Method of inspecting a wafer

Abstract

Provided is a wafer inspection method capable of simplifying inspection of a wafer subjected to laser processing. The wafer inspection method includes: a first image acquisition step of capturing the front surface side of a wafer for image acquisition before laser processing is performed, and acquiring a plurality of first images; a second image acquisition step of capturing a region corresponding to the first images in the front surface side of a wafer for image acquisition on which a modified layer is formed, and acquiring a plurality of second images; a processing mark image acquisition step of acquiring a processing mark image corresponding to the difference between the first images and the second images; a marking step of marking the second images with processing marks with reference to the second images and the processing mark image; and a learning step of configuring a machine learning unit that determines the presence or absence of a processing mark by machine learning using the marked second images.

Description

Technical Field

[0001] This invention relates to a method for inspecting wafers on which a modified layer has been formed by irradiation with a laser beam. Background Technology

[0002] In the manufacturing process of device chips, a wafer in which devices are formed in multiple regions divided by a grid of predetermined dicing lines (spacers) arranged in a grid pattern is used. By dicing the wafer along the predetermined dicing lines, multiple device chips, each containing a device, can be obtained. Device chips are used in various electronic devices such as mobile phones and personal computers.

[0003] In wafer dicing, a cutting device is used, which cuts the wafer using a ring-shaped cutting tool. On the other hand, in recent years, there has been interest in techniques for dicing wafers using laser processing. For example, a method has been put into practical use where a modified region (modified layer) is formed inside the wafer along a predetermined dicing line by irradiating a wafer with a laser beam that is transparent to the wafer (see Patent Documents 1 and 2). The region where the modified layer is formed on the wafer becomes more brittle than other regions. Therefore, when an external force is applied to the wafer with the modified layer formed, the wafer breaks from the modified layer, thereby being diced along the predetermined dicing line.

[0004] Patent Document 1: Japanese Patent Application Publication No. 2002-192370

[0005] Patent Document 2: Japanese Patent Application Publication No. 2005-184032

[0006] In the case of dividing a wafer using laser processing as described above, a laser beam is irradiated along intersecting predetermined dividing lines. Furthermore, depending on the wafer's thickness and material, multiple layers of modifiers are sometimes formed along each predetermined dividing line in the wafer's thickness direction. Therefore, when a laser beam is irradiated toward the wafer, the laser beam sometimes irradiates the modifier layers already formed on the wafer, resulting in laser beam scattering (diffuse reflection), a phenomenon known as sputtering.

[0007] When laser beam scattering occurs, there is a possibility that areas on the front side of the wafer that deviate from the predetermined dicing line may be accidentally processed, damaging the device. Therefore, after the modifier layer is formed on the wafer, the front side of the wafer is photographed to check whether there are any processing marks (laser marks) caused by laser beam scattering remaining on the front side of the wafer. Moreover, after confirming that there are no processing marks in areas that are significantly deviated from the modifier layer, wafer processing continues.

[0008] However, sometimes foreign matter such as dust adheres to the front side of the wafer even before laser processing. Therefore, even if the front side of the wafer is photographed after laser processing and the resulting image contains dot-like images, it is sometimes difficult to determine whether the image represents foreign matter that was originally attached to the wafer or processing marks formed by the scattering of the laser beam.

[0009] Therefore, the following operation is performed: an image of the front side of the wafer is obtained before laser processing, and a processing mark image is generated that represents the difference between the wafer images before and after laser processing. This processing mark image does not show patterns that existed on the wafer before laser processing; only patterns newly formed by laser processing are shown. Therefore, by observing the processing mark image, it is easy to determine whether processing marks formed due to laser beam scattering are present.

[0010] However, using the above method requires the following steps for each wafer inspection: taking multiple images of the wafer's entire front side before laser processing (pre-processing image); taking multiple images of the wafer after laser processing to obtain multiple images corresponding to the pre-processing image (post-processing image); and performing image processing on the pre-processing and post-processing images to generate a processing mark image. Therefore, wafer inspection is time-consuming and labor-intensive, and may contribute to reduced wafer processing efficiency. Summary of the Invention

[0011] The present invention was made in view of the above-mentioned problems, and its object is to provide a method for inspecting wafers that simplifies the inspection of wafers that have undergone laser processing.

[0012] According to one aspect of the present invention, a method for inspecting a wafer is provided to check for processing marks caused by laser beam scattering in a laser-processed wafer. The wafer has devices formed on its front side in multiple regions divided by predetermined dividing lines. The laser processing involves irradiating the wafer from its back side with a laser beam of a wavelength transmissible to the wafer focused at a region inside the wafer corresponding to the predetermined dividing lines to form a modified layer. The wafer inspection method includes the following steps: a first image acquisition step, which takes a picture of the front side of the wafer before laser processing to obtain multiple first images; and a second image acquisition step, which acquires an image of the wafer with the modified layer formed. The process involves: a first image acquisition step, which captures an image of the area corresponding to the first image on the front side of the wafer; a second image acquisition step, which acquires an image of the processing marks corresponding to the difference between the first image and the second image; a marking step, which marks the second image based on the processing marks by referring to the second image and the processing mark image; a learning step, which constructs a machine learning unit to determine whether the processing marks are present by using machine learning on the marked second image; an inspection image acquisition step, which captures an image of the front side of the wafer on which the modified layer is formed; and a determination step, which inputs the inspection image into the machine learning unit, determines whether the processing marks are present in the inspection image, and outputs the determination result.

