Overlay estimation based on optical inspection and machine learning

Abstract

One or more optical images of a portion of a semiconductor wafer are obtained. The one or more optical images exhibit a first structure in a first process layer and a second structure in a second process layer. The one or more optical images are provided to a machine learning model trained to estimate a superposition offset between the first structure and the second structure. An estimated superposition offset between the first structure and the second structure is obtained from the machine learning model.

Description

[0001] Related Applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 420,683, filed on October 31, 2022, the entire contents of which are incorporated herein by reference for all purposes. Technical Field

[0003] The present disclosure relates to measuring overlay (ie, overlay shift) of semiconductor devices, and more particularly, to measuring overlay shift using optical inspection and machine learning. Background Art

[0004] In semiconductor manufacturing, misalignment between process layers causes an undesired shift in the position of structures in one process layer as compared to the position of structures in another process layer. This shift (i.e., displacement) is called overlay or equivalently, overlay offset. Measuring these shifts is called overlay metrology. Accurate overlay metrology is important for measuring process drift and therefore for establishing and maintaining process control, especially during the ramp-up period of new semiconductor processes and / or devices. As feature sizes (e.g., line widths) on semiconductor wafers shrink, accurate overlay metrology becomes increasingly important because a reduction in feature size correspondingly reduces the available overlay budget; as features shrink, less overlay error can be accommodated.

[0005] Scanning electron microscopes (SEMs) are useful tools for superposition metrology because they have high resolution and can image structures regardless of their shape or size. However, scanning electron microscopes are slow and therefore have low throughput. Therefore, it is desirable to use optical inspection for superposition metrology. But optical inspection does not provide the accuracy that SEMs provide for superposition metrology. For example, optical inspection may not be able to resolve structures used in superposition metrology because the size of the structures is below the diffraction limit. Therefore, superposition cannot usually be measured simply by observing images produced by optical inspection or by applying object-recognition algorithms to such images. However, optical images still contain information about the superposition encoded therein, and algorithms exist to extract this information from the images. But the accuracy of measurements made using these classical algorithms is limited. Summary of the invention

[0006] To improve the accuracy of optical inspection-based overlay metrology, machine learning is used to analyze optical images of semiconductor structures.

[0007] In some embodiments, a method includes obtaining one or more optical images of a portion of a semiconductor wafer, the one or more optical images showing a first structure in a first process layer and a second structure in a second process layer. The one or more optical images are provided to a machine learning model trained to estimate an overlay offset between the first structure and the second structure. An estimated overlay offset between the first structure and the second structure is obtained from the machine learning model.

[0008] In some embodiments, a non-transitory computer-readable storage medium stores one or more programs executed by one or more processors. The one or more programs include instructions for executing the above method.

[0009] In some embodiments, a system includes one or more optical inspection tools for optically inspecting semiconductor wafers, one or more processors, and one or more programs stored for execution by the one or more processors. The one or more programs include instructions for executing the above method, wherein one or more optical images are obtained from the one or more optical inspection tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] For a better understanding of the various described embodiments, reference should be made to the following detailed description in conjunction with the following drawings.

[0011] Figure 1 is a cross-sectional side view of a portion of a semiconductor wafer to be subjected to optical imaging to perform overlay metrology and including a first structure in a first process layer and a second structure in a second process layer according to some embodiments.

[0012] Figure 2 is a cross-sectional side view of a portion of a semiconductor wafer to be subjected to optical imaging to perform overlay metrology and including a first structure in a first process layer, a second structure in a second process layer, and a third structure in a third process layer according to some embodiments.

[0013] Figure 3 is a flow chart illustrating a method of measuring overlay offset using machine learning according to some embodiments.

[0014] Figure 4 is a flow chart illustrating a method of measuring overlay shifts of optical images from multiple optical inspection tools using machine learning, according to some embodiments.

[0015] Figure 5 is a flow chart showing a method for measuring stacking offset using an ensemble of machine learning models according to some embodiments.

[0016] Figure 6 is a block diagram of a semiconductor inspection system according to some embodiments.

[0017] Throughout the drawings and description, like reference numerals refer to corresponding parts. DETAILED DESCRIPTION

[0018] Reference will now be made in detail to various embodiments, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the various described embodiments. However, it will be apparent to one of ordinary skill in the art that the various described embodiments may be practiced without these specific details. In other instances, well-known methods, processes, components, circuits, and networks are not described in detail to avoid unnecessarily obscuring aspects of the embodiments.

[0019] Figure 1 1 is a cross-sectional side view of a portion 100 of a semiconductor wafer to be subjected to optical imaging to perform overlay metrology according to some embodiments. The semiconductor wafer may be fully or partially processed. Portion 100 includes a first structure 104 in a first process layer and a second structure 110 in a second process layer. Figure 1 In an example of , first structure 104 is a first grating including a plurality of grating elements 106, and second structure 110 is a second grating including a plurality of grating elements 112. The plurality of grating elements 106 may be periodically arranged in first structure 104, wherein consecutive grating elements 106 are separated by the same distance and thus have the same pitch. Similarly, the plurality of grating elements 112 may be periodically arranged in second structure 110, wherein consecutive grating elements 112 are separated by the same distance and thus have the same pitch. The pitch of grating elements 112 in second structure 110 may be the same as the pitch of grating elements 106 in first structure 104.

