Overlay estimation based on optical inspection and machine learning

Abstract

One or more optical images are acquired for a portion of a semiconductor wafer. These one or more optical images show a first structure within a first process layer and a second structure within a second process layer. These one or more optical images are fed into a machine learning model trained to estimate the overlay offset between the first and second structures. The estimated overlay offset between the first and second structures is obtained from the machine learning model.

Description

Technical Field

[0001] [Related Application] In this application, priority is claimed based on U.S. Provisional Patent Application No. 63 / 420,683 filed on October 31, 2022, and the entire content thereof is incorporated herein by reference for any purpose.

[0002] This disclosure relates to the measurement of overlay (i.e., overlay offset) in semiconductor devices, and more specifically to overlay offset measurement using optical inspection and machine learning.

Background Art

[0003] Process layer misalignment in semiconductor manufacturing causes an unwanted shift in the position of structures within a process layer compared to the position of structures within other process layers. Such shifts (i.e., displacements) are called overlay, and are also called overlay offset. Measuring these shifts is called overlay metrology. Accurate overlay metrology is important for measuring process drift, and thus for establishing and maintaining process control, especially during the startup period of new semiconductor processes and / or devices. Accurate overlay metrology has become increasingly important as the feature size (e.g., line width) on semiconductor wafers has decreased; because as the feature size decreases, the available overlay budget correspondingly decreases, making overlay errors less tolerable as the feature size shrinks.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] Scanning electron microscopes (SEMs) are a potentially useful tool for overlay metricing due to their high resolution and ability to image structures regardless of their shape or size. However, SEMs are slow and have low throughput. Therefore, optical inspection is preferable for overlay metricing. However, optical inspection does not provide the same accuracy as SEM-based overlay metricing. For example, optical inspection cannot resolve the structures used in overlay metricing because their size is below the diffraction limit. In other words, overlays cannot usually be simply measured by viewing images generated by optical inspection or by applying object recognition algorithms to such images. Nevertheless, overlay-related information is encoded within optical images, and algorithms exist to extract this information from these images. However, the accuracy of measurement results obtained using such classical algorithms is limited. [Means for solving the problem]

[0006] To improve the accuracy of overlay metricing, which is based on optical inspection, we will use machine learning to analyze optical images of semiconductor structures.

[0007] In certain embodiments of the method, one or more optical images are acquired of a portion of a semiconductor wafer, showing a first structure within a first process layer and a second structure within a second process layer. These one or more optical images are fed into a machine learning model trained to estimate the overlay offset between the first and second structures. The estimated overlay offset between the first and second structures is obtained from the machine learning model.

[0008] In certain embodiments, one or more programs executed by one or more processors are stored in a non-temporary computer-readable storage medium. These one or more programs include instructions that cause the above-described method to be executed.

[0009] A system in certain embodiments comprises one or more optical inspection tools for optically inspecting a semiconductor wafer, one or more processors, and a memory containing one or more programs executed by the one or more processors. The one or more programs include instructions that cause the above-described method to be performed on one or more optical images acquired from the one or more optical inspection tools.

[0010] For a better understanding of the various implementations described, the following diagrams should be referenced in conjunction with the detailed descriptions provided later. [Brief explanation of the drawing]

[0011] [Figure 1] This is a longitudinal cross-sectional view of a portion of a semiconductor wafer subjected to optical imaging for the purpose of performing overlay metricing, relating to a certain embodiment, and this portion contains a first structure within a first process layer and a second structure within a second process layer. [Figure 2] This is a longitudinal cross-sectional view of a portion of a semiconductor wafer subjected to optical imaging for the purpose of performing overlay metricing, relating to a certain embodiment, and this portion contains a first structure within the first process layer, a second structure within the second process layer, and a third structure within the third process layer. [Figure 3] This flowchart illustrates a method for measuring overlay offset using machine learning, relating to a certain embodiment. [Figure 4] This flowchart illustrates a method, according to a certain embodiment, for measuring overlay offsets for optical images from multiple optical inspection tools using machine learning. [Figure 5] This flowchart illustrates a method for measuring overlay offset using a machine learning model ensemble, relating to a certain embodiment. [Figure 6] This is a block diagram of a semiconductor inspection system according to a certain embodiment. [Modes for carrying out the invention]

[0012] Corresponding parts are referenced using the same reference number throughout the drawings and specifications.

[0013] The following detailed descriptions refer in detail to various embodiments illustrated in the attached drawings. The following detailed descriptions provide numerous specific details to facilitate a general understanding of the various embodiments described. However, the various embodiments described can also be implemented without such specific details. On the other hand, well-known methods, procedures, components, circuits, and networks are omitted to avoid unnecessary ambiguity in the various aspects of the embodiments.

[0014] Figure 1 is a longitudinal cross-sectional view of a portion 100 of a semiconductor wafer subjected to optical imaging for overlay metricing, relating to a certain embodiment. The semiconductor wafer may be fully processed or partially processed. The portion 100 contains a first structure 104 within a first process layer and a second structure 110 within a second process layer. In the example of Figure 1, the first structure 104 is a first grid having a plurality of grid elements 106, and the second structure 110 is a second grid having a plurality of grid elements 112. By periodically arranging the plurality of grid elements 106 within the first structure 104, the grid elements 106 can be separated by an equal distance from each other, that is, their pitch can be made the same. Similarly, by periodically arranging the plurality of grid elements 112 within the second structure 110, the grid elements 112 can be separated by an equal distance from each other, that is, their pitch can be made the same. The pitch of the inner lattice elements 112 of the second structure 110 may be the same as the pitch of the inner lattice elements 106 of the first structure 104.

