applied
sciences
Article
MEMS High Aspect Ratio Trench Three-Dimensional
Measurement Using Through-Focus Scanning Optical
Microscopy and Deep Learning Method
Guannan Li 1 , Junkai Shi 1, * , Chao Gao 1,2 , Xingjian Jiang 1 , Shuchun Huo 1 , Chengjun Cui 1 , Xiaomei Chen 1
and Weihu Zhou 1,2
Citation: Li, G.; Shi, J.; Gao, C.; Jiang,
X.; Huo, S.; Cui, C.; Chen, X.; Zhou,
W. MEMS High Aspect Ratio Trench
Three-Dimensional Measurement
Using Through-Focus Scanning
Optical Microscopy and Deep
Learning Method. Appl. Sci. 2022, 12,
8396. https://doi.org/10.3390/
app12178396
Academic Editor: Andrea Li Bassi
Received: 26 July 2022
Accepted: 22 August 2022
Published: 23 August 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Appl. Sci. 2022, 12, 8396. https://doi.org/10.3390/app12178396
1 Institute of Microelectronics of Chinese Academy of Sciences, Beijing 100029, China
2 University of Chinese Academy of Sciences, Beijing 100049, China
* Correspondence: shijunkai@ime.ac.cn
Abstract: High-aspect-ratio structures have become increasingly important in MEMS devices. In
situ, real-time critical dimension and depth measurement for high-aspect-ratio structures is critical
for optimizing the deep etching process. Through-focus scanning optical microscopy (TSOM) is a
high-throughput and inexpensive optical measurement method for critical dimension and depth
measurement. Thus far, TSOM has only been used to measure targets with dimension of 1 µm or
less, which is far from sufficient for MEMS. Deep learning is a powerful tool that improves the TSOM
performance by taking advantage of additional intensity information. In this work, we propose a
convolutional neural network model-based TSOM method for measuring individual high-aspect-ratio
trenches on silicon with width up to 30 µm and depth up to 440 µm. Experimental demonstrations
are conducted and the results show that the proposed method is suitable for measuring the width and
depth of high-aspect-ratio trenches with a standard deviation and error of approximately a hundred
nanometers or less. The proposed method can be applied to the semiconductor field.
Keywords: optical micro-nano metrology; three-dimensional non-destructive measurement;
high-aspect-ratio structure; through-focus scanning optical microscopy; convolutional neural network
1. Introduction
Micro-electro-mechanical systems (MEMS) comprise both mechanical and electronic
components manufactured on a common silicon substrate, and have gained significant
attention in the last three decades. Due to the miniaturized size, low cost, high manufac-
turing throughput, and expectance of superior performance, MEMS are used in a wide
range of consumer, automotive, medical, and industrial products [1,2]. With the increase
in the integration density and the reduction in the feature size, high-aspect-ratio (HAR)
structures in silicon play an increasingly important role in micro-sensors and actuators [3,4].
Most MEMS HAR structures are fabricated by the dry etching process, especially deep
reactive-ion etch (DRIE) with the Bosch process [5,6]. The precisely controlled profile
of HAR structures is required for better device performance and higher resolution and
reliability [7]. Due to aspect-ratio-dependent etching and microloading effects, there is a
complex interplay during etching between the device design (topology dimensions and
shape) and etch parameters [8]. Therefore, in situ, real-time critical dimension and depth
measurement for HAR structures is critical for optimizing the Bosch deep etching process.
To date, various methods have been used to measure the HAR structures of MEMS [9,10].
Scanning electron microscopy (SEM) is a commonly used tool with high resolution and
high magnification, but it is costly, inefficient, and not suitable for online real-time detec-
tion. Optical methods, such as scatterometry and infrared reflectometry, are reported to
measure repeated structures with high throughput, but are limited in their ability for the
dimensional analysis of isolated HAR structures. Moreover, a silicon oxide film usually
https://www.mdpi.com/journal/applsci
Appl. Sci. 2022, 12, 8396 2 of 9

