Optical circuit measurement system and method

Abstract

The present disclosure provides an optical waveguide measurement system and a measurement method which have high throughput, and suppress a positional displacement or an increase in the data capacity of a captured image by applying processing trained by machine learning to a low magnification image of a wide imaging field of view. A measurement system according to the present disclosure comprises: an imaging unit which images an object to be imaged; a data set generation unit which implements pattern matching processing and trimming processing for a plurality of images having different magnifications, and generates a training image data set; a machine-learning unit which includes a neural network, is trained, on the basis of the training image data set, to implement ultra-resolution processing in the neural network, and generates a super-resolution-processed image on the basis of the image and a trained network configuration; and a measurement unit which measures, on the basis of the super-resolution-processed image, at least a portion of the object to be measured. In addition, the present disclosure provides an optical waveguide measurement method comprising training processing and execution processing.

Description

[Technical Field]

[0001] The present disclosure relates to a measurement system and a measurement method for optical waveguide devices, and more particularly to a measurement system and a calculation method for measuring at least a portion of an optical circuit formed on a semiconductor or insulator wafer with high accuracy and high throughput. [Background technology]

[0002] In recent years, with the further increase in capacity of wavelength-division multiplexed optical communications, there has been active research and development into optical waveguide devices, such as optical wavelength multiplexing / demultiplexing circuits and optical switch circuits, that support this. In many cases, the components of these optical circuits include multiple optical signal paths with different optical path lengths and multiplexing / demultiplexing elements, which realize wavelength multiplexing / demultiplexing and switching functions by utilizing the interference of light waves (see, for example, Non-Patent Documents 1 and 2).

[0003] The interference characteristics of light waves depend on the optical path length difference between the paths of optical signals. The effective refractive index, which determines the optical path length, is primarily determined by the refractive index of the material that makes up the optical waveguide and the dimensions of the optical waveguide core, such as its height and width. Therefore, the optical characteristics of optical circuits fabricated on a wafer vary depending on the distribution of the optical constants of the material that makes up the optical waveguide within the wafer surface and the dimensional variations in the optical waveguide structure. In particular, the width of the optical waveguide core has a large contribution to the optical characteristics of optical waveguide devices, and the variations in the optical characteristics of fabricated optical waveguide devices depend heavily on the accuracy of the optical waveguide patterning process.

[0004] FIG. 1 shows a conventional method for manufacturing an optical waveguide. A typical manufacturing process is described here using a silica-based planar lightwave circuit made of silica-based glass as an example. First, in the lower clad deposition process (1), a glass film that will become the lower clad 12 is deposited on a silicon substrate (wafer) 11. The lower clad 12 is made of SiO2 doped with P2O5 or B2O3, deposited by flame hydrolysis deposition (FHD). The soot-like glass particles deposited by the FHD process are heated at a high temperature of 1000°C or higher to obtain the transparent lower clad 12. Next, in the core deposition process (2), a thin-film glass that will become the core 13, which has a higher refractive index than the lower clad 12, is deposited using the same FHD process. The desired refractive index can be achieved by adding GeO2 to the SiO2 during the deposition of the core 13. As in the lower clad deposition process (1), the transparent core 13 is formed by heating at a high temperature of 1000°C or higher.

[0005] In step 3: photoresist film formation, a photoresist film 14 is formed on the substrate by spin coating. Next, in step 4: circuit pattern exposure, the photoresist film is irradiated with UV light 16 through a photomask 15 to expose a circuit pattern corresponding to the mask pattern. Then, in step 5: photoresist development, the circuit pattern of the photoresist film is developed to obtain a photoresist pattern 17.

[0006] Next, in step 6: etching, the photoresist pattern 17 is transferred to the core by reactive ion etching (RIE) to obtain a core pattern 18. Furthermore, in step 7: resist removal, the photoresist remaining on the core is removed by ashing. Finally, in step 8: upper cladding deposition, an upper cladding 19 is deposited by the same method as the lower cladding deposition in step 1: lower cladding deposition described above.