[0015] The first process layer is a layer in which the first structure 104 is located, and is on the substrate 102 of the semiconductor wafer. Alternatively, the first process layer may be located above other process layers. The lattice elements 106 of the first structure 104 are separated from each other by a filler (e.g., dielectric). The lattice elements 112 of the second structure 110 may be separated from each other by a filler (e.g., dielectric) or a filler (e.g., dielectric); which depends on how far away the semiconductor wafer has been processed. A filler layer 108 (e.g., interlayer dielectric) separates the second process layer from the first process layer. In certain embodiments, the lattice elements 106 and / or lattice elements 112 are metal lines. In overlay metric terminology, the second process layer above the first process layer is called the "outer layer," and the first process layer below the second process layer is called the "inner layer."

[0016] Figure 1 shows both the assumed and actual placements of the inner grid elements 112 of the second structure 110 relative to the grid elements 106 of the first structure 104. The assumed placements are shown with solid lines and filled patterns, while the actual placements are shown with dashed lines. There is an assumed (i.e., planned) overlay offset 114 between the second structure 110 and the first structure 104. The assumed overlay offset 114 is the overlay offset that would exist if the inner grid elements 112 of the second structure 110 were created precisely at the intended locations of the inner grid elements 106 of the first structure 104. In the example in Figure 1, the assumed overlay offset 114 is non-zero. Alternatively, the assumed overlay offset may be set to zero.

[0017] In reality, due to the imprecision and variability of photolithography, neither the internal grid elements 112 of the second structure 110 nor the internal grid elements 106 of the first structure 104 can generally be created in the intended locations. The difference between the actual and assumed relative positions is an additional and unexpected (i.e., error-based) overlay offset 116, which can be positive or negative. The total overlay offset 118 is equal to the sum of these assumed overlay offsets 114 and unexpected overlay offsets 116.

[0018] Portion 100 may be a test pattern on a semiconductor wafer. The test pattern may be separated from dies on the semiconductor wafer (e.g., it may be in a scribe line) or may be within a die on the semiconductor wafer. By having a plurality of test patterns similar to portion 100 provided at various locations on the semiconductor wafer, overlay metrology can be performed at multiple locations on the wafer. Alternatively, portion 100 may be a part of the circuit of the die itself.

[0019] FIG. 2 is a longitudinal cross-sectional view of a portion 200 of a semiconductor wafer subjected to optical imaging to perform overlay metrology, according to certain embodiments. The semiconductor wafer may be fully processed or partially processed. Portion 200 includes a first structure 104 in a first process layer, a second structure 110 in a second process layer, and a third structure 202 in a third process layer. The third structure 202 in the example of FIG. 2 is a third grating and has a plurality of grating elements 204. By arranging the plurality of grating elements 204 periodically within the third structure 202, the grating elements 204 can be separated from each other by an equal distance, i.e., their pitch can be made the same, and it can also be made the same as the pitch of the grating elements 106 in the first structure 104 and the pitch of the grating elements 112 in the second structure 110.

[0020] The grating elements 204 of the third structure 202 can be separated from each other by a filler (e.g., a dielectric) or air; which one it is depends on how far the processing of the semiconductor wafer has progressed. The grating elements 112 of the second structure 110 within portion 200 are separated from each other by a filler (e.g., a dielectric). A filler layer 206 (e.g., an interlayer dielectric) separates the third process layer from the second process layer. In certain embodiments, the grating elements 204 are made of metal lines (e.g., similar to the grating elements 106 and / or the grating elements 112).

[0021] In addition to showing the assumed and actual arrangements of the inner grid elements 112 of the second structure 110, Figure 2 also shows both the assumed and actual arrangements of the inner grid elements 204 of the third structure 202. The assumed arrangements are shown with solid lines and filled patterns, while the actual arrangements are shown with dashed lines. There is an assumed (i.e., planned) overlay offset 208 between the third structure 202 and the second structure 110. The assumed overlay offset 208 is the overlay offset that would exist if the inner grid elements 204 of the third structure 202 were created precisely at the intended locations of the inner grid elements 112 of the second structure 110. In the example in Figure 1, the assumed overlay offset 208 is non-zero. Alternatively, the assumed overlay offset may be set to zero. However, since neither the third structure 202 inner lattice element 204 nor the second structure 110 inner lattice element 112 can be created in the intended location, there will be an actual (i.e., total) overlay offset 210 that is different from the assumed overlay offset 208. The difference between the total overlay offset 210 and the assumed overlay offset 208 is the unexpected (i.e., error-related) overlay offset, which can be positive or negative.

[0022] When performing overlay weighing between the third structure 202 and the second structure 110 using part 200, the third process layer is above the second process layer, so the third process layer becomes the outer layer and the second process layer becomes the inner layer. When performing the second structure 110 and first structure 104 using part 200, the second process layer becomes the outer layer and the first process layer becomes the inner layer. It is also possible to perform overlay weighing between the third structure 202 and the first structure 104 using part 200, with the third process layer as the outer layer and the first process layer as the inner layer.

[0023] Part 200 may be a test pattern on a semiconductor wafer, as described with respect to Part 100 (Figure 1), or it may be a part of the circuit of the die itself.

[0024] Figure 3 is a flowchart illustrating a method 300 for measuring overlay offset using machine learning, relating to a certain embodiment. This method 300 can be performed by one or more computer systems (e.g., the computer system of the semiconductor inspection system 600 in Figure 6).

