2019 Thailand-Japan MicroWave (TJMW2019)

Auto-Encoder/Decoder for Planar Filter Analysis/Synthesis

Kosuke KATAYAMA† and Takaaki BABA‡
†Graduate School of Information, Production and Systems, Waseda University, Kitakyushu-shi, Fukuoka, 808-0135 Japan
‡Research Center of Information, Production and Systems, Waseda University, Kitakyushu-shi, Fukuoka, 808-0135 Japan
E-mail: †kkatayama@aoni.waseda.jp, ‡tbaba@waseda.jp

Abstract In this paper, we propose the first study which estimates the frequency response directly from the geometry of a
planar filter, and which also synthesizes the planar filter geometry directly from the given frequency response using a
convolutional neural network (CNN) based auto-encoder/decoder. we also explain the way to generate an accurate and massive
dataset for training the auto-encoder/decoder. In our experiments, the frequency response is estimated in 1.5 msec and the filter
geometry is synthesized in 2.7 msec, respectively.
Keyword Convolutional neural network, Edge model de-embedding/embedding, Electromagnetic simulation, F-parameter,
Microwave

1. I NTRODUCTION 2. D ATASET G ENERATION

In a microwave region, planar filters are u tilized. High-speed calculation of the frequency response
Since the planar filter is composed of distributed is vital to generate a huge amount of dataset. The
elements, element values of a filter theory cannot be frequency response of a planar filter can be calculated
used. Therefore, filter designers must tune the filter quickly by a cascade production of F-parameters [1,
geometry using an electromagnetic (EM) simulator 2]. However, the accuracy of the calculation is not
until the frequency response meets their wants. Since enough because uniformity of current distribution is
this task takes long time due to the long duration of corrupted at the edge where the two lines are
EM simulations, several neural networks have been connected as shown in Fig. 2. The accuracy of the
proposed for the reduction of design time. However, frequency response guarantees the consistency
these proposals need the design parameters of the between the filter geometry and its frequency
filter geometries and the neural networks tune the
response which is necessary for the convergence of
design parameters such as line lengths, width s and so
the neural network [3]. Since the uniformity of the
on. Extraction of the design parameters gives another
current distribution is corrupted at the edge of the
task to designers. Moreover, a designe r must set
constraints of the design parameters and convert the wires, the total F-parameter F total is not a simple
obtained design parameter into an actual geometry product of F-parameters (F left · F ri ght ) which is
although the neural networks are employed for the extracted under the conditions of uniform current
reduction of tasks. On the other hands, there is a distributions (Fig. 2). If we succeeded to extract the
convolutional neural network (CNN) called the auto- edge model F edge , the accuracy of F total would be
encoder/decoder (Fig. 1). This auto-encoder/decoder improved as F left ·F ed ge ·F ri ght . Following part of this
converts an image into a similar image via a compact section describes how to extract the F ed ge (de-
vector containing a feature of the image. Since the embedding) and implant the F edge (embedding). Since
planar filter has a two -dimensional geometry in its the F ed ge is functions of signal line widths ( W 1 , W 2 ),
surface, the auto-encoder/decoder can deal it with the all combinations of W 1 and W 2 should be examined as
image. illustrated in Fig. 3. This combination table contains
In this paper, we mapped the frequency response information about the edge models at the part of W 1 <
into the feature vector of the auto -encoder/decoder. W 2 and W 1 > W 2 with lead lines of W 1 and W 2 . This
As a result, auto-encoder outputs the frequency combination table also contains straight lines when
response from the planer filter image. Conversely, W 1 equals to W 2 and they are used for de-embedding
auto-decoder output the planer filter image from the the lead lines. Since these lines (W 1 = W 2 ) are two
desired frequency response. This is a first paper which times longer than lead lines of de -embedding targets
converts bidirectionally between the filter geometry (W 1 < W 2 , W 1 > W 2 ), they should be halved using
and frequency response not to be detouring the design equations in Fig. 4. In this case, the scaling factor n
parameters. In Section 2, we consider the dataset is 0.5. Figure 5(a) shows the edge model de -
generation for training the auto-encoder/decoder
embedding. Since the non-diagonal part of the table
because the quantity and quality of dataset are very
in Fig. 3 contains lead lines ( W 1 and W 2 ) and the edge
important for training as well as other neural networks.
model F edge , the lead lines are de-embedded using
 2019 Thailand-Japan MicroWave (TJMW2019)

diagonal part of the table (W 1 = W 2 ) after halving their lengths (L n ) are generated and connected u ntil the
lengths. The edge model F edge is a cascade product of accumulated length reaches the half of the total length
the inverse function of corresponding lead line F 1 , de- L total . Conversely, entire geometry is a unfold image
embedding target, and the inverse function of F 2 . As over vertically and horizontally. The frequency
a result, F ed ge represents the non-uniformity of the response of the randomly generated geometry is
current distribution as frequency response and its calculated and stored in a dataset with it s quarter-
length is zero because total lengths of the de - sized image. Depending on the geometry the line
embedding target (L 1 + L 2 ) is coincident with the total lengths are scaled as shown in Fig. 4. Edge models are
subtraction lengths (L 1 , L 2 ). Figure 6 shows a random embedded depending on the line widths as shown in
generation of a filter geometry. Since the filter Fig. 5(b). Since the edge model has no length, the
geometry has a symmetrical structure over vertically embedding does not affect the total length and affects
and horizontally, the quarter size of the geometry is only the frequency response.
generated as an image. Random widths ( W n ) and

