Inspection device, peeling device, and trained model generation method

Abstract

To detect a crack generated in an ingot or wafer.SOLUTION: An inspection unit 30 comprises: a temporary mount table 31 which exposes a first surface 201 of a wafer 220 to hold it; a light source 41 which irradiates the first surface 201 of the wafer 220 held by the temporary mount table 31 with light 44 at a prescribed incident angle 43; an imaging unit 42 which condenses reflection light 45 reflected on the first surface 201 of the wafer 220 to perform imaging and forms a projection image 300 in which irregularities generated on the first surface 201 due to a crack extending in the wafer 220 are highlighted; and a control unit 100 which determines the state of the wafer 220. The control unit 100 includes an arithmetic processing device and a storage device and has a learned model 400 which is configured by machine learning so as to output a determination result indicating whether or not the crack 223 is formed in the wafer 220 when the projection image acquired by the imaging unit 42 is input.SELECTED DRAWING: Figure 14

Description

[Technical Field]

[0001] The present invention relates to an inspection device that inspects whether cracks have occurred in a wafer using a trained model that creates cracks in the wafer from images obtained from reflected light. [Background technology]

[0002] Semiconductor devices are manufactured by forming devices on wafers made from semiconductor ingots and then dividing them into individual vise chips. For example, wafers made of silicon are cut from silicon ingots using a wire saw or similar tool and polished to form devices, but wafers made of SiC are a very hard material, so cutting them from ingots with a wire saw takes a much longer time than silicon.

[0003] To address this issue, the so-called KABRA (registered trademark) process was devised, in which a laser beam is applied to form a modified layer inside the ingot and cracks extending from the modified layer form a delamination layer, and wafers are then peeled off from the delamination layer (see, for example, Patent Document 1). After wafers have been peeled off from the ingot by the KABRA process, the modified layer remaining on the surface is ground and polished, and the surface is again irradiated with a laser beam to form a delamination layer inside, and the next wafer is then peeled off. This processing method makes it possible to slice wafers from an ingot in a short time. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-188117 Summary of the Invention [Problem to be solved by the invention]

[0005] However, cracks may occasionally occur in the ingots or delaminated wafers after delamination, which may result in defective wafers that are unsuitable for forming devices. For this reason, operators visually inspect each ingot before laser beam irradiation and each delaminated wafer, which requires labor and costs.

[0006] Detecting cracks from images of ingots and wafers was also considered, but a special camera was required to photograph internal cracks that do not appear on the wafer surface, and it was difficult to distinguish defective cracks from the modified layer that remained on the wafer. In particular, the shape of cracks has little regularity, making it difficult to identify cracks using methods that set thresholds for values ​​calculated from images.

[0007] An object of the present invention is to provide an inspection device, a peeling device, and a method for generating a trained model that can detect cracks that have occurred in an ingot or wafer. [Means for solving the problem]

[0008] In order to solve the above-mentioned problems and achieve the object, an inspection apparatus of the present invention is an inspection apparatus for inspecting an ingot or wafer having one surface and another surface, the one surface being a polished surface, and includes: a holding table that holds the polished surface of the ingot or wafer exposed; a light source that irradiates the polished surface of the ingot or wafer held on the holding table with light at a predetermined angle of incidence; an imaging unit that collects and photographs the light reflected on the polished surface of the ingot or wafer, and forms a projected image in which irregularities caused by cracks extending into the ingot or wafer are emphasized; and a determination unit that compares the formed projected image with preset conditions to determine the state of the ingot or wafer, the determination unit having a processor and memory, and a trained model configured by machine learning so as to output a determination result indicating whether or not the crack has formed inside the ingot or wafer when the projected image acquired by the imaging unit is input. The trained model uses a cascade classifier that trains the image of the crack using sample images consisting of the projection image including the crack and the projection image without the crack. It is characterized by the following.

[0011] In the inspection device, the wafer may be a delamination wafer manufactured by irradiating the polished surface of an ingot with a laser beam to form a delamination layer consisting of a modified layer and delamination cracks extending from the modified layer along the crystal, and then delaminating the wafer from the ingot starting from the delamination layer.

[0012] The peeling device of the present invention is an exposed Polished A delamination apparatus for manufacturing wafers from an ingot on which a delamination layer consisting of a modified layer and delamination cracks extending from the modified layer along the crystal is formed by irradiating a laser beam onto a polishing surface, the delamination apparatus comprising: a holding table for holding the ingot; a delamination unit for delaminating a wafer from the ingot held on the holding table; a temporary placement table for exposing and temporarily placing and holding the polished surface of the wafer delaminated by the delamination unit; and an inspection unit for inspecting the wafer held on the temporary placement table, the inspection unit inspecting the polished surface of the wafer held on the temporary placement table. The wafer polishing system comprises a light source that irradiates light at a predetermined angle of incidence, an imaging unit that collects and photographs the light reflected from the polished surface of the wafer, and forms a projection image that emphasizes the irregularities caused on the polished surface by cracks extending inside the wafer, and a judgment unit that compares the formed projection image with preset conditions to judge the state of the wafer, wherein the judgment unit has a processor and memory, and is equipped with a trained model constructed by machine learning so as to output a judgment result indicating whether or not the crack has formed inside the wafer when the projection image acquired by the imaging unit is input. In the peeling device, the trained model may utilize a cascade classifier that trains the image of the crack using sample images consisting of the projection image including the crack and the projection image without the crack. In the peeling device, the trained model may be a neural network including an input layer to which the projected image is input and an output layer to which the determination result is output.

[0013] The method for generating a trained model of the present invention includes: an inspection apparatus for inspecting an ingot or wafer having one surface and another surface, the one surface being a polished surface, the inspection apparatus comprising: a holding table for holding the polished surface of the ingot or wafer exposed; a light source for irradiating the polished surface of the ingot or wafer held on the holding table with light at a predetermined angle of incidence; an imaging unit for collecting and photographing the light reflected on the polished surface of the ingot or wafer, and forming a projected image in which irregularities caused by cracks extending into the ingot or wafer are emphasized; and a determination unit for comparing the formed projected image with preset conditions to determine the state of the ingot or wafer, the determination unit having a processor and memory, the inspection apparatus comprising a trained model configured by machine learning so as to output a determination result indicating whether or not the crack has formed inside the ingot or wafer when the projected image acquired by the imaging unit is input; Ingot or Applicable A trained model generation method for generating a trained model for determining a wafer state, The inspection device has the ingot or the wafer placed on the holding table, and irradiates light from the light source onto the ingot or the wafer on the holding table. Irradiating the polished surface with light from a light source at a predetermined angle of incidence; With the imaging unit The light reflected from the polished surface is collected and photographed. Applicable Ingot or Applicable an image acquisition step of acquiring a plurality of first learning images corresponding to images of the sample including cracks and a plurality of second learning images corresponding to images of the sample not including cracks by photographing a wafer as a sample; a machine learning unit to which the second learning image and information indicating the absence of the crack are input, and to which the first learning image and information indicating the position of the crack are input, and a learning step of generating a trained model through machine learning using the first training image and the second training image, which, when an image of an ingot or wafer is input, outputs a determination result indicating whether or not a crack has formed in the ingot or wafer.

