WO2024224567A1 - Defect inspecting device - Google Patents

Abstract

This defect inspecting device for inspecting a sample for a defect on the basis of light from the sample comprises one or more sensors for detecting light from the sample, and a signal processing device for processing input signals from the sensors, wherein the signal processing device: calculates a similarity between an observation vector based on the input signals from the sensors and a model vector relating to the defect; calculates a detection signal obtained by non-linearly changing the intensity of the input signal corresponding to the observation vector on the basis of the similarity; and determines a defect of the sample on the basis of the detection signal.

Description

Defect Inspection Equipment

The present invention relates to a defect inspection device that inspects sample surfaces and outputs the location, type, dimensions, etc. of defects.

In order to improve the yield of products such as semiconductor substrates and thin film substrates, the semiconductor substrates and thin film substrates produced on the production line are used as samples, and the sample surfaces are inspected for defects such as foreign bodies and dents. A defect inspection device used for this inspection is known that simultaneously detects scattered light from the sample surface with multiple sensors at different positions, and obtains detailed data on the position, shape, size, etc. of defects (see Patent Document 1, etc.).

Patent No. 5564807

For example, extremely small foreign matter of 15 nm or less may be mixed into various solutions used in semiconductor manufacturing processes. In recent years, there has been an increasing demand for defect inspection devices to be able to detect such extremely small defects. The defect inspection device of Patent Document 1 simultaneously detects scattered light from an illumination spot using multiple detection systems with different orientations relative to the illumination spot, thereby obtaining a large amount of data on defects. Because the amount of scattered light from a defect is proportional to approximately the sixth power of the defect size, the amount of detected light from the defect drops sharply as the defect being inspected becomes smaller.

Since the scattered light from tiny critical defects of 15 nm or less is weak, the roughness scattered light caused by minute irregularities on the sample surface can be an obstacle to the detection of these tiny critical defects. The roughness scattered light obtained by shining illumination light on a polished mirror wafer has statistical variation, and is observed as shot noise when detected by a sensor. As a result, the detection light from the defect is lost in the noise, making high-speed inspection difficult.

In response to this, the technology in Patent Document 1 discloses a method of scanning a sample with laser light, removing low-frequency components other than those generated by defects from the detection signal of light from the illumination spot on the sample surface, and calculating the size of the defect on the sample surface. Specifically, a specified monitoring time width is calculated based on the half-width of the defect detection signal obtained from the rotation speed of the sample stage, etc., and a signal whose intensity is above a threshold for a period of time equal to or greater than the specified monitoring time width is extracted as a defect detection signal, and other low-level noise signals, etc. are removed.

However, this prior art does not take into consideration the statistical variation of the signal itself due to a decrease in the amount of scattered light. The amount of scattered light from extremely small defects is often lower than the amount of roughness scattered light from the sample surface. The amount of light from roughness on the sample surface is also small, and the signal-to-noise ratio of the shot noise detected by the sensor is proportional to the square root of the amount of detected light, so the amount of detected light from minute defects is noisy. In the prior art method of discriminating defect detection signals by simply comparing signal intensity with a threshold value, the detection signal to be discriminated needs to be bright enough to be able to be determined to be a defect detection signal, and it is difficult to discriminate tiny, noisy defects whose brightness difference with the roughness scattered light is unclear.

The object of the present invention is to provide a defect inspection device that can detect minute and noisy defects with high sensitivity.

In order to achieve the above object, the present invention provides a defect inspection device that inspects a sample for defects based on light from the sample, the defect inspection device comprising one or more sensors that detect the light from the sample, and a signal processing device that processes an input signal from the sensor, the signal processing device calculates a similarity between an observation vector based on the input signal from the sensor and a model vector related to the defect, calculates a detection signal by nonlinearly changing the intensity of the input signal corresponding to the observation vector based on the similarity, and determines defects in the sample based on the detection signal.

The present invention makes it possible to detect minute and noisy defects with high sensitivity.

1 is a schematic diagram of a configuration example of a defect inspection apparatus according to a first embodiment of the present invention; FIG. 4 is a schematic diagram showing an example of a scanning trajectory of a sample by a scanning device provided in the defect inspection device according to the first embodiment of the present invention; FIG. 5 is a schematic diagram showing another example of the scanning trajectory of the sample by the scanning device provided in the defect inspection device according to the first embodiment of the present invention; FIG. 1 is a schematic diagram showing an attenuator included in a defect inspection apparatus according to a first embodiment of the present invention; FIG. 1 is a schematic diagram showing the positional relationship between the optical axis of illumination light guided obliquely to a sample surface by a scattered light illumination system provided in a defect inspection apparatus according to a first embodiment of the present invention and the illumination intensity distribution shape, in a cross section of the sample cut at the incident plane of the illumination light incident on the sample. FIG. 1 is a schematic diagram showing the positional relationship between the optical axis of illumination light guided obliquely to a sample surface by a scattered light illumination system provided in a defect inspection apparatus according to a first embodiment of the present invention, and the illumination intensity distribution shape, in a cross section of the sample cut by a plane that is perpendicular to the plane of incidence of the illumination light incident on the sample and includes the normal to the sample surface. FIG. 2 is a top view of an opening for collecting scattered light by a scattered light detection system provided in the defect inspection apparatus according to the first embodiment of the present invention; FIG. 1 is a configuration diagram of a scattered light detection system provided in a defect inspection apparatus according to a first embodiment of the present invention, which irradiates light onto a sample from a normal direction and collects scattered light from the sample surface. A top view of FIG. 8. A top view of the light intensity distribution of scattered light generated by oblique incidence illumination of a micro defect. A top view of the light intensity distribution of roughness scattered light from a sample surface caused by oblique incidence illumination of the sample surface. FIG. 1 is a processing block diagram of a main part of a signal processing device provided in a defect inspection device according to a first embodiment of the present invention; FIG. 1 is an explanatory diagram of the similarity with a model vector in the signal processing device according to the first embodiment of the present invention. FIG. 11 is a diagram showing an example of a gain based on the similarity to a model vector in the signal processing device according to the first embodiment of the present invention. 1 is a processing block diagram of a main part of a signal processing device provided in a signal processing device according to a second embodiment of the present invention; FIG. 11 is an explanatory diagram of the similarity with a model vector in the signal processing device according to the second embodiment of the present invention. 11 is a processing block diagram of a main part of a signal processing device according to a third embodiment of the present invention; 13 is a processing block diagram of a main part of a signal processing device provided in a signal processing device according to a fourth embodiment of the present invention; FIG. 13 is a schematic diagram of a scattered light imaging detection unit provided in a defect inspection apparatus according to a fifth embodiment of the present invention; A diagram illustrating the change in image spread of a scattered light imaging detection system accompanying a change in the working distance in a defect inspection device. An explanatory diagram of the change in the spread of the image of illumination light on the sample surface due to the change in the working distance in a defect inspection device. 13 is a processing block diagram of a main part of a signal processing device provided in a signal processing device according to a fifth embodiment of the present invention; FIG. 13 is an explanatory diagram of overlap scanning of illumination according to the sixth embodiment of the present invention and a model vector based on the overlap scanning. 13 is a processing block diagram of a main part of a signal processing device provided in a signal processing device according to a sixth embodiment of the present invention; 13 is a processing block diagram of a main part of a signal processing device provided in a signal processing device according to a sixth embodiment of the present invention; A processing block diagram of a main part of a signal processing device provided in a signal processing device according to a seventh embodiment of the present invention. A processing block diagram of a main part of a signal processing device provided in a signal processing device according to a seventh embodiment of the present invention. FIG. 23 is an explanatory diagram of the similarity with a model vector according to the seventh embodiment of the present invention. FIG. 13 is a schematic diagram of a configuration example of a defect inspection apparatus according to an eighth embodiment of the present invention; FIG. 13 is an explanatory diagram of an inspection target sample according to the eighth embodiment of the present invention. 13 is a processing block diagram of a main part of a signal processing device according to an eighth embodiment of the present invention. FIG. 23 is an explanatory diagram of the similarity with a model vector in the signal processing device according to the eighth embodiment of the present invention.

The following describes an embodiment of the present invention using the drawings.

The defect inspection device described in the following embodiments as an application of the present invention is used for defect inspection of the surface of a sample (wafer) performed during the manufacturing process of, for example, semiconductors. The defect inspection device according to each embodiment is suitable for quickly detecting minute defects and acquiring data on the number, position, size, and type of defects. The present invention can be applied to defect inspection devices of various types, such as an illumination spot scanning type laser scattering type, an imaging type laser scattering type, or a scanning electron microscope.

First Embodiment
- Defect inspection equipment -
FIG. 1 is a schematic diagram of a configuration example of a defect inspection apparatus according to a first embodiment of the present invention. The defect inspection apparatus 100 according to this embodiment inspects a sample 1, and detects and inspects defects such as foreign matter and dents on the sample 1, particularly defects of a type according to the inspection purpose, based on light from the surface of the sample 1 (hereinafter referred to as the sample surface). A representative example of the sample 1 is a disk-shaped semiconductor silicon wafer having a flat surface on which no pattern is formed. The defect inspection apparatus 100 includes a stage ST, a scattered light illumination system A, a plurality (n) of scattered light detection systems B1-Bn, a signal processing device D, a control device E1, a user interface E2, a monitor E3, and a secondary storage device DB. The scattered light detection systems B1-Bn include sensors C1P-CnP and C1S-CnS, respectively. When the scattered light detection system Bi is described, it refers to the i-th scattered light detection system Bi. When the sensor CiP is described, it refers to a sensor that detects P-polarized light of the scattered light detection system Bi. Similarly, when sensor CiS is mentioned, it refers to a sensor that detects S-polarized light of scattered light detection system Bi.

-stage-
The stage ST includes a sample stage ST1 and a scanning device ST2. The sample stage ST1 is a stage that supports the sample 1. The scanning device ST2 is a device that drives the sample stage ST1 to change the relative position between the sample 1 and the scattered light illumination system A, and includes a translation stage, a rotation stage, and a Z stage, which are not shown in detail. The translation stage supports the rotation stage via the Z stage, and the sample stage ST1 is supported on the rotation stage. The translation stage translates in the horizontal direction together with the rotation stage, and the rotation stage rotates around an axis that extends vertically. The Z stage serves to adjust the height of the sample surface.