exists in etched MEMS device, which makes it more problematic for SEM and scatterometry.
Through-focus scanning optical microscopy (TSOM) is a novel fast and low-cost optical
measurement method with nanometer sensitivity, and is achieved with bright-field optical
microscopy [11]. The conventional TSOM method is a model-based metrology method
analyzing the scattering intensity reflected from the target and the intensity change as the
target performs through-focus scans along the optical axis. The dimensional information of
the target is extracted by matching the intensity information with the simulation results
in the database. For improving the performance, great attention has been paid to mini-
mizing the dissimilarities between the intensity information acquisition and simulation
conditions [12–19]. In the meantime, model-based TSOM utilizes only the optical inten-
sity range and the mean-square intensity information of the TSOM image; hence, large
amounts of intensity distribution information correlated to the dimensional information of
the targets is ignored. To date, TSOM has only been used to measure targets with critical
dimensions and depth of 1 µm or less, which is far from sufficient for MEMS.
Deep learning (DL) is a powerful artificial intelligence technology and has made impor-
tant progress in pattern recognition, computer vision, automatic speech recognition, natural
language processing, and other fields. Training DL algorithms directly with measurement
data is a new approach to TSOM, an alternative to establishing a simulation database. It
improves the TSOM performance by taking advantage of additional intensity information
and avoiding the adverse effect of the dissimilarities between information acquisition and
simulation conditions [20]. A convolutional neural network (CNN), as one of the most
representative DL algorithms, can automatically learn the deep correlations between the
scattering intensity information and dimensional information of the targets [21], and is
often used in the field of image processing.
In this paper, we describe a CNN learning model-based TSOM method for measuring
the individual HAR trench on a silicon substrate with width up to 30 µm and depth up
to 440 µm. The TSOM image is mapped to a normalized image of appropriate size, and
the width and depth measured by SEM are taken as labels for iterative training to obtain
a DL regression model that is capable of accurately predicting the width and depth. The
well-trained model can predict the width and depth of the HAR trench on the silicon
substrate with a silicon oxide film, and the standard deviation and error of a few tens of
nanometers are achieved. These results indicate that the DL-based TSOM method has great
application prospects for the micro-nano three-dimensional measuring field.
The structure of the rest of this paper is as follows: the second section introduces the
materials and methods, including the experimental setup, the acquisition of the dataset,
and the structure of the CNN model. The third section introduces our experimental results
and discussion. The fifth section introduces our conclusions.

2. Materials and Methods

2.1. TSOM Setup
A schematic of the TSOM setup is shown in Figure 1a. The TSOM setup is mainly com-
posed of a commercial bright-field optical microscope with a Kohler illumination system
(Shanghai Guangmi, Shanghai, China). A halogen lamp is used as the light source, which
has an illumination peak wavelength of 589 nm with a 3 dB spectral bandwidth of 160 nm.
A Piezoelectric Transducer (PZT, Coremorrow, Harbin, China) is placed under the sample
stage to achieve the through-focus scanning. The scattered light intensity distribution
information is acquired by a charge-coupled device (CCD) camera (Thorlabs, Newton, NJ,
USA) with a resolution of 1280 px × 1024 px and pixel size of 4.65 µm × 4.65 µm. Figure 1b
provides a photograph of the TSOM setup.
Figure 2 shows the MEMS trench section diagram and its geometric parameters that
require metrology. The HAR trench samples are etched on the silicon substrate by DRIE
with the Bosch process, and the silicon oxide film is retained. Owing to the characteristics
of DRIE, there is a width difference between the silicon oxide film and the silicon layer. In
this paper, width is defined as the width at the 50% thickness position of the silicon oxide
 Newton, NJ, USA) with a resolution of 1280 px × 1024 px and pixel size of 4.65 μm × 4.65
μm. Figure 1b provides a photograph of the TSOM setup.

Appl.Sci.
Appl. Sci.2022, 12,x 8396
2022,12, FOR PEER REVIEW 3 of3 10
of 9

film, and depth

distribution is defined
information as the distance
is acquired from the surface
by a charge-coupled of the
device silicon
(CCD) oxide(Thorlabs,
camera film to the
Newton, NJ, USA) with a resolution of 1280 px × 1024 px and pixel size of 4.65 μm ×of
bottom of the HAR trench. With the prepared TSOM setup, the width and depth the
4.65
trench can be obtained easily.
μm. Figure 1b provides a photograph of the TSOM setup.

Figure 1. (a) Schematic and (b) photograph of TSOM setup. CCD, charge-coupled device; PZT, Pie-
zoelectric Transducer.

Figure 2 shows the MEMS trench section diagram and its geometric parameters that
require metrology. The HAR trench samples are etched on the silicon substrate by DRIE
with the Bosch process, and the silicon oxide film is retained. Owing to the characteristics
of DRIE, there is a width difference between the silicon oxide film and the silicon layer. In
this paper, width is defined as the width at the 50% thickness position of the silicon oxide
film, and depth is defined as the distance from the surface of the silicon oxide film to the
bottom of the HAR trench. With the prepared TSOM setup, the width and depth of the
Figure 1.1.(a)can
trench
Figure Schematic
(a) andand
be obtained
Schematic (b) photograph
easily.
(b) of TSOM
photograph setup.
of TSOM CCD,CCD,
setup. charge-coupled device;
charge-coupled PZT, Pie-
device; PZT,
zoelectric Transducer.
Piezoelectric Transducer.

Figure 2 shows the MEMS trench section diagram and its geometric parameters that
require metrology. The HAR trench samples are etched on the silicon substrate by DRIE
with the Bosch process, and the silicon oxide film is retained. Owing to the characteristics
of DRIE, there is a width difference between the silicon oxide film and the silicon layer. In
this paper, width is defined as the width at the 50% thickness position of the silicon oxide
film, and depth is defined as the distance from the surface of the silicon oxide film to the
bottom of the HAR trench. With the prepared TSOM setup, the width and depth of the
trench can be obtained easily.