[0014] In the method for generating a trained model, the trained model may utilize a cascade classifier that has trained the crack image using the first training image and the second training image.

[0015] In the method for generating a trained model, the trained model may be a neural network including an input layer to which the first training image and the second training image are input, and an output layer to which the judgment result is output. [Effects of the Invention]

[0016] The present invention has an effect of being able to detect cracks occurring in an ingot or wafer. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a perspective view showing an example of the configuration of a peeling device according to the first embodiment. [Figure 2] FIG. 2 is a side view of an ingot to be processed by the peeling apparatus shown in FIG. [Figure 3]FIG. 3 is a plan view of the ingot shown in FIG. [Figure 4] FIG. 4 is a perspective view of the ingot shown in FIG. [Figure 5] FIG. 5 is a perspective view of a wafer produced from the ingot shown in FIG. [Figure 6] FIG. 6 is a perspective view schematically showing a state in which a separation layer is formed on the ingot shown in FIG. [Figure 7] FIG. 7 is a side view schematically showing a state in which a separation layer is formed on the ingot shown in FIG. [Figure 8] FIG. 8 is a plan view of the ingot on which the peeling layer is formed. [Figure 9] FIG. 9 is a cross-sectional view taken along line IX-IX in FIG. [Figure 10] FIG. 10 is a plan view schematically showing an ingot and a wafer in which a crack has occurred. [Figure 11] FIG. 11 is a side view showing a configuration example of the ultrasonic wave applying unit of the peeling unit of the peeling device shown in FIG. [Figure 12] FIG. 12 is a side view showing a configuration example of a wafer separation unit of the separation apparatus shown in FIG. [Figure 13] FIG. 13 is a side view showing a state in which the wafer separation unit shown in FIG. 12 has separated the wafer from the ingot. [Figure 14] FIG. 14 is a side view schematically illustrating an example of the configuration of the inspection unit of the peeling apparatus shown in FIG. [Figure 15] FIG. 15 is a diagram showing an example of a projected image captured by the imaging unit of the inspection unit of the peeling apparatus shown in FIG. [Figure 16] FIG. 16 is a diagram showing an example of a first projection image including a crack for generating a trained model provided in the control unit of the inspection unit of the peeling apparatus shown in FIG. [Figure 17]FIG. 17 is a diagram showing another example of a first projection image including a crack for generating a trained model provided in the control unit of the inspection unit of the peeling apparatus shown in FIG. [Figure 18] FIG. 18 is a diagram showing another example of a first projection image including a crack for generating a trained model provided in the control unit of the inspection unit of the peeling apparatus shown in FIG. [Figure 19] FIG. 19 is a diagram showing yet another example of a first projection image including a crack for generating a trained model provided in the control unit of the inspection unit of the peeling apparatus shown in FIG. 1. [Figure 20] FIG. 20 is a diagram showing an example of a second projection image without cracks for generating a trained model provided in the control unit of the inspection unit of the peeling apparatus shown in FIG. [Figure 21] FIG. 21 is a diagram showing another example of a second projection image without cracks for generating a trained model provided in the control unit of the inspection unit of the peeling apparatus shown in FIG. [Figure 22] FIG. 22 is a diagram illustrating the configuration of a trained model provided in the control unit of the inspection unit of the peeling device shown in FIG. [Figure 23] FIG. 23 is a diagram illustrating the configuration of a strong classifier of the trained model shown in FIG. [Figure 24] FIG. 24 is a flowchart showing the flow of a trained model generation method according to the first embodiment. [Figure 25] FIG. 25 is a diagram illustrating the configuration of a trained model provided in the control unit of the inspection unit of the delamination apparatus according to the second embodiment. [Figure 26] FIG. 26 is a side view schematically showing an example of the configuration of an inspection unit of a delamination device according to a modification of the first and second embodiments of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0018] Modes (embodiments) for carrying out the present invention will be described in detail with reference to the drawings. The present invention is not limited to the contents described in the following embodiments. Furthermore, the components described below include those that can be easily imagined by a person skilled in the art and those that are substantially the same. Furthermore, the configurations described below can be combined as appropriate. Furthermore, various omissions, substitutions, or modifications of the configuration can be made within the scope of the gist of the present invention.

[0019] [Embodiment 1] An inspection device and a delamination device according to a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a perspective view showing an example of the configuration of the delamination device according to the first embodiment. FIG. 2 is a side view of an ingot to be processed by the delamination device shown in FIG. 1. FIG. 3 is a plan view of the ingot shown in FIG. 2. FIG. 4 is a perspective view of the ingot shown in FIG. 2. FIG. 5 is a perspective view of a wafer produced from the ingot shown in FIG. 2. FIG. 6 is a perspective view schematically showing a state in which a delamination layer is formed on the ingot shown in FIG. 2. FIG. 7 is a side view schematically showing a state in which a delamination layer is formed on the ingot shown in FIG. 2. FIG. 8 is a plan view of the ingot on which a delamination layer has been formed. FIG. 9 is a cross-sectional view taken along line IX-IX in FIG. 8. FIG. 10 is a plan view schematically showing an ingot and a wafer in which a crack has occurred.

[0020] (ingots and wafers) The delamination apparatus 1 shown in Fig. 1 according to the first embodiment is an apparatus for manufacturing a wafer 220 (also referred to as a delaminated wafer) shown in Fig. 5 from an ingot 200 (also referred to as a semiconductor ingot) shown in Fig. 2, Fig. 3, and Fig. 4. The ingot 200 shown in Fig. 2, Fig. 3, and Fig. 4, which is the object to be processed by the delamination apparatus 1 according to the first embodiment, is made of SiC (silicon carbide) and is formed in a cylindrical shape as a whole. In the first embodiment, the ingot 200 is a hexagonal single crystal ingot.