[0020] The first process layer where the first structure 104 is located is on the substrate 102 of the semiconductor wafer. Alternatively, the first process layer can be above other process layers. The grating elements 106 of the first structure 104 are separated from each other by a filling material (e.g., a dielectric). The grating elements 112 of the second structure 110 can be separated from each other by a filling material (e.g., a dielectric) or by air, depending on how far the processing of the semiconductor wafer has been carried out. A filling layer 108 (e.g., an interlayer dielectric) separates the second process layer from the first process layer. In some embodiments, the grating elements 106 and / or the grating elements 112 are metal lines. In the nomenclature of overlay metrology, the second process layer located above the first process layer is referred to as an "outer layer" and the first process layer located below the second process layer is referred to as an "inner layer".

[0021] Figure 1Both the intended and actual positioning of grating elements 112 in second structure 110 relative to grating elements 106 of first structure 104 are shown. The intended positioning is shown with solid lines and fill patterns, while the actual positioning is shown with dashed lines. There is an intended (i.e., programmed) overlay offset 114 between second structure 110 and first structure 104. The intended overlay offset 114 is the overlay offset that would exist if grating elements 112 in second structure 110 were manufactured exactly at their intended locations relative to grating elements 106 in first structure 104. Figure 1 In the example of , the expected overlay offset 114 is non-zero. Alternatively, the expected overlay offset may be zero.

[0022] In practice, due to inaccuracies and variability in the photolithography process, the grating elements 112 in the second structure 110 or the grating elements 106 in the first structure 104 are typically not manufactured precisely at the intended location. The difference between the actual relative position and the intended relative position is an additional, unintended (i.e., erroneous) overlay offset 116, which may be positive or negative. The total overlay offset 118 is equal to the sum of the intended overlay offset 114 and the unintended overlay offset 116.

[0023] Portion 100 may be a test pattern on a semiconductor wafer. The test pattern may be separate from a die on a semiconductor wafer (e.g., may be in a scribe street) or may be within a die on a semiconductor wafer. A semiconductor wafer may have multiple test patterns like portion 100 at various locations so that overlay metrology may be performed at multiple locations on the wafer. Alternatively, portion 100 may be part of the circuitry of the die itself.

[0024] Figure 2 2 is a cross-sectional side view of a portion 200 of a semiconductor wafer to be subjected to optical imaging to perform overlay metrology according to some embodiments. The semiconductor wafer may be fully or partially processed. The 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. Figure 2 In the example of , the third structure 202 is a third grating including a plurality of grating elements 204. The plurality of grating elements 204 may be periodically arranged in the third structure 202, wherein consecutive grating elements 204 are separated by the same distance and thus have the same pitch, which is the same pitch 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.

[0025] The grating elements 204 of the third structure 202 may be separated from each other by a filling material (e.g., a dielectric) or by air, depending on how far the processing of the semiconductor wafer has progressed. The grating elements 112 of the second structure 110 in the portion 200 are separated from each other by a filling material (e.g., a dielectric). A filling layer 206 (e.g., an interlayer dielectric) separates the third process layer from the second process layer. In some embodiments, the grating elements 204 are metal lines (e.g., as are the grating elements 106 and / or the grating elements 112).

[0026] In addition to showing both the intended and actual positioning of the grating elements 112 in the second structure 110, Figure 2 Also shown are both the intended and actual positioning of the grating elements 204 in the third structure 202. The intended positioning is shown with solid lines and fill patterns, while the actual positioning is shown with dashed lines. There is an intended (i.e., programmed) overlay offset 208 between the third structure 202 and the second structure 110. The intended overlay offset 208 is the overlay offset that would exist if the grating elements 204 in the third structure 202 were manufactured exactly at their intended locations relative to the grating elements 112 in the second structure 110. Figure 1 In the example of , the expected overlay offset 208 is non-zero. Alternatively, the expected overlay offset may be zero. However, since the grating elements 204 in the third structure 202 and the grating elements 112 in the second structure 110 are generally not manufactured at the expected positions accurately, there is an actual (i.e., total) overlay offset 210 that is different from the expected overlay offset 208. The difference between the total overlay offset 210 and the expected overlay offset 208 is an unexpected (i.e., erroneous) overlay offset that may be positive or negative.

[0027] When the portion 200 is used to perform additive metrology between the third structure 202 and the second structure 110, the third process layer is an outer layer and the second process layer is an inner layer, wherein the third process layer is located above the second process layer. When the portion 200 is used to perform additive metrology between the second structure 110 and the first structure 104, the second process layer is an outer layer and the first process layer is an inner layer. The portion 200 can also be used to perform additive metrology between the third structure 202 and the first structure 104, wherein the third process layer is an outer layer and the first process layer is an inner layer.

[0028] For example, for part 100 ( Figure 1 ), portion 200 may be a test pattern on a semiconductor wafer, or may be part of the circuitry of the die itself.

[0029] Figure 3 is a flow chart of a method 300 for measuring stacking offset using machine learning according to some embodiments. Figure 6The method 300 is performed by using a computer system of a semiconductor inspection system 600.