[0025] In certain embodiments, a machine learning model is trained using a training set of optical images showing the first and second structures so that the method 300 can estimate the overlay offset between the first structure in the first process layer and the second structure in the second process layer (302). For example, the first structure is the first structure 104 (Figure 1), and the second structure is the second structure 110 (Figure 1). In other embodiments, the first and second structures are individual components of a separate periodic or non-periodic grid; one of these components may be a test pattern on a semiconductor wafer (e.g., one provided in a scribe line). In yet another embodiment, the first and second structures are individual components of separate test patterns (e.g., a box-in-box); these components may be provided in a scribe line. In yet another embodiment, the first and second structures are individual components of a semiconductor die circuit created in whole or in part on a semiconductor wafer.

[0026] Optical images within the training set are annotated with individual predetermined overlay offset values ​​(e.g., total overlay offset value and / or unexpected overlay offset value): these predetermined overlay offset values ​​are stored in association with the optical image corresponding to them. The individual predetermined overlay offset values ​​for the optical images within the training set are pre-measured, for example, using scanning electron microscopy: parts of the semiconductor wafer that have been optically inspected to generate the optical images within the training set are also inspected using a scanning electron microscope (SEM) (e.g., SEM634 in Figure 6) or multiple SEMs, and the overlay offset values ​​are measured based on the resulting SEM images. These predetermined overlay offset values ​​then act as ground truth during training 302. Examples of machine learning models, but not limited to these, include neural network models, gradient boosting models, support vector machines, and principal component regression models.

[0027] To perform training 302, a training set consisting of optical images is acquired along with a predetermined overlay offset value. The machine learning model is initialized (e.g., with randomization). The optical images in the training set are fed to the machine learning model as input, and the output produced by the machine learning model is compared with the predetermined overlay offset value for the individual image, i.e., the expected output. The optical images in the training set may be fed directly to the machine learning model, or feature extraction may be performed and the extracted features (e.g., in the form of feature vectors for the individual optical images) may be fed to the machine learning model. The machine learning model is adjusted based on the difference between the output produced (i.e., actual output) and the expected output, and the training process continues until the output produced converges to the expected output. Once convergence occurs, the machine learning model can be tested and then deployed for use in semiconductor manufacturing. The deployment destination of the machine learning model may be one or more semiconductor manufacturing facilities, commonly known as fabs, or a computer system connected to one or more fabs.

[0028] Alternatively, training of a machine learning model is omitted from Method 300. For example, the machine learning model used to perform Method 300 is assumed to be already trained to estimate the overlay offset between the first and second structures and to be deployed for use in semiconductor manufacturing. Here again, the first and second structures can be the first structure 104 (Figure 1) and the second structure 110 (Figure 1), respectively; they can be individual components of a different periodic or non-periodic grid, either of which can be a test pattern on a semiconductor wafer (e.g., one provided in a scribe line); they can be individual components of separate test patterns (e.g., box-in-box) provided in a scribe line; or they can be individual components of a semiconductor die circuit created entirely or partially on a semiconductor wafer.

[0029] One or more optical images are acquired (304) showing a portion of a semiconductor wafer, specifically the first structure within the first process layer and the second structure within the second process layer (i.e., showing the first and second structures of a specific instance (example) at a specific location within a particular semiconductor die). These one or more optical images differ from the optical images in the training set. For example, none of these one or more optical images were included in the training set. The first and second structures of that instance in these one or more optical images may therefore differ from the first and second structures of that instance in individual optical images within the training set (e.g., they may be located in a different location, on a different semiconductor die, and / or on a different semiconductor wafer). In certain embodiments, when acquiring these one or more optical images, the relevant portion of the semiconductor wafer is optically inspected (e.g., using an optical inspection tool 632 in Figure 6) (306). Alternatively, these one or more optical images are acquired from computer memory (e.g., a database) containing optical images pre-generated by optical inspection.

[0030] The one or more optical images are supplied to the machine learning model (e.g., a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model) (314). In certain embodiments, the one or more optical images are supplied directly to the machine learning model (e.g., according to deep learning), and the machine learning model receives the one or more optical images as input. Alternatively, the one or more optical images are supplied indirectly to the machine learning model. For example, feature extraction is performed on the one or more images, and the extracted features (e.g., in the form of feature vectors) are supplied as input to the machine learning model.

[0031] An estimated overlay offset between the first and second structures (e.g., relating to the first and second structures of a particular instance) is obtained from the machine learning model (322): The machine learning model determines (i.e. calculates) the estimated overlay offset using the one or more optical images mentioned above, and the estimated overlay offset is output. This estimated overlay offset may be the estimated total overlay offset (e.g., total overlay offset 118 in Figure 1) or the estimated unexpected overlay offset (e.g., unexpected overlay offset 116 in Figure 1). This estimated overlay offset can be received from the machine learning model as the one or more optical images are supplied to the machine learning model. The machine learning model's determination of the estimated overlay offset may be based solely on the one or more optical images, or it may be based on the one or more optical images and additional data (e.g., metadata relating to the one or more optical images and / or additional metric data relating to the first and second structures).

[0032] In certain embodiments, the aforementioned one or more optical images are a single optical image showing a first structure within a first process layer and a second structure within a second process layer. This single optical image is acquired (308) and supplied directly or indirectly to the machine learning model (316). Using the machine learning model, an estimated overlay offset is calculated based on the single optical image (324).