[0013] Furthermore, the machine learning unit preferably has a neural network comprising an input layer, an output layer, and a hidden layer. In the learning process, the neural network performs deep learning using multiple second images input to the input layer. In the determination process, the neural network performs calculations based on the image to be examined input to the input layer and outputs the determination result from the output layer. Additionally, the neural network preferably performs the determination through semantic segmentation.

[0014] In one aspect of the wafer inspection method of the present invention, a machine learning unit is configured to determine the presence or absence of processing marks by using machine learning to capture an image (second image) obtained by photographing the front side of a wafer for image acquisition on which a modified layer is formed. Furthermore, an image (inspection image) obtained by photographing the front side of a wafer to be inspected that has undergone laser processing is input into the machine learning unit, thereby determining whether or not processing marks are present in the inspection image.

[0015] When using the aforementioned wafer inspection method, after the machine learning department completes its learning, it can accurately determine the presence or absence of machining marks without obtaining an image of the wafer before laser processing. This eliminates the need for multiple images of the wafer before laser processing each time an inspection is performed, making wafer inspection extremely simple. Attached Figure Description

[0016] Figure 1 This is a three-dimensional view of the chip.

[0017] Figure 2 (A) is a perspective view showing the wafer in the first image acquisition process. Figure 2 (B) is an image diagram showing the first image obtained in the first image acquisition process.

[0018] Figure 3 This is a three-dimensional view showing a laser processing device.

[0019] Figure 4 It is a magnified cross-sectional view showing a portion of a wafer with a modified layer formed on it.

[0020] Figure 5 (A) is a perspective view showing the wafer in the second image acquisition process. Figure 5 (B) is an image diagram showing the second image obtained in the second image acquisition process.

[0021] Figure 6 It is an image diagram showing the machining mark image obtained in the machining mark image acquisition process.

[0022] Figure 7 This is a block diagram showing the arithmetic unit in the learning process.

[0023] Figure 8 (A) is a perspective view showing the wafer in the image acquisition process of the inspected image. Figure 8 (B) is an image diagram showing the image to be inspected acquired in the image acquisition process.

[0024] Figure 9 This is a block diagram showing the calculation unit in the determination process.

[0025] Label Explanation

[0026] 11: Wafer; 11a: Front (side 1); 11b: Back (side 2); 13: Partition line (spacer); 15: Device; 17: Electrode pad; 19: Circuit; 21: Strip; 23: Frame; 23a: Opening; 25: Modified layer (altered layer); 27: Crack; 29: Processing mark (laser mark); 31: Wafer; 31a: Front (side 1); 31b: Back (side 2); 33: Partition line (spacer); 35: Device; 37: Crack; 39: Electrode pad; 41: Circuit; 43: Processing mark (laser mark); 2: Photograph 1. Unit (image capturing component); 2. Control unit (control unit); 3. Calculation unit; 4. Storage unit; 5. Image (first image, image before processing); 6. Laser processing apparatus; 7. Chuck table (holding table); 8. Laser irradiation unit; 9. Laser beam; 10. Control unit (control unit); 11. Image (second image, image after processing); 12. Image (processing mark image, difference image); 13. Image (image under inspection); 14. Machine learning unit; 15. Neural network; 16. Input layer; 17. Output layer; 18. Hidden layer (intermediate layer). Detailed Implementation

[0027] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. First, an example of the structure of a wafer that can be used in the wafer inspection method of this embodiment will be described. Figure 1 This is a perspective view of chip 11.

[0028] The wafer 11 is, for example, a disk-shaped semiconductor wafer formed of silicon or the like, having a front side (first side) 11a and a back side (second side) 11b that are substantially parallel to each other. The wafer 11 is divided into multiple rectangular regions by multiple predetermined dividing lines (spacers) 13 arranged in a grid pattern in an intersecting manner. Furthermore, devices 15 such as ICs (Integrated Circuits), LSIs (Large Scale Integrations), LEDs (Light Emitting Diodes), and MEMSs (Micro Electro Mechanical Systems) are formed on the front side 11a of the regions divided by the predetermined dividing lines 13.

[0029] Furthermore, there are no restrictions on the material, shape, structure, or size of the wafer 11. For example, the wafer 11 can be any shape and size wafer formed from semiconductors other than silicon (GaAs, InP, GaN, SiC, etc.), sapphire, glass, ceramics, resin, metal, etc. Additionally, there are no restrictions on the type, quantity, shape, structure, size, or arrangement of the devices 15.

[0030] By processing the wafer 11 and dividing it along the predetermined dividing line 13, multiple device chips, each having a device 15, can be obtained. For example, the wafer 11 can be divided into multiple device chips by laser processing as described later.