[0030] In some embodiments, method 300 includes training (302) a machine learning model to estimate an overlay offset between the first structure and the second structure using a training set of optical images showing a first structure in a first process layer and a second structure in a second process layer. For example, the first structure is first structure 104 ( Figure 1 ) and the second structure is the second structure 110 ( Figure 1 ). In other examples, the first structure and the second structure are respective components of another periodic grating or a non-periodic grating, either of which may be a test pattern on a semiconductor wafer (e.g., as implemented in a scribe line street). In still other examples, the first structure and the second structure are respective components of a different test pattern (e.g., a box-in-box) that may be implemented in a scribe line street. In still other examples, the first structure and the second structure are respective components of a circuit system on a semiconductor die as fabricated in whole or in part on a semiconductor wafer.

[0031] The optical images of the training set are annotated with respective predetermined overlay offset values ​​(e.g., total overlay offset values ​​and / or unexpected overlay offset values): the predetermined overlay offset values ​​are stored in association with their respective optical images. The respective predetermined overlay offset values ​​of the optical images of the training set have been previously measured using, for example, a scanning electron microscope: the portions of the semiconductor wafers that were optically inspected to generate the optical images in the training set were also measured using a scanning electron microscope (SEM) (e.g., Figure 6 The SEM 634 of FIG. 304 is used or multiple SEMs are used for inspection, and the overlay offset values ​​are measured from the resulting SEM images. These predetermined overlay offset values ​​are used as real data for training 302. 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.

[0032] To perform training 302, a training set of optical images and associated predetermined overlay offset values ​​are obtained. A machine learning model is initialized (e.g., in a manner involving randomization). The optical images of the training set are provided to the machine learning model as input, and the resulting output of the machine learning model is compared to the predetermined overlay offset value of the corresponding image (which is the expected output). The optical images of the training set may be provided directly to the machine learning model, or feature extraction may be performed and the extracted features (e.g., in the form of feature vectors of the corresponding optical images) may be provided to the machine learning model. The machine learning model is adjusted based on the difference between the resulting output (i.e., the actual output) and the expected output, and the training process continues until the resulting output converges with the expected output. Once this convergence occurs, the machine learning model may be tested and then deployed for semiconductor manufacturing. The machine learning model may be deployed to one or more semiconductor manufacturing facilities, referred to as a wafer fab, or a computer system communicatively coupled to one or more wafer fabs.

[0033] Alternatively, training the machine learning model is omitted from method 300. For example, method 300 is performed using a machine learning model that has been trained to estimate the stacking offset between the first structure and the second structure and has been deployed for semiconductor manufacturing. Again, the first structure and the second structure may be the first structure 104 ( Figure 1 ) and the second structure 110( Figure 1 ), may be a corresponding component of another periodic grating or a non-periodic grating, either of which may be a test pattern on a semiconductor wafer (e.g., as implemented in a scribe line); may be a corresponding component of a different test pattern (e.g., a box-in-box) that may be implemented in a scribe line; or may be a corresponding component of a circuit system on a semiconductor die that is fabricated in whole or in part on a semiconductor wafer.

[0034] One or more optical images of a portion of a semiconductor wafer are obtained (304), the one or more optical images showing a first structure in a first process layer and a second structure in a second process layer (i.e., specific instances of the first and second structures shown at specific locations in a specific semiconductor die). The one or more optical images are different from a training set of optical images. For example, none of the one or more optical images were included in the training set. Thus, the instances of the first and second structures in the one or more optical images may be different from the instances of the first and second structures in corresponding optical images of the training set (e.g., may be at different locations, on different semiconductor dies, and / or on different semiconductor wafers). In some embodiments, obtaining the one or more optical images includes optically inspecting (306) the portion of the semiconductor wafer (e.g., using Figure 6Alternatively, one or more optical images are obtained from a computer memory (e.g., a database) storing optical images previously generated by optical inspection.

[0035] The one or more optical images are provided (314) to a machine learning model (e.g., a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model). In some embodiments, the one or more optical images are provided directly to the machine learning model (e.g., according to deep learning) such that the machine learning model receives the one or more optical images as input. Alternatively, the one or more optical images are provided 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 provided as input to the machine learning model.

[0036] Obtain (322) an estimated overlay offset between the first structure and the second structure (e.g., for a particular instance of the first and second structures) from the machine learning model: The machine learning model determines (i.e., calculates) an estimated overlay offset using the one or more optical images and provides the estimated overlay offset as an output. The estimated overlay offset may be an estimated total overlay offset (e.g., Figure 1 The total stacking offset 118) or the estimated unexpected stacking offset (e.g., Figure 1 The estimated stacking offset may be received from the machine learning model in response to providing the one or more optical images to the machine learning model. The machine learning model may determine the estimated stacking offset based solely on the one or more optical images, or based on the one or more optical images and further based on additional data (e.g., metadata for the one or more optical images and / or additional metrology data for the first and second structures).

[0037] In some embodiments, the one or more optical images are a single optical image showing a first structure in a first process layer and a second structure in a second process layer. The single optical image is obtained (308) and provided (316) directly or indirectly to a machine learning model. The machine learning model is used to calculate (224) an estimated overlay offset based on the single optical image.

[0038] In some other embodiments, the 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. The first and second optical images are obtained (310) and provided (318) directly or indirectly to a machine learning model. The first and second optical images may be obtained, for example, using respective first and second cameras in an optical inspection system, wherein the first and second cameras have a known alignment. A stitching process may be performed on the first and second images to align them according to the known alignment. The machine learning model is used to calculate (326) an estimated overlay offset based on the first and second optical images.