[0033] In other embodiments, the aforementioned one or more optical images include a first optical image focused on a first process layer and a second optical image focused on a second process layer. These first and second optical images are acquired (310) and supplied directly or indirectly to the machine learning model (318). The first and second optical images may be acquired, for example, by using corresponding first and second cameras in the optical inspection system, after setting the first and second cameras to a known alignment. By performing a stitching process on these first and second images, they can be aligned according to their known alignment. Using the machine learning model, an estimated overlay offset is calculated based on these first and second optical images (326).

[0034] In certain embodiments, the one or more optical images described above include multiple optical images obtained by optically inspecting the relevant portion of the semiconductor wafer using multiple optical modes (312). Each of these multiple optical modes inspects the relevant portion of the semiconductor wafer using a separate set of optical conditions, where one or more optical conditions within the set differ from other sets of optical conditions for those multiple optical modes. Examples of optical conditions, but not limited to, include light sources, wavelength ranges, polarization, focus, transmission (transmittance) distribution within the illumination aperture, transmission (transmittance) distribution within the focusing aperture, and phase shift distribution within the focusing aperture. In addition to or instead of this, individual optical modes may be associated with individual Müller components and / or individual harmonic signals originating from the imaging Müller ellipsometer. Each of the multiple optical images described above is obtained using the corresponding optical mode among the multiple optical modes. Using each optical mode, it is possible to generate a single optical image showing both the first and second structures, or to generate multiple optical images (e.g., a first image focused on the first process layer and a second image focused on the second process layer). These multiple optical images are supplied directly or indirectly to the machine learning model (320). Using the machine learning model, an estimated overlay offset is calculated based on these multiple optical images (328).

[0035] The estimated overlay offset obtained by this method 300 can be used directly to monitor and control the semiconductor manufacturing process (e.g., for statistical process control). Alternatively, the estimated overlay offset may be used as one of several inputs, and the inter-structure overlay offset between the first and second structures may be determined later using these inputs. For example, the estimated overlay offset can be combined with one or more other overlay offset values ​​calculated for the first and second structures using another technique (e.g., a technique that does not involve machine learning), for example, by averaging them.

[0036] Method 300 can be extended to determine the overlay offset between a second structure (e.g., second structure 110 in Figure 2) and a third structure within the third process layer (e.g., third structure 202 in Figure 2). For example, the aforementioned one or more optical images are further shown showing the third structure within the third process layer (i.e., the third structure of a specific instance is shown). The aforementioned machine learning model is referred to as the first machine learning model, and Method 300 is further enhanced by incorporating a second machine learning model (e.g., a neural network model, gradient boosting model, support vector machine, or principal component regression model) trained to estimate the overlay offset between the second and third structures. The second machine learning model may be trained as part of Method 300 by applying training 302, or it may be pre-trained and deployed in preparation for use in semiconductor manufacturing. The aforementioned one or more optical images are supplied directly or indirectly to the first machine learning model and the second machine learning model in step 314. In one example, the one or more optical images are a single optical image showing the first, second, and third structures. This single optical image is supplied to the first machine learning model and the second machine learning model in step 316. In another example, the one or more optical images include a first optical image focused on the first process layer, a second optical image focused on the second process layer, and a third optical image focused on the third process layer. The first and second optical images are supplied to the first machine learning model in step 318, and the second and third optical images are supplied to the second machine learning model. Other examples may exist. The estimated overlay offset between the second and third structures (e.g., for the second and third structures of a particular instance) is obtained from the second machine learning model: the estimated overlay offset between the second and third structures is calculated using the second machine learning model. Estimated overlay offsets include the total overlay offset (e.g., the total overlay offset of 210 in Figure 2) and unexpected overlay offsets (e.g., the difference between the total overlay offset of 210 and the expected overlay offset of 208 in Figure 2).

[0037] In addition to or instead of using a second machine learning model to obtain the estimated overlay offset between the second and third structures, a third machine learning model (e.g., a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model) may be used to obtain the estimated overlay offset between the first structure (e.g., the first structure 104 in Figure 2) and the third structure (e.g., the third structure 202 in Figure 2). The third machine learning model may be trained as part of Method 300 by applying training 302, or it may be pre-trained and deployed in preparation for use in semiconductor manufacturing. One or more of the aforementioned optical images are supplied directly or indirectly to the first machine learning model and the third machine learning model (and, according to certain embodiments, to the second machine learning model) in step 314. In one example, the one or more optical images are a single optical image showing the first, second, and third structures. The single optical image is supplied to the first machine learning model and also to the third machine learning model (and, according to certain embodiments, to the second machine learning model) in step 316. In another example, the one or more optical images include a first optical image focused on the first process layer, a second optical image focused on the second process layer, and a third optical image focused on the third process layer. The first and second optical images are supplied to the first machine learning model in step 318, and the first and third optical images are supplied to the third machine learning model (and, according to certain embodiments, the second and third optical images are supplied to the second machine learning model). An estimated overlay offset between the first and third structures (e.g., relating to the first and third structures of a particular instance) is obtained from the third machine learning model: the estimated overlay offset between the first and third structures is calculated using the third machine learning model. Possible estimated overlay offsets include the total overlay offset and unexpected overlay offsets.

[0038] The first, second, and / or third machine learning models may be combined to form a single machine learning model with multiple outputs. This single machine learning model is trained to estimate the offset overlays between the first and second structures, between the second and third structures, and / or between the first and third structures. Of its multiple outputs, the first output provides the estimated offset overlay between the first and second structures, the second output provides the estimated offset overlay between the second and third structures, and the third output provides the estimated offset overlay between the first and third structures.