[0021] 2, 3, and 4, the ingot 200 has a first surface 201 (corresponding to one surface) that is circular and polished, a second surface 202 (corresponding to the other surface) that is circular and located on the back side of the first surface 201, and a peripheral surface 203 that connects the outer edges of the first surface 201 and the second surface 202. The ingot 200 also has, on the peripheral surface 203, a first orientation flat 204 that indicates the crystal orientation of the ingot 200, and a second orientation flat 205 that is perpendicular to the first orientation flat 204 and indicates the crystal orientation of the ingot 200. The orientation flats 204 and 205 are flat surfaces that form straight lines when viewed from above the ingot 200. The length of the first orientation flat 204 is longer than the length of the second orientation flat 205.

[0022] The ingot 200 also has a C-axis 208 tilted at an off angle α in a tilt direction 207 toward the second orientation flat 205 with respect to a normal 206 to the first surface 201, and a c-plane 209 perpendicular to the C-axis 208. The c-plane 209 is tilted at an off angle α with respect to the first surface 201 of the ingot 200. The tilt direction 207 of the C-axis 208 from the normal 206 is perpendicular to the extension direction of the second orientation flat 205 and is parallel to the first orientation flat 204.

[0023] An infinite number of c-planes 209 are set in the ingot 200 at the molecular level of the ingot 200. In the first embodiment, the off-angle α is set to 1°, 4°, or 6°, but in the present invention, the ingot 200 can be manufactured by freely setting the off-angle α within the range of, for example, 1° to 6°.

[0024] After the first surface 201 of the ingot 200 is ground by a grinding device, the ingot 200 is polished by a polishing device to form the first surface 201 into a mirror finish. A portion of the ingot 200 on the side of the first surface 201 is peeled off, and the peeled portion is manufactured into the wafer 220 shown in FIG.

[0025] 5 is obtained by separating a portion including the first surface 201 of the ingot 200 as the wafer 220, and grinding, polishing, etc., the separated surface 221 (corresponding to the other surface) separated from the ingot 200. For this reason, the wafer 220 has the first surface 201 and the separated surface 221. After the wafer 220 is separated from the ingot 200, devices are formed on the surface.

[0026] In the first embodiment, the device is a MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor), a MEMS (Micro Electro Mechanical Systems), or an SBD (Schottky Barrier Diode), but in the present invention, the device is not limited to a MOSFET, a MEMS, or an SBD. Note that the same parts of the wafer 220 as those of the ingot 200 are denoted by the same reference numerals, and their description will be omitted.

[0027] 2, 3, and 4, a separation layer 211 shown in Figures 6, 7, 8, and 9 is formed, and then a portion of the ingot 200, i.e., a wafer 220 to be manufactured, is separated and separated starting from the separation layer 211. The separation layer 211 is formed by irradiating the ingot 200 with a pulsed laser beam 217 (shown in Figures 6 and 7) having a wavelength that is transparent to the ingot 200, with a focal point 218 of the laser beam 217 positioned at a depth 213 (shown in Figure 7) corresponding to a thickness 222 (shown in Figure 5) of the wafer 220 to be manufactured from the first surface 201 of the ingot 200, while the ingot 200 is moved along a second orientation flat 205 relative to the laser beam 217.

[0028] 9, when the ingot 200 is irradiated with the laser beam 217, SiC separates into Si (silicon) and C (carbon), and the subsequently irradiated pulsed laser beam 217 is absorbed by the previously formed C, causing the SiC to separate into Si and C in a chain reaction, forming a modified layer 214 inside the ingot 200 along the second orientation flat 205, and also generating a delamination crack 215 extending from the modified layer 214 along the c-plane 209. In this way, when the ingot 200 is irradiated with the pulsed laser beam 217 of a transparent wavelength, a delamination layer 211 including the modified layer 214 and the delamination crack 215 formed from the modified layer 214 along the c-plane 209 (corresponding to a crystal) is formed inside the ingot 200.

[0029] Once the exfoliation layer 211 is formed over the entire length of the second orientation flat 205, the ingot 200 is moved a predetermined distance 219 along the first orientation flat 204 relative to the laser beam 217 (hereinafter referred to as index feed), and then the focal point 218 of the laser beam 217 is positioned at the aforementioned depth 213. The ingot 200 is then moved along the second orientation flat 205 relative to the laser beam 217, and is irradiated with the laser beam 217 to form the exfoliation layer 211. The irradiation of the laser beam 217 while moving the ingot 200 along the second orientation flat 205 relative to the laser beam 217 and the index feed are alternately repeated until the exfoliation layer 211 is formed over the entire lower surface of the first surface 201, and as shown in FIGS.

[0030] In this way, the exposed first surface 201 of the ingot 200 is irradiated with a laser beam 217, and a delamination layer 211 consisting of a modified layer 214 and delamination cracks 215 extending from the modified layer 214 along the c-plane 209 is formed all over the area below the first surface 201. After a portion of the ingot 200, i.e., a wafer 220, is delaminate starting from the delamination layer 211, the delamination surface from which the wafer 220 has been delaminate is formed into a mirror surface by grinding and polishing, and the delamination surface is formed on the first surface 201, and again a delamination layer 211 is formed, and the wafer 220 is delaminate.

[0031] In this way, the thickness of the ingot 200 decreases as the wafer 220 is peeled off, and a peeling layer 211 is formed until the ingot 200 reaches a predetermined thickness, and the wafer 220 is peeled off. The wafer 220 is a peeled wafer manufactured by irradiating the first surface 201 of the ingot 200 with a laser beam 217 to form a peeling layer 211 consisting of a modified layer 214 and peeling cracks 215 extending from the modified layer 214 along the crystal, and then peeling the wafer 220 from the ingot 200 starting from the peeling layer 211. In the present invention, the ingot 200 may have an off angle α of zero degrees, and may be an ingot formed from a material other than SiC, such as GaN (gallium nitride), single crystal diamond, or single crystal Si.

[0032] As shown schematically in Fig. 10, cracks 223 may form inside the ingot 200 and wafer 220 described above. The cracks 223 are minute spaces formed inside the ingot 200 and wafer 220. For this reason, the mirror-finished first surface 201 is flat at positions where no cracks 223 have formed and at positions overlapping in the thickness direction with the cracks 223, but at positions overlapping in the thickness direction with the internal cracks 223, minute protrusions that are not visible to the naked eye are formed. The planar shape of these minute protrusions is generally the same as the planar shape of the internal cracks 223.