Figure 2 is a schematic diagram showing the scanning trajectory of the sample 1 by the scanning device ST2. As will be described later, the illumination spot BS formed on the sample surface by the illumination light emitted from the scattered light illumination system A has an illumination intensity distribution that is long in one direction as shown in the figure. The long axis direction of the illumination spot BS is s2, and the direction intersecting the long axis (for example, the short axis direction perpendicular to the long axis) is s1. The sample 1 rotates with the rotation of the rotating stage, and the illumination spot BS is scanned in the s1 direction relative to the sample surface, and the sample 1 moves in the horizontal direction with the translation stage translation, and the illumination spot BS is scanned in the s2 direction relative to the sample surface. By moving the sample 1 while rotating due to the operation of the scanning device ST2, the illumination spot BS moves in a spiral trajectory from the center to the outer edge of the sample 1 as shown in Figure 2, and the entire surface of the sample 1 is scanned. The illumination spot BS moves in the s2 direction by a distance less than the length of the illumination spot BS in the s2 direction during one rotation of the sample 1.

It is also possible to apply a scanning device having a configuration in which, instead of a rotation stage, another translation stage whose movement axis extends in a direction intersecting with the movement axis of the translation stage in a horizontal plane is provided. In this case, as shown in FIG. 3, the illumination spot BS scans the sample surface by folding over a linear trajectory instead of a spiral trajectory. Specifically, the first translation stage is driven in translation at a constant speed in the s1 direction, the second translation stage is driven in the s2 direction by a predetermined distance (for example, a distance equal to or less than the length of the illumination spot BS in the s2 direction), and then the first translation stage is turned back in the s1 direction and driven in translation again. As a result, the illumination spot BS repeats linear scanning in the s1 direction and movement in the s2 direction to scan the entire surface of the sample 1. Compared to this scanning method, the spiral scanning method shown in FIG. 2 does not involve reciprocating motion, and is therefore advantageous in performing sample inspection in a short time.

- Scattered light illumination system -
The scattered light illumination system A shown in Fig. 1 is configured to include a group of optical elements for irradiating a desired illumination light onto a sample 1 placed on a sample stage ST1. As shown in Fig. 1, this scattered light illumination system A includes a laser light source A1, an attenuator A2, an emitted light adjustment unit A3, a beam expander A4, a polarization control unit A5, a focusing optical unit A6, reflection mirrors A7-A10, etc.

Laser light source The laser light source A1 is a unit that emits a laser beam as illumination light. When the defect inspection device 100 detects minute defects near the sample surface, a laser light source A1 that emits a high-power laser beam with an output of 2 W or more in ultraviolet or vacuum ultraviolet with a short wavelength (wavelength of 355 nm or less) that does not easily penetrate into the inside of the sample 1 is used. The diameter of the laser beam emitted by the laser light source A1 is typically about 1 mm. When the defect inspection device 100 detects defects inside the sample 1, a laser light source A1 that emits a visible or infrared laser beam with a long wavelength that easily penetrates into the inside of the sample 1 is used.

Attenuator FIG. 4 is a schematic diagram showing the attenuator A2. The attenuator A2 is a unit that attenuates the light intensity of the illumination light from the laser light source A1. In this embodiment, the attenuator A2 is a combination of a first polarizing plate A2a, a half-wave plate A2b, and a second polarizing plate A2c. The half-wave plate A2b is configured to be rotatable around the optical axis of the illumination light. The illumination light incident on the attenuator A2 is converted into linearly polarized light by the first polarizing plate A2a, and then the polarization direction is adjusted to the slow axis azimuth angle of the half-wave plate A2b and passes through the second polarizing plate A2c. The light intensity of the illumination light can be attenuated at an arbitrary ratio by adjusting the azimuth angle of the half-wave plate A2b. If the linear polarization degree of the illumination light incident on the attenuator A2 is sufficiently high, the first polarizing plate A2a can be omitted. Note that the attenuator A2 is not limited to the configuration illustrated in FIG. 4, but can also be configured using an ND filter having a gradation density distribution, and can also be configured in such a way that the attenuation effect can be adjusted by combining multiple ND filters with different densities.

Emitted light adjustment unit The emitted light adjustment unit A3 shown in FIG. 1 is a unit that adjusts the angle of the optical axis of the illumination light attenuated by the attenuator A2, and in this embodiment, it is configured to include multiple reflecting mirrors A3a and A3b. The reflecting mirrors A3a and A3b sequentially reflect the illumination light, but in this embodiment, the incident and exit surfaces of the illumination light to the reflecting mirror A3a are configured to be perpendicular to the incident and exit surfaces of the illumination light to the reflecting mirror A3b. The incident and exit surfaces are surfaces that include the optical axis incident on the reflecting mirror and the optical axis emitted from the reflecting mirror. For example, if a three-dimensional XYZ orthogonal coordinate system is defined and the illumination light is incident on the reflecting mirror A3a in the +X direction, the illumination light is redirected in the +Y direction by the reflecting mirror A3a and then in the +Z direction by the reflecting mirror A3b, although this is different from the schematic FIG. 1. In this example, the incident and exit surfaces of the illumination light to the reflecting mirror A3a are the XY plane, and the incident and exit surfaces to the reflecting mirror A3b are the YZ plane. The reflecting mirrors A3a and A3b are provided with a mechanism for translating and tilting the reflecting mirrors A3a and A3b, respectively, although not shown. The reflecting mirrors A3a and A3b translate, for example, in the incident or outgoing direction of the illumination light with respect to themselves, and tilt around the normal to the incident and outgoing surfaces. This allows, for example, the offset amount and angle in the XZ plane and the offset amount and angle in the YZ plane of the optical axis of the illumination light emitted in the +Z direction from the outgoing light adjustment unit A3 to be independently adjusted. In this example, a configuration using two reflecting mirrors A3a and A3b is illustrated, but a configuration using three or more reflecting mirrors may be used.

Beam Expander The beam expander A4 is a unit that expands the diameter of the luminous flux of the incident illumination light, and has a plurality of lenses A4a and A4b. An example of the beam expander A4 is a Galilean type that uses a concave lens as the lens A4a and a convex lens as the lens A4b. The beam expander A4 is provided with a mechanism for adjusting the distance between the lenses A4a and A4b (zoom mechanism), and the expansion rate of the luminous flux diameter changes by adjusting the distance between the lenses A4a and A4b. The expansion rate of the luminous flux diameter by the beam expander A4 is, for example, about 5-10 times. In this case, if the beam diameter of the illumination light emitted from the laser light source A1 is 1 mm, the beam system of the illumination light is expanded to about 5-10 mm. If the illumination light incident on the beam expander A4 is not a parallel luminous flux, collimation (quasi-parallelization of the luminous flux) is also possible in addition to the luminous flux diameter by adjusting the distance between the lenses A4a and A4b. However, the collimation of the light beam may be performed by disposing a collimating lens upstream of the beam expander A4 separately from the beam expander A4.

Beam expander A4 is installed on a translation stage with two or more axes (two degrees of freedom) and is configured so that its position can be adjusted so that its center coincides with the incident illumination light. Beam expander A4 also has a swing angle adjustment function with two or more axes (two degrees of freedom) so that the incident illumination light coincides with the optical axis.

Polarization control unit The polarization control unit A5 is an optical system that controls the polarization state of the illumination light, and is configured to include a 1/2 wavelength plate A5a and a 1/4 wavelength plate A5b. For example, when performing oblique incidence illumination by inserting a reflecting mirror A7 described later into the optical path, the amount of scattered light from defects on the sample surface is increased by making the illumination light P-polarized by the polarization control unit A5 compared to polarization other than P polarization. If scattered light (called haze) from minute irregularities on the surface of the sample itself hinders the detection of minute defects, the haze can be reduced by making the illumination light S-polarized compared to polarization other than S polarization. The polarization control unit A5 can also make the illumination light circularly polarized or 45-degree polarized, which is intermediate between P polarization and S polarization.

Reflecting Mirror As shown in Fig. 1, the reflecting mirror A7 can be moved in parallel in the direction of the arrow by a driving mechanism (not shown) to enter and exit the optical path of the illumination light toward the sample 1, thereby switching the incidence path of the illumination light to the sample 1. By inserting the reflecting mirror A7 into the optical path, the illumination light emitted from the polarization control unit A5 is reflected by the reflecting mirror A7 and enters the sample 1 obliquely via the focusing optical unit A6 and the reflecting mirror A8 as described above. On the other hand, when the reflecting mirror A7 is removed from the optical path, the illumination light emitted from the polarization control unit A5 enters the sample 1 perpendicularly via the reflecting mirrors A9 and A10, the polarization control unit B2', the reflecting mirror B1', and the scattered light detection system B3. The polarization control unit B2' includes a 1/2 wavelength plate Ba' and a 1/4 wavelength plate Bb', similar to the polarization control unit A5.

Figures 5 and 6 are schematic diagrams showing the positional relationship between the optical axis of the illumination light guided obliquely to the sample surface by the scattered light illumination system A and the illumination intensity distribution shape. Figure 5 shows a schematic cross-section of the sample 1 cut at the plane of incidence of the illumination light incident on the sample 1. Figure 6 shows a schematic cross-section of the sample 1 cut at a plane that is perpendicular to the plane of incidence of the illumination light incident on the sample 1 and includes the normal to the sample surface. The plane of incidence is a plane that includes the optical axis OA of the illumination light incident on the sample 1 and the normal to the sample surface. Note that Figures 5 and 6 show only a portion of the scattered light illumination system A, and for example, the exit light adjustment unit A3 and the reflecting mirrors A7 and A8 are not shown.

When the reflecting mirror A7 is inserted into the optical path, the illumination light emitted from the laser light source A1 is collected by the focusing optical unit A6, reflected by the reflecting mirror A8, and obliquely incident on the sample 1. In this way, the scattered light illumination system A is configured to allow illumination light to be incident on the sample 1 from a direction oblique to the normal line of the sample surface. This oblique incidence illumination has its light intensity adjusted by the attenuator A2, its light beam diameter adjusted by the beam expander A4, and its polarization adjusted by the polarization control unit A5, so that the illumination intensity distribution is uniform within the incident surface. As shown in the illumination intensity distribution (illumination profile) LD1 in Figure 5, the illumination spot formed on the sample 1 has a Gaussian light intensity distribution in the s2 direction, and the length of the beam width l1 defined at 13.5% of the peak is, for example, about 25 μm to 4 mm.

In the plane perpendicular to the incident surface and the sample surface, the illumination spot has a light intensity distribution with weak intensity at the periphery relative to the center of the optical axis OA, as shown in the illumination intensity distribution (illumination profile) LD2 in FIG. 6. Specifically, the intensity distribution is a Gaussian distribution reflecting the intensity distribution of the light incident on the focusing optical unit A6, or an intensity distribution similar to a first-order Bessel function of the first kind or a sinc function reflecting the aperture shape of the focusing optical unit A6. The length l2 of the illumination intensity distribution in the plane perpendicular to the incident surface and the sample surface is shorter than the beam width l1 shown in FIG. 5 in order to reduce haze generated from the sample surface, and is set to, for example, about 1.0 μm to 20 μm. This length l2 of the illumination intensity distribution is the length of the area having an illumination intensity of 13.5% or more of the maximum illumination intensity in the plane perpendicular to the incident surface and the sample surface.