[0031] In the wafer inspection method of this embodiment, firstly, the front side 11a of the wafer 11 before laser processing is performed is photographed to obtain multiple images (first image, image before processing) (first image acquisition process). Figure 2 (A) is a perspective view showing the wafer 11 in the first image acquisition process.

[0032] In the first image acquisition process, the wafer 11 is used as the wafer for image acquisition, and the front side 11a of the wafer 11 is imaged using the imaging unit (imaging member) 2. For example, the imaging unit 2 is a camera that includes imaging elements such as CCD (Charged-Coupled Devices), CMOS (Complementary Metal-Oxide-Semiconductor), and optical elements such as lenses, and the imaging unit 2 is disposed above the front side 11a of the wafer 11.

[0033] The shooting unit 2 is connected to a control unit (control unit) 4 that controls the shooting unit 2. For example, the control unit 4 is composed of a computer that includes a processor such as a CPU (Central Processing Unit) and various memories that function as main storage devices, auxiliary storage devices, etc. Furthermore, in Figure 2 In (A), the functional structure of the control unit 4 is represented by blocks. Specifically, the control unit 4 includes an arithmetic unit 6 that performs calculations and a storage unit 8 that stores various information (data, programs, etc.) used for the calculations of the arithmetic unit 6.

[0034] For example, the wafer 11 is held by a chuck stage (not shown) with its front side 11a facing upwards. Furthermore, the imaging unit 2 takes multiple pictures of the front side 11a of the wafer 11 to obtain multiple partial images of the front side 11a of the wafer 11.

[0035] Figure 2 Image (B) is an image diagram showing image 10 (first image, pre-processing image) obtained in the first image acquisition process. Image 10 shows the dicing predetermined line 13 and device 15, etc., arranged on the front side 11a of the wafer 11. In addition, the device 15 includes structural elements such as electrode pads 17 and circuits 19, and sometimes a portion of them is shown in image 10.

[0036] In the first image acquisition process, while changing the positional relationship between the wafer 11 and the imaging unit 2, the imaging unit 2 is used to take multiple pictures of the wafer 11, thereby obtaining multiple different images 10. For example, a ball screw type moving mechanism (not shown) that moves the chuck table in the horizontal direction and a rotation drive source (not shown) such as an electric motor that rotates the chuck table about a rotation axis that is approximately parallel to the vertical direction are connected to the chuck table.

[0037] Then, while changing the position of the chuck stage via a moving mechanism or changing the rotation angle of the chuck stage via a rotation drive source, the imaging unit 2 takes multiple pictures of the front side 11a of the chip 11. Thus, multiple (e.g., more than 500) images 10 are obtained using a single chip 11.

[0038] Multiple images 10 acquired by the imaging unit 2 are output to the control unit 4 and stored in the storage unit 8. In addition, the storage unit 8 stores positional information (such as the coordinates of the chuck stage, rotation angle, etc.) showing the positional relationship between the wafer 11 and the imaging unit 2 when the wafer 11 is photographed by the imaging unit 2.

[0039] Next, laser processing (laser processing step) is performed on the wafer 11. In the laser processing of the wafer 11, a laser processing apparatus is used to process the wafer 11 by irradiation with a laser beam. Figure 3 This is a perspective view showing the laser processing apparatus 12.

[0040] When the wafer 11 is processed using the laser processing apparatus 12, the wafer 11 is supported by an annular frame 23 to facilitate the handling (holding, transporting, etc.) of the wafer 11. Specifically, a circular strip 21 with a diameter larger than that of the wafer 11 is attached to the front side 11a (device 15 side) of the wafer 11. As a result, the device 15 is covered and protected by the strip 21.

[0041] As shown in section 21, a sheet material is used, which has a substrate formed into a circular film and an adhesive layer (paste layer) disposed on the substrate. For example, the substrate is formed of a resin such as polyolefin, polyvinyl chloride, or polyethylene terephthalate, and the adhesive layer is formed of an epoxy-based, acrylic-based, or rubber-based adhesive. Alternatively, the adhesive layer may be made of a UV-curable resin that is cured by ultraviolet light.

[0042] The outer periphery of the belt 21 is attached to a ring-shaped frame 23 made of metal or the like. A circular opening 23a, larger in diameter than the wafer 11, is provided in the center of the frame 23, and the wafer 11 is attached to the center of the belt 21, positioned inside the opening 23a. Thus, the wafer 11 is supported by the frame 23 via the belt 21. Furthermore, the wafer 11, supported by the frame 23, is transported to the laser processing apparatus 12.

[0043] The laser processing apparatus 12 includes a chuck stage (holding stage) 14 for holding a wafer 11 and a laser irradiation unit 16 for irradiating a laser beam 18. A ball screw-type moving mechanism (not shown) and a rotary drive source (not shown) such as an electric motor are connected to the chuck stage 14. The moving mechanism moves the chuck stage 14 along the X-axis (machining feed direction, first horizontal direction) and the Y-axis (indexing feed direction, second horizontal direction). The rotary drive source rotates the chuck stage 14 about a rotation axis that is substantially parallel to the Z-axis (vertical direction, up-down direction).

