Method for automatically measuring thickness of oxide layer of semiconductor FinFET structure based on semantic segmentation model

Abstract

The invention discloses a semiconductor FinFET structure oxide layer thickness automatic measurement method based on a semantic segmentation model, and belongs to the technical field of semiconductor manufacturing and detection. Aiming at the problems of low manual measurement efficiency of the existing TEM image and insufficient calculation precision of the thickness of the bent oxide layer, an oxide layer mask image is extracted through a semantic segmentation model, a backbone path of a center line of the oxide layer is extracted by combining skeletonization and an image structure optimization technology, and a normal direction is dynamically calculated by adopting a principal component analysis method after equidistant sampling along the path; and pixel-level thickness measurement of the bending area is realized. According to the method, boundary errors are eliminated through mirror image filling, segmentation precision is improved by using a deep learning model, skeleton branches are optimized, interference is eliminated, and finally high-precision automatic measurement of the thickness of the nanoscale oxide layer is realized. Compared with a traditional manual method, the method has the advantages of being high in measurement efficiency, high in objectivity, visual and visual and the like, and is suitable for oxide layer quality monitoring in the FinFET manufacturing process.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing processes, and particularly to an automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model. Background Art

[0002] In recent years, with the rapid development of semiconductor technology, the size of transistors has been continuously reduced. In order to address problems such as short-channel effects and increased leakage current that occur in traditional planar transistors, three-dimensional FinFETs (fin field-effect transistors) have been widely used in high-performance integrated circuits. By surrounding the fin structure with a gate on three sides, FinFETs effectively improve the control ability of channel current, reduce leakage current, increase switching speed, and at the same time reduce power consumption. As the transistor size is further reduced, higher requirements are put forward for the thickness, uniformity, and quality of the oxide layer.

[0003] The oxide layer is one of the core components of FinFETs. Among them, the gate oxide isolates the gate from the channel and enhances electric field control, such as high-k materials, hafnium oxide (HfO2). The interfacial oxide layer optimizes the interface characteristics between the channel and the gate oxide and improves carrier mobility, such as silicon dioxide (SiO2). In actual manufacturing, the thickness of the oxide layer is usually at the nanometer level (1 - 5 nm). A slight thickness deviation or uneven distribution may lead to a decline in transistor performance or even the failure of the entire chip.

[0004] Currently, the measurement of the oxide layer thickness is mainly carried out by a transmission electron microscope (TEM). By projecting a high-energy electron beam onto the sample, recording the diffraction and scattering patterns of the electrons, generating an image with atomic-level resolution, and then using a DigitalMicrograph tool for manual measurement. There are problems such as low density of thickness measurement, slow speed, high influence of human subjectivity, and high cost.

[0005] Therefore, there is a need to find a fast, high-precision, and automated method for measuring the thickness of the oxide layer.

[0006] Traditional transmission electron microscope (TEM) image analysis mainly relies on manual observation and measurement, with problems such as low efficiency and high influence of human subjectivity. To achieve automatic measurement, it is necessary to rely on a high-precision semantic segmentation model such as Unet, Unet++, etc. to segment the area to be measured to obtain the corresponding mask image, and then use an automatic measurement algorithm based on the mask image to measure the thickness of the oxide layer. The oxide layers such as silicon dioxide and hafnium oxide in the FinFET structure are in a curved shape, and the layer thickness varies. It is difficult to measure at the curved parts with the automatic measurement algorithm based on the mask image. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes an automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model, which can automatically measure the thickness of curved irregular curves and has the characteristics of high efficiency, high accuracy, strong objectivity, and good visualization effect.

[0008] The technical solution of this application is as follows:

[0009] An automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model specifically includes the following steps:

[0010] Obtain the TEM (transmission electron microscope) image of the semiconductor product to be measured, and calculate the scale parameter of the image according to the actual size and pixel size of the image.

[0011] Label the area to be measured in the obtained TEM image, divide the training set and the validation set, and train the semantic segmentation model.

[0012] Input the TEM image of the semiconductor product to be detected into the semantic segmentation model for segmenting the area to be measured to obtain a mask image of the area to be measured.

[0013] After mirror filling 50 pixels in the lower, left, and right directions of the mask image, perform skeletonization, and crop the filled area in the result after skeletonization to obtain a skeleton image of the same size as the original segmentation image.

[0014] Connect the skeleton structure with a graph to obtain a graph structure, delete the small branches of the skeletonization caused by thickness problems, obtain an optimized graph structure, and generate an optimized skeleton image.

[0015] Sample all the points between the first end point and the second end point of the points in the optimized graph structure according to the required measurement density to obtain a set of skeleton points to be measured after sampling.

[0016] Iteratively take a skeleton point to be measured from the set of skeleton points to be measured after sampling. Taking the corresponding point of this skeleton point to be measured in the optimized skeleton image as the center, take a local area of 11*11 in size with a length of 5 pixels upward, downward, left, and right, and use the principal component analysis method to obtain the normal direction. Taking this skeleton point to be measured as the center, find a boundary point of the mask image in the positive and negative directions of the normal direction respectively to obtain the positive direction point A and the negative direction point B, calculate the Euclidean distance between points A and B, and obtain the thickness of the skeleton point to be measured at this place. Repeat this process until all the skeleton points to be measured are measured.