[0007] The optical waveguide obtained by the above-mentioned manufacturing process is subjected to various inspections, such as an inspection of the pattern size and optical characteristics of the optical circuit. For the evaluation of optical characteristics, for example, a method is used to evaluate parameters correlated with the amplitude and phase of light in each path with different optical path lengths related to interference. A known technique for measuring this phase information is to input one branch of light from a tunable light source into an optical circuit and the other into an optical signal path with a known optical path length, cause them to interfere on the output side, output them to a reference port, and evaluate the interference pattern (see, for example, Patent Document 1 and Non-Patent Document 3).

[0008] However, when evaluating the loss and phase information of an optical circuit using such an optical circuit evaluation device, there is a problem in that the dimensional accuracy of the optical waveguide cannot be evaluated until the optical waveguide device has been manufactured. Furthermore, it is necessary to construct an apparatus consisting of a circuit to be measured and a path with a known optical path length, which creates the problem that slight changes in the environment (temperature, humidity, etc.) or the adhesion of foreign matter can reduce the measurement accuracy. Furthermore, conventional optical circuit evaluation devices require the evaluation device to be incorporated into the optical circuit device, which increases the circuit area.

[0009] In response to these challenges, the present disclosure aims to provide a technology for evaluating the dimensional accuracy of an optical waveguide during the manufacture of an optical waveguide device, and for measuring the optical characteristics of the optical waveguide device without reducing measurement accuracy due to environmental influences and without increasing the circuit area. As described above, the core width of the optical waveguide in particular contributes significantly to the optical characteristics of the optical waveguide device, and the variation in the optical characteristics of the fabricated optical device depends on the accuracy of the optical waveguide patterning process. Therefore, by measuring the dimensions of the core pattern with high accuracy, it is possible to estimate the optical characteristics of the optical waveguide device and predict the variation in those optical characteristics during the manufacturing process.

[0010] Conventionally, there have been methods for estimating optical characteristics during the manufacture of optical waveguide devices without the need for increasing the path area and without the need for deterioration in accuracy due to environmental influences. A typical example of such a method is to measure the dimensions of optical waveguides and other elements using a microscope during the manufacture of optical waveguide devices.

[0011] However, in order to estimate the optical characteristics of an optical waveguide device with high accuracy, it is necessary to measure the core width of the optical waveguide with high accuracy. Typically, the core width of an optical waveguide that constitutes an optical circuit ranges from 0.1 μm to 100.0 μm, so in order to measure the core width with high accuracy, it is necessary to use a microscope to image the optical waveguide at high magnification (for example, 500 times or more). However, this method requires that all optical waveguides within the wafer surface be imaged using a high-magnification microscope, and that the distribution of core width within the wafer surface be calculated using image analysis, etc. With this method, since the imaging field of a high-magnification microscope image is narrow, (1) Imaging position deviation due to insufficient positioning accuracy of the stage (2) Decrease in throughput due to an increase in the number of images taken (3) Increased storage capacity due to increased number of images taken This is a manufacturing issue. [Prior art documents] [Patent documents]

[0012] [Patent Document 1] Japanese Patent Application Laid-Open No. 2004-126398 [Non-patent literature]

[0013] [Non-Patent Document 1] Inoue, Y.; Kaneko, A.; Hanawa, F.; Takahashi, H.; Hattori, K.; Sumida, S.: 'Athermal silica-based arrayed-waveguide grating multiplexer', Electronics Letters, 1997, 33, (23), p. 1945-1947, DOI: 10.1049 / el:19971317 [Non-patent document 2] T. Miya, "Silica-based planar lightwave circuits: passive and thermally active devices," in IEEE Journal of Selected Topics in Quantum Electronics, vol. 6, no. 1, pp. 38-45, Jan.-Feb. 2000, doi: 10.1109 / 2944.826871 [Non-patent document 3] T. Saida, T. Tanaka, T. Shibata, H. Takanashi, and Y. Hida, “Arrayed-Waveguide Grating With Built-In Interferometer for Measuring Amplitude and Phase Distribution,” IEEE Photon. Technol. Lett. VOL. 17, NO. 8, AUGUST 2005. Summary of the Invention

[0014] The present disclosure has been made in view of the above-mentioned problems, and its purpose is to provide a system and method for measuring optical waveguides that suppresses positional deviations in captured images and increases in data volume, and has high throughput, by applying processing trained by machine learning to images with a wide field of view and low magnification.