[0044] The upper surface of the chuck stage 14 forms a holding surface for holding the wafer 11. The holding surface of the chuck stage 14 is a flat surface approximately parallel to the X and Y axes, and may be circular, for example, corresponding to the shape of the wafer 11. However, the shape and size of the holding surface of the chuck stage 14 can be appropriately varied depending on the wafer 11. Furthermore, the holding surface of the chuck stage 14 is connected to a suction source (not shown) such as an ejector via flow paths (not shown), valves (not shown), etc., formed inside the chuck stage 14.

[0045] A laser irradiation unit 16 is provided above the chuck stage 14 to irradiate the wafer 11 held by the chuck stage 14 with a laser beam 18. The laser irradiation unit 16 has a laser oscillator such as a YAG laser, YVO4 laser, or YLF laser that pulses to emit laser light, and a focusing lens that focuses the laser light pulsed from the laser oscillator.

[0046] The irradiation conditions of the laser beam 18 are set so that the area of ​​the wafer 11 irradiated by the laser beam 18 is modified (degraded) through multiphoton absorption. Specifically, the wavelength of the laser beam 18 is set so that the laser beam 18 is transmissible relative to the wafer 11. Therefore, at least a portion of the laser beam 18 that is transmissible to the wafer 11 is irradiated from the laser irradiation unit 16 onto the wafer 11. In addition, other irradiation conditions of the laser beam 18 (output, pulse width, spot diameter, repetition frequency, etc.) are also appropriately set so that the wafer 11 is appropriately modified.

[0047] Furthermore, the laser processing apparatus 12 has a control unit (control unit) 20 connected to each structural element of the laser processing apparatus 12 (chuck table 14, laser irradiation unit 16, etc.). For example, the control unit 20 is configured as a computer including a processor and memory, and generates control signals to control the operation of the structural elements of the laser processing apparatus 12. Thus, the operation of the laser processing apparatus 12 is controlled.

[0048] in addition, Figure 2The imaging unit 2 shown can also be installed in the laser processing apparatus 12. Furthermore, when the imaging unit 2 is used to image the wafer 11, the wafer 11 can also be held by the chuck stage 14. That is, the first image acquisition process described above can also be performed using the laser processing apparatus 12. In this case, the control unit 4 (see reference 12) connected to the imaging unit 2 can also be used. Figure 2 The control unit 20 is assembled in the laser processing device 12, and the functions of the control unit 4 are realized through the control unit 20.

[0049] Along the predetermined dividing line 13 (see reference) by the laser processing device 12 Figure 1 The wafer 11 is then processed. Specifically, firstly, the wafer 11 is held using a chuck stage 14. For example, the wafer 11 is positioned on the chuck stage 14 with its front side 11a (band 21 side) facing the holding surface of the chuck stage 14 and its back side 11b side exposed upwards. In this state, when a negative pressure from a suction source is applied to the holding surface, the front side 11a of the wafer 11 is attracted and held by the chuck stage 14 through the band 21.

[0050] Next, the chuck table 14 is rotated so that a predetermined dividing line 13 (refer to...) is drawn. Figure 1 The length direction of the laser beam 18 is aligned with the X-axis direction. Furthermore, the position of the chuck stage 14 in the Y-axis direction is adjusted so that the focusing point of the laser beam 18 is positioned on the extension of a predetermined dividing line 13. Furthermore, the height of the focusing point of the laser beam 18 is aligned with the height of the interior of the wafer 11 (below the back surface 11b and above the front surface 11a).

[0051] Then, while irradiating the laser beam 18 from the laser irradiation unit 16, the chuck stage 14 is moved along the X-axis (processing feed), causing the chuck stage 14 and the laser irradiation unit 16 to move relative to each other along the X-axis. Thus, with the focal point of the laser beam 18 positioned in the region inside the wafer 11 corresponding to the dicing predetermined line 13, the laser beam 18 is irradiated from the back side 11b of the wafer 11 along the dicing predetermined line 13.

[0052] The region inside wafer 11 that was irradiated by laser beam 18 is modified (remodeled) through multiphoton absorption. As a result, a modified layer (remodeled layer) 25 is formed inside wafer 11 along the predetermined dividing line 13.

[0053] Then, the same steps are repeated to form the modified layer 25 along the other predetermined dividing lines 13. Moreover, when the modified layer 25 is formed along all the predetermined dividing lines 13, a wafer 11 with the modified layer 25 formed in a lattice pattern can be obtained. Alternatively, depending on the thickness and material of the wafer 11, two or more modified layers 25 can be formed in the thickness direction of the wafer 11.

[0054] Figure 4 This is a magnified cross-sectional view showing a portion of the wafer 11 to which the modified layer 25 is formed. When the modified layer 25 is formed by irradiation with a laser beam 18, cracks 27 are generated in the modified layer 25. These cracks 27 tend to propagate toward a side (front side 11a) opposite to the side on which the laser beam 18 is incident (back side 11b).