[0039] In some embodiments, the one or more optical images include a plurality of optical images obtained (312) by optically inspecting a portion of a semiconductor wafer using a plurality of optical modes. Each optical mode of the plurality of optical modes has a set of different optical conditions for inspecting a portion of a semiconductor wafer, wherein one or more optical conditions in the set are different from other sets of optical conditions of the plurality of optical modes. Examples of optical conditions include, but are not limited to, light source, wavelength range, polarization, focus, transmission distribution in an illumination aperture, transmission distribution in a light collection aperture, and phase shift distribution in a light collection aperture. The respective optical modes may also or alternatively correspond to respective Mueller elements from an imaging Mueller ellipsometer and / or respective harmonic signals from an imaging ellipsometer. Each optical image of the plurality of optical images is obtained using a respective optical mode of the plurality of optical modes. Each optical mode may be used to generate a single optical image showing both a first structure and a second structure or to generate a plurality of optical images (e.g., a first image focused on a first process layer and a second image focused on a second process layer). The plurality of optical images are provided (320) directly or indirectly to a machine learning model. The machine learning model is used to calculate (328) an estimated stack offset based on the plurality of optical images.

[0040] The estimated overlay offset obtained in method 300 may be used directly to monitor and control the semiconductor manufacturing process (e.g., for statistical process control). Alternatively, the estimated overlay offset may be one of a plurality of inputs subsequently used to determine the overlay offset between the first and second structures. For example, the estimated overlay offset may be averaged or otherwise combined with one or more other overlay offset values ​​calculated for the first and second structures using a different technique (e.g., a technique not involving machine learning).

[0041] The method 300 may be extended to determine a second structure (e.g., Figure 2 The second structure 110) and the third structure in the third process layer (for example, Figure 2The method 300 further includes a first machine learning model and a second machine learning model (e.g., a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model) trained to estimate the superposition offset between the second structure and the third structure. Similar to training 302, the second machine learning model may be trained as part of the method 300, or may have been previously trained and deployed for semiconductor manufacturing. According to step 314, the one or more optical images are provided directly or indirectly to the first machine learning model and to the second machine learning model. In an example, the one or more optical images are a single optical image showing the first structure, the second structure, and the third structure. According to step 316, the single optical image is provided to the first machine learning model and is also provided to the second machine learning model. 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. According to step 318, the first and second optical images are provided to the first machine learning model, and the second and third optical images are provided to the second machine learning model. Other examples are possible. An estimated overlay offset between the second structure and the third structure (e.g., for a specific instance of the second and third structures) is obtained from the second machine learning model: The second machine learning model is used to calculate the estimated overlay offset between the second structure and the third structure. The estimated overlay offset can be a total overlay offset (e.g., Figure 2 Total stacking offset 210) or unexpected stacking offset (e.g., Figure 2 The difference between the total stacking offset 210 and the expected stacking offset 208).

[0042] In addition to or as an alternative to using the second machine learning model to obtain the estimated stacking offset between the second structure and the third structure, 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) can be used to obtain the first structure (e.g., Figure 2 The first structure 104) and the third structure (for example, Figure 2Similar to training 302, the third machine learning model may be trained as part of method 300, or may have been previously trained and deployed for semiconductor manufacturing. According to step 314, the one or more optical images are provided directly or indirectly to the first machine learning model, and to the third machine learning model (and, in accordance with some embodiments, to the second machine learning model). In an example, the one or more optical images are a single optical image showing the first structure, the second structure, and the third structure. According to step 316, the single optical image is provided to the first machine learning model, and to the third machine learning model (and, in accordance with some embodiments, to the second machine learning model). In another example, the one or more optical images include a first optical image focused on a first process layer, a second optical image focused on a second process layer, and a third optical image focused on a third process layer. According to step 318, the first and second optical images are provided to the first machine learning model, and the first and third optical images are provided to the third machine learning model (and, in accordance with some embodiments, the second and third optical images are provided to the second machine learning model). Obtaining an estimated stacking offset between the first structure and the third structure (e.g., for a specific instance of the first and third structures) from a third machine learning model: The third machine learning model is used to calculate an estimated stacking offset between the first structure and the third structure. The estimated stacking offset can be a total stacking offset or an unexpected stacking offset.

[0043] The first, second and / or third machine learning models may be combined into a single machine learning model having multiple outputs. The single machine learning model is trained to estimate the offset stack between the first and second structures, between the second and third structures, and / or between the first and third structures. The multiple outputs include a first output to provide an estimated offset stack between the first and second structures, a second output to provide an estimated offset stack between the second and third structures, and a third output to provide an estimated offset stack between the first and third structures.

[0044] The first, second and third structures may be respectively the first structure 104 ( Figure 2 )、Second structure 110( Figure 2 ) and the third structure 202( Figure 2 ), may be a corresponding component of another periodic grating or a non-periodic grating, either of which may be a test pattern on a semiconductor wafer (e.g., as implemented in a scribe line); may be a corresponding component of a different test pattern (e.g., a box-in-box) that may be implemented in a scribe line; or may be a corresponding component of a circuit system on a semiconductor die that is fabricated in whole or in part on a semiconductor wafer.