[0015] In response to the above-mentioned problems, the present disclosure provides a measurement system for an optical waveguide device, comprising: an imaging unit that captures an image of the appearance of an object to be measured and generates image data of the image; a dataset generation unit that implements pattern matching processing and cropping processing on the image data of each of a plurality of images captured by the imaging unit at different magnifications and generates a training image dataset based on the image data of each of the plurality of images at different magnifications; a machine learning unit that includes a neural network, trains the neural network to implement super-resolution processing based on the training image dataset, generates a trained network setting that describes the conditions of the trained neural network, and generates a super-resolution processed image based on the trained network setting and the image; and a measurement unit that measures at least a portion of the object to be measured based on the super-resolution processed image.

[0016] Furthermore, the present disclosure provides a measurement method for an optical waveguide device, comprising a training step and an execution step, wherein the training step comprises: capturing a first training image representing the appearance of a training measurement object at any coordinates and a second training image having a magnification lower than that of the first training image, and generating image data for each; implementing pattern matching processing and cropping processing on the first training image and the second training image, and generating a training image dataset; and training a neural network using the training image dataset, and generating a trained network setting that describes the conditions of the trained neural network; and the execution step comprises: acquiring a measurement image representing the appearance of the measurement object at any coordinates and having a magnification lower than that of the first training image; generating a super-resolution processed image based on the measurement image using a neural network that reflects the trained network setting; and measuring the length or area of ​​at least a portion of the super-resolution processed image and outputting the measurement results. [Brief explanation of the drawings]

[0017] [Figure 1] 1A to 1C are diagrams illustrating a method for manufacturing an optical waveguide according to the prior art. [Figure 2] FIG. 1 is a block diagram illustrating a measurement system for an optical waveguide device according to the present disclosure. [Figure 3] 1 is a flowchart illustrating a method for measuring an optical waveguide device according to the present disclosure. [Figure 4] 10A and 10B are diagrams illustrating verification results of measurement throughput when super-resolution processing according to the present disclosure is used. [Figure 5] 10A and 10B are diagrams illustrating verification results of measurement accuracy when super-resolution processing according to the present disclosure is used. [Figure 6] 6A and 6B are diagrams schematically illustrating the configuration of an AWG that is an object to be measured in an embodiment of the present disclosure, where FIG. 6A is a top view and FIG. 6B is a cross-sectional view taken along the line VIb-VIb. [Figure 7] FIG. 1 conceptually illustrates training a super-resolution processing network, according to one embodiment of the present disclosure. [Figure 8] 1 is a flowchart illustrating a procedure for training a super-resolution processing network according to one embodiment of the present disclosure. [Figure 9] 1 is a flowchart illustrating a procedure for super-resolution processing using a trained super-resolution processing network and measuring the width of a core in one embodiment of the present disclosure. [Figure 10] 1 is a flowchart illustrating a procedure for performing super-resolution processing using a trained super-resolution processing network and measuring foreign matter in an embodiment of the present disclosure. [Figure 11] FIG. 10 conceptually illustrates measuring the diameter and area of ​​a foreign object according to one embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0018] Various embodiments of the present disclosure will be described in detail below with reference to the drawings. The same or similar reference numerals indicate the same or similar elements, and redundant description may be omitted. The following description is an example, and some configurations may be omitted or modified, or additional configurations may be added, as long as they do not deviate from the gist of one embodiment of the present disclosure.

[0019] 2 is a block diagram showing an optical waveguide device measurement system 20 according to the present disclosure. The optical waveguide device measurement system 20 according to the present disclosure includes a database 21 that transmits and receives data to and from each component of the system (e.g., an imaging unit 22 and a machine learning unit 24, which will be described later) and temporarily stores the received data; the imaging unit 22 that captures an external image of the object to be measured; a dataset generation unit 23 that generates a training image dataset based on the images captured by the imaging unit 22; the machine learning unit 24 that generates a trained network configuration by machine learning based on the training image dataset and generates a super-resolution processed image based on the trained network configuration and image data of the object to be measured (image data of a "measurement image," which will be described later); and a measurement unit 25 that at least partially measures the object to be measured based on the super-resolution processed image. The database 21 is connected to each of the imaging unit 22, the dataset generation unit 23, the machine learning unit 24, and the measurement unit 25 so as to be able to communicate with each other, thereby enabling data transmission and reception.

