled plasma mass spectrometry (ICP-MS) is one of the most sensitive methods for measuring metal impurity concentrations. Using this chemical analysis method, the concentration of non-electrolyte metal impurities such as Ni or Cu in silicon crystals can be measured. However, its detection sensitivity is affected by the apparatus, conditions, and the composition of the sample; for example, it is difficult to detect concentrations as low as 1 × 10⁻⁶. 10 atoms / cm 3 The following is Ni.

[0003] Furthermore, defects formed by metallic impurities are also an important indicator in the evaluation of metal contamination. For example, Ni forms silicides in silicon crystals, creating Ni defects associated with silicides. Details will be described later; Patent Document 1 discloses that when the number of Ni-origin defects (Ni defects) present on each semiconductor substrate is less than 50, the Ni contamination concentration is estimated to be 1 × 10⁻⁶. 10 atoms / cm 3 Therefore, Ni defects are an effective indicator for evaluating trace Ni contamination.

[0004] Regarding defects on the surface of semiconductor substrates, the number of defects, in-plane defect distribution, defect size, defect shape, and defect composition are important evaluation indicators. For the number of defects, in-plane defect distribution, and defect size, surface inspection equipment such as the SP-3 manufactured by KLA-Tencor can be used for evaluation. Furthermore, for the evaluation of defect shape or composition, scanning electron microscopy (SEM) with energy dispersive X-ray analysis (EDX) is generally used. Therefore, by combining a surface inspection device with SEM-EDX, defects on semiconductor crystalline substrates can be evaluated efficiently.

[0005] Furthermore, in prior art patent document 1, a combined surface inspection device and SEM-EDX are disclosed, which evaluate the concentration of trace metal contamination on a semiconductor substrate based on the metal contamination state and the shape, size, and number of metal defects. Specifically, Figure 14 shows that the number of defects can be used to infer 1×10⁻⁶... 10atoms / cm 3 The following are the Ni contamination concentrations.

[0006] Existing technical documents

[0007] Patent Literature

[0008] Patent Document 1: Japanese Patent Application Publication No. 2015-220296 Summary of the Invention

[0009] (a) Technical problems to be solved

[0010] In the evaluation of trace Ni contamination, identifying Ni defects is crucial, and EDX (Electrode X-ray Diffusion) is required for Ni detection. Therefore, as the number of defects on the semiconductor substrate surface increases, the time required for EDX also increases. Furthermore, while existing devices and evaluation methods can classify defect shapes, they cannot differentiate between defects originating from metals and semiconductor crystallization defects.

[0011] The present invention was made in view of the problems of the prior art described above, and its purpose is to provide a simple method for evaluating Ni defects on the surface of a semiconductor substrate.

[0012] (II) Technical Solution

[0013] To address the aforementioned problems, the defect evaluation method of the present invention is a defect evaluation method for semiconductor substrates, comprising the following steps: a first step, detecting defects on the surface of multiple semiconductor substrates; a second step, acquiring microscopic images of the defects; a third step, performing component analysis to determine whether the defects are Ni defects; a fourth step, classifying the types of defects based on the microscopic images and the results of the component analysis; a fifth step, using image classification techniques to perform machine learning on the microscopic images of the various defects classified in the fourth step; a sixth step, applying the image classification techniques performed in the fifth step to the microscopic images of unknown defects to infer the types of the unknown defects; and a seventh step, stratifying the types of unknown defects inferred in the sixth step into Ni defects and non-Ni defects and integrating them.

[0014] Thus, in this invention, considering the difference in defect shape between Ni defects and non-Ni defects, the types of defects are pre-classified based on the results of microscopic images of multiple defects and compositional analysis to determine whether a defect is a Ni defect, according to the first to fifth steps. For each classification, image classification techniques are used to perform machine learning on the microscopic images of the defects, thereby enabling the type of defect to be inferred solely from the microscopic images. Then, after the sixth step, by applying the machine learning-based image classification technique to the microscopic images of unknown defects, the type of unknown defects can be inferred and determined to be either Ni or non-Ni defects. Therefore, when inferring the type of unknown defects, only the microscopic images of the unknown defects are needed, eliminating the need for time-consuming compositional analysis. This reduces the evaluation time and allows for a simpler evaluation of Ni defects on the surface of semiconductor substrates.