The angle of incidence of the oblique incidence illumination on the sample 1 (the inclination angle of the incident optical axis with respect to the normal to the sample surface) is adjusted to an angle suitable for detecting minute defects by adjusting the positions and angles of the reflecting mirrors A7 and A8. The angle of the reflecting mirror A8 is adjusted by the adjustment mechanism A8a. For example, the larger the angle of incidence of the illumination light on the sample 1 (the smaller the illumination elevation angle, which is the angle between the sample surface and the incident optical axis), the weaker the haze that becomes noise in the scattered light from minute foreign objects on the sample surface, making it more suitable for detecting minute defects. From the viewpoint of suppressing the effect of haze on the detection of minute defects, it is preferable to set the incidence angle of the illumination light to, for example, 75 degrees or more (elevation angle 15 degrees or less). On the other hand, in the case of oblique incidence illumination, the smaller the illumination incidence angle, the greater the absolute amount of scattered light from minute foreign objects, so from the viewpoint of aiming to increase the amount of scattered light from defects, it is preferable to set the incidence angle of the illumination light to, for example, 60 degrees or more and 75 degrees or less (elevation angle 15 degrees or more and 30 degrees or less).

- Scattered light detection system -
The scattered light detection systems B1-Bn are units that collect and detect scattered light from the illumination spot BS on the sample surface, and are configured to include a plurality of optical elements including a collecting lens (objective lens). This optical system detects scattered light from the sample surface and performs laser scattering detection. The n in the scattered light detection systems Bn represents the number of scattered light detection systems, and the defect inspection device 100 of this embodiment will be described with an example in which 13 sets of scattered light detection systems are provided (n=13). The i-th scattered light detection system Bi is configured to include an objective lens Bi1, a half-wave plate Bi2 that can be rotated around the optical axis of the scattered light detection system Bi by a rotation mechanism not shown, and a polarizing beam splitter Bi3. The polarization direction is controlled by rotating the half-wave plate Bi2, and the polarizing beam splitter Bi3 separates the light into two desired lights whose polarization directions are orthogonal to each other. The P-polarized light that travels straight through the polarizing beam splitter Bi3 is detected by a sensor CiP, and the S-polarized light that is reflected by the polarizing beam splitter Bi3 is detected by a sensor CiS.

FIG. 7 shows the openings through which the scattered light detection system B1-B13 collects scattered light as viewed from above, and corresponds to the arrangement of the objective lenses in the scattered light detection system B1-B13. In the following explanation, the incident direction of the oblique incidence illumination on the sample 1 is used as a reference, and the direction of travel of the incident light with respect to the illumination spot BS on the sample surface as viewed from above (to the right in FIG. 7) is treated as the forward direction, and the opposite direction (to the left in FIG. 7) is treated as the backward direction. Therefore, the lower side in FIG. 7 is the right side and the upper side is the left side with respect to the illumination spot BS.

Each objective lens of the scattered light detection system B1-B13 is arranged along the upper half of a hemisphere (celestial sphere) centered on the illumination spot BS on the sample 1. This hemisphere is divided into 13 regions, apertures L1-L6, H1-H6, and V, and the scattered light detection systems B1-B13 each collect and focus the scattered light at the corresponding aperture.

Aperture V is an area that overlaps the zenith and is located directly above the illumination spot BS formed on the sample surface.

Apertures L1-L6 are equal divisions of a ring-shaped region that surrounds 360 degrees around the illumination spot BS at a low angle, and are lined up in the order of apertures L1, L2, L3, L4, L5, and L6 in a counterclockwise direction from the direction of incidence of the oblique incidence illumination when viewed from above. Of these apertures L1-L6, apertures L1-L3 are located to the right of the illumination spot BS, aperture L1 is located to the rear right of the illumination spot BS, aperture L2 is located to the right, and aperture L3 is located to the front right. Apertures L4-L6 are located to the left of the illumination spot BS, aperture L4 is located to the front left of the illumination spot BS, aperture L5 is located to the left, and aperture L6 is located to the rear left.

The remaining apertures H1-H6 are equal divisions of a ring-shaped region that surrounds the illumination spot BS 360 degrees at high angles (between apertures L1-L6 and aperture V), and are arranged in the order of apertures H1, H2, H3, H4, H5, and H6 in a counterclockwise direction from the direction of incidence of the oblique incidence illumination as viewed from above. Compared to the low-angle apertures L1-L6, the high-angle apertures H1-H6 are positioned 30 degrees apart as viewed from above. Of the apertures H1-H6, aperture H1 is located behind the illumination spot BS, and aperture H4 is located in front. Apertures H2 and H3 are located to the right of the illumination spot BS, aperture H2 is located to the rear right of the illumination spot BS, and aperture H3 is located to the front right. Apertures H5 and H6 are located to the left of the illumination spot BS, aperture H5 is located to the front left of the illumination spot BS, and aperture H6 is located to the rear left.

In Figure 1, scattered light incident on the scattered light detection system Bi is collected and guided to the corresponding sensors CiP, CiS. When comparing Figure 1 with Figure 7, for example, the scattered light detection system B1 in Figure 1 can be treated as an example of an optical system that collects scattered light at the aperture L4 in Figure 7, the scattered light detection system B2 at the aperture L6, and the scattered light detection system B3 at the aperture V.

Figure 8 is a schematic diagram of the scattered light detection system B3 into which the scattered light emitted from the sample 1 in the normal direction is incident, and Figure 9 is a plan view of Figure 8 viewed from above. The scattered light detection system B3 is composed of a condenser lens (objective lens) B3a and an imaging lens B3b, and the scattered light collected by the condenser lens B3a is detected by sensors C3P, C3S via the imaging lens B3b, 1/2 wavelength plate B32, and polarizing beam splitter B33. This is the same as the other scattered light detection systems B1, B2, B4, etc. The scattered light detection system B3 differs from the other scattered light detection systems in that a reflecting mirror B1' is placed at the position of its own pupil between the condenser lens B3a and the imaging lens B3b. As mentioned above, during epi-illumination, illumination light is incident on the sample 1 from the normal direction via the reflecting mirror B1'. Therefore, the condenser lens B3a of the scattered light detection system B3 also serves as a condenser lens that guides the epi-illumination to the sample 1.

As mentioned above, the illumination spot BS has a long linear intensity distribution in the s2 direction. As shown in Figure 9, the reflecting mirror B1' is longer than the illumination spot BS in the minor axis direction (s1 direction) of the linear illumination spot BS when viewed from the side of the sensors C3P, C3S, and shorter than the illumination spot BS in the major axis direction (s2 direction) of the illumination spot BS. The reflecting mirror B1' is positioned at the pupil position of the focusing lens B3a, and the directly reflected light incident on the focusing lens B3a from the sample 1 is reflected by the reflecting mirror B1', and the scattered light that is not reflected here is guided to the imaging lens B3b.

-Sensor-
The sensors C1P-CnP and C1S-CnS are sensors that detect light from the sample 1, and are single-pixel point sensors that convert the scattered light collected through the corresponding aperture into an electrical signal and output it. For the sensors C1P-CnP and C1S-CnS, a photomultiplier tube, SiPM (silicon photomultiplier tube), or the like that photoelectrically converts a weak signal with high gain can be used. Typically, compared to a photomultiplier tube, a SiPM is compact and robust against magnetic noise, while a photomultiplier tube is superior in terms of signal linearity. Therefore, both of these may be applied simultaneously. In this embodiment, a SiPM is applied to the sensor C1P-CnP, and a photomultiplier tube is applied to the sensor C1S-CnS. The polarization direction of the light incident on the sensors CiP and CiS can be adjusted by setting the rotation angle of the half-wave plate Bi2. For example, the types of defects that require highly sensitive detection differ depending on the inspection process, but by adjusting the rotation angle of the polarizing beam splitter Bi2, it is possible to guide light in a polarization direction that requires highly sensitive detection to the sensor CiP. The output signals of the sensors C1P-CnP and C1S-CnS are input to the signal processing device D as needed.

-Control device-
The control device E1 is a computer that controls the defect inspection device 100, and includes a ROM, a RAM, and other memories, as well as a CPU, an FPGA, a timer, and the like. The control device E1 is connected to the monitor E30 and the signal processing device D by wire or wirelessly. The control device E1 is connected to a device E2 that allows a user to input various operations, and various input devices such as a keyboard, a mouse, and a touch panel are appropriately connected to E2. The control device E1 receives the encoders of the rotation stage and the translation stage, and the inspection conditions input from the input device E2 in response to the operation of the operator. The inspection conditions include, for example, the type, size, shape, material, illumination conditions, and defect judgment conditions of the sample 1. The control device E1 also outputs a command signal that commands the operation of the stage ST and the scattered light illumination system A, etc., in response to the inspection conditions, and outputs coordinate data of the illumination spot BS synchronized with the defect detection signal to the signal processing device D. The control device E1 also displays and outputs the output of the signal processing device D (such as the defect inspection result of the sample 1) to the monitor E3. As shown in Fig. 1, the control device E1 is connected to a network, and the input of inspection condition data and the output of inspection results can be performed via the network. It is also possible to output the inspection results to an inspection/measurement device connected to this network. As an example, a DR-SEM (Defect Review-Scanning Electron Microscope), which is an electron microscope for defect inspection, is connected. In this case, it is also possible to transmit the inspection results from the control device E1 as data on defect inspection results from the DR-SEM, observe the defects with the DR-SEM, determine the type of the defects, and then receive the results by the control device E1.

--Example of scattered light distribution--
Figure 10 shows the distribution of scattered light from a foreign particle on the sample surface when illuminated with oblique incidence from the left side of the figure, viewed from the wafer normal direction. Light scattered in the direction of direct reflection of the illumination is called forward scattering, and light scattered in the direction of incidence of the illumination is called back scattering, and are labeled "front" and "back" respectively in the figure. Light scattered to the left and right of the incidence of the illumination is called "left side scattering" and "right side scattering", respectively, and are labeled "left" and "right" respectively in the figure. Scattering from a foreign particle on the sample surface is isotropic, and the amount of scattered light at high angles is weaker than the amount of scattered light at low angles. The distribution of scattered light changes depending on the shape of the defect, and the scattering is not always isotropic.

Figure 11 shows the distribution of roughness scattered light from arbitrary coordinates on the sample surface when a polished wafer with a polished surface is used as the sample and illuminated with oblique incidence in the same way. Roughness scattered light generates shot noise in the sensor, worsening the defect detection sensitivity. As shown in Figure 11, roughness scattered light is strongly backscattered and weakly forward scattered, so good sensitivity can be obtained with a front or side sensor. In epitaxial wafers created by vapor-phase growth of silicon single crystals on the surface of a polished wafer, directional roughness occurs on the surface due to the orientation of the silicon single crystals on the surface.