Convolution
+ Convolution
Sigmoid +
42 Sigmoid Full-connection
+
2 90 Sigmoid
1
100 3 6
3 14
50 200
6
Auto-encoder (CNN)/EM simulation (human)

Frequency response
1
61
Auto-decoder (CNN)/Artwork (human)
Deconvolution
+ Deconvolution Full-connection
Sigmoid + +
Sigmoid Sigmoid
42

2 90
1 1
100 3 6
3 14
50 200
6

1
Fig. 1. Auto-encoder and decoder of a planar filter using convolutional neural networks (CNNs). Since the surface of a planar filter has
a two-dimensional geometry, it can be assumed as an image. The system of an auto-encoder and decoder is a special case of a
neural network which converts an image to its similar image via a compact vector which contains a feature of the images. The
compact vector is mapped to a frequency response of a planar filter in this case. The path from an image to its feature vector is
called auto-encoder. Auto-decoder is a path from a feature vector to its related image. Since the images have higher-order
vectors than the feature vector, convolution, and de-convolution are utilized for down- and up-sampling in the auto-encoder
and decoder paths. A couple of full-connected neural networks are the bridges of a convoluted image to the feature vector or
the deconvoluted image from the feature vector. Non-linear sigmoidal functions are inserted properly to extract logical
relationships and to normalize the values. This auto-decoder and encoder paths are also human mimic tasks. When designing a
planar filter, the human creates the filter geometry from the desired frequency. Afterword, the frequency response is confirmed
using an electromagnetic (EM ) simulator to ensure its response meets the desired frequency.
 2019 Thailand-Japan MicroWave (TJMW2019)

Fleft Fright Fedge

W1 W1
W2 W2
x x =
L1 L1 L2 L2 0

(a) Edge model (Fedge ) de-embedding.

Ftotal W1 W1
W2 W2
Ftotal ≠ Fleft · Fright Ftotal ≈ Fleft · Fedge · Fright x x =
L1 0 L2 L1 L2
Fig. 2. The current density of connected transmission lines (TLs)
whose widths are different. At the edge where the two TLs (b) Edge model (Fedge ) embedding.
are connected, the current distribution is non-uniform.
Fig. 5. Edge model Fedge de-embedding and embedding. Edge
Therefore, the simple product of left- and right-side F-
model de-embedding is calculated from the inverse
parameters (Fleft, Fright) is different from the total one (Ftotal).
functions of TLs whose length and width are coincided
For the accurate calculation of the Ftotal, the non-uniformity
with the those of target ones. Conversely, Fedge is
of the connection (Fedge ) should be considered.
embedded between TLs when the line widths are different.

W1 W2 W1
W2 Ltotal
Port1

Port2

Port1

Port2

L1 L2 L3/2
W2
W2/2

port2
port1

W1 = W2 W1 < W2

W1/2 unfold
W3/2
W1
unfold

W1 > W2 W1 = W2 Fig. 6. Random geometry generation of a filter. TL length (L1, L2

and L3/2) and width (W1/2, W2/2 and W3/2) are generated
W2 randomly until the sum of the length (L1 + L2 + L3/2)
W1 W1 W2
Port1

Port2
Port1

Port2

reaches the half of the total length (Ltotal/2). The geometry

is mirrored vertically and horizontally.

Fig. 3. A combination table of transmission lines. This table is

generated by changing the line width W1 (port 1 side) and
W2. It contains all variations of a straight line (W1 = W2)
and all combinations of line width (W1 < W2). Note that the
W1 > W2 is generated by the port swappingof W1 < W2. Port2
1.8 mm thickness
Copper
W2

W1
W1
L1 0.8mm thikness
Polytetrafluoroethylene
Port1 (PTFE)
W1 er = 2.06, tand = 0.0002
Fig. 7. The appearance of the EMPro, which is an electromagnetic
L2 (EM ) simulator provided by Agilent technology Inc. Signal
line of a 1.8 mm copper foil is on a 0.8 mm
polytetrafluoroethylene (PTFE) bulk which is shield with a
Fig. 4. Length scaling of a transmission line (TL). Since the perfect electrical conductor (PEC). Signal line widths of W1
propagation constant g and characteristic impedance Z0 are and W2 are swept to generate the combination table of Fig.
not depending on the length of the TL, scaled F-parameter 3. The width sweeping and S-parameter extraction are
(F2) is calculated from F1 using the scaling factor n. performed automatically by a built-in Python script.
 2019 Thailand-Japan MicroWave (TJMW2019)