[0015] Preferably, the semiconductor substrate is a silicon single crystal substrate that has undergone mirror polishing.

[0016] If such a semiconductor substrate is used, the defect evaluation method for semiconductor substrates of the present invention can easily evaluate the Ni defects on the surface of the semiconductor substrate.

[0017] Additionally, preferably, the plurality of semiconductor substrates have at least one of the Ni defects and the non-Ni defects.

[0018] If a semiconductor substrate has such a defect, the defect evaluation method of the semiconductor substrate of the present invention can be used to easily evaluate whether it is a Ni defect or a non-Ni defect.

[0019] In addition, preferably, the Ni defects and the non-Ni defects are classified into at least one level based on the results of the microscopic images and the component analysis.

[0020] Therefore, based on the corresponding classification levels, machine learning can be performed on the microscopic images of defects using image classification methods.

[0021] Furthermore, preferably, in the sixth step, if the image classification method that has undergone machine learning cannot infer the type of the unknown defect from the microscopic image of the unknown defect, it is classified as unclassified.

[0022] Therefore, even when the types of unknown defects cannot be predicted, they can still be accurately classified.

[0023] (III) Beneficial Effects

[0024] By using the defect evaluation method for semiconductor substrates of the present invention, Ni defects and non-Ni defects on the surface of semiconductor substrates can be identified without performing additional component analysis, thereby reducing the time required for Ni defect evaluation. Attached Figure Description

[0025] Figure 1 A flowchart illustrating an example of the present invention.

[0026] Figure 2 This is an example of the experimental results (analysis time) of the embodiments and comparative examples of the present invention. Detailed Implementation

[0027] The following describes embodiments of the present invention, but the present invention is not limited to these embodiments.

[0028] Please refer to the attached diagram for further explanation.

[0029] Figure 1 A flowchart illustrating an example of the present invention.

[0030] First, multiple silicon single-crystal substrates are prepared as semiconductor substrates for machine learning and evaluation. The surface of the substrate to be evaluated is mirror-finished by mirror polishing. Additionally, the multiple semiconductor substrates have at least one of Ni defects and non-Ni defects.

[0031] (First process)

[0032] The first step is to inspect the surface defects of multiple semiconductor substrates. While not specifically limited, this first step can be performed on multiple semiconductor substrates with and without Ni defects, for example, using a surface inspection device. In this case, the surface inspection device is used to obtain the defect coordinates of each silicon single-crystal substrate under any measurement conditions recommended by the manufacturer. Furthermore, the upper limit for the number of semiconductor substrates is not specifically limited; for example, it can be set to 1000 or less.

[0033] (Second process)

[0034] The second step is to obtain microscopic images of the defects. While not specifically limited in this second step, for example, a SEM can be used to obtain microscopic images of the defects detected in the inspection step. In this case, an SEM is used, with any measurement conditions recommended by the manufacturer, to obtain microscopic images of the defects on each silicon single-crystal substrate using the defect coordinates obtained in the first step.

[0035] (Third process)

[0036] The third step is to perform compositional analysis to determine whether the defect is a Ni defect. While not specifically limited in this third step, EDX can be used for compositional analysis, for example. Alternatively, SEM-EDX can be used in conjunction with the second and third steps to obtain microscopic images of the defects detected in the inspection step and perform compositional analysis. In this case, EDX is used under any measurement conditions recommended by the manufacturer, utilizing the defect coordinates obtained in the first step to perform compositional analysis of defects on each silicon single-crystal substrate.