[0020] Furthermore, the database 21 includes a plurality of ports 211 for inputting and outputting data to be transmitted and received, a memory 212 for temporarily storing data, and a processing unit 213 for executing the processes of transmitting, receiving, and storing data.

[0021] The imaging unit 22 also includes a microscope 221 that acquires an external image of the object to be measured, a camera 222 that captures the external image acquired by the microscope 221 as image data, a memory 223 that temporarily stores the image data, and a port 224 that outputs the image data to the database 21. Note that various types of microscopes can be used as the microscope 221, such as an optical microscope, a laser microscope, an electron microscope, an X-ray microscope, and an ultrasonic microscope.

[0022] The dataset generation unit 23 also includes multiple ports 231 for transmitting and receiving data (image data, training image datasets, etc.) to and from the database 21, a memory 232 for temporarily storing received image data, and a processing unit 233 for generating a training image dataset based on the image data. As will be described later, the processing unit 233 previously includes algorithms for implementing pattern matching processing and trimming processing based on two sets of image data captured at the same coordinates but with different magnifications. In addition, the processing unit 233 of the dataset generation unit 23 may detect the angle of a pattern in the first training image and the second training image and correct the angle of the pattern by referring to a preset axis.

[0023] The machine learning unit 24 also includes multiple ports 241 for transmitting and receiving data (training image datasets and super-resolution processed image data) to and from the database 21, a memory 242 for temporarily storing the received training image dataset, and a super-resolution processing network 243 for performing training based on the received training image dataset. As will be described later, the super-resolution processing network 243 is a neural network with training images captured at low magnification as an input layer and super-resolution processed images used for measurement as an output layer. This super-resolution processing network 243 may be a generative adversarial network or a convolutional neural network.

[0024] The measurement unit 25 also includes multiple ports 251 for sending and receiving data (super-resolution processed image data and measurement result data) to and from the database 21, a memory 252 for temporarily storing the super-resolution processed image data, and a processing unit 253 for implementing measurement of the object to be measured based on the super-resolution processed image data.

[0025] The optical waveguide device measurement system 20 according to the present disclosure, having such a configuration, differs from conventional technology in that it includes a machine learning unit 24 that trains the super-resolution processing network 243. Then, by using the trained super-resolution processing network 243 to process low-magnification images (hereinafter referred to as super-resolution processing), it becomes possible to achieve measurement accuracy that is equal to or higher than that when measuring high-magnification images. A measurement method using the measurement system 20 according to the present disclosure will be described in detail below.

[0026] 3 is a flowchart showing a measurement method 30 for an optical waveguide device according to the present disclosure. The measurement method for an optical waveguide device according to the present disclosure includes a training step 31 for training a super-resolution processing network 243 in the machine learning unit 24, and an execution step 32 for implementing super-resolution processing using the trained super-resolution processing network 243 and performing measurement of an object to be measured based on a super-resolution processed image.

[0027] Furthermore, the training step 31 includes a step 311 in which the imaging unit 22 captures a first training image and a second training image with different magnifications at arbitrary coordinates of the training measurement object, and stores image data of the first training image and the second training image in a training image directory in the memory 212 of the database 21; and a step 312 in which the database 21 transmits the first training image and the second training image to the data set generation unit 23, and the data set generation unit 23 generates a training image data set based on the first training image and the second training image. The method includes a step 312 of storing the training image dataset in a directory of training image datasets in the memory 212 of the database 21, and a step 313 of the database 21 transmitting the training image dataset to the machine learning unit 24, which uses the training image dataset to train the super-resolution processing network 243 in the machine learning unit 24, generating trained network configuration data, and storing the trained network configuration data in a directory of trained network configuration data in the memory 212 of the database 21.