[0039] The first, second, and third structures can be the first structure 104 (Figure 2), the second structure 110 (Figure 2), and the third structure 202 (Figure 2), respectively; they can be individual components of different periodic or non-periodic grids, one of which can be a test pattern on a semiconductor wafer (e.g., one provided within a scribe line); they can be individual components of separate test patterns (e.g., box-in-box) and provided within a scribe line; or they can be individual components of a circuit on a semiconductor die, either entirely or partially, created on a semiconductor wafer.

[0040] Figure 4 is a flowchart illustrating a method 400 for measuring overlay offset with respect to optical images from multiple optical inspection tools, according to a certain embodiment. This method 400 can be performed by one or more computer systems (e.g., the computer system of the semiconductor inspection system 600 in Figure 6). Method 300 (Figure 3) can be incorporated into this method 400, or it can be performed as part of this method 400.

[0041] In certain embodiments, a machine learning model is trained using a training set of optical images showing the first and second structures so that it can estimate the overlay offset between the first structure in the first process layer (e.g., the first structure in method 300 in Figure 3) and the second structure in the second process layer (e.g., the second structure in method 300 in Figure 3) (402). The optical images in the training set are annotated with individual predetermined overlay offset values ​​(e.g., total overlay offset value and / or unexpected overlay offset value): these predetermined overlay offset values ​​are stored in association with the optical images corresponding to them. The individual predetermined overlay offset values ​​for the optical images in the training set are measured, for example, using a scanning electron microscope (e.g., as described with respect to training 302 in method 300 in Figure 3). Examples of machine learning models include, but are not limited to, neural network models, gradient boosting models, support vector machines, and principal component regression models. The optical images forming the training set can be obtained from multiple optical inspection tools: by optically inspecting individual semiconductor wafers using individual optical inspection tools among these multiple optical inspection tools, individual optical images within the training set can be generated. Training 402 can be performed as described with respect to training 302 in method 300 (Figure 3).

[0042] Alternatively, training a machine learning model may be omitted from Method 400. For example, the machine learning model used to perform Method 400 may already be trained to estimate the overlay offset between the first and second structures (e.g., the first and second structures in Method 300 in Figure 3) and deployed for use in semiconductor manufacturing.

[0043] Multiple sets of one or more optical images are obtained for various parts of various semiconductor wafers from multiple optical inspection tools (e.g., optical inspection tool 632 in Figure 6) (404). Each set shows a first and second structure (e.g., the first and second structures of an individual instance at a specific location within a specific semiconductor die on a specific semiconductor wafer are shown). Each set is generated by an individual optical inspection tool among those multiple optical inspection tools. The one or more optical images differ from the optical images in the training set. For example, none of the one or more optical images were included in the training set. The first and second structures of the individual instances in those multiple sets may therefore differ from each other and from the first and second structures of the corresponding instances in the individual optical images in the training set (e.g., they may be located in different locations, on different semiconductor dies, and / or on different semiconductor wafers). Those multiple inspection tools may be the same as the multiple inspection tools from which the training set was obtained, or they may be a subset or superset thereof. One or more optical images in Method 300 (Figure 3) can be used to represent individual sets within those sets.

[0044] In certain embodiments, when acquiring one or more optical images, parts of various semiconductor wafers are optically inspected using the multiple optical inspection tools (406). For example, if the multiple inspection tools include two tools, parts of some semiconductor wafers are inspected using the first tool, thereby acquiring some sets, and parts of the remaining semiconductor wafers are inspected using the second tool, thereby acquiring the remaining sets. If the multiple inspection tools include three or more tools, parts of some semiconductor wafers are inspected using the first tool, thereby acquiring some sets, parts of other semiconductor wafers are inspected using the second tool, thereby acquiring other sets, parts of yet another semiconductor wafer are inspected using the third tool, thereby acquiring yet another set, and so on.

[0045] The aforementioned sets are supplied to the aforementioned machine learning model (e.g., a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model) (414). For example, the sets may be supplied sequentially to the machine learning model, with each set being supplied successively or as it becomes available. As described with respect to step 314 of Method 300 (Figure 3), each set may be supplied directly or indirectly to the machine learning model.

[0046] For each of these sets, an individually estimated overlay offset between the first and second structures (e.g., relating to the first and second structures of an individual instance) is obtained from the machine learning model (418): the individually estimated overlay offset is calculated using the machine learning model. Possible individually estimated overlay offsets include the estimated total overlay offset (e.g., the total overlay offset 118 in Figure 1) and the estimated unexpected overlay offset (e.g., the unexpected overlay offset 116). The machine learning model determines the individually estimated overlay offset using one or more optical images within the set, and the estimated overlay offset is supplied as output. The individually estimated overlay offset can be received from the machine learning model as the set is supplied to the machine learning model. The determination of the estimated overlay offset by the machine learning model may be based solely on one or more optical images within the set, or it may be based on one or more optical images within the set and additional data (e.g., metadata relating to the set and / or additional metric data relating to the first and second structures of the instance that correspond to the set).

[0047] In certain embodiments, when supplying the multiple sets to the machine learning model (414), the machine learning model is explicitly informed of the individual optical inspection tools from which each set was obtained (e.g., the source of each set in the multiple sets) (416). The machine learning model can then know which optical inspection tool generated which set and use that information in conjunction with the sets themselves to calculate the estimated overlay offset. For example, one or more optical images forming a first set are obtained from a first optical inspection tool among the multiple optical inspection tools mentioned above. The first set is supplied to the machine learning model as indicated in the notification that the first set is from the first optical inspection tool (e.g., along with the notification). The machine learning model calculates the individual estimated overlay offset for the first set based on the first set and further based on the fact that the first set is from the first optical inspection tool.