[0055] The regions of wafer 11 with the modified layer 25 or cracks 27 become more brittle than other regions of wafer 11. Therefore, when an external force is applied to wafer 11, wafer 11 is divided starting from the modified layer 25 or cracks 27. That is, the modified layer 25 and cracks 27 function as the starting point (the beginning of the division) when dividing wafer 11.

[0056] When the modified layer 25 is appropriately formed by irradiation with laser beam 18, such as Figure 4 As shown, cracks 27 generated in the modified layer 25 propagate and reach the front side 11a of the wafer 11. As a result, a lattice-like pattern of cracks 27 corresponding to the modified layer 25 appears on the front side 11a of the wafer 11 (see reference). Figure 5 (A)). Moreover, after the laser processing is completed, the strip 21 is peeled off and removed from the front side 11a of the wafer 11.

[0057] Next, the area corresponding to image 10 on the front side 11a of the wafer 11 on which the modified layer 25 is formed is photographed to obtain multiple images (second image, processed image) (second image acquisition process). Figure 5 (A) is a perspective view showing the wafer 11 in the second image acquisition process.

[0058] For example, the wafer 11 is held by a chuck stage (not shown) with its front side 11a facing upwards. Furthermore, the front side 11a of the wafer 11 is photographed multiple times using the imaging unit 2, following the same steps as in the first image acquisition process.

[0059] Furthermore, in the second image acquisition process, the imaging unit 2 is used to image the front side 11a of the chip 11 with the image 10 (see reference). Figure 2 The area corresponding to (B) (the area shown in image 10) is captured. Specifically, the control unit 4 adjusts the position of the wafer 11 or the imaging unit 2 according to the position information stored in the storage unit 8, so as to realize the positional relationship between the wafer 11 and the imaging unit 2 when the wafer 11 is captured by the imaging unit 2 in the first image acquisition process. In this state, the front side 11a of the wafer 11 is captured by the imaging unit 2, thereby obtaining an image showing the area in the wafer 11 after laser processing that corresponds to image 10.

[0060] Figure 5 (B) is an image diagram showing image 22 (second image, processed image) obtained in the second image acquisition process. (Refer to image 10) Figure 2 Similar to (B), image 22 includes a pre-segmentation line 13 and a device 15, etc. In addition, image 22 shows a crack 27 formed along the pre-segmentation line 13 (modified layer 25) and reaching the front side 11a of the wafer 11.

[0061] Furthermore, image 22 sometimes includes processing marks (laser marks) 29 formed on wafer 11 during the laser processing described above. These processing marks 29 are formed on the front side 11a of wafer 11 due to the laser beam 18 (see reference 18). Figure 3 It was formed by the scattering of light by the unintentional processing.

[0062] Specifically, during laser processing, the laser beam 18 sometimes irradiates the modified layer 25 or cracks 27 already formed on the wafer 11. For example, when a modified layer 25 is formed along a predetermined dividing line 13, and the laser beam 18 is irradiated along other predetermined dividing lines 13 that intersect the predetermined dividing line 13, the laser beam 18 irradiates the modified layer 25 or cracks 27 in the intersection region of the predetermined dividing lines 13. Furthermore, when multiple modified layers 25 are formed in the thickness direction of the wafer 11, the laser beam 18 sometimes irradiates the modified layer 25 or cracks 27 already formed on the wafer 11 during the formation of the second or subsequent modified layers 25.

[0063] When the laser beam 18 irradiates the modified layer 25 or the crack 27, scattering (diffuse reflection) of the laser beam 18, known as sputtering, is generated inside the wafer 11. Moreover, when the scattered laser beam 18 reaches the front side 11a of the wafer 11, the area on the front side 11a of the wafer 11 that was irradiated by the laser beam 18 is accidentally processed, forming a processing mark 29.

[0064] Furthermore, the number and location of the processing marks 29 vary depending on the degree of scattering of the laser beam 18. For example, when the scattering of the laser beam is low, the number of processing marks 29 is small, and the processing marks 29 tend to form closer to the modified layer 25 (such as inside the predetermined dividing line 13). On the other hand, when the scattering of the laser beam is high, sometimes multiple processing marks 29 are formed in areas significantly deviated from the modified layer 25 (such as outside the predetermined dividing line 13). In this case, the scattered laser beam 18 may irradiate the device 15 and damage it.

[0065] Thus, the processing marks 29 contained in image 22 become useful information for evaluating whether laser processing was performed appropriately. Then, the multiple images 22 acquired by the imaging unit 2 are output to the control unit 4 and stored in the storage unit 8.

[0066] As described above, when the first image acquisition process and the second image acquisition process are performed, multiple sets of images 10 showing the same area on the front side 11a of the wafer 11 before and after laser processing can be obtained (see reference). Figure 2 (B) and Image 22 (refer to) Figure 5 (B)) of (pair).

[0067] Next, an image corresponding to the difference between image 10 and image 22 (processing mark image, difference image) is obtained (processing mark image acquisition process). In the processing mark image acquisition process, firstly, the arithmetic unit 6 of the control unit 4 accesses the storage unit 8 and reads the set of images 10 and 22 stored in the storage unit 8. Then, the arithmetic unit 6 subtracts image 10 from image 22, thereby generating an image corresponding to the difference between image 10 and image 22. Then, the same image processing is performed on all sets of images 10 and 22.