[0037] (Fourth process)

[0038] The fourth step is to classify defects based on microscopic images and compositional analysis results. In the fourth step, based on the microscopic images obtained in the second step and the compositional analysis results obtained in the third step, defects are classified into a total of 9 levels, such as the following: 3 levels for those with detected Ni and thus Ni defects: "Ni-pit", "Ni-PID (Polishing Induced Defect)", and "Ni-PID + pit"; 3 levels for those without detected Ni and thus non-Ni defects: "pit", "PID", and "PID + pit"; and 3 levels for "particle", "scratch", and "no defect", for a total of 6 levels.

[0039] Here, "Ni-pit" and "pit" refer to microscopic images of one or more concave defects; "Ni-PID" and "PID" refer to microscopic images of one or more convex defects; "Ni-PID+pit" and "PID+pit" refer to microscopic images of defects having one or more convex and one or more concave defects within a single field of view; "particle" refers to a microscopic image of surface deposits from grinding particles or organic matter; "scratch" refers to a microscopic image of one or more concave defects that can be identified as originating from grinding because its aspect ratio is sufficiently large compared to "Ni-pit" and "pit"; and "no defect" refers to a microscopic image of suspected defects or discoloration (Japanese: ステイン). Thus, by classifying defects in detail based on their shapes, the accuracy of machine learning in the next process can be improved.

[0040] In this case, "Ni defects" and "non-Ni defects" differ at least in their defect shapes. Furthermore, for "Ni-pits" and "pits," "Ni-PIDs" and "PIDs," and "Ni-PID+pits" and "PID+pits"—defects with similar shapes that are difficult to distinguish—there may be slight differences in defect shape based on experience. For example, compared to "pits," "Ni-pits" are closer to circular, and their inner surfaces are smoother. Additionally, compared to "PIDs," "Ni-PIDs" typically have rougher defect surfaces. "Ni-PID+pits" exhibit characteristics of either "Ni-pits" or "Ni-PIDs."

[0041] Thus, since the stratified "Ni defects" and "non-Ni defects" contain defects of at least one different shape, the accuracy of machine learning in the next process can be improved by further subdividing "Ni defects" and "non-Ni defects" into at least one level. Furthermore, the upper limit of the classification levels is not specifically limited; for example, it can be set to 30 levels or less.

[0042] (Fifth process)

[0043] The fifth step involves using image classification techniques to perform machine learning on the microscopic images of various defects classified in the fourth step. While not specifically limited, this fifth step can utilize commercially available image classification software to perform machine learning on the classified microscopic images of various defects. In this case, commercially available image classification software is used to perform machine learning on the shapes of various defects classified in the fourth step. In image-based machine learning, the image is processed as numerical data, and the features associated with the image are calculated as numerical data, i.e., feature values. Therefore, the difference in defect shape between "Ni defects" and "non-Ni defects" can be reflected in the feature values.

[0044] (Sixth process)

[0045] The sixth step involves applying the image classification method, which has undergone machine learning in the fifth step, to the microscopic image of the unknown defect to infer the type of the unknown defect. Alternatively, in the sixth step, after the machine learning in the fifth step, image classification software is used to infer the defect category based on the shape of the microscopic image of the unknown defect, classifying the defect by its shape. In this case, if no similar shape exists and the defect category cannot be inferred, it can be classified as "unclassified." In this situation, image classification software that has undergone pre-processed machine learning is used to infer which of the nine levels the defect should be classified into based on the shape of the microscopic image of the unknown defect of the measured object. Using the feature values ​​calculated in the previous step, the type of defect can be inferred by applying the k-nearest neighbor (KNN) method or a neural network. In this case, defects whose shape is dissimilar to any of the nine levels and whose type cannot be inferred are classified as "unclassified."

