Inspecting apparatus, peeling apparatus, and learned model generating method

1 . An inspecting apparatus for inspecting a semiconductor ingot having one surface and another surface, the one surface being a polished surface resulting from polishing, the inspecting apparatus comprising:

a holding table configured to hold the semiconductor ingot with the polished surface of the semiconductor ingot exposed;

a light source configured to irradiate the polished surface of the semiconductor ingot held by the holding table with light at a predetermined incidence angle;

an imaging unit configured to condense and photograph reflected light reflected by the polished surface of the semiconductor ingot, and form a photographed image emphasizing unevenness produced on the polished surface by a crack extending in the semiconductor ingot; and

a determining section configured to determine a state of the semiconductor ingot by comparing the formed photographed image with a preset condition;

the determining section including a processor and a memory, and including a learned model configured by machine learning to output, when the photographed image obtained by the imaging unit is input, a determination result indicating whether or not the crack is formed in the semiconductor ingot;

wherein the learned model utilizes a cascade classifier comprising a plurality of strong discriminators each configured to include a plurality of weak discriminators, each of the plurality of weak discriminators being configured to determine whether a feature quantity is present in the photographed image and assign an output value to the feature quantity; and

wherein, the feature quantity is determined by identifying an amount of light received by a particular pixel among a plurality of pixels composing the photographed image and comparing the amount of light received with a threshold value, a result of the comparison indicating a presence of the crack in the semiconductor ingot.

2 . The inspecting apparatus according to claim 1 , wherein

the cascade classifier is made to learn an image of the crack on a basis of sample images including a photographed image including the crack and a photographed image without the crack.

3 . The inspecting apparatus according to claim 1 , wherein

the learned model is a neural network including an input layer to which the photographed image is input and an output layer that outputs the determination result.

4 . The inspecting apparatus according to claim 1 , wherein

the learned model further includes a determiner configured to calculate a total value based on the output value provided by each of the plurality of weak discriminators and a coefficient generated by the learned model, and to analyze the total value to determine whether the total value exceeds a second threshold.

5 . A peeling apparatus for manufacturing, with a peel-off layer as a starting point, a wafer from a semiconductor ingot in which the peel-off layer including a modified layer and a crack extending from the modified layer along a c-plane is formed by irradiating an exposed polished surface of the semiconductor ingot with a laser beam, the peeling apparatus comprising:

a holding table configured to hold the semiconductor ingot;

a peeling unit configured to peel off the wafer from the semiconductor ingot held on the holding table;

a temporary placement table on which the wafer peeled off by the peeling unit is temporarily placed; and

an inspecting unit configured to inspect the wafer held on the temporary placement table;

the inspecting unit including

a light source configured to irradiate the polished surface of the semiconductor ingot held by the holding table with light at a predetermined incidence angle,

an imaging unit configured to condense and photograph reflected light reflected by the polished surface of the semiconductor ingot, and form a photographed image emphasizing unevenness produced on the polished surface by a crack extending in the semiconductor ingot, and

a determining section configured to determine a state of the semiconductor ingot by comparing the formed photographed image with a preset condition,

the determining section including a processor and a memory, and including a learned model configured by machine learning to output, when the photographed image obtained by the imaging unit is input, a determination result indicating whether or not the crack is formed in the semiconductor ingot,

wherein the learned model utilizes a cascade classifier comprising a plurality of strong discriminators each configured to include a plurality of weak discriminators, each of the plurality of weak discriminators being configured to determine whether a feature quantity is present in the photographed image and assign an output value to the feature quantity, and

wherein, the feature quantity is determined by identifying an amount of light received by a particular pixel among a plurality of pixels composing the photographed image and comparing the amount of light received with a threshold value, a result of the comparison indicating a presence of the crack in the semiconductor ingot.

6 . A learned model generating method for generating a learned model configured to determine a state of a semiconductor ingot having one surface and another surface, the one surface being a polished surface resulting from polishing,

the learned model generating method being performed by an inspecting apparatus including a holding table configured to hold the semiconductor ingot with the polished surface of the semiconductor ingot exposed;

a light source configured to irradiate the polished surface of the semiconductor ingot held by the holding table with light at a predetermined incidence angle;

an imaging unit configured to condense and photograph reflected light reflected by the polished surface of the semiconductor ingot, and form a photographed image emphasizing unevenness produced on the polished surface by a crack extending in the semiconductor ingot; and

a determining section configured to determine a state of the semiconductor ingot by comparing the formed photographed image with a preset condition, the determining section having a processor and a memory,

the learned model generating method comprising:

an image obtaining step of obtaining a plurality of first learning images corresponding to an image of a sample including a crack and a plurality of second learning images corresponding to an image of the sample including no crack, by photographing the semiconductor ingot as the sample by irradiating the polished surface with light from a light source at a predetermined incidence angle and condensing and photographing the light reflected by the polished surface; and

a learning step of generating, by machine learning using the first learning images and the second learning images, the learned model configured to output, when an image of the semiconductor ingot is input, the determination result indicating whether or not the crack is formed in the semiconductor ingot,

wherein the learned model utilizes a cascade classifier comprising a plurality of strong discriminators each configured to include a plurality of weak discriminators, each of the plurality of weak discriminators configured to determine whether a feature quantity is present in the photograph of the semiconductor ingot and to assign an output value to the feature quantity, and

wherein, the feature quantity is determined by identifying an amount of light received by a particular pixel among a plurality of pixels composing the photographed image and comparing the amount of light received with a threshold value, a result of the comparison indicating a presence of the crack in the semiconductor ingot.

7 . The learned model generating method according to claim 6 , wherein

the cascade classifier is made to learn an image of the crack on a basis of the first learning images and the second learning images.

8 . The learned model generating method according to claim 7 , wherein

the learned model is a neural network including an input layer to which the first learning images and the second learning images are input and an output layer that outputs the determination result.