[0068] Figure 6 This is an image diagram showing the image (processing mark image, difference image) 24 obtained in the process of obtaining the processing mark image. Image 24 is an image showing the difference between image 10 of the wafer 11 before laser processing and image 22 of the wafer 11 after laser processing.

[0069] Here, sometimes dust or other foreign matter adheres to the front side 11a of the wafer 11 before laser processing is performed. In this case, in the image 22 of the wafer 11 after laser processing (see reference) Figure 5 In (B), in addition to the image corresponding to the machining mark 29, there is sometimes an image corresponding to the foreign object. Moreover, when the two are similar in size and shape, it is difficult to distinguish between the foreign object and the machining mark 29 even when referring only to the image 22, and it is sometimes difficult to evaluate the machining mark 29.

[0070] On the other hand, in image 24, the image shared by images 10 and 22 is not displayed; only the elements newly formed on the wafer 11 by laser processing (cracks 27, processing marks 29, etc.) are displayed. Therefore, by referring to image 24, foreign objects and processing marks 29 can be easily distinguished, and the number and location of processing marks 29 contained in image 22 can be accurately determined.

[0071] Next, refer to image 22 (reference). Figure 5 (B) and Image 24 (refer to) Figure 6 The multiple images 22 are marked (classified) based on the processing marks 29 (marking process). For example, in the marking process, the images 22 are classified according to the presence or absence of processing marks 29.

[0072] Specifically, by referring to image 22 (reference) Figure 5 Image 24 corresponding to (B) (refer to) Figure 6 The image corresponding to the machining mark 29 is identified from the images contained in image 22. Then, based on whether image 22 contains an image corresponding to the machining mark 29, image 22 is classified into images with machining marks and images without machining marks.

[0073] The above-described classification operation based on the presence or absence of processing marks 29 is performed on multiple images 22. Alternatively, the classification can be performed by a person (operator) through visual observation of images 22 and 24, or it can be performed automatically by using image processing of images 22 and 24.

[0074] Next, a machine learning unit (learning process) is constructed to determine whether there are processing marks 29 by using machine learning on the labeled image 22. Figure 7 This is a block diagram showing the arithmetic unit 6 in the learning process.

[0075] The computation unit 6 includes a machine learning unit 30, which performs machine learning using the labeled image 22 as learning data. The machine learning unit 30 is configured to, when inputting the image 22, output a determination result regarding whether the image 22 contains a processing mark 29. Hereinafter, as an example, the machine learning unit 30 will be described with a neural network 32. However, the structure of the machine learning unit 30 is not limited as long as it enables machine learning.

[0076] For example, neural network 32 is a hierarchical neural network, comprising an input layer 34 for input data, an output layer 36 for output data, and multiple hidden layers (intermediate layers) 38 disposed between the input layer 34 and the output layer 36. The input layer 34, output layer 36, and hidden layers 38 each contain multiple nodes (units, neurons). Nodes in the input layer 34 are connected to nodes in the first hidden layer 38, and nodes in the output layer 36 are connected to nodes in the final hidden layer 38. Furthermore, nodes in the hidden layers 38 are connected to nodes in the input layer 34 or the previous hidden layer 38, and to nodes in the output layer 36 or the next hidden layer 38.

[0077] The number of nodes in the input layer 34, output layer 36, and hidden layer 38, as well as the activation function of each node, can be freely set. Furthermore, there is no limit to the number of hidden layers 38. In particular, a neural network 32 containing two or more hidden layers 38 can be called a deep neural network (DNN). Furthermore, the learning process of a deep neural network can be called deep learning.

[0078] In the learning process, multiple labeled images 22 are used as learning data for deep learning of the neural network 32. Specifically, teacher-assisted learning is performed, using images 22 as learning images (learning data) and the presence or absence of processing marks 29 (the result of labeling) as positive solution labels (teacher data). For example, backpropagation can be used as the learning algorithm. However, there are no restrictions on the learning method of the neural network 32.

[0079] Through learning by neural network 32, the weights and biases of the nodes in input layer 34, output layer 36, and hidden layer 38 are updated. Thus, neural network 32 is configured such that when image 22 is input to input layer 34, it outputs a result from output layer 36 that determines whether there are processing marks 29 in image 22.

[0080] In the learning process described above, it can be used in the second image acquisition process (see reference). Figure 5 The machine learning unit 30 learns from multiple images 22 obtained from (A). Thus, a machine learning unit 30 capable of classifying images 22 with high accuracy can be constructed. Furthermore, after learning is complete, the values ​​of various parameters (weights, biases, etc.) applied to the input layer 34, output layer 36, and hidden layer 38 are stored in the storage unit 8 (see reference 8). Figure 2 wait).