[0046] (Seventh process)

[0047] Finally, the seventh step is to categorize and integrate the unknown defects predicted in the sixth step into Ni defects and non-Ni defects. Alternatively, in the seventh step, defects classified into three levels—"Ni-pit," "Ni-PID," and "Ni-PID+pit"—can be integrated into "Ni defects," and defects classified into six levels—"pit," "PID," "PID+pit," "particles," "scratches," and "no defects"—can be integrated into "non-Ni defects." In this case, in order to distinguish between Ni defects and non-Ni defects, the defects classified in the fourth and sixth processes that are classified into the three levels of "Ni-pit", "Ni-PID" and "Ni-PID+pit" are layered and integrated into "Ni defects", and the defects classified into the three levels of "pit", "PID" and "PID+pit", along with the six levels of "particle", "scratch" and "no defect", are layered and integrated into "non-Ni defects".

[0048] Regarding the speculation on defect classification in the sixth process, in order to improve the evaluation and accuracy, the "compliance rate", "reproducibility rate" and "accuracy rate" were confirmed.

[0049] The compliance rate is the percentage of defects that are presumed to be classified into a particular category and that match the category determined by the operator.

[0050] The recurrence rate is the percentage of defects that match the operator's predicted category out of the total number of defects classified as a specific category by the operator.

[0051] Accuracy rate is the percentage of defects whose predicted category matches the category determined by the operator, relative to the total number of defects predicted for each category.

[0052] Example

[0053] The present invention will be further illustrated by the following examples.

[0054] [Example 1]

[0055] The present invention will be further described below based on embodiments, but these embodiments are illustrative in nature and should not be interpreted as limiting.

[0056] First, 145 silicon single-crystal substrates with a diameter of 300 mm and mirror polishing were prepared using the Tchaikovsky method as samples for machine learning and evaluation.

[0057] Defects on the surface of 145 silicon single crystal substrates were detected using a surface inspection device.

[0058] Next, using SEM-EDX, microscopic images of 3,514 defects on the surface of 145 silicon single-crystal substrates were obtained and their composition was analyzed.

[0059] Next, the operators categorized the 3,514 defects into a total of 9 levels: three levels where Ni was detected by EDX: “Ni-Pit”, “Ni-PID”, and “Ni-PID+Pit”; three levels where Ni was not detected: “Pit”, “PID”, and “PID+Pit”; and “Particle”, “Scratch”, and “No Defect”.

[0060] Next, commercially available image classification software was used to perform machine learning on microscopic images of 1840 defects found on 132 of the 145 wafers. The operator-configurable settings within the image classification software were adjusted to ensure proper identification of defect shapes in the microscopic images. The machine learning methods employed were KNN and Convolutional Neural Networks (CNNs). The CNNs included parameters, functions, and algorithms that the operator could set within the image classification software, all using the manufacturer's recommended settings.

[0061] For the 1674 defects found on 13 wafers out of 145 wafers that were not used for machine learning, image classification software with machine learning capabilities was used to predict which of the nine levels each defect would be classified into. For a given defect category, if the predictions derived by KNN and CNN match, the prediction is determined. Conversely, if the predictions derived by KNN and CNN do not match, the operator can choose which result to use; in this embodiment, the prediction result from CNN is used. Furthermore, defects that neither KNN nor CNN can predict are classified as "unclassified." The results are shown in Table 1.

[0062] [Table 1]

[0063]

[0064] Of the 1674 defects, 1052 were classified, and 622 were unclassified. Among the 1052 defects whose classifications were inferred, the defect layers classified as "Ni-pit," "Ni-PID," and "Ni-PID+pit" (indicating Ni detection) were consolidated into "Ni defects." The defect layers classified as "pit," "PID," and "PID+pit" (indicating no Ni detection), along with "particles," "scratches," and "no defects" (a total of six levels), were consolidated into "non-Ni defects." The results are shown in Table 2.