[0028] In step 311, the magnification of the first training image is set to a value comparable to that of the prior art, i.e., an arbitrary magnification suitable for dimensional measurement for evaluating the optical properties of an optical waveguide device. For example, in dimensional measurement of a typical optical waveguide, it is preferable that the standard deviation σ of the measurement be within 1σ. From this perspective, a magnification of 500x is expected if the object to be measured is a silica-based waveguide, and a magnification of 1000x is expected if the object to be measured is a silicon-based waveguide. On the other hand, the magnification of the second training image is set to an arbitrary magnification that is at least lower than that of the first training image. For example, if the magnification of the first training image is M and the magnification of the second training image is N, the ratio (M / N) is expected to be 2 or more and 1000 or less. Furthermore, the range (field of view) of the first training image must be included within the second training image. It is preferable that the microscope 221 of the imaging unit 22 that captures the first training image and the second training image be of the same type (for example, if the first training image is captured with an optical microscope, it is preferable that the second training image is also captured with an optical microscope), but this is not limited to this.

[0029] In step 312, the training image dataset generated is a set of image data in which captured first and second training images are correlated with each other. The training image dataset is generated by implementing pattern matching and cropping processes on the second training images, converting them into images that represent the same coordinates as the first training images, and then correlating them with the first training images. In addition, the processing unit 233 of the dataset generation unit 23 may detect angles of patterns in the first and second training images and correct the angles of the patterns with reference to a preset axis.

[0030] In step 313, the trained network configuration data is data describing the conditions of the super-resolution processing network 243, trained in the form of any neural network. Note that to evaluate the training accuracy of the super-resolution processing network 243, the training in step 313 may be implemented using a portion of the training image dataset as test data.

[0031] On the other hand, the execution process 32 includes a process 321 in which the imaging unit 22 captures a measurement image at a single magnification at any coordinate of the object to be measured and stores image data of the measurement image in a directory for measurement images in the memory 212 of the database 21; a process 322 in which the database 21 transmits the measurement image and a pre-generated trained network setting to the machine learning unit 24, the machine learning unit 24 implements super-resolution processing based on the trained network setting, and stores image data of the generated super-resolution processed image in a directory for super-resolution processed images in the memory 212 of the database 21; and a process 323 in which the database 21 transmits the image data of the super-resolution processed image to the measurement unit 25, the measurement unit 25 measures at least a portion of the super-resolution processed image, and stores the measurement result data in a directory for measurement result data in the memory 212 of the database 21.

[0032] In step 323, the measurement results can be output as lengths (e.g., widths of rectangular patterns or diameters of circular patterns) or areas in the super-resolution processed image. Length measurements can be obtained by measuring the distance between multiple different contrasts in the image, while area measurements can be obtained by measuring multiple lengths in a specific range and integrating the length measurements.

[0033] FIG. 4 illustrates the results of a test of measurement throughput when using super-resolution processing according to the present disclosure. Specifically, the graph shows the relationship between the number of images taken and the number of objects under two conditions: when multiple high-magnification images are taken at arbitrary coordinates of the object under measurement (conventional technology), and when super-resolution processing according to the present disclosure is applied to low-magnification images. Here, as an example, the object under measurement is an arrayed waveguide grating (AWG), which is used in optical wavelength filters, etc. Furthermore, the high-magnification images taken using the conventional technology and the super-resolution processing according to the present disclosure have a magnification four times that of the low-magnification images. When comparing the measurement according to the conventional technology and the measurement according to the present disclosure under these conditions, as shown in FIG. 4, it can be seen that the measurement according to the present disclosure requires fewer images. Quantitatively, the number of images taken in the measurement according to the present disclosure was 1 / 16 of that in the measurement according to the conventional technology. In other words, it can be said that the measurement according to the present disclosure has higher throughput.

[0034] FIG. 5 illustrates the results of verifying measurement accuracy when super-resolution processing according to the present disclosure is used. Specifically, the graph plots measurement error for the same image under two conditions: when super-resolution processing according to the present disclosure is applied and when it is not. As described above, the measurement was performed by measuring the distance (length) between multiple different contrasts within the image. The graph also shows the standard deviation σ of the measurement error when super-resolution processing according to the present disclosure is applied and when it is not. As shown in FIG. 5, it can be seen that the variation in measurement error is smaller when super-resolution processing is applied than when super-resolution processing is not applied. This can be understood from the fact that the standard deviation σ is 0.06 when super-resolution processing according to the present disclosure is not applied, but is 0.03 when super-resolution processing is applied.