[0045] Figure 44 is a flow chart showing a method 400 for measuring an overlay offset of optical images from a plurality of optical inspection tools according to some embodiments. The method may be performed by one or more computer systems (e.g., Figure 6 The method 400 may be performed by using a computer system of a semiconductor inspection system 600. The method 400 may include the method 300 ( Figure 3 ), method 300 may be performed as part of method 400.

[0046] In some embodiments, a first structure in a first process layer (eg, Figure 3 300) and a second structure in a second process layer (eg, Figure 3 The method of claim 300 is used to train (402) a machine learning model to estimate an overlay offset between the first structure and the second structure using a training set of optical images of the first structure and the second structure. The optical images of the training set are annotated with corresponding predetermined overlay offset values ​​(e.g., total overlay offset values ​​and / or unexpected overlay offset values): the predetermined overlay offset values ​​are stored in association with their corresponding optical images. The corresponding predetermined overlay offset values ​​of the optical images of the training set are measured, for example, using a scanning electron microscope (e.g., as for 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. A training set of optical images may be obtained from a plurality of optical inspection tools: a corresponding optical inspection tool in the plurality of optical inspection tools may generate a corresponding optical image of the training set by optically inspecting a corresponding semiconductor wafer. Training 402 may be performed as described for method 300 ( Figure 3 ) is performed as described in training 302 in ).

[0047] Alternatively, training the machine learning model is omitted from method 400. For example, method 400 is performed using a machine learning model that has been trained to estimate the first structure and the second structure (e.g., Figure 3 The method 300 is an embodiment of the present invention which is a method for manufacturing semiconductor devices. The method 300 is an embodiment of the present invention which is a method for manufacturing semiconductor devices.

[0048] From multiple optical inspection tools (e.g. Figure 6Multiple sets of one or more images of a portion of a semiconductor wafer are obtained (404) using an optical inspection tool (632) of the plurality of optical inspection tools. Each set shows a first structure and a second structure (e.g., shows corresponding examples of the first and second structures at a specific location in a specific semiconductor die on a specific semiconductor wafer). Each set is generated by a corresponding optical inspection tool of the plurality of optical inspection tools. One or more optical images are different from a training set of optical images. For example, one or more optical images are not included in the training set. Therefore, the corresponding examples of the first and second structures of the plurality of sets may be different from each other and from the examples of the first and second structures in the corresponding optical images of the training set (e.g., may be at different locations, on different semiconductor dies, and / or on different semiconductor wafers). The plurality of inspection tools may be the same as the plurality of inspection tools from which the training set was obtained, or may be a subset or superset of the plurality of inspection tools. Method 300 ( Figure 3 ) may be a corresponding set of multiple sets.

[0049] In some embodiments, obtaining the one or more optical images includes optically inspecting (406) portions of the semiconductor wafer using a plurality of optical inspection tools. For example, if the plurality of inspection tools includes two tools, portions of some semiconductor wafers are inspected using a first tool to obtain some sets and portions of remaining semiconductor wafers are inspected using a second tool to obtain remaining sets. If the plurality of inspection tools includes three or more tools, portions of some semiconductor wafers are inspected using a first tool to obtain some sets, portions of other semiconductor wafers are inspected using a second tool to obtain other sets, portions of still other semiconductor wafers are inspected using a third tool to obtain still other sets, and so on.

[0050] The plurality of sets are provided (414) to a machine learning model (e.g., a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model). For example, the sets are provided to the machine learning model in succession, where each set is provided to the machine learning model in turn or as the set becomes available. As described with respect to method 300 ( Figure 3 ), each episode may be provided directly or indirectly to a machine learning model.

[0051] For each of the plurality of sets, a respective estimated stacking offset between the first structure and the second structure (e.g., for respective instances of the first and second structures) is obtained (418) from the machine learning model: the machine learning model is used to calculate the respective estimated stacking offset. The respective estimated stacking offset may be an estimated total stacking offset (e.g., Figure 1The machine learning model determines a respective estimated overlay offset using the one or more optical images of the set and provides the estimated overlay offset as an output. The respective estimated overlay offset may be received from the machine learning model in response to providing the set to the machine learning model. The machine learning model may determine the estimated overlay offset based solely on the one or more optical images of the set, or based on the one or more optical images of the set and further based on additional data (e.g., metadata for the set and / or additional metrology data for instances of the first and second structures corresponding to the set).

[0052] In some embodiments, providing (414) the plurality of sets to the machine learning model includes specifying (416) to the machine learning model the respective optical inspection tool from which the respective set was obtained (e.g., from which each set in the plurality of sets was obtained). Thus, the machine learning model knows which optical inspection tool produced which set, and can use this information as well as the set itself to calculate an estimated overlay offset. For example, a first set of one or more optical images is obtained from a first optical inspection tool of a plurality of optical inspection tools. The first set is provided to the machine learning model as is (e.g., together with) a specification that the first set is from the first optical inspection tool. The machine learning model calculates a respective estimated overlay offset for the first set based on the first set and further based on the first set being from the first optical inspection tool.

[0053] Alternatively, information about the optical inspection tool from which the set was acquired is not provided to the machine learning model, which calculates the estimated overlay offset without knowledge of this information.