[0017] An automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure implemented by the present invention may also have the following additional technical features:

[0018] Further, before inputting the TEM image of the semiconductor product to be detected into the semantic segmentation model for semantic segmentation, the method further includes:

[0019] Creating a training data set, where the training data set includes several samples of the TEM image of the semiconductor product and the annotation data of the required automatic measurement area corresponding thereto;

[0020] Further, after constructing the training data set, the method further includes training a neural network model based on deep learning using the training data set to obtain the semantic segmentation model, including:

[0021] Constructing a deep learning neural network model for the semantic segmentation task;

[0022] Inputting the sample of the TEM image into the deep learning neural network model for the semantic segmentation task to obtain a mask map of the area to be measured;

[0023] Calculating the loss based on the segmentation area mask map and the annotation data, and iteratively training the deep learning neural network model according to the loss calculation result to complete convergence to finally obtain the semantic segmentation model.

[0024] Beneficial effects:

[0025] (1) The present invention extracts the oxide layer mask map through the semantic segmentation model, combines the skeletonization and graph structure optimization technologies to extract the center line backbone path of the oxide layer, and dynamically calculates the normal direction by using the principal component analysis method after equidistant sampling along the path, so as to realize the pixel-level thickness measurement of the curved area.

[0026] (2) The method of the present invention eliminates the boundary error through mirror filling, improves the segmentation accuracy by using the deep learning model, optimizes the skeleton branches to eliminate interference, and finally realizes the high-density automatic measurement of the nano-scale oxide layer thickness.

[0027] (3) Compared with the traditional manual method, the present invention has the advantages of high measurement efficiency, strong objectivity, intuitive visualization, etc., and is suitable for the quality monitoring of the oxide layer in the FinFET manufacturing process.

[0028] (4) The present invention can automatically measure the thickness of curved irregular curves, and has the characteristics of high efficiency, high accuracy, strong objectivity, good visualization effect, etc. Description of the drawings

[0029] Figure 1 Flowchart of the method for automatically measuring the oxide layer thickness of the semiconductor FinFET structure based on the semantic segmentation model in the embodiment of the present invention.

[0030] Figure 2It is a schematic diagram of a TEM image of a semiconductor product according to an embodiment of the present invention.

[0031] Figure 3 This is a mask diagram of a TEM image schematic diagram of a semiconductor product according to an embodiment of the present invention, wherein the gray-white color is a schematic diagram of a mirror-filled area.

[0032] Figure 4 It is a skeleton diagram of a mask diagram of a schematic diagram of a TEM image of a semiconductor product according to an embodiment of the present invention, wherein the gray-white color is a schematic diagram of a mirror-filled area.

[0033] Figure 5 It is a schematic diagram of measurement results of a schematic diagram of a TEM image of a semiconductor product according to an embodiment of the present invention.

[0034] Figure 6 Comparison pictures before (left) and after (right) skeleton optimization. Specific implementation methods

[0036] The following is a detailed description of an embodiment of the present invention in conjunction with the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0037] like Figure 1 As shown, a method for automatically measuring the oxide layer thickness of a semiconductor FinFET structure based on a semantic segmentation model specifically includes the following steps:

[0038] S1. Obtain a TEM image of the semiconductor product to be inspected.

[0039] It can be understood that the automatic measurement method for the oxide layer thickness of the semiconductor FinFET structure based on the semantic segmentation model proposed in this embodiment is applicable to the TEM image of the semiconductor FinFET structure to be detected, which is collected by the transmission electron microscope equipment in the production process in the semiconductor manufacturing plant. Figure 2 As shown, and the scale parameters are calculated.

[0040] S2. Input the semiconductor product image to be inspected into the semantic segmentation model for semantic segmentation to obtain a mask image of the area to be measured.

[0041] Before inputting the semiconductor product image to be inspected into the semantic segmentation model for semantic segmentation, it includes: creating a training data set, wherein the training data set includes several samples of the semiconductor product TEM image and the corresponding annotation data of the required area to be measured.

[0042] After creating the training dataset, it is necessary to use the training dataset to train a deep learning neural network model; construct a deep learning neural network model for semantic segmentation tasks; input the image sample to be measured and the corresponding labeled data of the region to be measured into the deep learning neural network for semantic segmentation tasks, and run the model to obtain the prediction result of the region to be measured; calculate the loss based on the prediction result of the region to be measured and the labeled data of the region to be measured, and iteratively train the deep learning neural network model according to the loss calculation result to complete convergence, and finally obtain the semantic segmentation model.

[0043] Through the implementation of step S2 above, the trained semantic segmentation model has adapted to the characteristics of semiconductor product images and can accurately segment the region to be measured when facing TEM images in the test set. Figure 3 The segmentation mask map obtained by the semantic segmentation model of the embodiment of the present invention is shown.