Figure
Figure 2. 2. MEMS
MEMS trench
trench section
section diagram
diagram and
and itsits geometric
geometric parameters.
parameters. W,W, width;
width; D, D, depth.
depth.

2.2. The Dataset

2.2. The Dataset
The traditional metrology method based on optical microscopy utilizes only the focal
The traditional metrology method based on optical microscopy utilizes only the focal
plane image, and the nonfocal plane images are discarded as useless information. For
plane image, and the nonfocal plane images are discarded as useless information. For ge-
geometric dimensions extracted from the focal plane image, the measurement accuracy
is ometric
limited bydimensions
the opticalextracted from
resolution andthe focal
focal plane
depth image,
of the the measurement
microscope. The light accuracy
intensity is
limited by the optical resolution and focal depth of the microscope.
information out of the focal plane is also related to the sample shape, and reflects The light intensity
the
information
geometric out of the focal
characteristics plane
on some is also
level. related
Figure to thethe
3 shows sample shape,
process and reflects
to construct the ge-
a TSOM
ometric
image. Thecharacteristics
TSOM setupon some level.
captures Figure
a series 3 shows
of images asthe
theprocess
sampleto construct
scans through a TSOM
the
Figure
focus2.ofMEMS trench section
the microscope diagram
along and itsaxis.
the optical geometric parameters.
The image W, width;
sequence D, depth.
consists of an in-focus
image and a series of defocused images of the sample. The acquired optical images are
2.2. The Dataset
stacked at their corresponding scan positions, and there is a slight difference between
anyThetwotraditional metrology
adjacent images. Themethod
opticalbased on optical
intensities microscopy
in a vertical utilizes
plane passingonly the focal
through the
plane
trench image,
on theand the can
target nonfocal plane
be easily imagesto
extracted are discarded
construct as useless
a TSOM information.
image. The X axes, For ge-
Y axes,
and grayscale represent the spatial position, the focus position, and
ometric dimensions extracted from the focal plane image, the measurement accuracy is the optical intensity,
respectively.
limited by theFigure
optical4 resolution
shows the SEM images
and focal of trench
depth of thesections with different
microscope. The lightwidths and
intensity
information out of the focal plane is also related to the sample shape, and reflects the ge-in
depths. The insets are the corresponding TSOM images, which show obvious differences
intensity
ometric distribution. The
characteristics widthlevel.
on some and depth
Figurecan be extracted
3 shows by establishing
the process the one-to-one
to construct a TSOM
 atattheir
theircorresponding
correspondingscan scanpositions,
positions,andandthere
thereisisaaslight
slightdifference
differencebetween
betweenany anytwo
two
adjacentimages.
adjacent images.TheTheoptical
opticalintensities
intensitiesininaavertical
verticalplane
planepassing
passingthrough
throughthethetrench
trenchon
on
thetarget
the targetcan
canbebeeasily
easilyextracted
extractedtotoconstruct
constructaaTSOM TSOMimage.
image.TheTheXXaxes,
axes,YYaxes,
axes,and
and
grayscalerepresent
grayscale representthethespatial
spatialposition,
position,the
thefocus
focusposition,
position,and
andthe
theoptical
opticalintensity,
intensity,re-
re-
Appl. Sci. 2022, 12, 8396 spectively. Figure 4 shows the SEM images of trench sections with different
spectively. Figure 4 shows the SEM images of trench sections with different widths and widths and
4 of 9
depths.The
depths. Theinsets
insetsare
arethe
thecorresponding
correspondingTSOM TSOMimages,
images,which
whichshow
showobvious
obviousdifferences
differences
ininintensity
intensitydistribution.
distribution.TheThewidth
widthand
anddepth
depthcancanbebeextracted
extractedby byestablishing
establishingthe
theone-to-
one-to-
onecorrespondence
one correspondencebetween
betweengeometric
geometriccharacteristics
characteristicsand andTSOM
TSOMimages.
images.Therefore,
Therefore,
correspondence between geometric characteristics and TSOM images. Therefore, TSOM
TSOMcan
TSOM can breakthethe limitsofofthe
the opticalresolution
resolutionand andfocal
focal depth.
can break break
the limits limits
of the opticaloptical
resolution and focal depth.depth.

Figure
Figure 3.Process
Figure3.3. Process
Processtoto construct
constructaaaTSOM
toconstruct TSOMimage.
TSOM image.
image.