[0054] In some embodiments, each set is (408) a single optical image showing the first structure and the second structure. Figure 3 ), each episode is provided to a machine learning model that calculates an estimated stacking offset for each episode based on a single optical image.

[0055] In some other embodiments, each set includes (410) a first optical image focused on a first process layer and a second optical image focused on a second process layer. Figure 3 ), each episode is provided to a machine learning model that calculates an estimated stacking offset for each episode based on the first and second optical images.

[0056] In some embodiments, each set includes (412) a plurality of optical images obtained by optically inspecting a portion of a semiconductor wafer using a plurality of corresponding optical modes. Figure 3), each episode is provided to a machine learning model that calculates an estimated stacking offset for each episode based on multiple optical images.

[0057] The respective estimated overlay offset obtained in method 400 may be used directly to monitor and control the semiconductor manufacturing process (e.g., for statistical process control). Alternatively, the respective estimated overlay offset may be one of a plurality of inputs subsequently used to determine the respective overlay offset between particular instances of the first and second structures. For example, the respective estimated overlay offset may be averaged or otherwise combined with one or more other respective overlay offset values ​​calculated for instances of the first and second structures using different techniques (e.g., techniques not involving machine learning).

[0058] Figure 5 is a flow chart showing a method for measuring stacking offset using a collection of machine learning models (i.e., multiple machine learning models) according to some embodiments. The method may be performed by one or more computer systems (e.g., Figure 6 The method 500 may be performed by using a computer system of a semiconductor inspection system 600 of the present invention. The method 500 may include the method 300 ( Figure 3 ), method 300 may be performed as part of executing method 500.

[0059] In some embodiments, a first structure in a first process layer (eg, Figure 3 300) and a second structure in a second process layer (eg, Figure 3 A set of distinct machine learning models is trained (502) using a training set of optical images of the first structure and the second structure of method 300 to estimate the overlay offset between the first structure and the second structure. The optical images of the training set are annotated with corresponding predetermined overlay offset values ​​(e.g., as measured using a scanning electron microscope). The method 300 ( Figure 3 ) step 302 and / or method 400 ( Figure 4 ) to train each machine learning model in the collection as described in step 402.

[0060] The set of machine learning models may be (504) multiple distinct machine learning models of the same type. For example, the machine learning models may all be neural network models, gradient boosting models, support vector machines, or principal component regression models, where the type is thus selected from the group consisting of neural network models, gradient boosting models, support vector machines, or principal component regression models. Training of each machine learning model in the set may involve randomization (e.g., randomized initialization). This randomization may result in each machine learning model in the set being distinct. Alternatively, different machine learning models in the set may be of different types.

[0061] Alternatively, training the machine learning model may be omitted from method 500. For example, method 500 may be performed using a collection of machine learning models (e.g., of the same type, such as a neural network model, a gradient boosting model, a support vector machine, or a principal component regression model) that have been trained to estimate the first structure and the second structure (e.g., Figure 3 The method 300 is an embodiment of the present invention which is a method for manufacturing semiconductor devices. The method 300 is an embodiment of the present invention which is a method for manufacturing semiconductor devices.

[0062] For example, for method 300 ( Figure 3 ), obtaining (304) one or more optical images of a portion of a semiconductor wafer, the one or more optical images showing a first structure in a first process layer and a second structure in a second process layer. In some embodiments, obtaining the one or more optical images includes optically inspecting (306) the portion of the semiconductor wafer (e.g., using Figure 6 Optical inspection tool 632).

[0063] The one or more optical images are provided (506) to a collection of different machine learning models. As described in method 300 ( Figure 3 ), one or more optical images may be provided to each machine learning model of the collection. For example, as described for method 300 ( Figure 3 ) and / or step 316 or step 318 of method 300 ( Figure 3 ), one or more optical images may be provided to each machine learning model in the collection.

[0064] A plurality of estimated stacking offsets between a first structure and a second structure (e.g., for specific instances of the first and second structures) are obtained (508) from a set of distinct machine learning models. Each estimated stacking offset of the plurality of estimated stacking offsets may be an estimated total stacking offset (e.g., Figure 1 The total stacking offset 118 of the estimated stacking offset may be an estimated total stacking offset, or may be an estimated unexpected stacking offset. Each estimated stacking offset of the plurality of estimated stacking offsets is a corresponding machine learning model from a set of distinct machine learning models. The method 300 ( Figure 3 Each estimated stacking offset is obtained as described in step 322 of method 300 ( Figure 3 ) and / or step 324 or step 326 of method 300 ( Figure 3 ) as described in step 328 of the embodiment of the present invention.

[0065] A stacking offset between a first structure and a second structure (e.g., for specific instances of the first and second structures) is determined (510) based at least in part on a plurality of estimated stacking offsets. For example, the stacking offset is determined by averaging the plurality of estimated stacking offsets (e.g., calculating an average of the plurality of estimated stacking offsets) or by otherwise combining the plurality of estimated stacking offsets. In some embodiments, the stacking offset is determined using the plurality of estimated stacking offsets and additional information. For example, the stacking offset is determined using the plurality of estimated stacking offsets and one or more additional stacking offset values ​​for the first and second structures determined without using a set of machine learning models. The determined stacking offset may be an estimated total stacking offset (e.g., Figure 1 The total stacking offset 118) or the estimated unexpected stacking offset (e.g., the unexpected stacking offset 116).