[0065] [Table 2]

[0066]

[0067] Of the 401 defects classified as "Ni defects" by the operator, 223 defects were correctly classified, and 162 were correctly classified as Ni defects. Of the 1273 defects classified as "non-Ni defects" by the operator, 829 defects were correctly classified, and 790 were correctly classified as "non-Ni defects." The accuracy rate for "Ni defects" was 80.6%, and the reproducibility rate was 72.6%. The accuracy rate for "non-Ni defects" was 92.8%, and the reproducibility rate was 95.3%. The overall accuracy rate was 90.5%. Therefore, although there are instances of incorrect classification, the defect evaluation method for semiconductor substrates of the present invention is useful for evaluating trace Ni contamination. Furthermore, by combining the present invention with prior art Patent Document 1, it is expected that even more useful trace Ni contamination evaluations can be performed.

[0068] [Comparative Example 1]

[0069] To demonstrate how easily the defect evaluation method for semiconductor substrates of the present invention can evaluate defects, a comparative example is used to compare the time required for predicting defect classification by using existing EDX methods for defect component analysis.

[0070] For 10, 25, 50, 100, 250, 500, 1000, 2500, and 5000 defects present on the surface of a silicon single-crystal substrate, the time required for EDX in the comparative example is compared with the time required for defect classification performed by the defect evaluation method on the semiconductor substrate according to the present invention. The time required for EDX is calculated by multiplying the analysis time for each defect by the aforementioned number of defects, which is set to 30 seconds. The time required for defect classification according to the present invention is defined as the time from the start of the transmission of the microscopic image to the image classification software that has undergone machine learning until the time when the image classification software finishes its prediction of defect classification.

[0071] The results are expressed as follows: Figure 2 As the number of defects in the analyzed object increases, the time required for the comparative example (EDX in the figure) and the inference of defect classification performed by the present invention (the present technology in the figure) both increase. However, the time required for the inference of defect classification performed by the present invention is shorter than the time required by EDX. For example, when the number of defects in the analyzed object is 10, 100, and 1000, the average processing time required for each defect by the present invention is 10 seconds, 2 seconds, and 0.4 seconds, respectively. In addition, the processing time required for each defect converges to 0.4 seconds when the number of defects is 1000 or more.

[0072] Therefore, the defect classification prediction made by the present invention can evaluate Ni defects in a shorter time compared to the EDX time of the comparative example, and its effectiveness is more obvious when the total number of defects in the object is larger.

[0073] As shown above, according to embodiments of the present invention, Ni defects on the surface of a semiconductor substrate can be easily evaluated.

[0074] Based on the above description, if the defect evaluation method for a semiconductor substrate of the present invention includes the first to seventh steps, it can easily evaluate Ni defects on the surface of the semiconductor substrate. The first step detects defects on the surfaces of multiple semiconductor substrates; the second step obtains microscopic images of the defects; the third step performs component analysis to determine whether the defects are Ni defects; the fourth step classifies the types of defects based on the microscopic images and the results of the component analysis; the fifth step uses image classification techniques to perform machine learning on the microscopic images of the various defects classified in the fourth step; the sixth step applies the image classification techniques used in the fifth step to the microscopic images of unknown defects to infer the types of the unknown defects; and the seventh step stratifies and integrates the types of unknown defects inferred in the sixth step into Ni defects and non-Ni defects.

[0075] In particular, by first going through the first to fifth processes, the types of defects are classified based on microscopic images of multiple defects and the results of component analysis to determine whether the defect is a Ni defect. Image classification techniques are then used to perform machine learning on the microscopic images of each defect category, allowing the type of defect to be inferred solely from the microscopic images. Next, after the sixth process, for the microscopic images of unknown defects, the machine learning-based image classification technique is applied to infer the type of the unknown defect and determine whether it is a Ni defect or not. Therefore, when inferring the type of an unknown defect, only the microscopic image of the unknown defect is needed; time-consuming component analysis is unnecessary. This reduces the evaluation time and allows for a simpler evaluation of Ni defects on the surface of a semiconductor substrate.