When sample 1 is a polished wafer, the difference in sensitivity due to aperture position is almost constant. For example, as shown in Figure 10, in the case of a foreign object on the sample surface, the scattered light from the defect is almost isotropically the same amount of light is obtained at the low-angle apertures L1-L6. In the distribution of roughness scattered light in Figure 11, the amount of forward scattered light is low, so even if the wafer is rotated for inspection, foreign objects on the sample surface can always be detected with good sensitivity at the sensors corresponding to apertures L3 and L4.

In contrast, when the roughness of the sample surface has directionality, such as in the case of an epitaxial wafer, the intensity of the roughness scattered light changes individually at each aperture position as the wafer rotates. Therefore, the sensor that provides good sensitivity changes from moment to moment as the angle of the sample 1 relative to the oblique incidence illumination changes.

- Signal processing device -
The signal processing device D is a computer that processes input signals from the sensors C1P-CnP and C1S-CnS, and like the control device E1, is configured to include a ROM, RAM, and other memories as well as a CPU, FPGA, a timer, and the like. As an example, the signal processing device D is assumed to be configured as a single computer that forms a unit with the device body of the defect inspection device 100 (stage, scattered light illumination system, scattered light detection system, etc.), but it may be configured as multiple computers. In this case, a server set apart from the device body can be used as one of the multiple computers, and this server can also be included as a component of the defect inspection device 100. For example, a computer attached to the device body can acquire defect detection signals from the device body, process the detection data as necessary and send it to a server, and the server can execute processes such as defect detection and classification.

In this embodiment, the signal processing device D includes a frequency separation unit D1, a model similarity calculation unit D2, a defect manifestation unit D3, and a defect determination unit D4. The frequency separation unit D1, the model similarity calculation unit D2, the defect manifestation unit D3, and the defect determination unit D4 may be virtually realized by software, or may be realized by hardware such as an electronic circuit. Some of the frequency separation unit D1, the model similarity calculation unit D2, the defect manifestation unit D3, and the defect determination unit D4 (particularly the upstream process of the frequency separation unit D1, etc.) can be configured with an FPGA or DSP. In addition, some or all of the functions of the frequency separation unit D1, the model similarity calculation unit D2, the defect manifestation unit D3, and the defect determination unit D4 can be configured to be executed by a server as the signal processing device D.

FIG. 12 is a schematic diagram showing an example of the processing blocks of the frequency separation unit D1, model similarity calculation unit D2, and defect manifestation unit D3. FIG. 12 shows a data processing unit that processes the signal of one specific sensor among sensors C1S-CnS and sensors C1P-CnP. If the total number of sensors is N, then there are the same number of data processing units in FIG. 12 in parallel as the number of sensors (i.e., N sets). The signal photoelectrically converted by each sensor is converted into a digital signal by an A/D converter and input to the frequency separation unit D1, which processes each sensor.

- Frequency separation section (Fig. 12) -
The frequency separation unit D1 separates the input signal from the corresponding sensor of the scattered light detection system B1-Bn into a high frequency component and a low frequency component. Specifically, the frequency separation unit D1 includes a low pass filter D11A and a difference calculation unit D11B. The low pass filter D11A extracts a low frequency signal from the input signal from the sensor, and the difference calculation unit D11B subtracts the low frequency signal from the input signal to obtain a detection signal S11, which is a time series signal of the high frequency component of the input signal.

-Model similarity calculation unit (Fig. 12)-
The model similarity calculation unit D2 calculates a similarity between an observation vector based on the detection signal S11 and a model vector related to a defect, based on the detection signal S11 obtained by the frequency separation unit D1. Specifically, the model similarity calculation unit D2 includes a model vector generation unit D21A and a similarity calculation unit D21B.

Here, a minute defect (e.g., a standard particle) of an ideal shape that is sufficiently small in size relative to the illumination wavelength or optical resolution and present on the sample surface before the pattern is formed is scanned with the illumination spot BS, and multiple samplings are performed while the illumination spot BS passes over the defect, thereby obtaining a time series signal including multiple detection signals while the defect crosses the illumination spot BS once. The time series signal expected to be detected in this case has a profile that correlates with the illumination profile LD2 (FIG. 6). Model data of the defect having this time series distribution is stored in the model vector generation unit D21A. The profile of this model data is determined by the intensity of the illumination light and the relative speed (scanning speed) of the defect with respect to the illumination spot BS, and may be given in advance in multiple values according to the scanning speed, or may be generated by an operator based on the scanning speed. The model vector is generated based on the model data, and correlates with the time series signal expected to be detected when light from an ideal minute defect (e.g., a standard particle) is detected by a sensor.

In addition, in this embodiment, as shown in FIG. 12, a window x1 having a time width corresponding to the width of the illumination profile LD2 is set for the time-series detection signal S11, and the time-series signal (detection signal set) sampled within this window x1 is converted into a vector signal to generate an observation vector. This observation vector is a vector whose components (feature quantities) are the time-series signals sampled at different timings, specifically the intensities of the detection signals sampled within the section of window x1, in other words, a vector whose components are the time-series signals obtained by sampling performed multiple times while the illumination spot BS crosses the point of interest.

For example, if the number of signal samplings (number of detection signals) in the section of window x1 is L, an L-dimensional feature space is assumed. The model vector generation unit 21A generates an L-dimensional model vector based on the model data corresponding to the illumination profile LD2, and transmits the model vector to the similarity calculation unit D21B. f1301 shown in FIG. 13 shows this feature space. In FIG. 13, the model vector V-M1 generated by the model vector generation unit 21A is represented by a dotted line, and the observation vector V-X1 of the micro defect is represented by a solid line. Although the detection signal strength is unknown, if the observation vector V-X1 observes a defect, it is known that the vectors V-X1 and V-M1 are correlated. Therefore, the similarity c1 of the observation vector V-X1 to the model vector V-M1 is calculated. For the similarity c1, a value that is invariant to the norm change of the vectors V-X1 and V-M1, for example, the angle θ between the vectors V-X1 and V-M1, can be used. In addition, values based on the angle θ, such as a modified Euclidean distance, cosine similarity, or values obtained by deep learning of these values, can also be used as the similarity c1. In the similarity calculation unit D21B, the window x1 is shifted sequentially in the time direction, and the similarity c1 is calculated for each window x1 and transmitted to the defect manifestation unit D3.

- Defect manifestation area (Fig. 12) -
In this embodiment, the defect revealing unit D3 performs a nonlinear operation between the observation vector V-X1 and the model vector V-M1 based on the similarity c1 calculated by the model similarity calculation unit D2, and outputs a detection signal that nonlinearly changes the intensity of the input signal corresponding to the observation vector V-X1 to reveal a defect. For example, the signal processing device D (defect revealing unit D3) acquires or calculates a gain that controls the intensity of the input signal corresponding to the observation vector V-X1 based on the similarity c1 between the observation vector V-X1 and the model vector V-M1, and calculates the detection signal based on the product of the input signal and the gain. The processing of the defect revealing unit D3 will be described below.

When the noise is caused by shot noise, the frequency of the noise is generally white, and the maximum SNR can be obtained by applying a filter with the same characteristics as the frequency of the signal profile. However, in the present invention, a nonlinear filter is applied as this filter. The optimal linear filter can be considered as a projection of the observation vector V-X1 between the model vector V-M1. In other words, it can be expressed as the inner product of the vectors V-X1 and V-M1. When the angle between the vectors V-X1 and V-M1 is 0, the square norm of the observation vector V-X1 is output. On the other hand, when this angle is π/2 rad, this output value is 0. In between, it can be interpreted that the gain of cosθ of the angle between the two vectors is multiplied by the square norm. The output value can be expressed by a linear expression of the vector elements.
In contrast, the filter applied in the present invention can be expressed as a nonlinear filter determined by the similarity c1.

Figure 14 shows an example of a non-linear gain. f1401 is a graph with similarity c1 on the X-axis and gain on the Y-axis. Gain is a numerical value between 0 and 1. Here, the angle θ between vectors V-X1 and V-M1 (Figure 13) is applied as similarity c1. Therefore, similarity c1 is highest when it is 0 (the characteristics of vectors V-X1 and V-M1 match), and decreases as it moves away from 0. In the example of Figure 14, the correspondence between similarity c1 and gain is set so that the gain corresponding to an arbitrary similarity c1 is larger than the gain corresponding to a similarity lower than the arbitrary similarity c1.

In Figure 14, gain curve P14-G1 shows cosθ, and represents the same output as a linear filter. In contrast, gain curves P14G2-P14G5 are set so that the gain is lower than gain curve P14G1 as c1 moves away from 0. This means that for detection signal waveforms that deviate greatly from the model vector V-M1, a small gain is given to the detection signal regardless of the signal strength (even if the signal strength is large), and noise is suppressed according to the similarity c1. The gain of gain curves P14G2 and P14G3 matches that of gain curve P14-G1, which represents a linear gain, only when the characteristics of vectors V-M1 and V-X1 match (θ = 0); otherwise, the gain is set lower than that of gain curve P14-G1. The gain curves P14G4 and P14G5 are set so that the gain is the same as that of the gain curve P14-G1 when the characteristics of the vectors V-M1 and V-X1 match (θ=0), and are set to be larger than that of P14G1 when they do not match but the similarity c1 is high (θ is near 0). The gain curves P14G4 and P14G5 are set so that the gain is not set low when the similarity c1 with the model is high, and the value drops sharply when the similarity c1 with the model falls below a preset allowable value. Since the waveform of the defect signal actually obtained changes depending on the adjustment state of the device, such a gain may be appropriate. P14G1 and P14G5 can be input to the defect revealing unit D3 as signal processing parameters, so that an appropriate gain curve can be obtained in actual operation. The gain curves in FIG. 14 can be prepared in advance, or can be calculated according to the scanning speed.

As an example, the defect revealing unit D3 multiplies the length (light intensity) of the observation vector V-X1 projected onto the model vector V-M1 by a gain based on the gain curve in FIG. 14. As another example of the processing of the defect revealing unit D3, the square norm of the observation vector V-X1 may be multiplied by a gain. Although it is expressed as a gain here, it is also possible to achieve a similar effect by other methods. For example, it is also possible to perform a similar calculation by defining a nonlinear kernel function that calculates a value from the observation vector V-X1 and the model vector V-M1. The output of the kernel function is required to be lower when the similarity is high than when the similarity is lower than that. As an example, considering a Gaussian kernel function used in conjunction with a support vector for machine learning, the maximum output can be obtained when the two vectors match. Therefore, after multiplying the gain so that the L2 norm of the model vector V-M1 is the same as the L2 norm of the observation vector V-X1, the difference between the vectors V-M1 and V-X1 is regarded as the similarity. Here, it can be interpreted that the closer the vector difference is to 0, the higher the similarity. It is also possible to configure a filter that calculates the output value of a Gaussian kernel for two vectors after gain correction, and outputs the product of this output value and the L2 norm of V-X1. In this configuration, the output value of the Gaussian kernel has the same effect as the gain.