[0048] Alternatively, the machine learning model may calculate the estimated overlay offset without providing any information about the optical inspection tool from which the data was obtained.

[0049] In certain embodiments, each set is a single optical image showing the first and second structures (408). As described with respect to steps 316 and 324 of Method 300 (Figure 3), each set is fed into its machine learning model, which calculates an estimated overlay offset for each set based on its single optical image.

[0050] In other specific embodiments, each set includes a first optical image focused on a first process layer and a second optical image focused on a second process layer (410). As described with respect to steps 318 and 326 of Method 300 (Figure 3), each set is fed into its machine learning model, which calculates an estimated overlay offset for each set based on the first and second optical images.

[0051] In certain embodiments, each set includes multiple optical images obtained by optically inspecting various parts of semiconductor wafers using multiple individual optical modes (412). As described with respect to steps 320 and 328 of Method 300 (Figure 3), each set is fed into its machine learning model, which calculates an estimated overlay offset for each set based on those multiple optical images.

[0052] The individually estimated overlay offsets obtained by this method 400 can be used directly to monitor and control the semiconductor manufacturing process (e.g., for statistical process control). Alternatively, the individually estimated overlay offsets can be used as one of several inputs, and these can later be used to determine the individual overlay offsets between the first and second structures of a particular instance. For example, these individually estimated overlay offsets can be combined with one or more other individual overlay offset values ​​calculated for the first and second structures of that instance using another technique (e.g., a technique that does not involve machine learning), for example, by averaging them.

[0053] Figure 5 is a flowchart illustrating a method for measuring overlay offset using an ensemble (i.e., multiple) machine learning models, relating to a certain embodiment. Method 500 can be executed by one or more computer systems (e.g., the computer system of the semiconductor inspection system 600 in Figure 6). Method 300 (Figure 3) can be incorporated into Method 500, and can also be executed as part of the execution of Method 500.

[0054] In certain embodiments, a group of separate machine learning models are trained using a training set of optical images showing a first structure within a first process layer (e.g., the first structure in Method 300 in Figure 3) and a second structure within a second process layer (e.g., the second structure in Method 300 in Figure 3) (502). The optical images in the training set are annotated with individual predetermined overlay offset values ​​(e.g., measured using scanning electron microscopy). Each machine learning model in the group can be trained as described with respect to step 302 of Method 300 (Figure 3) and / or step 402 of Method 400 (Figure 4).

[0055] The group of machine learning models may consist of multiple distinct machine learning models of the same type (504). For example, all of these machine learning models may be neural network models, gradient boosting models, support vector machines, or principal component regression models; that is, they may be selected from a group consisting of neural network models, gradient boosting models, support vector machines, and principal component regression models. The training of each machine learning model within the group may be performed with randomization (e.g., random initialization). This randomization makes each machine learning model within the group distinct. Alternatively, the distinct machine learning models within the group may be of different types.

[0056] Alternatively, the training of the machine learning model may be omitted from Method 500. For example, the group of machine learning models used to perform Method 500 (e.g., those of the same type, such as neural network models, gradient boosting models, support vector machines, or principal component regression models) may be pre-trained to estimate the overlay offset between the first and second structures (e.g., the first and second structures in Method 300 in Figure 3) and deployed for use in semiconductor manufacturing.

[0057] One or more optical images showing a portion of a semiconductor wafer, including a first structure within a first process layer and a second structure within a second process layer, are acquired as described with respect to method 300 (Figure 3) (304). In certain embodiments, the portion of the semiconductor wafer is optically inspected (e.g., using an optical inspection tool 632 in Figure 6) when acquiring the one or more optical images (306).

[0058] The one or more optical images are supplied to the aforementioned separate group of machine learning models (506). As described in relation to step 314 of Method 300 (Figure 3), the one or more optical images may be supplied to each machine learning model in that group. For example, as described in relation to step 316 or step 318 of Method 300 (Figure 3) and / or step 320 of Method 300 (Figure 3), the one or more optical images may be supplied to each machine learning model in that group.

[0059] Multiple estimated overlay offsets between the first and second structures (e.g., those relating to the first and second structures of a particular instance) are obtained from the separate group of machine learning models (508). Among these multiple estimated overlay offsets, each estimated overlay offset can be an estimated total overlay offset (e.g., total overlay offset 118 in Figure 1) or an estimated unexpected overlay offset (e.g., unexpected overlay offset 116). For example, all of these multiple estimated overlay offsets may be estimated total overlay offsets, or they may be estimated unexpected overlay offsets. Each estimated overlay offset among these multiple estimated overlay offsets comes from an individual machine learning model within the separate group of machine learning models. Each estimated overlay offset can be obtained as described in step 322 of method 300 (Figure 3). For example, each estimated overlay offset can be obtained as described with respect to step 324 or step 326 of Method 300 (Figure 3) and / or step 328 of Method 300 (Figure 3).

[0060] The overlay offset between the first and second structures (e.g., relating to the first and second structures of a particular instance) is determined at least partially based on the multiple estimated overlay offsets (510). For example, the overlay offset is determined by averaging the multiple estimated overlay offsets (e.g., calculating their average) or by combining the multiple estimated overlay offsets in other ways. In certain embodiments, the overlay offset is determined using the multiple estimated overlay offsets and additional information. For example, the overlay offset is determined using the multiple estimated overlay offsets and one or more additional overlay offset values ​​relating to the first and second structures that are determined without using the machine learning model set. Possible overlay offsets that can be determined include the estimated total overlay offset (e.g., total overlay offset 118 in Figure 1) and the estimated unexpected overlay offset (e.g., unexpected overlay offset 116).