[0081] Furthermore, as described above, the machine learning unit 30 determines the image 22 based on whether it contains processing marks 29. Therefore, it is preferable that the machine learning unit 30 (neural network 32) performs the determination by semantic segmentation, which classifies the pixels contained in the image 22 individually. For example, a semantic segmentation network (SegNet) with an encoder / decoder structure can be used as the model for the neural network 32.

[0082] The semantic segmentation network comprises: an encoder network containing convolutional layers and pooling layers (max pooling layers); and a decoder network containing convolutional layers and upsampling layers (unpooling layers). Furthermore, the location information of the max pooling in the encoder is recorded and used for upsampling in the decoder. The layers of the semantic segmentation network are installed as hidden layers 38 of a neural network 32. Additionally, the softmax function is used as the activation function for the output layer 36.

[0083] The semantic segmentation network can be learned using teacher-guided learning of image 22. Specifically, firstly, image 24 (referring to image 24) acquired during the process of obtaining the processing mark image is used... Figure 6The image 24 is binarized, classifying each pixel into pixels representing crack 27 or processing mark 29 and other pixels. Then, using image 22 as input data and the binarized image 24 as the correct label (teacher data), the semantic segmentation network is trained. The result constitutes the machine learning unit 30, which infers from the image of the processed wafer 11 and outputs a two-dimensional mapping map that extracts the crack 27 and processing mark 29 contained in the image.

[0084] Next, the wafer to be inspected (wafer to be inspected) with the modified layer formed is photographed to obtain an image of the object to be inspected (image to be inspected) (image to be inspected acquisition process). Figure 8 (A) is a perspective view showing the wafer 31 in the process of acquiring the image being inspected.

[0085] The wafer 31 is the wafer that is inspected in the image acquisition process, such as a product wafer used in the manufacture of an actual device chip. Furthermore, the wafer 31 has the same features as the image acquisition wafer 11 used in the first image acquisition process (see reference 11). Figure 2 The same structure as (A).

[0086] Specifically, wafer 31 has a front side (first side) 31a and a back side (second side) 31b that are substantially parallel to each other. Wafer 31 is divided into multiple rectangular regions by a plurality of predetermined dividing lines (spacers) 33 arranged in a grid pattern in an intersecting manner. Furthermore, devices 35 are formed on the front side 31a of each region divided by the predetermined dividing lines 33. The material of wafer 31 is the same as that of wafer 11. The types of devices 35 formed on wafer 31 are also the same as those formed on device 15 on wafer 11.

[0087] Wafer 31 is processed by laser processing apparatus 12 in the same manner as wafer 11 (see reference). Figure 3 As a result, a modified layer is formed inside wafer 31 along the predetermined dividing line 33 (see reference). Figure 4 The modified layer 25). In addition, cracks 37 generated in the modified layer develop and reach the front side 31a of the wafer 31. As a result, a lattice-shaped pattern of cracks 37 corresponding to the modified layer appears on the front side 31a of the wafer 31.

[0088] The wafer 31 with the modified layer formed is held by a chuck stage (not shown), for example, with its front side 31a exposed laterally upwards. Then, it is processed through a process similar to the first image acquisition step (see reference). Figure 2 Following the same steps as (A), the imaging unit 2 is used to image the front side 31a of the wafer 31. The result is a partial image (inspected image) of the front side 31a of the processed wafer 31.

[0089] Figure 8 Image (B) is an image diagram showing the image (inspected image) 26 acquired during the image acquisition process. Image 26 shows the pre-defined dividing lines 33 and devices 35 disposed on the front side 31a of the wafer 31. Furthermore, the devices 35 include structural elements such as electrode pads 39 and circuits 41, and sometimes portions of these are shown in image 26. Moreover, the image 26 acquired by the imaging unit 2 is output to the control unit 4 and stored in the storage unit 8.

[0090] Additionally, image 26 shows a crack 37 formed along the predetermined dividing line 33 (modified layer) and reaching the front side 31a of the wafer 31. Furthermore, image 26 sometimes includes machining marks (laser marks) 43 formed on the front side 31a of the wafer 31. The principle of forming machining marks 43 is similar to that of machining mark 29 (see reference). Figure 5 (B) is the same.

[0091] Next, image 26 is input into machine learning unit 30, which determines whether there are processing marks 29 in image 26 and outputs the determination result (determination process). Figure 9 This is a block diagram showing the calculation unit 6 in the determination process.

[0092] In the judgment process, from storage unit 8 (refer to...) Figure 8 Image 26, acquired during the image acquisition process, is read from (A) and input into the machine learning unit 30. Then, the machine learning unit 30 determines whether image 26 contains a processing mark 43 through inference from the neural network 32. Specifically, image 26 is input into the input layer 34, and operations using image 26 as input data are sequentially performed in the input layer 34, hidden layer 38, and output layer 36. Finally, the determination result regarding the presence or absence of a processing mark 43 is output from the output layer 36.

[0093] The determination result output from the machine learning unit 30 is displayed on a display unit such as a monitor (not shown). This notifies the operator whether a machining mark 43 has been formed on the wafer 31.