[0076] The present invention includes the following methods. [1]:

[0078] A method for evaluating defects in a semiconductor substrate includes the following steps:

[0079] The first step is to inspect the surface defects of multiple semiconductor substrates;

[0080] The second step is to obtain a microscopic image of the defect;

[0081] The third step is to perform a compositional analysis to determine whether the defect is a Ni defect.

[0082] The fourth step is to classify the types of defects based on the results of the microscopic images and the component analysis.

[0083] The fifth step involves using image classification techniques to perform machine learning on the microscopic images of the various defects classified in the fourth step.

[0084] The sixth step involves applying the image classification method, which has undergone machine learning in the fifth step, to the microscopic image of the unknown defect to infer the type of the unknown defect; and

[0085] The seventh step involves classifying the unknown defects predicted in the sixth step into Ni defects and non-Ni defects and integrating them. [2]:

[0087] As described above [1] in the defect evaluation method for semiconductor substrates, wherein:

[0088] The semiconductor substrate is a mirror-polished silicon single crystal substrate. [3]:

[0090] The defect evaluation method for semiconductor substrates as described in [1] or [2] above, wherein:

[0091] The plurality of semiconductor substrates have at least one of the Ni defects and the non-Ni defects. [4]:

[0093] As described above [3] in the defect evaluation method for semiconductor substrates, wherein:

[0094] The Ni defects and the non-Ni defects are classified into at least one level based on the results of the microscopic images and the component analysis. [5]:

[0096] As described in any of [1] to [4] above, the defect evaluation method for semiconductor substrates includes:

[0097] In the sixth step, if the image classification method that has undergone machine learning cannot infer the type of the unknown defect from the microscopic image of the unknown defect, it is classified as unclassified.

[0098] Furthermore, the present invention is not limited to the embodiments described above. The embodiments described above are merely illustrative, and any technical solution having a structure substantially the same as the technical concept described in the claims of the present invention and achieving the same effect is included within the technical scope of the present invention.

Claims

1. A method for evaluating defects in a semiconductor substrate, characterized in that, It includes the following processes: The first step is to inspect the surface defects of multiple semiconductor substrates; The second step is to obtain a microscopic image of the defect; The third step is to perform a compositional analysis to determine whether the defect is a Ni defect. The fourth step is to classify the types of defects based on the results of the microscopic images and the component analysis. The fifth step involves using image classification techniques to perform machine learning on the microscopic images of the various defects classified in the fourth step. The sixth step involves applying the image classification method, which has undergone machine learning in the fifth step, to the microscopic image of the unknown defect in order to infer the type of the unknown defect. and The seventh step involves classifying the unknown defects predicted in the sixth step into Ni defects and non-Ni defects and integrating them.

2. The defect evaluation method for a semiconductor substrate as described in claim 1, characterized in that, The semiconductor substrate is a mirror-polished silicon single crystal substrate.

3. The defect evaluation method for a semiconductor substrate as described in claim 1, characterized in that, The plurality of semiconductor substrates have at least one of the Ni defects and the non-Ni defects.

4. The defect evaluation method for a semiconductor substrate as described in claim 2, characterized in that, The plurality of semiconductor substrates have at least one of the Ni defects and the non-Ni defects.

5. The defect evaluation method for a semiconductor substrate as described in claim 3, characterized in that, The Ni defects and the non-Ni defects are classified into at least one level based on the results of the microscopic images and the component analysis.

6. The defect evaluation method for a semiconductor substrate as described in claim 4, characterized in that, The Ni defects and the non-Ni defects are classified into at least one level based on the results of the microscopic images and the component analysis.

7. The defect evaluation method for a semiconductor substrate as described in any one of claims 1 to 6, characterized in that, In the sixth step, if the image classification method that has undergone machine learning cannot infer the type of the unknown defect from the microscopic image of the unknown defect, it is classified as unclassified.

Images