The signal processing device D performs the above process while moving the window x1 in the time direction, thereby calculating a time series signal S21 (Figure 21), which is a detection signal in which noise has been suppressed and defects have been made apparent. The signal obtained through this process is integrated with signals made apparent by similar processing in other sensors, and input to the defect determination unit D4. Then, in the defect determination unit D4, defects in the sample are determined based on the made-up detection signal.

-effect-
In recent years, defect inspection devices are being required to have the ability to detect extremely small defects. The inventors of the present application have focused on the fact that the scattered light intensity typically changes in proportion to the sixth power of the defect size, and that the scattered light intensity increases or decreases rapidly with a slight change in the defect size. Therefore, by evaluating the similarity between the model vector and the observation vector (for example, the angle between the vectors) as an index that is not easily affected by the intensity of the defect detection signal, and further applying a nonlinear weighting coefficient (gain) based on this index, it is possible to make the detection signal of a small defect hidden among the roughness scattered light, etc. apparent (time-series signal S21 in FIG. 12). This makes it possible to detect small and noisy defects quickly and with high sensitivity.

Second Embodiment
15 is a processing block diagram of the main parts of a signal processing device D provided in a signal processing device 100 according to a second embodiment of the present invention. In the second embodiment, the hardware is the same as in the first embodiment, and only the signal processing device D is different from the first embodiment. In this embodiment, simultaneous processing of input signals from a plurality of sensors is performed for the similarity determination process.

- Frequency separation section (Fig. 15) -
The frequency separation unit D1 of this embodiment includes low-pass filters D11A and D12A, and difference calculation units D11B and D12B. The low-pass filter D11A and the difference calculation unit D11B function similarly in this embodiment. The low-pass filter D12A performs the same process as the low-pass filter D11A, and the difference calculation unit D12B performs the same process as the difference calculation unit D11B.

-Model Similarity Calculation Unit (FIG. 15)-
The model similarity calculation D2 of this embodiment includes a model vector generation unit D21A-2 and a similarity calculation unit D21B-2. The model vector generation unit D21A-2 stores model data M1 and M2 of the illumination light profile, similar to the model vector generation unit D21A. The model data M1 corresponds to the input signal in1 from the first sensor, and the model data M2 corresponds to the input signal in2 from the second sensor. The first sensor and the second sensor are any sensors selected from the sensors C1P-CnP and C1S-CnS. Typically, this time series data is determined only by the profile in the s1 direction of the illumination spot BS, so the model data M1 and M2 are basically the same profile. Therefore, the model data M1 and M2 can be the same.

The similarity calculation unit D21B-2 corresponds to the similarity calculation unit D21B in the first embodiment. The scattered light from a defect may not be scattered isotropically depending on the defect type. If the scattering is not isotropic, the signal strength from the defect may differ between the detection signal S11 by the first sensor and the detection signal S12 by the second sensor. However, even if the scattered light from the defect is not isotropic, there is a high possibility that a profile correlated with the model data M1 or the model data M2 will be observed at the time the defect is scanned. Focusing on this point, in this embodiment, a similarity determination is performed based on input signals from multiple sensors.

FIG. 16 is an explanatory diagram of the similarity with the model vector in this embodiment. f1601 is a feature space, and has the same dimension as the sum of the sampling numbers of windows x1 and x2 shown in FIG. 15. Note that since the first sensor and the second sensor detect different scattered light from the same coordinates, it is necessary to normalize the noise distribution in advance so that it is similar before forming the feature space.

In the feature space of FIG. 16, V16-1 and V16-2 represent model vectors based on model data M1 and M2, respectively. V16-X represents an observation vector of a waveform observed by scanning. The observation vector V16-X is a vector whose components are signals (amount of light from the sample detected simultaneously at multiple different positions in the far field) of multiple sensors (first sensor and second sensor). Although the signal strength of the defect detected by the first sensor and the second sensor is unknown, if defect information based on the defect model is detected, the observation vector V16-X will be detected within a plane p1601 that includes the two vectors V16-1 and V16-2. Therefore, the similarity c2 of the observation vector V16-X to the model can be found by calculating the angle θ2 between the observation vector V16-X and the plane p1601. The angle θ2 is an angle within a plane that includes the observation vector V16-X and is perpendicular to the plane p1601. The similarity calculation unit D21B-2 calculates the similarity c2 based on θ2.

- Defect manifestation area (Fig. 15) -
In this embodiment, defect revealing parts D31A-2 and D32A-2 corresponding to the input signals in1 and in2, respectively, are included. The defect revealing parts D31A-2 and D32A-2 estimate light intensity by projecting an observation vector V16-X onto model vectors V16-1 and V16-2 which are models of the input signals in1 and in2, respectively, and multiply the light intensity by a gain (FIG. 14) based on the similarity c2 in the same manner as in the first embodiment, thereby suppressing noise with low similarity c2 to the model, and obtaining a detection signal S21 of the first sensor and a detection signal S22 of the second sensor in which a defect signal is revealed.

In all other respects, this embodiment is similar to the first embodiment.

According to this embodiment, the similarity c2 is determined using the outputs of multiple sensors, which allows for stronger model constraints than in the first embodiment, and a higher SNR can be expected compared to the first embodiment.

In the above explanation of this embodiment, an example using two sensors has been described for ease of understanding, but it is also possible to use input signals from three or more sensors, and it is also possible to use the output of all sensors implemented in the defect inspection device 100. According to this embodiment, by taking advantage of the physical characteristics of the defect inspection device 100, which can simultaneously obtain signal waveforms for the same coordinates using multiple sensors, it is possible to set stronger constraints and suppress noise.

Third Embodiment
FIG. 17 is a processing block diagram of the main part of the signal processing device D provided in the signal processing device 100 according to the third embodiment of the present invention. In the third embodiment, the hardware is the same as in the first embodiment, and only the signal processing device D is different from the first embodiment. As in the first embodiment, the data processing units in FIG. 17 are arranged in parallel and the number of the data processing units is the same as the number of the sensors. The detection signal D11B-out shown in FIG. 17 is a two-dimensional map of the signal intensity obtained by the spiral scanning described in FIG. 2. The horizontal axis of the detection signal D11B-out represents time, and the vertical axis represents the spiral position (or r coordinate). In the spiral scanning as in FIG. 2, the spiral scanning pitch in the S2 direction is typically set to 1/2 or less of the beam size in the longitudinal direction (i.e., S2 direction) defined by the intensity of exp(-2) of the peak light intensity of the illumination spot BS. At this time, in the map of the detection signal D11B-out, the intensity profile of the defect signal obtained by detecting a typical micro defect is almost correlated with the intensity profile of the illumination spot BS. In the third embodiment, this intensity profile is applied as a model.

- Frequency separation section (Fig. 17) -
The frequency separation unit D1 of this embodiment includes a low-pass filter D11A, a difference calculation unit D11B, and a memory unit D11C. The low-pass filter D11A and the difference calculation unit D11B perform the same processes as those in the first embodiment. The memory unit D11C holds high-frequency components of the detection light obtained by multiple spiral scans.

-Model similarity calculation unit (FIG. 17)-
The model similarity calculation unit D2 includes a model vector generation unit D21A-3, a model vector generation unit D21A-3, and a similarity calculation unit D21B-3.

The model vector generation unit D21A-3 holds data on the beam profile of the illumination spot BS, and generates a model vector corresponding to the window area x3 of the map in D11B-out based on the scanning speed in the S1 direction of the sample to be inspected and the pitch of the spiral scan in the S2 direction. As mentioned above, the pitch of the spiral scan in the S2 direction is typically coarse, about half the size of the illumination spot BS, so as shown in D21A-3-out, multiple different model vectors are generated by combining the illumination light and the assumed position of the defect. D21A-3-out shows multiple (four) examples of one-dimensional time series signals (data strings) only in the S2 direction. Considering that there is a difference in the timing of sampling by the sensor that is less than the sampling period, multiple types of time series signals with different sampling timings (different profiles) can be prepared. D21A-3-out shows an example of a one-dimensional data string in only the S2 direction, but the model vector generation unit D21A-3 generates a two-dimensional data string by combining the S2 direction profile with the S1 direction profile, and transmits it as a model vector to the similarity calculation unit D21B-3. The similarity calculation unit D21B-3 generates an observation vector from the signal sequence in the window x3, and calculates the similarity between the observation vector and the model vector. As a calculation method, a method of calculating the angle between the observation vector and the model vector, as in the first embodiment, cosine similarity, Euclidean distance, or a method using deep learning can be applied.

- Defect manifestation (Fig. 17) -
The defect revealing unit D31A-3 calculates the nonlinear gain for the observation vector calculated from the window for evaluating the nonlinear gain described in FIG. 14 as in the first embodiment. Here, the similarity calculation unit D21B-3 calculates a gain for each of the multiple similarities based on the nonlinear gain in order to calculate a plurality of similarities based on the illumination spot BS and the assumed defect position. Next, the observation vector is projected onto the model vector and multiplied by the gain. As another example, the square norm of the observation vector may be multiplied by the gain. By performing processing while moving the window x3 two-dimensionally in the time direction and the spiral pitch direction, a two-dimensional map D31-A3-out in which the noise is suppressed and the defect is revealed is calculated. This obtained signal is integrated with signals revealed by similar processing in other sensors, and the defect is judged in the judgment unit D4.

In other respects, this embodiment is similar to the first embodiment.

In the first embodiment, the data from a specific spiral scan was targeted, so if a defect is scanned at a position that is off in the S2 direction from the center of the illumination spot BS, the strength of the defect signal decreases, and the similarity with the model must be evaluated for defect signals with a weak amount of scattered light, and even signals that detect minute defects may be calculated with a low similarity. In this embodiment, the similarity is calculated based on all scattered light intensities obtained from multiple spiral scans, making the system robust against this issue.

Fourth Embodiment
FIG. 18 is a processing block diagram of the main part of the signal processing device D provided in the signal processing device 100 according to the fourth embodiment of the present invention. In the fourth embodiment, the hardware is the same as in the first embodiment, and only the signal processing device D is different from the first embodiment. The fourth embodiment is a combination of the second and third embodiments. In the fourth embodiment, similar to the second embodiment, a similarity with strong constraints is calculated by utilizing the structural characteristics of the defect inspection device 100 that simultaneously observes the same defect using the outputs of multiple sensors, thereby achieving a high SNR. In FIG. 18, for the sake of simplicity, a case in which the similarity is calculated based on input signals from two sensors is described, but similar to the second embodiment, the outputs of up to N sensors can be used.