[0033] (peeling device) The delamination apparatus 1 shown in FIG. 1 according to the first embodiment is an apparatus for manufacturing a wafer 220 by irradiating an exposed first surface 201 with a laser beam 217 to delaminate a wafer 220 from an ingot 200 having a delamination layer 211 formed thereon, the delamination layer 211 being composed of a modified layer 214 and delamination cracks 215 extending from the modified layer 214 along the crystal, starting from the delamination layer 211. FIG. 11 is a side view showing an example of the configuration of an ultrasonic wave application unit of the delamination unit of the delamination apparatus shown in FIG. 1. FIG. 12 is a side view showing an example of the configuration of a wafer delamination unit of the delamination unit of the delamination apparatus shown in FIG. 1. FIG. 13 is a side view showing a state in which the wafer delamination unit shown in FIG. 12 has delaminated a wafer from the ingot. FIG. 14 is a side view schematically showing an example of the configuration of an inspection unit of the delamination apparatus shown in FIG. 1. FIG. 15 is a diagram showing an example of a projected image captured by an imaging unit of the inspection unit of the delamination apparatus shown in FIG. 1.

[0034] As shown in FIG. 1 , the peeling device 1 includes a holding table 10, a peeling unit 20, and an inspection unit 30, which is an inspection device. The holding table 10 has a holding surface 11 on which a second surface 202 of an ingot 200 is placed, and the holding table 10 suction-holds the ingot 200 placed on the holding surface 11. The holding surface 11 of the holding table 10 is disk-shaped and formed from porous ceramic or the like along the horizontal direction, and is connected to a vacuum suction source (not shown) via a vacuum suction path (not shown). The holding surface 11 of the holding table 10 is sucked by the vacuum suction source, and the holding table 10 suction-holds the ingot 200 placed on the holding surface 11. The holding table 10 is rotated by a rotation drive source (not shown) around an axis that passes through the center of the holding surface 11 and is parallel to the vertical direction.

[0035] The separation unit 20 separates the wafer 220 from the ingot 200 held on the holding table 10. The separation unit 20 includes an ultrasonic wave applying unit 21, a wafer separation unit 25, and a moving unit 27.

[0036] The moving unit 27 includes a horizontal moving unit 28 that moves the ultrasonic wave application unit 21 and the wafer peeling unit 25 in the horizontal direction, and a lifting unit (not shown) that lifts and lowers each of the ultrasonic wave application unit 21 and the wafer peeling unit 25. The moving unit 27 can position the ultrasonic wave application unit 21 and the wafer peeling unit 25 above the holding table 10 by using the horizontal moving unit 28. The moving unit 27 can also position the wafer peeling unit 25 above the temporary placement table 31 of the inspection unit 30.

[0037] 11 , the ultrasonic wave application unit 21 includes an ultrasonic wave oscillator 22 located above the ingot 200 held on the holding table 10 and facing a first surface 201 of the ingot 200 with a gap therebetween, and a liquid supply unit 23 that supplies liquid between the ultrasonic wave oscillator 22 and the first surface 201 of the ingot 200. The ultrasonic wave oscillator 22 includes an ultrasonic vibrator (not shown) that expands and contracts when AC power is applied, generating ultrasonic vibrations on an opposing surface 24 that faces the first surface 201 of the ingot 200.

[0038] In the ultrasonic wave application unit 21, the ultrasonic wave oscillator 22 is positioned above the ingot 200 held on the holding table 10 by the horizontal movement unit 28 of the movement unit 27, and is then lowered by the lifting unit of the movement unit 27 so that the ultrasonic wave oscillator 22 faces the first surface 201 of the ingot 200 with a gap between them. In the ultrasonic wave application unit 21, the liquid supply unit 23 supplies liquid between the first surface 201 of the ingot 200 and the ultrasonic wave oscillator 22, immersing the ultrasonic wave oscillator 22 in the liquid on the first surface 201 of the ingot 200. The ultrasonic wave application unit 21 applies AC power to the ultrasonic vibrator of the ultrasonic wave oscillator 22 for a predetermined time to ultrasonically vibrate the opposing surface 24.

[0039] The ultrasonic wave applying unit 21 transmits ultrasonic vibrations from the opposing surface 24 to the first surface 201 of the ingot 200 via the liquid 51, and applies ultrasonic waves to the first surface 201 of the ingot 200 held on the holding table 10. Then, ultrasonic waves from the ultrasonic oscillator 22 stimulate the peeling layer 211, dividing the ingot 200 starting from the peeling layer 211, and separating the wafers 220 to be manufactured from the ingot 200.

[0040] The ultrasonic wave applying unit 21 applies AC power to the ultrasonic vibrator of the ultrasonic oscillation part 22 for a predetermined time to ultrasonically vibrate the opposing surface 24, and when the wafer 220 to be manufactured is separated from the ingot 200, the ultrasonic wave applying unit 21 stops applying AC power to the ultrasonic vibrator of the ultrasonic oscillation part 22 and stops the supply of liquid from the liquid supply part 23. The ultrasonic wave applying unit 21 is raised by the lifting unit of the movement unit 27, and is retracted from above the ingot 200 held on the holding table 10 by the horizontal movement unit 28.

[0041] The wafer separation unit 25 suction-holds the first surface 201 of the ingot 200, i.e., the wafer 220, held on the holding surface 11 of the holding table 10. The lower surface 26 of the wafer separation unit 25 is connected to a vacuum suction source (not shown) via a vacuum suction path (not shown). The lower surface 26 of the wafer separation unit 25 is sucked by the vacuum suction source, and the first surface 201 of the ingot 200, i.e., the wafer 220, which is in contact with the lower surface 26, is held by suction.

[0042] The ultrasonic oscillator 22 of the wafer separation unit 25 is positioned above the ingot 200 held on the holding table 10 by the horizontal movement unit 28 of the movement unit 27, and is then lowered by the lifting unit of the movement unit 27 until the lower surface 26 comes into contact with the first surface 201 of the ingot 200, as shown in Fig. 12. The wafer separation unit 25 is sucked by the vacuum suction source, and the first surface 201 of the ingot 200, i.e., the wafer 220, is suction-held on the lower surface 26.

[0043] The wafer separation unit 25 is raised by the lifting unit of the moving unit 27, and separates the wafer 220 held by suction on the lower surface 26 from the ingot 200, as shown in Fig. 13. The wafer separation unit 25 is positioned above the temporary placement table 31 of the inspection unit 30 by the horizontal movement unit 28 of the moving unit 27, and is lowered by the lifting unit of the moving unit 27 to place the wafer 220 held by suction on the temporary placement table 31. After the wafer separation unit 25 stops holding the wafer 220 by suction, it is raised by the lifting unit of the moving unit 27 and is retracted from above the temporary placement table 31 by the horizontal movement unit 28.