[0061] Using the overlay offset determined by methods 300, 400, and / or 500 (Figures 3-5), process drift can be measured, and process control can be established and maintained. For example, one or more process parameters (e.g., photolithography exposure and / or irradiation dose) can be modified at least partially based on the determined overlay offset. Establishing and maintaining process control in this manner is particularly important during the start-up period for new semiconductor processes and / or devices, but is also important for maintaining yield during continued production.

[0062] Figure 6 is a block diagram of a semiconductor inspection system 600 according to one embodiment. This semiconductor inspection system 600 can be used for overlay metricing (e.g., to perform methods 300, 400 and / or 500 in Figures 3 to 5). The semiconductor inspection system 600 has one or more (e.g., more) optical inspection tools 632, one or more SEMs 634, and a computer system, the computer system having one or more processors 602 (e.g., CPU and / or GPU), a user interface 606, memory 610, and a communication bus(s) 604 interconnecting these components. In one embodiment, the optical inspection tools 632 and the SEM(s) 634 are communicate-coupled to the computer system via one or more wired and / or wireless networks 630. The computer system can be provided with one or more network interfaces to enable communication with the optical inspection tool 632, the SEM(s) 634, and other remote devices through one or more networks 630.

[0063] The user interface 606 may include a display 607 and one or more input devices 608 (e.g., a keyboard, mouse, contact sensing surface of the display 607, etc.). The display 607 can display various results, including the overlay offset.

[0064] Memory 610 has volatile and / or non-volatile memory. Memory 610 (e.g., non-volatile memory within memory 610) has a non-temporary computer-readable storage medium. Memory 610 optionally has one or more storage devices remotely located relative to the processor 602, and / or a non-temporary computer-readable storage medium that can be inserted into and removed from the computer system of the semiconductor inspection system 600.

[0065] In certain embodiments, the following modules and data, or subsets or supersets thereof, are stored in memory 610 (e.g., a non-temporary computer-readable storage medium provided in memory 610): an operating system 612 incorporating procedures for handling various basic system services and performing hardware-dependent tasks; optical images 614 taken by an optical inspection tool(s) 632; SEM images 620 taken by an SEM(s) 634; a training module 622; one or more machine learning models 624 (e.g., a group of machine learning models); an overlay decision module 626; and a notification module 628 for notifying the results obtained from one or more machine learning models 624 and / or the overlay decision module 626.

[0066] The analysis images 616 from the optical images 614 will be supplied to one or more machine learning models 624. These one or more machine learning models 624 are executable and can determine estimated overlay offsets using the analysis images 616. The optical images 614 may also include a training set 618 that will be supplied to a training module 622 to train one or more machine learning models 624. The SEM images 620 contain images with the same structure as the optical images in the training set 618. Overlay offsets related to these structures are measured according to the SEM images 620, and these are used to annotate the optical images in the training set 618. Consequently, the overlay offsets measured according to the SEM images 620 are associated with individual optical images in the training set 618 and supplied to the training module 622.

[0067] The overlay determination module 626 is executable and can determine the overlay offset based at least partially on the estimated overlay offset from the machine learning model(s) 624. The overlay offset determined by the overlay determination module 626 can be supplied to the notification module 628. In certain embodiments, the overlay determination module 626 is omitted, and the estimated overlay offset from the machine learning model(s) 624 is supplied to the notification module 628.

[0068] Instructions for executing all or part of methods 300, 400, and / or 500 (Figures 3 to 5) can be placed in memory 610 (e.g., a non-temporary computer-readable storage medium provided in memory 610). One or more functions described herein can be executed by executing a set of instructions corresponding to each module (e.g., machine learning model(s) 624) stored in memory 610 using one or more processors 602. Separate modules do not need to be implemented as separate software programs. Modules and various subsets of modules can be combined or rearranged in other ways. In certain embodiments, a subset or superset of the above-described modules and / or data structures is stored in memory 610.

[0069] The aim of Figure 6 is to provide a functional description of the various features that may exist within a semiconductor inspection system, not to provide a structural schematic. For example, the functions of the computer system within the semiconductor inspection system 600 may be shared among multiple devices. Instead of doing so, some of the modules stored in memory 610 may be stored in one or more other computer systems that are communicably coupled to the computer system of the semiconductor inspection system 600 via one or more networks. For example, the training module 622, the training set 618, and the SEM images 620 may be stored on the first computer system that trains the machine learning model(s) 624. The machine learning model(s) 624 may then be deployed to a second computer system and stored there together with the analysis images 616, the overlay decision module 626, and / or the notification module 628.

[0070] The above description is for illustrative purposes only and is based on specific embodiments. However, the illustrative discussion above is not intended to be exhaustive, nor is it intended to limit the technical scope of the claims to the disclosed forms themselves. Many modifications and alterations can be considered in light of the above teachings. The intention of selecting the embodiments is to best illustrate the principles underlying the claims and their practical applications, so that other persons skilled in the art can best use the various modified embodiments suited to their intended use.

Claims

1. It is a method, One or more optical images are obtained showing a portion of a semiconductor wafer, where the first structure within the first process layer and the second structure within the second process layer are visible. The one or more optical images are supplied to a machine learning model trained to estimate the overlay offset between the first and second structures, and The estimated overlay offset between the first structure and the second structure is obtained from the machine learning model. method.