- Frequency separation section (Fig. 18) -
The frequency separation unit D1 of this embodiment includes low-pass filters D11A and D12A, difference calculation units D11B and D12B, and memory units D11C and D12C. The low-pass filters D11A and D12A and the difference calculation units D11B and D12B have the same functions as those of the second embodiment. The memory units D11C and D12C, like the memory unit D11C of the third embodiment, hold high-frequency components of input signals obtained by multiple revolutions of spiral scanning for each sensor.

-Model Similarity Calculation Unit (FIG. 18)-
The model similarity calculation D2 of this embodiment includes a model vector generation unit D21A-4 and a similarity calculation unit D21B-2. The model vector generation unit D21A-4 holds data on the beam profile of the illumination spot BS, similar to the model vector generation unit D21A-3 of the third embodiment, and generates model data M3 and M4 based on the scanning speed of the illumination spot BS in the S1 direction and the pitch of the spiral scan in the S2 direction. The model data M3 is a beam profile detected by the first sensor, and the model data M4 is a beam profile detected by the second sensor, but since both sensors detect scattered light from the same illumination spot BS, no significant difference occurs, and the model data M3 and M4 can be the same.

The similarity calculation unit D21B-4 evaluates the similarity by utilizing the fact that the detection signals from the two sensors are similar to the model vector at the same timing. The basic processing algorithm is the same as that of the second embodiment (FIG. 16), but here, as explained in the third embodiment, the difference is that there are multiple types of model vectors (four in FIG. 17), for example, as explained in FIG. 17 for D21A3-3-out. There are also two types of model vectors corresponding to the two sensors. When determining the similarity between multiple model vectors for each of the windows x4-1 and x4-2, a plane corresponding to the plane p1601 in FIG. 16 is created depending on the contribution of the first and second sensors in each determination. For the observation vectors obtained in the two windows x4-1 and x4-2, multiple planes (four in this example) in the multidimensional feature space determined by the two model vectors are generated depending on the positional relationship between the assumed illumination spot BS and the defect, and the deviation (θ) of the observation vector with respect to these multiple planes is output as multiple similarities (four in this example). The number of planes is not limited to four, and can be set arbitrarily, for example, to two or eight. In this embodiment, the similarity can be calculated using the angle θ between the plane and the vector, as well as Euclidean distance, cosine similarity, and similar features output by deep learning.

- Defect manifestation area (Fig. 18) -
The defect revealing unit D31A-4 plays a role of revealing a defect signal for the output signal of the first sensor, and the defect revealing unit D32A-4 plays a role of revealing a defect signal for the output signal of the second sensor. The defect revealing unit D31A-4 projects an observation vector obtained by vectorizing the data of the window x4-1 of the detection signal D11B-out output by the difference calculation unit D11B onto a plurality of model vectors (four in this example) generated based on the model data M3, multiplies the observation vector by a gain based on the similarity calculated by the similarity calculation unit D21B-4, and obtains outputs in the same number as the number of model vectors generated from one window x4-1. This process is performed while moving the window x4-1 in the detection signal D11B-out in the time direction and the spiral number direction, thereby obtaining a two-dimensional map of the time series signal D31A-4-out, which is a detection signal in which a defect signal is revealed. The defect revealing section D32A-4 also executes the same process as the defect revealing section D31A-4 to obtain a two-dimensional map of the time-series signal D32A-4-out. These two maps are transferred to the defect determining section D4, where the defect determining process is executed.

In all other respects, this embodiment is similar to the other embodiments.

This embodiment achieves the combined effects of the second and third embodiments.

Fifth Embodiment
- Oblique imaging detection system -
19 is a schematic diagram of a scattered light detection system B1-Bn provided in a defect inspection apparatus 100 according to a fifth embodiment of the present invention. In the fifth embodiment, the scattered light detection system B1-Bn is partially different from the embodiments described above. In the first to fourth embodiments, the scattered light detection system B1-Bn is a light collection detection system, and the sensors C1P-CnP, C1S-CnS are point sensors, but in this embodiment, the scattered light detection system B1-Bn is an imaging detection system B", and the sensors C1P-CnP, C1S-CnS are line sensors CP", CS". As for other hardware, this embodiment is similar to the first embodiment.

When detecting the illumination spot BS on the sample surface from an oblique direction, the working distance between the sample surface and the objective lens varies depending on the distance from the optical axis of the objective lens. Therefore, if the light receiving surface of the sensor is perpendicular to the optical axis of the detection system, a focus shift occurs. Therefore, the light receiving surface of the sensor is tilted with respect to the optical axis so that it is conjugate with the sample surface. This point is explained in Figure 19.

The oblique imaging detection system B" comprises an objective lens B1", a half-wave plate B2", a polarizing beam splitter B3", half-wave plates B4P" and B4S", and imaging lenses B5P" and B5S".

The half-wave plate B2" can be rotated by a rotation mechanism (not shown), and the polarization direction of the light detected by the objective lens B1" can be rotated to the desired direction. The polarizing beam splitter B3" splits the detected light into two optical paths, one for P polarization and one for S polarization. The optical path for P polarization is lined with the half-wave plate B4P", the imaging lens B5P", and the line sensor CP". Similarly, the optical path for S polarization is lined with the half-wave plate B4S", the imaging lens B5S", and the line sensor CS". The half-wave plates B4P", B4S" are each set on a rotation stage (not shown), and the polarization direction is rotated so as to optimize the detection efficiency of the line sensors CP", CS".

The line sensors CP'', CS'' are inclined with respect to the optical axis of the oblique imaging detection system B'' so as to be conjugate with the illumination spot BS on the sample surface.
Since the focal depth of the imaging detection system is shallower than that of the light collection detection system, in addition to the optical system provided in the first embodiment, a Z sensor is provided that measures the working distance between the sample surface and the detection optical system using the principle of an optical lever. A Z stage is also added to the stage ST so that the height between the top surface of the sample and the detection optical system can be changed. The control device E1 controls the working distance between the detection system and the sample surface to be constant on the order of micrometers based on the output value of the Z sensor.

In the high-speed inspection mode, the spiral scanning of the stage ST has a scanning speed of several tens of meters per second, making it difficult to keep the working distance constant as expected. When the working distance fluctuates, the imaging detection system B" becomes out of focus and the imaging position shifts. This is explained in Figure 20. In the figure, L1P, L3P, L4P, L6P, H2P, H3P, H5P, and H6P respectively indicate how the light incident on the apertures L1, L3, L4, L6, H2, H3, H5, and H6 (Figure 7) and passing through the polarizing beam splitter B3" expands in the S2 direction on the line sensor when the working distance shifts ΔZ from the designed position (PSF). When the working distance shift ΔZ from the designed position (the center of each figure in the ΔZ direction) increases, the image expands while shifting in the s2 direction on each sensor surface. It can be seen that the amount of shift in the s2 direction relative to the deviation ΔZ is greater for light scattered at low angles (L1P, L3P, L4P, L6P) than for light scattered at high angles (H2P, H3P, H5P, H6P).

- Depth of focus of the illumination system -
Fig. 21 shows the change in the spread of the image of the illumination spot BS on the sample surface in the S1 direction due to the shift ΔZ in the working distance. In the fifth embodiment, the width of the illumination spot BS in the S1 direction is reduced to a linear shape. As shown in Fig. 21, unlike the image on the sensor surface shown in Fig. 20, the illumination spot BS does not shift but expands in response to ΔZ.

- Frequency separation section (Fig. 22) -
22 is a processing block diagram of the main parts of the signal processing device D provided in the signal processing device 100 according to this embodiment. The frequency separation unit D1 of this embodiment includes a low-pass filter D11A-2 and a difference calculation unit D11B-2. The low-pass filter D11A-2 is applied to the input signal D11A-2IN from the sensor in the S1 direction, i.e., in the time direction, to extract a low-frequency signal. The low-pass filter D11A-2 does not apply a filter in the S2 direction. The difference calculation unit D11B-2 subtracts the low-frequency signal obtained by the low-pass filter D11A-2 from the input signal D11A-2IN to obtain a detection signal D11B-2-out of high-frequency components from which low frequencies near the DC component have been removed.

-Model Similarity Calculation Unit (FIG. 22)-
The model similarity calculation unit D2 of this embodiment includes a model vector generation unit D21A-5 and a similarity calculation unit D21B-5.

The model vector generating unit D21A-5 holds model data M5-M7 of the defect profile obtained when the working distance ΔZ changes. The S2-direction profile of the model data M5-M7 is determined based on the PSF (point spread function) in FIG. 20, which shows the beam spread in the S2 direction for each sensor with respect to ΔZ, and the pixel size of the sensor. Typically, the S2-direction profile of the model data M5-M7 is determined by convolution of the PSF with a square wave representing the sensor pixel. On the other hand, the S1-direction profile of the model data M5-M7 is calculated based on the profile of the illumination spot BS (FIG. 21), which changes according to the working distance deviation ΔZ, and the exposure time of the sensor pixel. The convolution of the time-direction profile converted from the profile of the illumination spot BS based on the scanning speed and the square wave representing the exposure time of the sensor pixel becomes the model data M5-M7 to be applied to the signal processing. Since the scanning speed changes depending on the spiral position during inspection, the model data M5-M7 to be applied to the signal processing is changed according to the change in the scanning speed.

The signal processing device D also receives the input signal D11A-2-IN from the sensor, along with the output value of the Z sensor at the time the signal was acquired. The model vector generation unit D21A-5 determines the deviation ΔZ from the design value of the working distance at the time of imaging of the input signal D11A-2-IN from the output value of the Z sensor, and determines which of the model data M5-M7 is appropriate. The model vector generation unit D21A-5 converts the profile of the determined model data into the scanning speed at that time, determines the model vector corresponding to the window x5, and transfers it to the similarity calculation unit D21B-5.

The similarity calculation unit D21B-5 calculates the similarity of the observation vector to the model vector based on the evaluation value (the angle between the observation vector and the model vector, as well as cosine similarity, Euclidean distance, or evaluation value based on deep learning) based on the observation vector set from the data string in the window x5 and the model vector obtained from the model vector generation unit D21A-5. The position of the window x5 is changed in the S1 direction and the S2 direction, and the similarity is calculated for all of the detection signals D11B-2-out obtained by scanning.