[0035] From the above, it can be seen that by applying the optical waveguide device measurement system 20 and measurement method 30 according to the present disclosure, it is possible to achieve high throughput while also achieving highly accurate measurements. As mentioned above, the measurement method 30 according to the present disclosure does not require capturing a large number of high-magnification images, and therefore it is clear that it is possible to suppress the accompanying image misalignment and increase in data volume. Therefore, it is thought that it is also possible to suppress the decrease in measurement accuracy due to misalignment when capturing an image of the object to be measured, and therefore it is possible to achieve measurements with higher accuracy than conventional techniques.

[0036] (First embodiment) A first embodiment of the present disclosure will be described in detail below with reference to the drawings. This embodiment relates to an example in which the core of an optical waveguide in an AWG is used as the object to be measured, and training in the super-resolution processing network 243 is performed by training two different types of networks in an adversarial manner.

[0037] FIG. 6 is a diagram schematically illustrating the configuration of an AWG 60, which is an object to be measured in one embodiment of the present disclosure. FIG. 6(a) illustrates a top view, and FIG. 6(b) illustrates a cross-sectional view taken along the VIb-VIb line. The AWG 60 includes a substrate 61, one or more input waveguides 62, one or more output waveguides 63, an input slab waveguide 64, and an output slab waveguide 65. Furthermore, the multiple waveguides between the input slab waveguide 64 and the output slab waveguide 65 in the AWG 60 each include a core 66, which serves as an optical path for signal light, and a cladding 67 surrounding the core. In the AWG 60 configured in this manner, wavelength-multiplexed light having multiple different wavelengths is input through the input waveguide 62. The signal light is demultiplexed between the input slab waveguide 64 and the output slab waveguide 65, and the demultiplexed light is output from the output waveguide 63. Because of these functions, the AWG 60 is generally put to practical use in optical wavelength filters and the like.

[0038] The central wavelength λ0 is one of the important parameters when evaluating the optical characteristics of such an AWG 60. The central wavelength λ0 is generally determined by Equation 1.

[0039]

number

[0040] where ΔL is the signal path difference between the waveguides, m is the diffraction order, and n array is the equivalent refractive index of the arrayed waveguide and is calculated by (Equation 2).

[0041]

number

[0042] where n core is the refractive index of the core, n clad is the refractive index of the cladding, T is the film thickness of the core, and w is the width of the core of the optical waveguide.

[0043] Therefore, it can be said that if the width w of the core of each optical waveguide is measured with high accuracy, it is possible to evaluate with high accuracy the optical characteristics of the AWG 60. In this embodiment, the width of each core of the AWG 60 is measured using the optical waveguide device measurement system 20 and measurement method 30.

[0044] 7 is a diagram conceptually illustrating training of a super-resolution processing network 243 according to an embodiment of the present disclosure. The super-resolution processing network 243 in this embodiment includes an image generator 71 and a classifier 72. These are the network layers that make up the super-resolution processing network 243, and as shown in FIG. 7, the classifier 72 is located at a higher layer than the image generator 71. In this embodiment, the image generator 71 and the classifier 72 are configured to be trained adversarially in step 313 shown in FIG. 3.

[0045] 8 is a flowchart showing a procedure for training the super-resolution processing network 243 according to one embodiment of the present disclosure. Training the super-resolution processing network 243 according to this embodiment includes step 81 of inputting second training images of a training image dataset generated in advance by the dataset generation unit 23 to the image generator 71, step 82 of inputting first training images of the training image dataset generated in advance by the dataset generation unit 23 to the classifier 72, step 83 of training the image generator 71 to generate super-resolution processed images similar to the first training images based on the second training images, and step 84 of the classifier 72 discriminating between the super-resolution processed images generated by the image generator 71 and the first training images.