[0044] In the first embodiment, the inspection unit 30 is an apparatus that inspects the wafer 220. In the first embodiment, the inspection unit 30 is an apparatus that inspects whether or not the wafer 220 has a crack 223 based on the principle of the devil's land. As shown in Fig. 14, the inspection unit 30 includes a temporary placement table 31 that is a holding table, a projection image forming unit 40, and a control unit 100 that is a determination unit.

[0045] The temporary placement table 31 temporarily places the wafer 220 separated by the separation unit 20, and holds the wafer 220 with the first surface 201 of the wafer 220 exposed. The temporary placement table 31 has a holding surface 32 parallel to the horizontal direction on which the separated surface 221 of the wafer 220 is placed, and holds the ingot 200 placed on the holding surface 32.

[0046] 14, the projection image forming unit 40 includes a light source 41 and an imaging unit 42. The light source 41 irradiates the first surface 201 of the wafer 220 held on the temporary placement table 31 with visible light 44 at a predetermined incident angle 43 (in the first embodiment, an angle smaller than 90 degrees).

[0047] The imaging unit 42 captures reflected light 45 of light 44 reflected by the first surface 201 of the wafer 220 held on the temporary placement table 31 by using a condenser lens 46, and forms a projected image 300, an example of which is shown in FIG. 15 . The imaging unit 42 is equipped with an imaging element that captures the reflected light. The imaging element is, for example, a CCD (Charge-Coupled Device) imaging element or a CMOS (Complementary MOS) imaging element. The projected image 300 formed by the imaging unit 42 capturing the reflected light 45 is a grayscale image in which the amount of light of the reflected light 45 is defined by multiple levels of gradation (for example, 256 levels).

[0048] As described above, when a crack 223 is formed inside the wafer 220, a minute convex portion is formed on the first surface 201 at a position that overlaps the crack 223 in the thickness direction. Furthermore, the first surface 201 of the wafer 220 is mirror-finished, and is flat at a position that does not overlap the crack 223 in the thickness direction. For this reason, light 44 from the light source 41 is reflected as light 45 by the convex portion at a position that overlaps the crack 223 on the first surface 201 in the thickness direction, and in the projection image 300 captured by the imaging unit 42, the reflected light 45 from the position that overlaps the crack 223 on the first surface 201 in the thickness direction is weaker than the reflected light 45 from a position that does not overlap the crack 223 on the first surface 201 in the thickness direction.

[0049] 15, positions that overlap with the cracks 223 on the first surface 201 in the thickness direction are darker than positions that do not overlap with the cracks 223 on the first surface 201 in the thickness direction. In this way, the projection image 300 formed by the imaging unit 42 capturing the reflected light 45 is an image in which the unevenness caused on the first surface 201 by the cracks 223 extending into the wafer 220 is emphasized. The imaging unit 42 outputs the formed projection image 300 to the control unit 100.

[0050] In the first embodiment, the imaging unit 42 captures reflected light 45 from within a 20 mm square range on the first surface 201, and forms a projected image 300 within the 20 mm square range on the first surface 201. Compared to irradiating a wide area such as the entire surface of a wafer, the light is irradiated only within a narrow range, so an image with sufficient contrast can be obtained with a small light source.

[0051] The inspection unit 30 according to the first embodiment also includes a table movement unit (not shown) that moves the temporary placement table 31 relatively to the inspection unit 30 in two directions that are parallel to the horizontal direction and perpendicular to each other. The inspection unit 30 according to the first embodiment uses the table movement unit to move the temporary placement table 31 relatively to the inspection unit 30 in two directions that are parallel to the horizontal direction and perpendicular to each other, thereby dividing and photographing the first surface 201 of the wafer 220 held on the temporary placement table 31 into a plurality of projection images 300.

[0052] (control unit) FIG. 16 is a diagram showing an example of a first projection image including a crack for generating a trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 17 is a diagram showing another example of a first projection image including a crack for generating a trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 18 is a diagram showing another example of a first projection image including a crack for generating a trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 19 is a diagram showing yet another example of a first projection image including a crack for generating a trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 20 is a diagram showing an example of a second projection image without a crack for generating a trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 21 is a diagram showing another example of a second projection image without a crack for generating a trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 22 is a diagram illustrating the configuration of the trained model provided by the control unit of the inspection unit of the separation device shown in FIG. 1. FIG. 23 is a diagram illustrating the configuration of a strong classifier for the trained model shown in FIG. 22.

[0053] The control unit 100 controls each of the above-mentioned constituent units that make up the delamination apparatus 1. That is, the control unit 100 causes the delamination apparatus 1 and the inspection unit 30 to perform processing operations on the ingot 200 and inspection operations on the wafer 220. The control unit 100 is a computer that includes an arithmetic processing device having a microprocessor such as a CPU (central processing unit), a storage device having memory such as a ROM (read only memory) or RAM (random access memory), and an input / output interface device.

[0054] The arithmetic processing device of the control unit 100 performs arithmetic processing in accordance with a computer program stored in the storage device, and outputs control signals for controlling the peeling device 1 to the above-mentioned components of the peeling device 1 via the input / output interface device. The control unit 100 is also connected to a display unit configured with a liquid crystal display device or the like that displays the status and images of the processing operation, an input unit used by the operator to register processing content information, and a notification unit that notifies the operator. The input unit is configured with at least one of a touch panel provided on the display unit and a keyboard, etc. The notification unit notifies the operator by emitting at least one of sound, light, and a message on the touch panel.

[0055] 14, the control unit 100 also includes a determination unit 101 that compares a projection image 300 formed by the imaging unit 42 with preset conditions to determine the state of the wafer 220. Specifically, the determination unit 101 of the control unit 100 has a trained model 400 stored in a storage device that is configured by machine learning so as to output a determination result indicating whether or not a crack 223 has formed inside the wafer 220 when the projection image 300 acquired by the imaging unit 42 is input, and determines whether or not a crack 223 is present in the projection image 300 formed by the imaging unit 42 using the trained model 400.

[0056] In embodiment 1, the trained model 400 is generated using a cascade classifier 401 (shown in Figure 22) that has trained an image of a crack 223 using sample images consisting of a projection image 300 (hereinafter indicated by the symbol 301, referred to as the first projection image, and corresponding to the first training image) including a crack 223 shown in Figures 16, 17, 18, and 19, and a projection image 300 (hereinafter indicated by the symbol 302, referred to as the second projection image, and corresponding to the second training image) without a crack shown in Figures 20 and 21.