[0066] Use Method 300, 400 and / or 500 ( Figures 3 to 5 ) The overlay shift determined by the method can be used to measure process drift and thus establish and maintain process control. For example, one or more process parameters (e.g., lithography exposure and / or dose) can be modified based at least in part on the determined overlay shift. Establishing and maintaining process control in this manner is particularly important during the ramp-up phase of new semiconductor processes and / or devices, but is also important for maintaining yield during ongoing production.

[0067] Figure 6 6 is a block diagram of a semiconductor inspection system 600 according to some embodiments. The semiconductor inspection system 600 may be used for overlay metrology (e.g., for performing Figures 3 to 5 The semiconductor inspection system 600 includes one or more (e.g., multiple) optical inspection tools 632, one or more SEMs 634, and a computer system having one or more processors 602 (e.g., CPUs and / or GPUs), a user interface 606, a memory 610, and a communication bus 604 interconnecting these components. In some embodiments, the optical inspection tools 632 and the SEMs 634 are communicatively coupled to the computer system through one or more wired and / or wireless networks 630. The computer system may include one or more network interfaces for communicating with the optical inspection tools 632, the SEMs 634, and other remote devices through the one or more networks 630.

[0068] The user interface 606 may include a display 607 and one or more input devices 608 (eg, a keyboard, a mouse, a touch-sensitive surface of the display 607, etc.) The display 607 may display the results including the superimposed offsets.

[0069] Memory 610 includes volatile and / or nonvolatile memory. Memory 610 (e.g., nonvolatile memory within memory 610) includes non-transitory computer-readable storage media. Memory 610 optionally includes one or more storage devices located remotely from processor 602 and / or non-transitory computer-readable storage media removably inserted into a computer system of semiconductor inspection system 600.

[0070] In some embodiments, the memory 610 (e.g., a non-transitory computer-readable storage medium of the memory 610) stores the following modules and data, or a subset or superset thereof: an operating system 612 including a system for handling various basic system services and for performing hardware-dependent tasks, an optical image 614 acquired with an optical inspection tool 632, a SEM image 620 acquired with a SEM 634, a training module 622, one or more machine learning models 624 (e.g., a collection of machine learning models), an overlay determination module 626, and a reporting module 628 for reporting results from the one or more machine learning models 624 and / or the overlay determination module 626.

[0071] The optical images 614 include images 616 for analysis to be provided to one or more machine learning modules 624. The one or more machine learning models 624 are executable to determine estimated overlay shifts using the images 616 for analysis. The optical images 614 may also include a training set 618 to be provided to a training module 622 to train the one or more machine learning models 624. The SEM images 620 include images having the same structures as the optical images in the training set 618. The overlay shifts of these structures are measured in the SEM images 620 and used to annotate the optical images in the training set 618, such that the overlay shifts measured in the SEM images 620 are provided to the training module 622 in association with the corresponding optical images in the training set 618.

[0072] The overlay determination module 626 is executable to determine an overlay offset based at least in part on the estimated overlay offset from the machine learning model 624. The overlay offset determined by the overlay determination module 626 can be provided to the reporting module 628. In some embodiments, the overlay determination module 626 is omitted and the estimated overlay offset from the machine learning model 624 is provided to the reporting module 628.

[0073] The memory 610 (eg, a non-transitory computer-readable storage medium of the memory 610) may include a computer program for executing the methods 300, 400, and / or 500 ( Figures 3 to 5) of all or a portion of the instructions of the processor 602. Each of the modules stored in the memory 610, including the machine learning model 624, corresponds to an instruction set that can be executed by the one or more processors 602 for performing one or more functions described herein. The separate modules are not necessarily implemented as separate software programs. The modules and various subsets of modules may be combined or otherwise rearranged. In some embodiments, the memory 610 stores a subset or superset of the modules and / or data structures identified above.

[0074] Figure 6 It is intended more as a functional description of various features that may be present in a semiconductor inspection system than as a block diagram. For example, the functionality of the computer system in the semiconductor inspection system 600 may be divided between multiple devices. A portion of the modules stored in the memory 610 may alternatively be stored in one or more other computer systems that are communicatively coupled to the computer system of the semiconductor inspection system 600 through one or more networks. For example, the training module 622, the training set 618, and the SEM image 620 may be stored on a first computer system that trains the machine learning model 624. The machine learning model 624 may then be deployed to a second computer system that stores the machine learning model 624 as well as the image 616 for analysis, the overlay determination module 626, and / or the reporting module 628.

[0075] For purposes of illustration, the foregoing description has been described with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the scope of the claims to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings. The described embodiments were chosen in order to best illustrate the principles underlying the claims and their practical application, to thereby enable one skilled in the art to best use the embodiments with various modifications as appropriate for the particular use contemplated.

Claims

1. A method comprising: obtaining one or more optical images of a portion of a semiconductor wafer, the one or more optical images showing a first structure in a first process layer and a second structure in a second process layer; providing the one or more optical images to a machine learning model trained to estimate an overlay shift between the first structure and the second structure; and An estimated stacking offset between the first structure and the second structure is obtained from the machine learning model. 2 . The method of claim 1 , wherein obtaining the one or more optical images comprises optically inspecting the portion of the semiconductor wafer.