Figure
Figure4.4.
Figure SEM
4.SEM
SEM images
images
imagesofofof
MEMS
MEMS
MEMS HAR
HARHAR trench
trench sections
sections
trench with
with
sections width
width
with andand
and
width depth
depth ofof(a)
depth (a) 1.98
of1.98 μmand
(a) μm
1.98 and169
µm 169
and
μm,
μm, (b)
(b) 10.3
10.3 μm
μm and
and 321
321 μm,
μm, (c)
(c) 30.3
30.3 μmμm and
and 443
443 μm,
μm, (d)
(d) 30.3
30.3 μm
μm and
and 37.3
37.3 μm,
μm, (e)
(e) 30.3μm
30.3μm
169 µm, (b) 10.3 µm and 321 µm, (c) 30.3 µm and 443 µm, (d) 30.3 µm and 37.3 µm, (e) 30.3µm and and
and 131
131
μm, (f)30.3
μm, 30.3 μmandand 443μm μm (inset:corresponding
correspondingTSOMTSOMimages).
images).
131 (f)
µm, (f) μm
30.3 µm 443
and 443(inset:
µm (inset: corresponding TSOM images).

For
For
Forthethesamples,
the samples,each
samples, eachtrench
each trench
trench isisis
55mm5mm
mm long.
long.
long. When
When
When collecting
collecting
collecting theTSOM
thethe
TSOM TSOM
image, image,
image, 1010po-
po-
10 positions
sitions
sitions are randomly
are randomly
are randomly selected
selected
selected over each over each
overtrench, trench,
each trench,
as shown as shown
as shown in Figure
in Figure
in Figure 5,
5, and5,theand
andpart the part
theofpart of
the of the
the
trench
trench
trench that
that isthat is contaminated
is contaminated
contaminated or close or close
ortoclose to the ends
to theshould
the ends should
ends should be avoided.
be avoided.
be avoided. Each position
Each position
Each position is scannedisis
scanned
scanned through-focus
through-focus
through-focus to obtaintotoaobtain
obtain
seriesaof aseries
seriesofofimages.
images. images. TheThescanning
The scanning scanning range
range
range is 250 µmisis250
250 μm
μm
with withaof
awith
step a
step
step
2 µm.ofof2Multiple
2μm.
μm.Multiple
Multiple lines
lines
lines are areselected
are
selected selected overeach
over over
each eachimage
image image withequal
with with
equal equal space,
space,
space, and andandconstruct
we wewecon-
con-
struct
struct multiple
multiple
multiple TSOM TSOM
TSOM images.
images.
images.
For the HAR trench etch, the reactive-ion etch lag leads to differences in trench depths
with different trench widths [6]. For example, the depths of 2 µm, 10 µm, and 30 µm
wide trenches etched simultaneously are 169 µm, 321 µm, and 443 µm, respectively. It
is difficult to obtain trenches of the same depth with different widths in a large range
(e.g., 2 µm~30 µm). To alleviate the impact of the depth difference, the trench sets for width
prediction consist of 4 trenches with widths around 2 µm, 10 µm, and 30 µm, respectively,
and the corresponding depths are approximately 169 µm, 321 µm, and 443 µm. The
depth prediction trench sets consist of 5 trenches with widths of 1.93 µm, 10.3 µm, and
30.3 µm, respectively. The corresponding depth ranges are from 24 µm to 169 µm, 34 µm to
321 µm, and 37 µm to 443 µm. The highest aspect ratios are approximately 88:1, 31:1, and
15:1, respectively.
Appl.
Appl.Sci.
Sci.2022,
2022,12,
12,x8396
FOR PEER REVIEW 5 of 109
5 of

Figure
Figure5.5.Process
Processto
tocollect
collectTSOM
TSOMimage
imagedataset.
dataset.