Fedge1 Fedge3 In the auto-decoder path, frequency response is fully

L1 L2 L3 L2 L1 connected to 14 x 3 x 6 matrix. The 14 x 3 x 6 matrix
is converted into 6 x 3 x 100 using 1 x 90

port2
port1

deconvolution kernels (stride is 4 x 2). The 6 x 3 x

W2
W2
F1 F2 F3 F4 F5 100 matrix is converted into 50 x 100 image using 42
W1 W3 W1 x 2 kernel (stride is 1 x 2). The calculation costs of
Fedge2 Fedge4 the auto-encoder/decoder are 302,652 and 7,323,372
(a) Geometry. in terms of the product-sum.
Ftotal = F1 ∙ F2 ∙ F3 ∙ F4 ∙ F5
Ftotal = F1 ∙ Fedge1 ∙ F2 ∙ Fedge2 ∙ F3 ∙ Fedge3 ∙ F4 ∙ Fedge4 ∙ F5 Image (quarter size of the geometry)
Frequency response
0
S21 S11

with edge model w/o edge model

-5 (2 msec) (1 msec)
S-parameter [dB]

discrepancy
(w/o edge model)
-10

EM simulation
(1.5 hours
-15 @ Core i5)
L1 = L2 = L3 = 4.0 mm
W1 = W3 = 0.4 mm
W2 = 3.2 mm
-20
0 5 10 15
Frequency [GHz]
(b) Frequency responses.
Fig. 8. Geometry of a planar filter and its frequency response. Dash Fig. 9. Part of a dataset. The dataset contains pairs of an image and
and solid-lines are S11 and S21 when the geometry is its frequency response. The image is a randomly generated
analyzed with an electromagnetic simulator. Symbols
quarter-sized planar filter (Fig. 6). Its frequency response is
denotes the reconstructed responses calculated from F-
calculated quickly and accurately taking account of the
parameters. Circle (triangle) symbols are case of ignoring
edge model (Fig. 5). Quantity and quality of a dataset are
(considering) the edge model.
important for the auto-encoder/decoder training (Fig. 1).

3. E XPERIMENTS
0.0105
Figure 1 also shows a structure of the auto -
encoder/decoder. The size of the target planer filter is
1 x 4 cm. Metal and bulk parts are converted into
Cost function/RMSE 2

white and black pixels. One pixel corresponds to 0.1 0.0100

x 0.1 mm and the quarter-sized image becomes 50 x
100. Since the frequency response is a magnitude of RMSE2
S 21 from DC to 30 GHz (500 MHz step), the number 0.0095
of points is 61. The pairs of image and frequency
response are stored in the dataset. The images are fed
to the auto-encoder and output of the auto -encoder is 0.0090
compared with the frequency response and residue are Cost function
fed back to train the network. The frequency
responses are fed to the auto-decoder and output of
0.0085
the auto-decoder is compared with the image and 0 100 200 300 400
residue are fed back to train the network. In the auto -
encoder path, an image is converted into 6 x 3 x 100 Epoch
matrix using 42 x 2 convolution ke rnels (stride is 4 x Fig. 10. Cost function and the square of RM S error (RM SE2) as a
2). The 6 x 3 x 100 matrix is converted into 14 x 3 x function of Epoch while input and output of auto-encoder
(Fig. 1) are images and their frequency responses (Fig.
6 using 1 x 90 kernels (stride is 1 x 2). The 14 x 3 x 6
9) , respectively.
matrix are fully connected to the frequency response.
 2019 Thailand-Japan MicroWave (TJMW2019)

To generate the table shown in Fig. 3, an EM 0.080

simulator called EMPro is utilized as shown in Fig. 7.
Figure 8(a) shows an example of a filter geometry.
Figure 8(b) shows the frequency responses of the

Cost function/RMSE2
0.075
geometry. The cascaded production of F-parameters
with edge models (triangles) exactly traces the
frequency response which is the EM simulation result
of the entire geometry (lines). On the other hands, the 0.070
one without edge models has a discrepancy. This RMSE2
result shows the accuracy using edge mo dels is
comparable to that of the EM simulation even though 0.065
the calculation is done in 2 msec while the EM Cost function
simulation takes 1.5 hours.
0.105 0.060
0 100 200 300 400
Epoch
0.100
RMS Error (Validation)

Fig. 13. Cost function and the square of RM S error (RM SE2) as a
function of Epoch while input and output of auto-decoder
(Fig. 1) are frequency responses and their images (Fig. 9),
0.095 respectively.