3. The method according to claim 2, wherein: The one or more optical images include a plurality of optical images; and Obtaining the plurality of optical images includes optically inspecting the portion of the semiconductor wafer using a plurality of optical modes; Each optical mode of the plurality of optical modes has a distinct set of optical conditions for inspecting the portion of the semiconductor wafer; and Each optical image of the plurality of optical images is obtained using a corresponding optical mode of the plurality of optical modes.

4. The method according to claim 1, wherein: Obtaining the one or more optical images includes obtaining a single optical image showing the first structures in the first process layer and the second structures in the second process layer; Providing the one or more optical images to the machine learning model comprises providing the single optical image to the machine learning model; and Obtaining the estimated overlay shift between the first structure and the second structure includes using the machine learning model to calculate the estimated overlay shift based on the single optical image.

5. The method according to claim 1, wherein: Obtaining the one or more optical images includes obtaining a first optical image focused on the first process layer and a second optical image focused on the second process layer; Providing the one or more optical images to the machine learning model includes providing the first optical image and the second optical image to the machine learning model; and Obtaining the estimated overlay shift between the first structure and the second structure includes using the machine learning model to calculate the estimated overlay shift based on the first optical image and the second optical image.

6. The method of claim 1, wherein the machine learning model is a first machine learning model in a set of distinct machine learning models trained to estimate a stacking shift between the first structure and the second structure, the method comprising: providing the one or more optical images to the collection of distinct machine learning models; and A plurality of estimated stacking offsets between the first structure and the second structure are obtained from the set of distinct machine learning models, wherein each of the plurality of estimated stacking offsets is a respective machine learning model from the set of distinct machine learning models. 7 . The method of claim 6 , further comprising determining an overlay offset between the first structure and the second structure based at least in part on the plurality of estimated overlay offsets.

8. The method of claim 6, wherein the set of machine learning models is a plurality of distinct machine learning models of the same type.

9. The method of claim 8, wherein the type is selected from the group consisting of a neural network model, a gradient boosting model, a support vector machine, and a principal component regression model.

10. The method of claim 1, wherein: The machine learning model is a first machine learning model; The one or more optical images further show a third structure in the third process layer; and The method comprises: providing the one or more optical images to the first machine learning model and to a second machine learning model trained to estimate an overlay shift between the second structure and the third structure; and An estimated stacking offset between the second structure and the third structure is obtained from the second machine learning model.

11. The method of claim 1 , wherein the one or more optical images are a first set of one or more optical images in a plurality of sets of one or more optical images of a portion of a semiconductor wafer, each set in the plurality of sets showing the first structure and the second structure, the method further comprising: obtaining the plurality of sets from a plurality of optical inspection tools, wherein each set in the plurality of sets is generated by a corresponding optical inspection tool in the plurality of optical inspection tools; providing the plurality of sets to the machine learning model; and Obtaining, from the machine learning model, a respective estimated stacking offset between the first structure and the second structure for each of the plurality of sets, includes calculating the respective estimated stacking offset using the machine learning model.

12. The method according to claim 11, wherein: said first set of one or more optical images being obtained from a first optical inspection tool of said plurality of optical inspection tools; The method further includes specifying to the machine learning model that the first set is from the first optical inspection tool; and Obtaining the respective estimated overlay offsets between the first structure and the second structure of the first set includes using the machine learning model to calculate the respective estimated overlay offsets based on the first set and further based on the first set being from the first optical inspection tool.

13. The method of claim 11, further comprising, prior to providing the plurality of sets to the machine learning model, training the machine learning model using a training set of optical images from the plurality of inspection tools, wherein: The training set of optical images is different from the plurality of sets of one or more optical images; and The optical images of the training set are annotated with corresponding predetermined overlay offset values.

14. The method of claim 13, wherein the corresponding predetermined overlay offset values ​​of the optical images of the training set are measured using a scanning electron microscope.

15. The method of claim 1, further comprising, prior to providing the one or more optical images to the machine learning model, training the machine learning model using a training set of optical images, wherein: The training set of optical images is different from the one or more optical images; and The optical images of the training set are annotated with corresponding predetermined overlay offset values.

16. The method of claim 15, wherein the corresponding predetermined overlay offset values ​​of the optical images of the training set are measured using a scanning electron microscope.

17. A non-transitory computer-readable storage medium storing one or more programs for execution by one or more processors, the one or more programs comprising instructions for: obtaining one or more optical images of a portion of a semiconductor wafer, the one or more optical images showing a first structure in a first process layer and a second structure in a second process layer; providing the one or more optical images to a machine learning model trained to estimate an overlay shift between the first structure and the second structure; and An estimated stacking offset between the first structure and the second structure is obtained from the machine learning model.

18. A system comprising: one or more optical inspection tools for optically inspecting semiconductor wafers; one or more processors; and a memory storing one or more programs for execution by the one or more processors, the one or more programs comprising instructions for: obtaining one or more optical images of a portion of a semiconductor wafer from the one or more optical inspection tools, the one or more optical images showing a first structure in a first process layer and a second structure in a second process layer; providing the one or more optical images to a machine learning model trained to estimate an overlay shift between the first structure and the second structure; and An estimated stacking offset between the first structure and the second structure is obtained from the machine learning model.

Images