2.3. The Structure of the CNN

For the HAR trench etch, the reactive-ion etch lag leads to differences in trench
depths Towith
reveal the correspondence
different trench widthsbetween geometricthe
[6]. For example, characteristics
depths of 2 μm,and 10
TSOM
μm, images,
and 30
we resize the TSOM image to 256 × 64 pixels, and then construct
μm wide trenches etched simultaneously are 169 μm, 321 μm, and 443 μm, respectively.a suitable CNN model
for training. CNN models usually consist of a convolution layer, pooling
It is difficult to obtain trenches of the same depth with different widths in a large range layer, activation
layer, and full connection layer. An improved CNN model is designed based on the
(e.g., 2 μm~30 μm). To alleviate the impact of the depth difference, the trench sets for
traditional CNN model, as shown in Figure 6. The TSOM image is a grayscale image,
width prediction consist of 4 trenches with widths around 2 μm, 10 μm, and 30 μm, re-
so the single-channel image can be convolved. There are 5 convolution layers (including
spectively, and the corresponding depths are approximately 169 μm, 321 μm, and 443 μm.
two 3 × 5 kernels and three 3 × 3 kernels), 4 pooling layers (being applied to the first
The depth prediction trench sets consist of 5 trenches with widths of 1.93 μm, 10.3 μm,
four convolution layers), and 4 full connection layers in the CNN model. All convolution
and 30.3 μm, respectively. The corresponding depth ranges are from 24 μm to 169 μm, 34
layers contain zero padding. The ReLU function is used as the activation function, and the
μm to 321 μm, and 37 μm to 443 μm. The highest aspect ratios are approximately 88:1,
maximum pooling function is used as the pooling layer. The parameters of the CNN model
31:1, and 15:1, respectively.
can be found in Figure 6. The mean-square loss is used as the final loss function to calculate
the mean-square error between the output values and the actual values. The loss function
2.3. The Structure of the CNN
equation is:
To reveal the correspondenceloss between
( xi , yi ) geometric
= ( xi − yi )characteristics
2 and TSOM images, (1)
we
Appl. Sci. 2022, 12, x FOR PEER REVIEW
resize the TSOM image to 256 x 64 pixels, and then construct a suitable CNN model
6 of 10
for training.
where CNN models
xi indicates the trueusually consist
value of of a convolution
the target layer,the
and yi indicates pooling layer,value
predicted activation
of the
layer, and full connection layer. An improved CNN model is designed based on the tra-
CNN model.
ditional CNN model, as shown in Figure 6. The TSOM image is a grayscale image, so the
single-channel image can be convolved. There are 5 convolution layers (including two 3 ×
5 kernels and three 3 × 3 kernels), 4 pooling layers (being applied to the first four convo-
lution layers), and 4 full connection layers in the CNN model. All convolution layers con-
tain zero padding. The ReLU function is used as the activation function, and the maximum
pooling function is used as the pooling layer. The parameters of the CNN model can be
found in Figure 6. The mean-square loss is used as the final loss function to calculate the
mean-square error between the output values and the actual values. The loss function
equation is:
𝑙𝑜𝑠𝑠(𝑥 , 𝑦 ) = (𝑥 − 𝑦 ) (1)
where xi indicates the true value of the target and yi indicates the predicted value of the
CNN model.
The training image set and test image set contain 3000 and 200 TSOM images, respec-
tively, which are generated from one trench set. SEM is used to measure the width and
depth by the cross-section of the trench center position, and the SEM result is taken as
labels. The training data set and SEM results are applied to train the CNN model for learn-
ing the correspondence of TSOM images and trench parameters. The trained model is
evaluated by the test data set.

Figure 6. Schematic diagram of improved CNN model.

Figure 6. Schematic diagram of improved CNN model.

3. Results and Discussion

To verify the parameter prediction performance of the trained CNN model, the posi-
tion close to the trench center was through-focus scanned continually 10 times. The center
lines of a series of microscope images were selected to construct prediction TSOM image
Appl. Sci. 2022, 12, 8396 6 of 9

The training image set and test image set contain 3000 and 200 TSOM images, re-
spectively, which are generated from one trench set. SEM is used to measure the width
and depth by the cross-section of the trench center position, and the SEM result is taken
as labels. The training data set and SEM results are applied to train the CNN model for
learning the correspondence of TSOM images and trench parameters. The trained model is
evaluated by the test data set.

3. Results and Discussion

To verify the parameter prediction performance of the trained CNN model, the position
close to the trench center was through-focus scanned continually 10 times. The center lines
of a series of microscope images were selected to construct prediction TSOM image sets,
and the average of 10 measurement results was considered as the predicted value.
Figure 7a–c show a comparison between the true (black triangles) and predicted (red
circles) widths of HAR trenches with widths of around 2 µm, 10 µm, and 30 µm, respectively.
Figure 7e shows a comparison between the actual (black triangles) and predicted (red circles)
depths of HAR trenches with widths of 1.93 µm, 10.3 µm, and 30.3 µm, respectively. The
comparison results show that the predicted value and true value match well. In order to
further analyze the results shown in Figure 4, the standard deviation and error were used
to evaluate the prediction performance. The definitions of the standard deviation and error
are given in Equations (2) and (3), respectively:
s
n  2
1
SD = ∑
n − 1 i =1
fi − f (2)

E = f − ft (3)
Appl. Sci. 2022, 12, x FOR PEER REVIEW 7 of 10
where fi is the predicted value, f is the average of the predicted values, ft is the true value,
and n is the number of measurements, which is 10 in this paper.

Figure
Figure 7.
7. Width
Width and
and depth
depth prediction
prediction results.
results. Comparison
Comparison between
between the
the actual
actual (black
(black triangles)
triangles) and
and
predicted (red circles) widths of HAR trenches with widths of around (a) 2 μm, (b) 10
predicted (red circles) widths of HAR trenches with widths of around (a) 2 µm, (b) 10 µm, and μm, and (c)
30 μm; comparison between the actual (black triangles) and predicted (red circles) depths of
(c) 30 µm; comparison between the actual (black triangles) and predicted (red circles) depths of HARHAR
trenches with widths of (d) 1.93 μm, (e) 10.3 μm, and (f) 30.3 μm.
trenches with widths of (d) 1.93 µm, (e) 10.3 µm, and (f) 30.3 µm.