64 batch size
0.090 400 epoch Figure 9 shows a part of the dataset. The dataset
80% for Training contains pars of randomly generated image and its
20% for Validation frequency response, which is calculated accurately
0.085 considering the edge models. Taking advantage of
1k
k 3k 30k
high-speed calculation, 30,000 pairs are stored in the
Dataset size dataset.
Fig. 11. RM S error of auto-encoder with respect to the dataset size. Figure 10 shows a learning curve of the auto-
Input image encoder. Figure 11 shows the root-mean-square
50 pixels (RMS) error with respect to the dataset size. For each
10 mm size, 80% (20%) of the dataset is used for training
200 pixels
(validation). The quantity of dataset improves the
40 mm
neural network. Figure 12 shows the input image and
(a) Input image and unfolded geometry.
frequency response of the dataset and output of the
auto-encoder when this image is applied. This pair of
1.2
image and frequency response are not used for
Dataset training and used only for validation. This result
1
shows the auto-encoder can estimate the frequency
response from the input image.
0.8 Figure 13 shows a learning curve of the auto -
S21 Magnitude

decoder. Figure 14 shows an example of filter

0.6 synthesis. Brick-wall filter response is applied to the
auto-decoder as a desired frequency response. The
0.4 auto-decoder synthesis the most possible filter
geometry. After the binarization of the image, the
0.2 auto-encoder estimates the frequency response from
the image. The sigmoidal function of the auto -decoder
Encoded
can be replaced with the binary-sigmoidal function to
0
0 10 20 30 obtain the binary image, but the binary-sigmoidal
function conceals the derivatives which used in the
Frequency [GHz]
training process. Since the planar filter cannot always
(b) Frequency responses. realize the arbitral frequency, some iterative
Fig. 12. Input image and frequency responses stored in a dataset optimizations seem to be required in a practical usage.
and that encoded from the image. This image and this The auto-decoder and encoder take 2.7 and 1.5 msec
frequency response are not used for the auto-encoder
using nVidia RTX 2080, respectively.
training (they are used only for validation).
 2019 Thailand-Japan MicroWave (TJMW2019)

1 200 pixels
Auto-decoder (CNN)
50 pixels

10 mm
0.8

binarization
S21 Magnitude

0.6
40 mm
200 pixels
0.4
50 pixels

10 mm
0.2 Auto-encoder (CNN)

0
0 10 20 30 40 mm
Frequency [GHz]
Fig. 14. An example of synthesized filter image (upper right) decoded from the desired frequency response (dash line in the graph)
and a frequency response (solid line in the graph) encoded from a binarized image (bottom right). Desired frequency
response is fed to the auto-decoder and a quarter-sized filter image is synthesized. Since the auto-decoder outputs gray-scale
image, it is binarized using a certain image library such as OpenCV. The binarized image is fed to the auto-encoder and its
frequency response is estimated. The estimated frequency response tends to be different from the desired frequency if it is
physically difficult to be realized as shown in this example of a brick-wall filter. Auto-decoder and encoder take 2.7 and 1.5
msec, respectively.

4. C ONCLUSION [3] Kosuke Katayama, Kiyoto Takahata, Takashi

In this paper, we propose the auto -encoder/decoder Ohsawa, Takaaki Baba, “Edge Model Embedding
which is used for a planar filter analysis/synthesis. Method for Planar Filter Synthesis,” RISP
The geometry of the planar filter is regarded as an International Workshop on Nonlinear Circuits,
image and the feature vector of the auto - Communications and Signal Processing (NCSP) ,
encoder/decoder is mapped to the frequency response. pp. 502-505, March 5-7, 2019.
We also describe the way to generate a massive and
accurate dataset to train the auto-encoder/decoder. We
demonstrate the auto-encoder can estimate the
frequency response from the input image and the auto -
decoder can synthesize the filter geometry from the
given frequency response. This is a first study which
converts directly and bidirectionally between filter
geometry and frequency response.

Acknowledgements
This work was supported by Japan Society for the
Promotion of Science (JSPS) through a Grant -in-Aid
for Scientific Research (No. 18K04155). This work
was also supported by VLSI Design and Education
Center (VDEC), The University of Tokyo with the
collaboration with Agilent Technologies.

References
[1] J. S. Hong and M. J. Lancaster, Microstrip Filters
for RF/Microwave Applications, John Wiley &
Sons, New York, 2001.
[2] K. Katayama, K. Takahata, T. Ohsawa and T.
Baba, “Training Dataset Synthesis for Planar
Filter Circuits Learning,” The 71th Joint Conf. of
Electrical, Electronics and Information Engineers
in Kyushu, pp. 264, Sept. 9, 2018.