- Defect manifestation (Fig. 22) -
As in the first embodiment, the defect revealing unit D31A-5 reveals a defect signal by multiplying the signal intensity of a vector within a window x5 of the detection signal D11B-2-out by a gain based on a nonlinear gain curve.

In other respects, this embodiment is similar to the first embodiment.

As in this embodiment, even in a defect inspection apparatus 100 that uses an imaging detection system B" for the scattered light detection systems B1-Bn, the same effects as in the first embodiment can be obtained.

Sixth Embodiment
In the sixth embodiment, the hardware is the same as in the fifth embodiment, and only the signal processing device D is different from the fifth embodiment. In the fifth embodiment, an example in which similarity is calculated for data of a specific revolution has been described, but in this embodiment, a change in the amount of detected light due to multiple revolutions of spiral scanning is added to the model data, and the constraints are strengthened to improve sensitivity.

FIG. 23 is an explanatory diagram of overlapping illumination scanning according to the sixth embodiment of the present invention and the model vector based on it. In this embodiment, the spiral pitch is set so that half of the illumination beam diameter defined by exp(-2) of the illumination peak overlaps in the S2 direction. This allows the same defect to be detected twice. In the graph of f2301, the horizontal axis is the S2 direction, p2301 represents the illumination light intensity in the i-th rotation, and p2302 represents the illumination light intensity in the i+1th rotation. In this example, for the same defect, scattered light p2301-a is detected in the i-th rotation scan, and scattered light p2301-b is detected in the i+1th rotation scan. In the line sensor, pixels are arranged in an array to form a light receiving section, and the expected detection light amount illum(i,r) for each r coordinate on the sample surface can be specified. i is the spiral number, and r is the r coordinate.

Feature space f2302 shown at the bottom of the figure illustrates model vector V23-a obtained by spiral scanning in the i-th rotation and model vector V23-b obtained in the i+1th rotation for a defect on the sample surface whose center is the scanning trajectory of a specific pixel. The observation vector obtained from an ideal defect has the same vector components as model vector V23-c, which is a combination of these two model vectors V23-a and V23-b. Therefore, the similarity between observation vector V23-d and model vector V23-c can be calculated based on the angle θ between the two vectors, cosine similarity, Euclidean distance, or an evaluation value obtained by deep learning, as in the respective embodiments described above. This series of processes will be explained using Figures 24 and 25.

24 and 25 are processing block diagrams of the main parts of the signal processing device D provided in the signal processing device 100 according to this embodiment. The low-pass filter D11A-2, the difference calculation unit D11B-2, and the model vector generation unit D21A-5 are the same as those in the fifth embodiment. However, the model vector generation unit D21A-5 transmits a model vector to the model vector projection unit D11D. The model vector projection unit D11D outputs the light intensity D11D-out obtained by projecting the observation vector defined in the window x5 set in the same manner as in the fifth embodiment onto the model vector for the detection signal D11B-2-out output by the difference calculation unit D11B-2, and stores this in the memory unit D11C-6. On the other hand, the signal intensity calculation unit D11E outputs the map D11E-out of the square norm of the observation vector obtained from the window x5, and stores this in the memory unit D11C-6. The data stored in the memory unit D11C-6 is processed by the similarity calculation unit D21B-6 and then further processed by the defect manifestation unit D31A-6.

FIG. 25 shows model vector projection maps D11C-6-a and D11C-6-b obtained by the i-th and i+1-th spiral scans stored in the memory unit D11C-6, and square norm maps D11C-6-c and D11C-6-d obtained by the i-th and i+1-th spiral scans. As in other embodiments, the similarity of the positions of each map is calculated from these maps using the vector angle θ, cosine similarity, Euclidean distance, or deep learning, and transmitted to the defect revealing unit D31A-6. As in other embodiments, the defect revealing unit D31A-6 sets a nonlinear gain based on the similarity and reveals defects.

In all other respects, this embodiment is similar to the other embodiments.

In this embodiment, the same effects can be obtained as in the other embodiments.

Seventh Embodiment
In the seventh embodiment, the hardware is the same as that in the fifth embodiment, and only the signal processing device D is different from the fifth embodiment. In this embodiment, the similarity with the model vector is calculated based on input signals from a plurality of sensors. A processing block diagram of the main parts of the signal processing device D provided in the signal processing device 100 according to the seventh embodiment of the present invention is shown in FIG.

-Model similarity calculation unit (FIG. 26)-
The frequency separation unit D1 of this embodiment includes low-pass filters D11A-2 and D12A, difference calculation units D11B-2 and D12B, model vector projection units D11D and D12D, and signal intensity calculation units D11E and D12E. The low-pass filters D11A-2 and D12A, difference calculation units D11B-2 and D12B, model vector projection units D11D and D12D, and signal intensity calculation units D11E and D12E execute the same processes as those of the sixth embodiment.

-Model similarity calculation unit (FIG. 26)-
The model similarity calculation D2 of this embodiment includes a model vector generation unit D21A-7 and a similarity calculation unit D21B-2. The model vector generation unit D21A-7 holds the profile of the detection system and the profile of the illumination system when the Z sensor is changed for each of the first and second sensors in the same manner as the model vector generation unit D21A-5. The illumination profile is generally observed in the same manner for each sensor, but the detection system detects from different orientations, and the working distance to the sample surface does not completely match in the adjusted state, so an assumed profile is prepared for each sensor. The model vector generation unit D21A-7 generates a model vector for each timing at which data is acquired based on the input Z sensor output and the scanning speed in the S1 direction at that time, and outputs the vector to the model vector projection units D11D and D12D. As in the sixth embodiment, the model vector projection units D11D and D12D output the result of projecting the observation vector onto the input model vector to the memory unit D11C-3. Similarly to the sixth embodiment, the signal intensity calculation units D11E and D12E calculate the square norm within the window and transmit it to the memory unit D11C-3. The memory unit D11C-3 holds a map of the square norm and a map of the projection onto the model vector obtained by two adjacent rotations of the spiral scan in each sensor. Similarly to the sixth embodiment, D21B-7-in-1 and D21B-7-in-2 are maps calculated based on the data acquired by the first sensor and the second sensor, respectively.

The similarity with the model vector is calculated based on the maps of D21B-7-in-1 and D21B-7-in-2. When calculating the similarity, first, the vector direction determined by the position in the S2 direction between the two spiral scans described in FIG. 23 is set for the first and second sensors. Next, the angle θ between the plane containing these two vectors and the observation vector is calculated. As in the previously described embodiment, the angle θ may be used as the similarity, or a value converted using a method such as cosine similarity, Euclidean distance, or deep learning may be used as the similarity.

- Defect manifestation area (Fig. 26) -
As in the embodiment described above, the defect revealing units D3A-7-1 and D3A-7-2 reveal defects by multiplying a map in which the output of each sensor is projected onto a model vector or a map of square norm by a nonlinear gain determined for each data constituting the map. A common value of the nonlinear gain is applied to the two sensors, but the signal revealing the SNR of each sensor is output independently.

In FIG. 26, the similarity is calculated for two sensors to suppress noise. The defect revealing units D3A-7-1 and D3A-7-2 calculate the similarity to the model from a combination of two sensors, but as in the previously described embodiment, the number of sensors is not limited to two. For example, it is also possible to calculate the similarity using the output of all sensors provided in the defect inspection device 100 and determine the nonlinear gain.

- Defect Judgment Section -
The defect determination unit will be described with reference to FIG. 27. The process of the defect determination unit D4 described below is not limited to this embodiment, and can be performed in other embodiments. The defect manifestation units D3A-7-1 to D3A-7-N output signals with suppressed noise for each sensor to the defect determination unit D4. In the defect determination unit D4, as illustrated in FIG. 28, the output signals from each sensor are mapped to a feature space with the same dimension as the number of sensors. In FIG. 28, V-f1 to N are output signals from each sensor. Here, a known defect scattering distribution is given as a model, and the gain is maintained for signals with a high similarity to the model, and the gain of defects that are not similar is suppressed, thereby improving the detection sensitivity of the defect to be detected. For example, when the purpose is to detect surface foreign matter, the low-angle detector emits scattered light almost isotropically. This feature is represented by a model vector f-D4. The angle between the observation vector X-28 and the model vector f-D4 is calculated, and the gain is reduced as the angle moves away from 0. Since the important defects to be detected differ depending on the inspection, the gain is made controllable by the inspection parameters.

In all other respects, this embodiment is similar to the other embodiments.

In this embodiment, the same effects can be obtained as in the other embodiments.

Eighth embodiment
FIG. 29 is a configuration diagram of a defect inspection apparatus 200 according to the eighth embodiment of the present invention.

The defect inspection device 200 of this embodiment is a device that captures an external image of the sample 10 and uses the external image to inspect defects in the sample 10. For convenience of description, the defect inspection device 200 and the signal processing device D-B are described separately below, but they can also be configured as an integrated device or can be interconnected via an appropriate communication line.

The sample 10 is an object to be inspected, such as a semiconductor wafer. The stage ST3 is composed of a combination of three stages that mount the sample 10 and move in a straight line in each of the X, Y and Z directions, and a rotating stage that rotates the wafer. A Z sensor (not shown) measures the height of the surface of the sample to be inspected, controls the stage that moves in a straight line in the Z direction, and keeps the working distance between the optical system and the sample surface constant.

Either of the illumination optical systems 231 and 232 irradiates illumination light obliquely onto the sample 10. The upper detection system 241 and the oblique detection system 242 form an image of the scattered light from the sample 10. The image sensors 261 and 262 receive the optical images formed by each detection system and convert them into image signals. A line sensor is used as the image sensor. As for the type of line sensor, a typical line sensor in which the light receiving parts are arranged one-dimensionally or a TDI sensor arranged two-dimensionally can be applied. In this embodiment, a TDI sensor is applied. An analyzer 252 is arranged in front of the image sensor 261. A two-dimensional image of the sample 10 can be obtained by detecting the scattered light while moving the stage ST3 in the horizontal direction.

The light source for the illumination optical systems 231 and 232 may be a laser or a lamp. The wavelength of each light source may be a single wavelength or may be broadband wavelength light (white light). When using short wavelength light, ultraviolet light (Ultra Violet Light) may be used to increase the resolution of the detected image (to detect minute defects). When using a laser as the light source, if it is a single wavelength laser, the illumination optical systems 231 and 232 may also be equipped with a means for reducing coherence.

By placing a wave plate (not shown) between each of the illumination optical systems 231, 232 and the sample 10, the polarization state of the illumination light incident on the sample 10 can be changed.

The control device E1 controls the overall operation of the defect inspection device 200, including each illumination optical system and each image sensor. The signal processing device D-B can be connected to the defect inspection device 200 via the control device E-1.