The standard deviation and error of the width prediction results are shown in Figure
8. As a comparison, the results predicted by the traditional machine learning (ML) method
demonstrated in Reference [22] are shown in order to verify the superiority of the CNN
model. The adopted ML model is a support vector regression (SVR) model combined with
a Histogram of Oriented Gradients (HOG) feature extractor [23–25]. The three trench sets
Appl. Sci. 2022, 12, 8396 7 of 9

The standard deviation and error of the width prediction results are shown in Figure 8.
As a comparison, the results predicted by the traditional machine learning (ML) method
demonstrated in Reference [22] are shown in order to verify the superiority of the CNN
model. The adopted ML model is a support vector regression (SVR) model combined
with a Histogram of Oriented Gradients (HOG) feature extractor [23–25]. The three trench
sets for width prediction were irregularly spaced, with width intervals of a few hundred
nanometers around 2 µm, 10 µm, and 30 µm. The standard deviation and error were
less than 10 nm for 2 µm and 10 µm wide trenches. The corresponding ML-predicted
results were 40 nm and 180 nm for 2 µm and 10 µm wide trenches, which were much
higher than the DL-predicted results, especially for 10 µm wide trenches. For 30 µm wide
trenches, the standard deviation and error were significantly increased, but still less than
60 nm and 80 nm, respectively. The increase in the standard deviation and error was
mainly due to the etching error, which increased as the width increased. Moreover, the
depth of the 30 µm wide trench was 443 µm. A large scanning range was necessary for a
deep trench to obtain the complete TSOM image. Incomplete TSOM images also led to
poorer performance. The corresponding ML-predicted results were 170 nm and 250 nm,
Appl. Sci. 2022, 12, x FOR PEER REVIEW 8 of 10
which were approximately three times the DL-predicted results. The comparison results
confirmed the superior prediction ability of the DL model for width prediction.

Figure8.8.Standard
Figure Standard deviation
deviation and
and error
error of
of width
width results
results for
for widths
widthsaround
around(a)
(a)22μm,
µm,(b)
(b)10
10μm,µm,and
and
(c)30
(c) 30µm,
μm,predicted
predictedby
by machine
machine learning
learning model
model (dot
(dot line)
line) and
anddeep
deeplearning
learningmodel
model(solid
(solidline).
line).ML,
ML,
machinelearning;
machine learning;DL,
DL, deep
deep learning;
learning; SD,
SD, standard
standard deviation.
deviation.

Figure99illustrates
Figure illustratesthe
thestandard
standard deviation
deviation andand error
error of the
of the depth
depth prediction
prediction results.
results. The
The 1.93 μm, 10.3 μm, and 30.3 μm wide trench sets for depth prediction were
1.93 µm, 10.3 µm, and 30.3 µm wide trench sets for depth prediction were also irregularly also irreg-
ularly spaced.
spaced. As theAs the depth
depth increased,
increased, the depth
the depth interval
interval alsoalso increased
increased from
from a few
a few tensof
tens
of micrometers
micrometers to more
to more than
than 200200 μm.The
µm. Thestandard
standarddeviation
deviation and
and error
error were
were 110
110nmnmandand
360nm
360 nmforfor 1.93
1.93 µmμm wide
wide trenches.
trenches. For
For 10.3
10.3 μm
µm wide
wide trenches,
trenches,the
theresults
resultswere
were9090nmnmandand
410nm.
410 nm.The
Thecorresponding
correspondingML-predicted
ML-predicted results
results were
were around
around a few
a few micrometers,
micrometers, and
and were
were approximately
approximately an order
an order of magnitude
of magnitude higherhigher
thanthan the DL-predicted
the DL-predicted results.
results. ForFor
30.330.3
µm
μm wide trenches, the standard deviation and error were 280 nm and 300
wide trenches, the standard deviation and error were 280 nm and 300 nm, respectively. The nm, respec-
tively. The corresponding
corresponding ML-predicted ML-predicted
results wereresults
2.4 µmwere
and 2.4
850 μm
nm,and
which850were
nm, much
which higher
were
much
than higher
the than the DL-predicted
DL-predicted results. results.
For ML depth prediction results, the predicted performance was degraded obviously
as the trench depth increased or width decreased. This can be understood by the fact that
less light reached the bottom and carried the depth information for narrow or deep trenches.
Less depth information collected resulted in a larger standard deviation and greater error.
Moreover, the scanning range was much smaller than the depth of the trench. This resulted
in incomplete TSOM images, which could lead to poorer performance. However, the DL
depth prediction performance did not change obviously as the trench depth or width
changed. It was proven that the DL model extracted more information from less input data
compared with the ML model. It was also noticed that the standard deviation and error
of the depth prediction results were obviously higher than those of the width prediction
results. This was mainly because the depth interval of the depth training trench set was
Figurelarger
much 9. Standard deviation
than the widthand error of
interval of depth results
the width for widths
training of (a)set,
trench 1.93 μm, (b)
which was10.3 μm,aand
only few
(c) 30.3 μm, predicted by machine learning model (dot line) and deep learning model (solid line).
ML, machine learning; DL, deep learning; SD, standard deviation.