2. A method according to claim 1, wherein the portion of the semiconductor wafer is optically inspected when acquiring one or more optical images.

3. The method according to claim 2, The aforementioned one or more optical images include multiple optical images, When acquiring the aforementioned multiple optical images, the aforementioned portion of the semiconductor wafer is optically inspected using multiple optical modes. Each of the aforementioned multiple optical modes has a separate set of optical conditions for inspecting the portion of the semiconductor wafer, and A method for acquiring each optical image from the aforementioned plurality of optical images using an individual optical mode from the aforementioned plurality of optical modes.

4. The method according to claim 1, When acquiring one or more optical images, a single optical image is acquired that shows the first structure in the first process layer and the second structure in the second process layer. When supplying one or more optical images to the machine learning model, the single optical image is supplied to the machine learning model, and A method for obtaining the estimated overlay offset between the first structure and the second structure, by using the machine learning model to calculate the estimated overlay offset based on the single optical image.

5. The method according to claim 1, When acquiring one or more optical images, a first optical image focused on the first process layer and a second optical image focused on the second process layer are acquired. When supplying the one or more optical images to the machine learning model, the first optical image and the second optical image are supplied to the machine learning model, and A method for obtaining the estimated overlay offset between the first structure and the second structure, using the machine learning model to calculate the estimated overlay offset based on the first optical image and the second optical image.

6. The method according to claim 1, wherein the machine learning model is a first machine learning model within a group of separate machine learning models trained to estimate the overlay offset between the first and second structures, The one or more optical images are supplied to the separate group of machine learning models, and Multiple estimated overlay offsets between the first structure and the second structure are obtained from the separate group of machine learning models, provided that each of these multiple estimated overlay offsets is obtained from an individual machine learning model within the separate group of machine learning models. method.

7. The method according to claim 6, further comprising determining the overlay offset between the first structure and the second structure, at least in part, based on the plurality of estimated overlay offsets.

8. The method according to claim 6, wherein the group of machine learning models consists of multiple separate machine learning models of the same type.

9. A method according to claim 8, wherein the type is selected from the group consisting of neural network models, gradient boosting models, support vector machines, and principal component regression models.

10. The method according to claim 1, The aforementioned machine learning model is the first machine learning model, The method further shows the third structure within the third process layer in the aforementioned one or more optical images, The one or more optical images are supplied to the first machine learning model and to the second machine learning model trained to estimate the overlay offset between the second and third structures, and The estimated overlay offset between the second structure and the third structure is obtained from the second machine learning model. method.

11. A method according to claim 1, wherein the one or more optical images are one or more optical images forming a first set of one or more optical images from a total of multiple sets of one or more optical images representing various parts of various semiconductor wafers, and the first structure and the second structure are shown in each of the multiple sets, further, The aforementioned multiple sets are obtained from multiple optical inspection tools, provided that each set within those multiple sets is generated by an individual optical inspection tool among those multiple optical inspection tools. The aforementioned multiple sets are supplied to the machine learning model, and For each of the multiple sets, the individually estimated overlay offset between the first structure and the second structure is obtained from the machine learning model, and the individually estimated overlay offset is calculated using the machine learning model. method.

12. The method according to claim 11, A method for acquiring one or more optical images of the first set from a first optical inspection tool among the plurality of optical inspection tools, further, The machine learning model is made aware that the first set is from the first optical inspection tool, and With respect to the first set, when obtaining the individually estimated overlay offset between the first structure and the second structure, the machine learning model is used to calculate the individually estimated overlay offset based on the first set, and further based on the fact that the first set is from the first optical inspection tool. method.

13. The method according to claim 11, further comprising training the machine learning model using a training set consisting of optical images from the plurality of inspection tools before supplying the plurality of sets to the machine learning model, If the optical image in the training set is different from one or more optical images in the set, A method in which the optical images within the training set are annotated with individual predetermined overlay offset values.

14. A method according to claim 13, wherein the individual predetermined overlay offset values ​​with respect to the optical images in the training set are measured using scanning electron microscopy.

15. The method according to claim 1, further comprising training the machine learning model using a training set of optical images before supplying the machine learning model with the one or more optical images, The optical images in the training set are different from the one or more optical images, and method. A method in which the optical images within the training set are annotated with individual predetermined overlay offset values.

16. A method according to claim 15, wherein the individual predetermined overlay offset values ​​with respect to the optical images in the training set are measured using scanning electron microscopy.

17. A non-temporary computer-readable storage medium containing one or more programs that are executed by one or more processors, A command to acquire one or more optical images showing the first structure within the first process layer and the second structure within the second process layer of a portion of a semiconductor wafer, A command to supply the one or more optical images to a machine learning model trained to estimate the overlay offset between the first and second structures, An instruction to obtain the estimated overlay offset between the first structure and the second structure from the machine learning model, A non-temporary computer-readable storage medium containing [something].

18. It is a system, One or more optical inspection tools for optically inspecting semiconductor wafers, One or more processors, A memory containing one or more programs executed by the aforementioned one or more processors, The program comprises, and the one or more programs are A command to acquire one or more optical images from one or more optical inspection tools showing the first structure in the first process layer and the second structure in the second process layer for a portion of a semiconductor wafer, A command to supply the one or more optical images to a machine learning model trained to estimate the overlay offset between the first and second structures, An instruction to obtain the estimated overlay offset between the first structure and the second structure from the machine learning model, A system that includes this.

Images