[0046] In training the super-resolution processing network 243 according to this embodiment, as configured in this manner, in step 84 shown in FIG. 8 , the classifier 72 is trained so that it can distinguish between the super-resolution processed image and the first training image with high accuracy. The discrimination is implemented based on a difference analysis between the super-resolution processed image and the first training image using a function such as PSNR (Peak Signal to Noise Ratio). This discrimination includes the steps of: calculating the error between the super-resolution processed image and the first training image on a pixel-by-pixel basis using a function such as PSNR; calculating the error gradient (dependence on error) for changeable parameters (weights, biases, etc.) in the super-resolution processing network using backpropagation, gradient descent, or the like, with reference to the calculation results; and changing the parameters to minimize the error. The network, whose parameters have been optimized through these steps, performs super-resolution processing on the first image. The super-resolution processing network 243 trained in this manner then generates a trained network setting and stores it in a directory of trained network settings in the memory 212 of the database 21.

[0047] 9 is a flowchart illustrating a procedure for performing super-resolution processing and measuring the core width using a trained super-resolution processing network 243 in one embodiment of the present disclosure. The procedure for super-resolution processing and measuring the core width in this embodiment includes step 91 of transmitting trained network settings from the database 21 to the machine learning unit 24 and reflecting them in the super-resolution processing network 243; step 92 of transmitting image data of a measurement image, which at least partially includes the core 66 of the AWG 60 and was previously captured by the imaging unit 22, from the database 21 to the image generator 71 of the super-resolution processing network 243; step 93 of the image generator 71 generating a super-resolution processed image and storing it in a directory of super-resolution processed image data in the memory 212 of the database 21; and step 94 of transmitting the super-resolution processed image data from the database 21 to the measurement unit 25 and measuring the width of the core 66 of the AWG 60. As described above, the width of the core 66 is determined by measuring the distance between different contrasts.

[0048] By measuring the width of the core 66 using this procedure, highly accurate measurements can be achieved with fewer images taken than with conventional techniques. This also makes it possible to evaluate the optical characteristics of the AWG 60 with high throughput while suppressing image misalignment and an increase in data volume.

[0049] (Second embodiment) A second embodiment of the present invention will be described in detail below with reference to the drawings. This embodiment relates to an example in which a convolutional neural network is used in the super-resolution processing network 243 to measure the size of foreign particles on a wafer with high accuracy and high throughput.

[0050] In optical waveguide devices, if foreign particles are scattered on a semiconductor wafer and are close to the optical waveguide formed on the semiconductor wafer, the foreign particles will interfere with the signal light propagating through the optical waveguide, resulting in a loss of signal strength. In other words, in the manufacturing process of optical waveguide devices, measuring the size of foreign particles on a semiconductor wafer with high precision in advance makes it possible to estimate the impact of the foreign particles on device characteristics, leading to improved manufacturing yields.

[0051] The optical waveguide device measurement system 20 and measurement method 30 according to the present disclosure can measure the size of such foreign particles on a semiconductor wafer with high accuracy and high throughput. For example, in this embodiment, a convolutional neural network is used in the super-resolution processing network 243 in the measurement system 20 to measure the diameter and area of ​​the foreign particle.

[0052] In training the super-resolution processing network 243 in this embodiment, a second training image containing a foreign object and a first training image at the same coordinates are captured in advance by the imaging unit 22, and a training image dataset based on both sets of image data is generated by the dataset generation unit 23. The second training image is then loaded into the super-resolution processing network 243 as an input layer and the first training image as a filter. Image features are reflected in the convolution layer by comparing the two images, and a super-resolution processed image is generated by pooling. By implementing such training, the super-resolution processing network 243 is optimized based on the correlation between the second training image and the first training image. The trained super-resolution processing network 243 generates trained network settings and stores them in a directory of trained network settings in the database 21, similar to the first embodiment.

[0053] 10 is a flowchart showing a procedure for performing super-resolution processing and measuring foreign matter using a trained super-resolution processing network 243 in one embodiment of the present disclosure. The procedure for super-resolution processing and measuring foreign matter in this embodiment includes step 101 of transmitting trained network settings from the database 21 to the machine learning unit 24, which then causes the machine learning unit 24 to reflect the settings in the super-resolution processing network 243, step 102 of transmitting image data of an image including a foreign matter that was previously captured by the imaging unit 22 from the database 21 to the super-resolution processing network 243, step 103 of the super-resolution processing network 243 generating a super-resolution processed image and storing it in a directory of super-resolution processed image data in the memory 212 of the database 21, and step 104 of transmitting the super-resolution processed image data from the database 21 to the measurement unit 25 and measuring the width and area of ​​the foreign matter. In step 101, the items that the machine learning unit 24 reflects on the super-resolution processing network 243 may be information regarding the configuration of the aforementioned network (e.g., the number of layers in the network, the configuration of each layer, information regarding the type of input and output images, etc.) and parameters (e.g., weights, biases, etc.).