[0094] Furthermore, when using a semantic segmentation network (SegNet) as the model for neural network 32, the machine learning unit 30 infers and outputs an image (two-dimensional map) containing the crack 37 and the machining mark 43 in image 26. Therefore, the operator can refer to the output image to confirm whether a machining mark 43 is formed on the front side 31a of the wafer 31 and to confirm the number and location of the machining mark 43.

[0095] Then, based on the determination of whether there is a machining mark 43, the subsequent processing of the wafer 31 is determined. For example, if the machine learning unit 30 determines that no machining mark 43 has been formed on the wafer 31, or if the location of the machining mark 43 is within a range specified from the crack 37 (e.g., inside the predetermined dividing line 33), the processing of the wafer 31 continues.

[0096] Specifically, an external force is applied to the wafer 31 to divide it along a predetermined dividing line 33. For example, a circular strip (expansion strip) that can be expanded by applying an external force is attached to the wafer 31. Then, an external force is applied to the wafer 31 by stretching the expansion strip outward in the radial direction.

[0097] Thus, wafer 31 is divided along predetermined dividing line 33, starting from the modified layer and crack 37. As a result, multiple device chips, each having a device 35, can be obtained. The device chips are picked up, for example, by a chuck (not shown), and mounted on a predetermined substrate (wiring substrate, etc.) by chip bonding or the like.

[0098] As described above, in the wafer inspection method of this embodiment, a machine learning unit 30 for determining whether there are processing marks 29 is constructed by using machine learning to capture an image 22 obtained by photographing the front side 11a side of the wafer 11 for image acquisition on which the modified layer 25 is formed. Furthermore, an image 26 obtained by photographing the front side 31a side of the wafer 31, which is the inspection target and has undergone laser processing, is input into the machine learning unit 30, thereby determining whether there are processing marks 43 in the image 26.

[0099] Therefore, after the machine learning unit 30 completes its learning, it can accurately determine whether there are processing marks 43 without obtaining an image of the wafer 31 before laser processing, which is to be inspected. As a result, the process of taking pictures of the wafer 31 before laser processing can be omitted every time an inspection is performed, making wafer inspection extremely simple.

[0100] Furthermore, the functions of the machine learning unit 30 can be implemented using either software or hardware. For example, the machine learning unit 30 can be implemented using a program that describes a series of operations performed by the machine learning unit 30. Specifically, when the machine learning unit 30 has a neural network 32, the operations in the input layer 34, output layer 36, and hidden layer 38 are described by a program, and this program is stored in the storage unit 8 of the control unit 4. Moreover, when determining whether there are machining marks, the program is read from the storage unit 8 and executed by the arithmetic unit 6.

[0101] Additionally, in the process of acquiring the image being inspected (refer to...) Figure 8In (A), multiple images 26 can also be obtained by taking multiple pictures of the chip 31 using the imaging unit 2. In this case, while the chip 31 is held by the chuck stage, the position of the chuck stage is changed by the moving mechanism, and the imaging unit 2 takes multiple pictures of the front 31a side of the chip 31.

[0102] Through the above steps, multiple images 26 representing different regions of the wafer 31 are obtained (for example, an image 26 of the entire area of ​​the front side 31a of the wafer 31). Then, in the determination process, the multiple images 26 are sequentially input into the machine learning unit 30, and the presence or absence of machining marks 43 is determined for each of the multiple images 26.

[0103] In addition, the structure and method of this embodiment can be implemented with appropriate modifications without departing from the purpose of the present invention.

Claims

1. A method for inspecting a wafer, comprising inspecting a laser-processed wafer for processing marks caused by laser beam scattering, wherein devices are formed on the front side of the wafer in multiple regions divided by predetermined dividing lines, and the laser processing is performed by irradiating the wafer from the back side with the laser beam focused at a point in a region inside the wafer corresponding to the predetermined dividing lines to form a modified layer. Its features are, The inspection method for this wafer includes the following steps: In the first image acquisition process, the front side of the wafer used for image acquisition is photographed before the laser processing is performed, and multiple first images are obtained. The second image acquisition process involves taking a picture of the area on the front side of the wafer on which the modified layer is formed, corresponding to the first image, to acquire multiple second images. The process of obtaining a machining mark image is to obtain a machining mark image corresponding to the difference between the first image and the second image; In the marking process, referring to the second image and the processing mark image, the second image is marked according to the processing mark; The learning process involves using machine learning on the labeled second image to construct a machine learning unit that determines whether or not the processing mark exists. In the process of acquiring the image under inspection, the front side of the wafer on which the modified layer is formed is photographed to acquire the image under inspection. as well as In the judgment process, the image to be inspected is input into the machine learning unit, which determines whether the image contains the processing mark and outputs the judgment result.

2. The wafer inspection method according to claim 1, characterized in that, The machine learning department has a neural network, which includes an input layer, an output layer, and hidden layers. In this learning process, the neural network performs deep learning using multiple of the second images input to the input layer. In this determination process, the neural network performs calculations based on the image to be inspected input to the input layer and outputs the determination result from the output layer.

3. The wafer inspection method according to claim 2, characterized in that, The neural network makes this determination through semantic segmentation.

Images