[0044] S3. Skeletonize the segmentation mask map using the Guo-Hall algorithm or the Zhang-Suen algorithm to obtain a skeleton map.

[0045] Specifically, in order to ensure that the ends of the skeletons at both ends of the curved irregular curve are not affected by end offsets, according to the TEM images with high magnification in the FinFET structure, the two ends of the region to be measured will stick to the boundaries of the image, especially the bottom and the left and right sides of the image. After mirror filling 50 pixels in the bottom and the left and right directions of the TEM image mask map, use the Guo-Hall algorithm or the Zhang-Suen algorithm for skeletonization, and crop the filled area in the result after skeletonization to obtain a skeleton map of the same size as the original segmentation map.

[0046] S4. Connect the skeleton map using an undirected graph to obtain a graph structure.

[0047] Specifically, convert the skeleton map into a graph structure, where each skeleton pixel is used as a node of the graph, and adjacent pixels are connected by edges.

[0048] S5. Optimize the graph structure and the skeleton map

[0049] Specifically, calculate the shortest paths between all endpoints (node degrees equal to 1), and select the longest one as the main path. The graph structure formed by this main path is the optimized graph structure, and draw all points on the graph to obtain the optimized skeleton map.

[0050] S6. Customize the measurement density

[0051] Specifically, according to the main path of the optimized graph structure in S5, perform equidistant sampling from the start point to the end point to obtain a set of skeleton points to be measured, so as to achieve the function of customizing the measurement density or display density.

[0052] S7. Width Measurement

[0053] Specifically, take a skeleton point to be measured in sequence from the set of skeleton points to be measured in S6. Taking the corresponding point of this skeleton point to be measured in the optimized skeleton graph as the center, take a local area of 11*11 size with a pixel length of 5 in each of the up, down, left, and right directions. Use the principal component analysis method to analyze, and take the direction perpendicular to the first principal component direction as the normal direction. Taking this skeleton point to be measured as the center, find a boundary point of the mask graph in each of the positive and negative directions of the normal direction to obtain the positive direction point A and the negative direction point B, and calculate the Euclidean distance between points A and B to obtain the thickness of the skeleton point to be measured at this location. Repeat this process until all skeleton points to be measured are completed.

[0054] S8. True Distance Conversion

[0055] Specifically, from the set of pixel point measurement thicknesses obtained in S7, multiply the pixel distance by the scale parameter according to the scale parameter to obtain the actual thickness.

[0056] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model, characterized in that, Specifically, it includes the following steps: S1. Obtain the transmission electron microscope (TEM) image of the semiconductor FinFET structure to be measured, and calculate the image scale parameter; S2. Input the TEM image into a pre-trained semantic segmentation model for segmentation to obtain a mask image of the area to be measured; S3. Perform mirroring and skeletonization on the mask image to generate a skeleton image; S4. Convert the skeleton image into a graph structure, which includes nodes and the connection relationships between adjacent nodes; S5. Optimize the skeleton image and the graph structure, and extract the main path; S6. Generate a set of skeleton points to be measured by equidistant sampling along the main path; S7. Perform thickness measurement on each skeleton point to be measured; S8. Convert the pixel distance into the actual thickness according to the scale parameter, and output the measurement result.

2. The automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, wherein, In step S2, the semantic segmentation model is a deep learning model based on the U-Net or U-Net++ architecture. During training, a TEM image dataset containing oxide layer annotations is used, and the model parameters are iteratively optimized through a loss function.

3. The automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, characterized in that, When performing skeletonization on the mask image in step S3, it includes: Perform mirror padding on the lower part and the left and right directions of the mask image, and the padding width is 50 pixels; Use the Guo-Hall algorithm or the Zhang-Suen algorithm for skeletonization; After cropping the mirror padding area, output a skeleton image with the same size as the original mask image.

4. The automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, wherein The optimization of the graph structure in step S5 includes: Calculate the minimum path between the longest endpoints as the main path, obtain the optimized graph structure, and draw the optimized skeleton image according to the optimized graph structure.

5. The automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, wherein The spacing of equidistant sampling in step S6 is dynamically adjusted according to the measurement density defined by the user.

6. The automatic measurement method for the thickness of the oxide layer of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, wherein When performing thickness measurement on each skeleton point to be measured in step S7, specifically: Intercept a local area centered on the current skeleton point, and calculate the normal direction through the principal component analysis method; Search for mask boundary points on both the positive and negative sides along the normal direction, and calculate the Euclidean distance between the two points as the thickness value; 7. The automatic measurement method for the oxide layer thickness of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, wherein In step S7, the size of the local area is 11×11 pixels, and the positive and negative search ranges in the normal direction are 5-pixel steps.

8. The automatic measurement method for the oxide layer thickness of a semiconductor FinFET structure based on a semantic segmentation model according to claim 1, characterized in that, The measurement result in step S7 makes the display density controllable by controlling the sampling spacing in step S6 and is visually superimposed on the original TEM image.

Images