[0054] FIG. 11 is a diagram conceptually illustrating measurement of the diameter and area of ​​a foreign particle according to an embodiment of the present disclosure. By measuring the distance between different contrasts in a portion of the super-resolution processed image that shows the foreign particle, the width of the foreign particle in the X direction can be measured. Furthermore, this measurement is performed by scanning in the Y direction perpendicular to the X direction, and multiple widths of the foreign particle in the X direction (L1, L2, . . . , L n In this embodiment, for example, the widths (L1, L2, . . . , L) of the foreign matter in the X direction are measured. n ) with the maximum width (L MAX ) is the diameter D of the foreign particle. Furthermore, the widths (L1, L2, . . . , L n ) is defined as the area S of the foreign object. From the diameter D and area S of the foreign object thus determined, it is possible to quantitatively estimate the size of the foreign object.

[0055] In FIG. 11, the length of the foreign matter to be measured is in the X direction, but this is not limitative and it is possible to measure the length in any direction on the XY plane. [Industrial Applicability]

[0056] The present invention is expected to be applicable to monitoring and inspection during the manufacture of optical circuits as a measurement system and method for optical waveguide devices with higher accuracy and throughput than conventional measurement systems and methods.

Claims

1. A measurement system for an optical waveguide device, comprising: an imaging unit that captures an image of an external appearance of the measurement object and generates image data of the image; A dataset generation unit that implements a pattern matching process and a trimming process on the image data of each of the plurality of images captured by the imaging unit and having different magnifications, and generates a training image dataset based on the image data of each of the plurality of images having different magnifications; a machine learning unit that includes a neural network, trains the neural network to implement super-resolution processing based on the training image dataset, generates a trained network configuration that describes conditions of the trained neural network, and generates super-resolution processed images based on the trained network configuration and the images; a measurement unit that measures at least a portion of the object to be measured based on the super-resolution processed image; A measurement system for an optical waveguide device comprising:

2. The measurement system for an optical waveguide device according to claim 1 , wherein the data set generation unit detects an angle of a pattern in the image and corrects the angle of the pattern by referring to a preset axis.

3. 3. The measurement system for an optical waveguide device according to claim 1, wherein the neural network of the machine learning section includes two different types of networks, and the two different types of networks are trained in an adversarial manner against each other.

4. 3. The optical waveguide device measurement system according to claim 1, wherein the neural network of the machine learning unit is a convolutional neural network.

5. A method for measuring an optical waveguide device, comprising a training step and an execution step, The training step comprises: A step of capturing a first training image representing an appearance of a training measurement object at an arbitrary coordinate and a second training image having a magnification lower than that of the first training image, and generating image data for each of the first training image and the second training image; implementing a pattern matching operation and a cropping operation on the first training images and the second training images to generate a training image data set; training a neural network using a training image data set to generate a trained network configuration describing conditions for the trained neural network; Equipped with The performing step includes: acquiring a measurement image showing an appearance of an object at an arbitrary coordinate and having a magnification lower than that of the first training image; generating a super-resolution processed image based on the measurement image using the neural network reflecting the trained network settings; measuring a length or an area of ​​at least a portion of the super-resolution processed image and outputting a measurement result; A method for measuring an optical waveguide device comprising the steps of:

6. 6. The method for measuring an optical waveguide device according to claim 5, wherein the step of generating the training image data set further comprises the steps of detecting an angle of a pattern in at least one of the first training image and the second training image, and correcting the angle of the pattern with reference to a preset axis.

7. 7. The method for measuring an optical waveguide device according to claim 5, wherein in the step of generating the super-resolution processed image, the training of the neural network comprises training two different types of networks in an adversarial manner against each other.

8. 7. The method for measuring an optical waveguide device according to claim 5, wherein in the step of generating the super-resolution processed image, the training of the neural network is training using a convolutional neural network.

Images