For ML depth prediction results, the predicted performance was degraded obviously
as the trench depth increased or width decreased. This can be understood by the fact that
 Figure 9 illustrates the standard deviation and error of the depth prediction results.
The 1.93 μm, 10.3 μm, and 30.3 μm wide trench sets for depth prediction were also irreg-
ularly spaced. As the depth increased, the depth interval also increased from a few tens
of micrometers to more than 200 μm. The standard deviation and error were 110 nm and
360 nm for 1.93 μm wide trenches. For 10.3 μm wide trenches, the results were 90 nm and
Appl. Sci. 2022, 12, 8396 410 nm. The corresponding ML-predicted results were around a few micrometers, and 8 of 9

were approximately an order of magnitude higher than the DL-predicted results. For 30.3
μm wide trenches, the standard deviation and error were 280 nm and 300 nm, respec-
tively. The
hundred corresponding
nanometers. ML-predicted
The prediction results were
performance 2.4 be
could μmfurther
and 850 nm, which
improved were
the proper
much higher
adoption than the DL-predicted
of parameter intervals. results.

Figure 9.
Figure 9. Standard
Standard deviation
deviation and
anderror
errorof
ofdepth
depthresults
resultsfor
forwidths
widthsofof(a)
(a)1.93
1.93μm,
µm,(b)
(b)10.3
10.3μm,
µm,and
and
(c) 30.3 μm, predicted by machine learning model (dot line) and deep learning model (solid line).
(c) 30.3 µm, predicted by machine learning model (dot line) and deep learning model (solid line). ML,
ML, machine learning; DL, deep learning; SD, standard deviation.
machine learning; DL, deep learning; SD, standard deviation.
For ML depth
Previously prediction
reported TSOM results, the predicted
systems performance
had achieved was degraded
a nanometer measurementobviously
resolu-
as the trench depth increased or width decreased. This can be understood
tion, which was much higher than the result shown in this paper. This was mainly by the factdue
thatto
lessaccuracy
the light reached
of the the
SEMbottom
values.andThecarried the depth
SEM value information
was taken forvalue,
as the true narrow ordirectly
and deep
trenches.the
affected Less depth information
prediction ability of thecollected
CNN resulted
model. in In aour
larger standard
study, the SEM deviation
results and
were
greater error. Moreover, the scanning range was much smaller than
rounded to the nearest 0.01 µm for dimensions less than 10 µm, 0.1 µm for dimensions fromthe depth of the
trench. This resulted in incomplete TSOM images, which could lead to
10 µm to 100 µm, and 1 µm for dimensions greater than 100 µm, and this was not sufficient poorer perfor-
mance.
for However,measurement
dimensional the DL depth withprediction performance
nanometer accuracy. didMoreover,
not changethe obviously
trainingas the
trench
trench depth or width changed. It was proven that the DL model extracted
set and trench set for SEM were etched in the same wafer, but with different positions, so more infor-
mation from
etching errorsless
alsoinput data compared
degraded the predictionwithability
the ML of model.
the CNN It was also noticed that the
model.
standard deviation and error of the depth prediction results were obviously higher than
those
4. of the width prediction results. This was mainly because the depth interval of the
Conclusions
depth Intraining
this work,trench
HARsettrenches
was much withlarger
widthsthan the2width
from µm tointerval
30 µm andof the width
depths training
from 20 µm
trench set, which was only a few hundred nanometers. The
to 440 µm were measured by a CNN model-based TSOM method. The TSOM image prediction performance could
was
be further
mapped toimproved
a normalizedthe proper
image of adoption of parameter
appropriate size, andintervals.
the width and depth measured by
SEM were taken as labels for iterative training to obtain a DL regression model that was
capable of accurately predicting the width and depth. The standard deviation and error
were around a hundred nanometers or less. The superiority of the convolutional neural
network model-based TSOM compared to machine learning-based TSOM was experimen-
tally investigated and discussed. This method has the potential to achieve nanometer
precision, and its application fields can be extended to integrated circuit metrology, such as
through-silicon via.

Author Contributions: Conceptualization, J.S. and X.C.; methodology, G.L. and J.S.; software, G.L.;
validation, G.L. and C.G.; formal analysis, S.H.; investigation, G.L.; resources, C.C.; data curation,
G.L.; writing—original draft preparation, G.L. and J.S.; writing—review and editing, J.S. and X.C.;
visualization, X.J.; supervision, W.Z.; project administration, X.C.; funding acquisition, W.Z. All
authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the National Key Research and Development Program
of China (No. 2019YFB2005600), National Key Laboratory of Precision Testing Techniques and
Instrument, Tsinghua University (No. TH20-01), National Natural Science Foundation of China
(No. 51905528), State Key Laboratory of Precision Measuring Technology and Instruments (Tianjin
University) (No. pilab2102), and Key Laboratory of Microelectronic Devices Integrated Technology,
Chinese Academy of Sciences.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available from the corresponding
author, J.S., upon reasonable request.
Appl. Sci. 2022, 12, 8396 9 of 9

Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wang, P.; Lu, Q. Evolutionary design optimization of MEMS a review of its history and state-of-the-art. Clust. Comput. 2019,
22, 9105. [CrossRef]
2. Shinde, P.; Shiurkar, U. MEMS for detection of environmental pollutants: A review pertains to sensors over a couple of decades in
21st century. Mater. Today 2021, 44, 615–624. [CrossRef]
3. Bosch, S.; Xu, J. Real time etch depth measurements of mems devices. J. Microelectromech. Syst. 2002, 11, 111–117. [CrossRef]
4. Xie, J. Fabrication challenges and test structures for high-aspect-ratio SOI MEMS devices with refilled electrical isolation trenches.
Microsyst. Technol. 2015, 21, 1719–1727. [CrossRef]
5. Kim, H.; Kim, D. Photo-assisted electrochemical etching of a nano-gap trench with high aspect ratio for MEMS applications.
J. Micromech. Microeng. 2006, 16, 906–913. [CrossRef]
6. Gerlt, M.; Laubli, N. Reduced etch lag and high aspect ratios by deep reactive ion etching (DRIE). Micromachines 2021, 12, 542.
[CrossRef]
7. Laermer, F.; Urban, A. MEMS at Bosch–Si plasma etch success story, history, applications, and product. Plasma Process. Polym.
2019, 16, 1800207. [CrossRef]
8. Baklykov, D.; Andronic, M. Self-Controlled Cleaving Method for Silicon DRIE Process Cross-Section Characterization.
Micromachines 2021, 12, 534. [CrossRef]
9. Attota, R.; Kavuri, P. Nanoparticle size determination using optical microscopes. Appl. Phys. Lett. 2014, 105, 163105. [CrossRef]
10. Attota, R.; Kang, H. Nondestructive shape process monitoring of three-dimensional, high-aspect-ratio targets using through-focus
scanning optical microscopy. Meas. Sci. Technol. 2018, 29, 125007. [CrossRef]
11. Attota, R.; Germer, T. Through-focus scanning-optical-microscope imaging method for nanoscale dimensional analysis. Opt. Lett.
2008, 33, 1990–1992. [CrossRef] [PubMed]
12. Ryabko, M.; Shchekin, A. Through-focus scanning optical microscopy (TSOM) considering optical aberrations: Practical imple-
mentation. Opt. Express 2015, 23, 32215–32221. [CrossRef] [PubMed]
13. Attota, R. Noise analysis for through-focus scanning optical microscopy. Opt. Lett. 2016, 41, 745–748. [CrossRef]
14. Attota, R. Fidelity test for through-focus or volumetric type of optical imaging methods. Opt. Express 2018, 26, 19100–19114.
[CrossRef]
15. Park, S.; Park, G. Through-focus scanning optical microscopy with the Fourier modal method. Opt. Express 2018, 26, 11649–11657.
[CrossRef] [PubMed]
16. Park, S.; You, B. Metrological sensitivity improvement of through-focus scanning optical microscopy by controlling illumination
coherence. Opt. Express 2019, 27, 1981–1990. [CrossRef] [PubMed]
17. Rim, M.; Agocs, E. Detecting nanoscale contamination in semiconductor fabrication using through-focus scanning optical
microscopy. J. Vac. Sci. Technol. B Nanotechnol. Microelectron. Mater. Process. Measur. Phenom. 2020, 38, 050602. [CrossRef]
18. Peng, R.; Jiang, J. Lateral movement and angular illuminating non-uniformity corrected TSOM image using Fourier transform.
Opt. Express 2020, 28, 6294–6305. [CrossRef]
19. Lee, J.; You, B. Motion-free TSOM using a deformable mirror. Opt. Express 2020, 28, 16352–16362. [CrossRef]
20. Qu, Y.; Hao, J. Machine-learning models for analyzing TSOM images of nanostructures. Opt. Express 2019, 27, 33978–33998.
[CrossRef]
21. Kwak, H.; Ryu, S. Non-destructive thickness characterisation of 3D multilayer semiconductor devices using optical spectral
measurements and machine learning. Light: Adv. Manuf. 2021, 2, 9–11. [CrossRef]
22. Li, G.; Shi, J. Through-focus scanning optical microscopy measurement based on machine learning. Chin. Opt. 2022, 15, 703–711.
(In Chinese)
23. Dalal, N.; Triggs, B. Histograms of oriented gradients for human detection. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, San Diego, CA, USA, 20–25 June 2005.
24. Basak, D.; Srimanta, P. Support Vector Regression. Neur. Inform. Proc. 2007, 11, 203.
25. Yan, G.; Zhu, Y. Parameters Selection Method for Support Vector Machine Regression. Comput. Eng. 2009, 35, 218. (In Chinese)