[0057] When a projection image 300 is input, the cascade classifier 401 of the trained model 400 determines whether or not a crack 223 is present in the projection image 300. The cascade classifier 401 of the trained model 400 includes a plurality of strong classifiers 402, as shown in Fig. 22. Each strong classifier 402 includes a plurality of weak classifiers 403 and a determiner 404, as shown in Fig. 23.

[0058] The weak classifier 403 determines whether the feature of the projection image 300 (in the first embodiment, the amount of light received by the pixel that has the weakest amount of reflected light among the pixels that make up the projection image 300) is below a threshold generated by machine learning, and if it is below, outputs a signal indicating 1 to the determiner 404, and if it is not below, outputs a signal indicating 0 to the determiner 404.

[0059] The determiner 404 calculates a value obtained by multiplying the signal output by each weak classifier 403 by a coefficient generated by machine learning, and calculates a value (hereinafter referred to as a total value) obtained by adding up the values ​​obtained by multiplying the signal output by each weak classifier 403 by the coefficient generated by machine learning. The determiner 404 determines whether the total value exceeds a second threshold generated by machine learning, and if it determines that the total value exceeds the second threshold value, it outputs a determination result that a crack 223 is present in the projection image 300, and if it determines that the total value does not exceed the second threshold value, it outputs a determination result that a crack 223 is not present in the projection image 300.

[0060] The cascade classifier 401 of the trained model 400 determines the presence or absence of a crack 223 in the projection image 300 one by one from the multiple strong classifiers 402. When one of the multiple strong classifiers 402 outputs a determination result that the crack 223 is not present in the projection image 300, the cascade classifier 401 of the trained model 400 outputs a determination result that the input projection image 300 does not contain a crack 223, without determining the presence or absence of a crack 223 from the strong classifier 402 that output the determination result that the crack 223 is not present. When all the strong classifiers 402 output determination results that the crack 223 is present in the projection image 300, the cascade classifier 401 of the trained model 400 outputs a determination result that the input projection image 300 contains a crack 223.

[0061] 22 shows three strong classifiers 402, but in the present invention, the cascade classifier 401 of the trained model 400 may include two, four or more strong classifiers 402. Furthermore, while FIG. 23 shows three weak classifiers 403, in the present invention, the strong classifier 402 may include two, four or more weak classifiers 403.

[0062] The control unit 100 also includes a machine learning unit 102. In the first embodiment, the machine learning unit 102 uses a cascade classifier 401 to perform machine learning of the first projection image 301 including the crack 223 and the second projection image 302 without the crack 223 using the first projection image 301 and the second projection image 302, thereby generating a trained model 400 and storing it in a storage device. The functions of the determination unit 101 and the machine learning unit 102 are realized by an arithmetic processing device performing arithmetic processing in accordance with a computer program stored in the storage device.

[0063] (How to generate a trained model) Next, a method for generating a trained model will be described. Fig. 24 is a flowchart showing the flow of the method for generating a trained model according to the first embodiment. The method for generating a trained model is a method for generating a trained model 400 that determines the state of a wafer having a first surface 201 and a second surface 202, where the first surface 201 is a polished surface. As shown in Fig. 24, the method for generating a trained model includes an image acquisition step 501 and a learning step 502.

[0064] The image acquisition step 501 is a step in which light 44 from the light source 41 is irradiated onto the first surface 201 at a predetermined angle of incidence, and the reflected light 45 reflected by the first surface 201 is collected and photographed, thereby photographing the wafer 220 as a sample, thereby obtaining first projection images 301 (shown in Figures 16, 17, 18 and 19) which are multiple first learning images corresponding to images of the sample including cracks 223, and second projection images 302 (shown in Figures 20 and 21) which are multiple second learning images corresponding to images of the sample not including cracks.

[0065] In the image acquisition step 501, the inspection unit 30 places the wafer 220 on the temporary placement table 31, irradiates light 44 from the light source 41 onto the wafer 220 on the temporary placement table 31, and captures the reflected light 45 using the imaging unit 42 to acquire a predetermined number of first projected images 301 and second projected images 302. In the first embodiment, in the image acquisition step 501, 7000 first projected images 301 and 3000 second projected images 302 are acquired. However, there is no limit to the number of projected images to be acquired.

[0066] The learning step 502 is a step of generating a trained model 400 that, when a projection image 300 (corresponding to an image) of the wafer 220 is input, outputs a determination result indicating whether or not a crack 223 is formed in the wafer 220, by machine learning using the first projection image 301 and the second projection image 302. In the learning step 502, the machine learning unit 102 of the control unit 100 performs machine learning on the projection images 301 and 302. Specifically, in the learning step 502, the second projection image 302 acquired in the image acquisition step 501 and information indicating the absence of a crack 223 are input to the machine learning unit 102 of the control unit, and the first projection image 301 acquired in the image acquisition step 501 and information indicating the position of the crack 223 in the first projection image 301 are input to the machine learning unit 102 of the control unit.

[0067] Then, in the learning step 502, the machine learning unit 102 of the control unit 100 generates thresholds and coefficients for each weak classifier 403 and a second threshold for each strong classifier 402, generates a trained model 400, and stores it in a storage device. In the first embodiment, the trained model 400 generated by the trained model generation method uses a cascade classifier 401 that has trained the first projection image 301 (corresponding to an image) of the crack 223 using the first projection image 301 and the second projection image 302.

[0068] Next, a description will be given of the processing operation of the delamination apparatus 1 according to embodiment 1. First, in the delamination apparatus 1, the control unit 100 receives and stores processing conditions. When the second surface 202 of the ingot 200 on which the delamination layer 211 has been formed is placed on the holding table 10 and the control unit 100 receives an instruction to start the processing operation from an operator, the delamination apparatus 1 starts an inspection operation.

[0069] In the processing operation, the control unit 100 of the peeling apparatus 1 sucks and holds the ingot 200 on the holding surface 11 of the holding table 10. The control unit 100 controls the peeling unit 20, the moving unit 27, etc. to peel the wafer 220 from the ingot 200 and place the peeled wafer 220 on the temporary placement table 31.

[0070] In the processing operation, the control unit 100 of the delamination apparatus 1 controls the inspection unit 30 to form a projection image 300 of the first surface 201 of the wafer 220 placed on the temporary placement table 31. In the processing operation, the determination unit 101 of the control unit 100 uses the trained model 400 to determine whether or not a crack 223 is present in each projection image 300, and determines whether or not a crack 223 is present in the wafer 220 to be inspected.