The signal processing device D-B has an image processing unit D-B1, a defect determination unit D-B2, an image storage unit D-B3, an extraction condition calculation unit D-B4, an extraction condition storage unit D-B5, and a display unit D-B6. Each of the above functional units is a function that is realized by the memory of the processor and by executing a program on the processor.

The image processing unit D-B1 acquires an external image of the sample 10 via the control unit E-1, and performs the processing described later in FIG. 31. The defect determination unit D-B2 extracts defects in the sample 10 based on the feature amounts of the external image, according to the extraction conditions described in the extraction condition data stored in the extraction condition storage unit D-B5. The image storage unit D-B3 stores a feature amount image representing the feature amounts of the sample 10, the determination results by the defect determination unit D-B2, etc. The extraction condition calculation unit D-B4 calculates new defect extraction conditions according to the procedure described later, and the defect determination unit D-B2 extracts defects using those conditions. The display unit D-B6 displays and outputs the processing results by the signal processing device D-B, such as the determination results by the defect determination unit D-B2, on the monitor E3 (FIG. 1).

Figure 30 is a top view showing an example of sample 10. When sample 10 is, for example, a semiconductor wafer, identical semiconductor chips (dies) 111 to 115 are formed on sample 10. Memory areas 1151-1, 1151-2, and 1151-3 and a peripheral circuit area 1152 are formed on semiconductor chip 115. Defect inspection device 100 acquires an external image while moving stage ST3 in a direction perpendicular to illumination line (scanning line) 1153, which is an illumination spot. Signal processing device D-B extracts defects by comparing the images of semiconductor chips 111 to 115 with each other. Details will be described later.

FIG. 31 is a configuration diagram of the internal calculation block provided in the signal processing device D-B. The signal processing device D-B has the internal calculation block of FIG. 31 for each detection system provided in the defect inspection device 200, and processes the appearance image detected by each detection system individually. Here, the internal calculation block 302a is shown as an example, which processes the appearance image detected by the first detection system (e.g., upper detection system 141). The other internal calculation blocks have the same configuration, so alphabetical suffixes are used when it is necessary to distinguish between them. The same applies to the drawings described later.

The internal calculation block 302a receives pixel values 301a of the appearance image of the sample 10 and stores them in the image memory 311 to generate an inspection target image (e.g., semiconductor chip 111) 302. Similarly, a comparison target image is generated. As the comparison target image, an adjacent image (e.g., semiconductor chips 112 to 115) 303-1 or a similar image within the same die 303-2 (e.g., memory areas 1151-1 to 1151-3 of the same design formed within the same semiconductor chip) is generated. Since the similar image within the die is close to the inspection target, an image that is more similar to the inspection target can be expected. Therefore, when a similar image within the same die similar to the inspection target area can be obtained, the similar image within the same die 303-2 is used as the comparison target image, and when such an image is not available, the adjacent image 303-1 is used as the comparison target image.

The position shift calculation unit 312 calculates the amount of position shift between the inspection target image 302 and the adjacent image 303, for example, by calculating the normalized correlation between the two images. The alignment unit 313 aligns the positions of the inspection target image 302 or the adjacent image 303 by moving them according to the amount of position shift. The reference image synthesizer 314-1 generates a reference image composed of, for example, the median value of the pixel values (brightness values) of multiple adjacent images 303. Meanwhile, the reference image synthesizer 314-2, which is independent of this, synthesizes a reference image using only the inspection target image 302. For example, semiconductor chips are often composed of locally repeated identical patterns, and the reference image is synthesized by taking the median value of the brightness values by utilizing the repetition of the same patterns.

Also, if the pattern is extremely fine relative to the resolution of the detection system, an image of uniform brightness obtained by taking the median value of the unresolved image area may be used as the reference image. Examples of this include line patterns and memory cell sections with a wiring pitch of several tens of nanometers. If a UV laser is used as the illumination light source for 131 and 132, the wavelength size is 200 nm or more, so line patterns and memory cell sections with a pitch of several tens of nanometers cannot be resolved and are generally imaged as extremely dark uniform areas. Reference image synthesizer 314-3 synthesizes the final reference image based on the reference images input from reference image synthesizers 314-1 and 314-2. If a reference image can be synthesized in the vicinity of the inspection target, an approximation image that is most appropriate as a normal part of the inspection target can be obtained, so the output of reference image synthesizer 314-2 is applied as the reference image, and if this cannot be obtained, the output of reference image synthesizer 314-1 is applied as the reference image. Also, although not shown, it is possible to configure the configuration so that the outputs of both reference image combiners 314-1 and 314-2 are applied as reference images, and two comparison image processes are performed for one inspection target image using different reference images. The difference calculation unit D1-B creates a difference image by calculating the difference between the inspection target image 302 and the reference image.

The model vector generation unit D21A-B stores defect model data M8-M10. When a TDI is used as the image sensor, the defect model vector is calculated by convolution of the PSF of the detection system and a rectangular pattern representing the pixel size for both the S1 and S2 directions. The deviation ΔZ from the ideal working distance is input from the Z sensor, and the model data is changed according to the deviation ΔZ to calculate the model vector.

The similarity calculation unit D21B-B calculates the similarity based on the angle between the observation vector obtained from the data string in the set window in the difference calculation unit D1-B and the model vector generated by the model vector generation unit D21A-B.

The processing of the defect revealing unit D3A-B will be explained using FIG. 32. In a semiconductor wafer with a pattern such as the sample 10, the noise distribution state changes depending on the circuit pattern formed on the surface. Since this noise distribution state changes depending on the position of the window, the noise distribution state is evaluated using the coordinates of the reference image pattern or the semiconductor wafer. In FIG. 32, V-f1 to V-f3 are the coordinate axes of the feature space, and f-D4 is the model vector generated by the model vector generating unit D21A-B. X-32 is the observation vector to be evaluated. Statistically, there are not an extremely large number of defects in the sample 10 to be inspected, and most are normal. Therefore, the variation in the feature space of this normal point group is calculated. Examples of the method of calculating the variation include a method using K times the standard deviation of the same design part, or a maximum value by Poisson distribution estimation. In addition, morphological processing such as taking the maximum value of N neighborhoods for the map subjected to these statistical processing may be adopted. This determines the variance for each pixel in the window, and then normalizes the variance so that the variance for all pixels is the same. As a result, the observation vector X-32 is transformed into X-32', and the model vector f-D4 is transformed into f-D4'. The distribution of normal data also changes. In this transformed feature space, the similarity can be calculated based on the angle θ between the model vector f-D4' and the observation vector X-32'.

Based on this similarity, the defect revealing unit D3A-B reveals the defects. The nonlinear gain explained in FIG. 14 is applied as the revealing gain. The gain can be applied to either the normalized observed data or the data projected onto the model vector.

In all other respects, this embodiment is similar to the other embodiments.

In this embodiment, the same effects can be obtained as in the other embodiments.

In this embodiment, processing of the output of a specific sensor has been described, but as in the other embodiments, it is also possible to use a method that uses the output of multiple image sensors, or to apply overlapping results to calculate similarity and gain.

The present invention is not limited to the above-described embodiments, and includes various modified examples. For example, the above-described embodiments have been described in detail to easily explain the present invention, and all of the configurations described are not necessarily required. It is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.
For example, an example has been described in which the invention is applied to a defect inspection device that detects light from a sample using multiple sensors, but the present invention can also be applied to a defect inspection device that has only one sensor that detects light from a sample.

The above configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, such as an integrated circuit. The above configurations, functions, etc. may be realized by software, with a processor interpreting and executing a program that realizes each function. Information on the programs, tables, files, etc. that realize each function may be stored in a recording device such as a memory, a hard disk, or an SSD (Solid State Drive), or on a recording medium such as a flash memory card or a DVD (Digital Versatile Disk).

In each embodiment, the control lines and information lines are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are interconnected.

1,10...sample, 100,200...defect inspection device, A...scattered light illumination system, BS...illumination spot, c1...similarity, C1P-CnP, C1S-CnS, CP", CS"...sensor, D...signal processing device, ST2...scanning device, V-X1, V16-X, V23-d, X-32...observation vector, V-M1, V23-a, V23-b, V23-c, f-D4...model vector, θ...angle

Claims (11)

1. A defect inspection apparatus for inspecting a sample for defects based on light from the sample, comprising:
one or more sensors that detect light from the sample;
a signal processing device for processing an input signal from the sensor,
The signal processing device includes:
Calculating a similarity between an observation vector based on an input signal from the sensor and a model vector related to a defect;
calculating a detection signal by nonlinearly changing the intensity of the input signal corresponding to the observation vector based on the similarity;
a defect inspection apparatus for determining defects in the sample based on the detection signal. 2. The defect inspection apparatus according to claim 1,
the signal processing device acquires or calculates a gain that controls an intensity of an input signal corresponding to the observation vector based on a similarity between the observation vector and the model vector;
The detection signal is calculated based on a product of the input signal and the gain;
A defect inspection apparatus comprising: a gain setting unit that sets a gain corresponding to an arbitrary similarity in correspondence with the similarity and the gain so as to be greater than a gain corresponding to a similarity lower than the arbitrary similarity. 2. The defect inspection apparatus according to claim 1,
the similarity is calculated from input signals from the one or more sensors;
A defect inspection apparatus, characterized in that the detection signal is calculated individually for each input signal from the one or more sensors. 2. The defect inspection apparatus according to claim 1,
The observation vector is a time series signal sampled at different times,
The defect inspection device according to claim 1, wherein the detection signal is a time series signal. 2. The defect inspection apparatus according to claim 1,
A defect inspection apparatus according to claim 1, wherein the observation vector is a vector whose components are time-series signals obtained by sampling performed a plurality of times while an illumination spot crosses a target point. 2. The defect inspection apparatus according to claim 1,
A defect inspection apparatus, wherein the observation vector is a vector whose components are the amounts of light from the sample detected simultaneously at a plurality of different positions in the far field. 2. The defect inspection apparatus according to claim 1,
A defect inspection apparatus characterized in that the model vector is a vector that correlates with a time series signal that is expected to be detected when the sensor detects light from a standard particle whose size is small relative to the illumination wavelength when the sample is scanned. 8. The defect inspection apparatus according to claim 7,
a plurality of model vectors are generated based on a plurality of time-series signals in which the difference in sampling timing by the sensor is equal to or less than a sampling period; 2. The defect inspection apparatus according to claim 1,
A defect inspection apparatus, wherein the similarity is a value that is invariant to a change in the norm of the model vector and the observation vector. 2. The defect inspection apparatus according to claim 1,
A defect inspection apparatus, wherein the similarity is an angle between the model vector and the observation vector or a value based thereon. 2. The defect inspection apparatus according to claim 1,
an illumination system that applies illumination to the sample to form an illumination spot;
a scanning device that scans the illumination spot on the sample.

Images