[0071] As described above, the delamination apparatus 1 and the inspection unit 30 serving as an inspection device according to the first embodiment determine the presence or absence of a crack 223 of an undefined shape by the determination unit 101 of the control unit 100 using the trained model 400 configured by machine learning. As a result, the delamination apparatus 1 and the inspection unit 30 serving as an inspection device according to the first embodiment exhibit the effect of being able to detect a crack 223 of an undefined shape that has occurred in the wafer 220.

[0072] Furthermore, in the peeling apparatus 1 and the inspection unit 30 serving as an inspection device according to the first embodiment, the light source 41 of the projection image forming unit 40 irradiates the first surface 201 of the wafer 220 held on the temporary placement table 31 with visible light 44 at a predetermined angle of incidence 43, and the imaging unit 42 collects and captures reflected light 45 from the first surface 201 of the wafer 220 held on the temporary placement table 31 using a condenser lens 46, thereby forming a projection image 300 based on the so-called magic mirror principle. As a result, the peeling apparatus 1 and the inspection unit 30 serving as an inspection device according to the first embodiment have the effect of being able to easily capture an image of a crack 223 inside the wafer 220 using the magic mirror principle.

[0073] The peeling device 1 and inspection unit 30, which is an inspection device, according to the first embodiment utilize a cascade classifier 401 in which a trained model 400 determines the presence or absence of a crack 223 in a projection image 300 from a plurality of strong classifiers 402 one by one in turn, and when one of the plurality of strong classifiers 402 outputs a determination result that a crack 223 is not present in the projection image 300, the cascade classifier 401 outputs a determination result that a crack 223 is not present in the input projection image 300, without determining the presence or absence of a crack 223 from the strong classifier 402 that output the determination result that a crack 223 is not present. This makes it possible to speed up processing when a crack 223 is not present in the projection image 300.

[0074] [Embodiment 2] A delamination apparatus according to a second embodiment of the present invention will be described with reference to the drawings. Fig. 25 is a diagram illustrating the configuration of a trained model provided in a control unit of an inspection unit of the delamination apparatus according to the second embodiment. The delamination apparatus 1 according to the second embodiment is the same as the first embodiment, except that the trained model 400-2 is different, which is generated by the machine learning unit 302 of the control unit 100, stored in a storage device, and outputs a determination result indicating whether or not a crack 223 has formed inside the wafer 220.

[0075] 25, the trained model 400-2 of the peeling apparatus 1 and the inspection unit 30 according to the second embodiment is a neural network including an input layer 405 to which the projection images 300, 301, and 302 are input, and an output layer 406 that outputs a determination result. The trained model 400 is a trained model trained by deep learning, and is composed of variables and a model (configuration information of the neural network) that defines the neural network that constitutes the classifier trained by deep learning.

[0076] Deep learning is a type of machine learning technique, and in a narrow sense, is composed of, for example, a multi-layer neural network. In the second embodiment, the trained model 400 is a fully connected neural network model, and includes an input layer 405, a hidden layer 407, and an output layer 406. When a projection image 300 is input to the input layer 405, the trained model 400 performs calculations in the hidden layer 407 and outputs a determination result indicating whether or not a crack 223 is present in the projection image 300 from the output layer 406. In FIG. 25, the trained model 400 has two hidden layers 407, but the number of hidden layers 407 is not limited to two and can be any number.

[0077] In the method for generating the trained model 400 in the second embodiment, in a learning step 502, the machine learning unit 102 of the control unit 100 inputs the second projection image 302 acquired in the image acquisition step 501 and information indicating the absence of cracks 223 to the input layer 405, and the machine learning unit 102 of the control unit 100 inputs the first projection image 301 acquired in the image acquisition step 501 and information indicating the position of the crack 223 in the first projection image 301 to the input layer 405. Then, in the learning step 502, the machine learning unit 102 of the control unit 100 generates variables and generates a trained model 400-1, which is stored in a storage device.

[0078] In the processing operation of the peeling apparatus 1 of embodiment 2, as in embodiment 1, the judgment unit 101 of the control unit 100 uses the trained model 400-2 to determine whether or not a crack 223 is present in each projection image 300, and determines whether or not a crack 223 is present in the wafer 220 to be inspected.

[0079] The peeling device 1 and the inspection unit 30, which is an inspection device, according to embodiment 2 have the effect of being able to detect cracks 223 that have occurred in the wafer 220 because the judgment section 101 of the control unit 100 judges the presence or absence of cracks 223 with an undefined shape using a trained model 400-2 constructed by machine learning.

[0080] Furthermore, the peeling device 1 and the inspection unit 30, which is an inspection device, according to the second embodiment, like the first embodiment, have the effect of being able to easily photograph the crack 223 inside the wafer 220 using the principle of the so-called magic mirror to form the projection image 300.

[0081] [Modification] A delamination device according to a modification of the first and second embodiments of the present invention will be described with reference to the drawings. Fig. 26 is a side view schematically showing an example of the configuration of an inspection unit of a delamination device according to a modification of the first and second embodiments of the present invention. In Fig. 26, the same parts as those in the first embodiment are denoted by the same reference numerals, and their description will be omitted.

[0082] The inspection unit 30 of the peeling device 1 of the modified example is the same as that of embodiment 1, except that the second surface 202 of the ingot 200 is placed on the temporary placement table 31, light is irradiated from the light source 41 onto the first surface 201 of the ingot 200, the reflected light 45 is collected and captured by the imaging unit 42 to form projected images 300, 301, and 302. That is, in the modified example, a temporary placement table 31 holds the first surface 201 of the ingot 200 exposed, a light source 41 irradiates light 44 onto the first surface 201 of the ingot 200 held on the temporary placement table 31 at a predetermined angle of incidence 43, an imaging unit 42 collects and photographs the reflected light 45 reflected by the first surface 201 of the ingot 200, and forms a projection image 300 in which the unevenness caused on the first surface 201 by a crack 223 extending inside the ingot 200 is emphasized, and a judgment unit 101 compares the formed projection image 300 with preset conditions to judge the state of the ingot 200, and the judgment unit 101 is equipped with a trained model 400 configured by machine learning so as to output a judgment result indicating whether or not a crack has formed inside the ingot 200 when the projection image 300 acquired by the imaging unit 42 is input.

[0083] In addition, in a modified example, the method for generating a trained model is a method for generating a trained model 400 that determines the state of an ingot 200, in which the image acquisition step 501 is a step of irradiating light 44 from a light source 41 onto the first surface 201 at a predetermined angle of incidence 43, collecting and photographing the reflected light 45 reflected by the first surface 201, and photographing the ingot 200 as a sample, thereby obtaining a plurality of first projection images 301 corresponding to images of the sample including cracks and a plurality of second projection images 302 corresponding to images of the sample not including cracks, and the learning step 502 is a step of generating a trained model 400 that, when a projection image 300 of the ingot 200 is input, outputs a determination result indicating whether or not a crack 223 has formed in the ingot 200 by machine learning using the first projection images 301 and the second projection images 302.

[0084] In the example shown in FIG. 26, the control unit 100 stores the trained model 400 of the first embodiment, but in the present invention, the control unit 100 may store the trained model 400-2 of the second embodiment.

[0085] The peeling device 1 and the inspection unit 30, which is an inspection device, relating to the modified example have the effect of being able to detect cracks 223 that have occurred in the ingot 200 because the judgment section 101 of the control unit 100 determines the presence or absence of cracks 223 with an undefined shape using a trained model 400 constructed by machine learning.

[0086] Furthermore, the peeling device 1 and the inspection unit 30, which is an inspection device, according to the modified example have the effect of being able to easily photograph the crack 223 inside the ingot 200 using the so-called magic mirror principle to form a projection image 300, similar to embodiment 1.

[0087] The present invention is not limited to the above-described embodiment, and can be implemented in various modifications without departing from the gist of the present invention. [Explanation of symbols]

[0088] 1 Peeling device 10 Holding table 20 Peeling unit 30 Inspection unit (inspection equipment) 31 Temporary table (holding table) 41 Light source 42 Imaging unit 43 Angle of incidence 44 light 45 Reflected light (light) 100 control unit 101 Judgment section 200 ingots (semiconductor ingots) 201 Side 1 (One side) 202 Second side (other side) 211 Peeling layer 214 Modified Layer 215 Peeling crack 217 Laser Beam 220 wafers (peeled wafers) 221 Peeling surface (other surface) 223 Crack 300 projected images 301 First projection image (first learning image) 302 Second projection image (second learning image) 400,400-2 trained model 401 Cascade Classifier 405 Input Layer 406 Output Layer 501 Image Acquisition Steps 502 Learning Steps

Claims

1. An inspection apparatus for inspecting an ingot or wafer having one surface and another surface, the one surface being a polished surface, comprising: a holding table that holds the polished surface of the ingot or wafer while exposing it; a light source that irradiates the polished surface of the ingot or wafer held on the holding table with light at a predetermined angle of incidence; an imaging unit that collects and captures light reflected from the polished surface of the ingot or wafer, and forms a projected image that emphasizes irregularities that have occurred on the polished surface due to cracks extending into the ingot or wafer; a determination unit that compares the formed projection image with preset conditions to determine the state of the ingot or wafer, The determination unit a trained model configured by machine learning, which has a processor and a memory, and which outputs a determination result indicating whether or not the crack has formed inside the ingot or wafer when the projection image acquired by the imaging unit is input; The inspection device is characterized in that the trained model utilizes a cascade classifier that has trained the image of the crack using sample images consisting of the projection image including the crack and the projection image without the crack.

2. 2. The inspection device according to claim 1, wherein the wafer is a delamination wafer manufactured by irradiating the polished surface of an ingot with a laser beam to form a delamination layer consisting of a modified layer and delamination cracks extending from the modified layer along the crystal, and then delaminating the wafer from the ingot starting from the delamination layer.

3. A delamination apparatus for manufacturing wafers from an ingot having a delamination layer formed thereon, the delamination layer comprising a modified layer and delamination cracks extending from the modified layer along the crystal, by irradiating a laser beam onto an exposed polished surface, the delamination apparatus comprising: a holding table for holding an ingot; a separation unit for separating a wafer from the ingot held on the holding table; a temporary placement table for temporarily placing and holding the wafer separated by the separation unit with the polished surface thereof exposed; and an inspection unit for inspecting the wafer held on the temporary placement table; The inspection unit comprises: a light source that irradiates the polishing surface of the wafer held on the temporary placement table with light at a predetermined incident angle; an imaging unit that collects and captures light reflected from the polished surface of the wafer, and forms a projected image that emphasizes irregularities that have occurred on the polished surface due to cracks extending into the wafer; a determination unit that compares the formed projection image with preset conditions to determine the state of the wafer, The determination unit A peeling device having a processor and memory, and equipped with a trained model constructed by machine learning so as to output a determination result indicating whether or not a crack has formed inside the wafer when the projection image acquired by the imaging unit is input.

4. The peeling device described in claim 3, characterized in that the trained model utilizes a cascade classifier that has trained the image of the crack using sample images consisting of the projection image including the crack and the projection image without the crack.

5. The peeling device according to claim 3 , wherein the trained model is a neural network including an input layer to which the projected image is input and an output layer to which the determination result is output.

6. an inspection apparatus for inspecting an ingot or wafer having one surface and another surface, the one surface being a polished surface, the inspection apparatus comprising: a holding table for holding the polished surface of the ingot or wafer exposed; a light source for irradiating the polished surface of the ingot or wafer held on the holding table with light at a predetermined angle of incidence; an imaging unit for collecting and photographing the light reflected on the polished surface of the ingot or wafer, and forming a projected image in which irregularities caused by cracks extending into the ingot or wafer are emphasized; and a determination unit for determining the state of the ingot or wafer by comparing the formed projected image with preset conditions, the determination unit having a processor and memory, the determination unit having a trained model configured by machine learning to output a determination result indicating whether or not the crack has formed inside the ingot or wafer when the projected image acquired by the imaging unit is input, the method comprising: generating a trained model for determining the state of the ingot or wafer, the inspection apparatus comprising: an image acquisition step in which the inspection device places the ingot or the wafer on the holding table, irradiates light from the light source onto the polished surface of the ingot or the wafer on the holding table at a predetermined angle of incidence, and collects and photographs the light reflected from the polished surface with the imaging unit, thereby acquiring a plurality of first learning images corresponding to images of the sample including cracks and a plurality of second learning images corresponding to images of the sample not including cracks; a learning step in which a machine learning unit that is input with the second learning image and information indicating the absence of the crack, and the first learning image and information indicating the position of the crack, uses machine learning using the first learning image and the second learning image to generate a trained model that, when an image of an ingot or wafer is input, outputs a determination result indicating whether or not a crack has formed in the ingot or wafer.

7. The method for generating a trained model described in claim 6, characterized in that the trained model utilizes a cascade classifier in which the crack image is trained using the first training image and the second training image.

8. The method for generating a trained model described in claim 6, characterized in that the trained model is a neural network including an input layer to which the first training image and the second training image are input, and an output layer to which the judgment result is output.

Images