2024 Asia-Pacific Microwave Conference (APMC)

Deep Learning-Enabled Rapid Optimization

for Microwave Filter Design
Sha Luo, Jiteng Ma, Shuping Dang, and Andrew Austin
School of Electrical, Electronic and Mechanical Engineering, University of Bristol, Bristol, United Kingdom
{cv22194, jiteng.ma, shuping.dang, a.austin}@bristol.ac.uk

Abstract—Conducting microwave simulations with traditional training data conflicts since a single design target might have
electronic design automation (EDA) software could be inefficient multiple design solutions. In this way, optimization algorithms
and time-consuming when the electromagnetic (EM) effects are become a feasible scheme for design optimization of mi-
complex. Deep learning (DL) models have been widely used to
improve the efficiency of microwave filter behavior prediction; crowave filters. Another crucial consideration is the significant
however, its simulation results cannot guarantee the same ac- impact of the utility of DL-based simulation model on design
curacy as physical-equation-based EDA software without a large model behavior [7]. Consequently, incorporating a fine-tuning
amount of high-quality training data. Many optimizer algorithms process using established EDA software would further improve
have been utilized to improve the DL model performance. The design optimization outcomes.
paper introduces a DL-enabled particle swarm optimization
Considering that particle swarm optimization (PSO) algo-
2024 Asia-Pacific Microwave Conference (APMC) | 979-8-3503-6354-8/24/$31.00 ©2024 IEEE | DOI: 10.1109/APMC60911.2024.10867510

(PSO) algorithm to solve the trade-off between efficiency and

accuracy of microwave filter behavior prediction. Additionally, rithm has been a popular optimization technique that is widely
the DL-enabled PSO algorithm combines the CST simulation, utilized in different fields [8], this paper introduces a DL
further optimizing the microwave circuit design. An example of model in conjunction with PSO for microwave filter behavior
a hairpin band-pass filter is discussed in the paper to verify the prediction. By integrating the PSO optimization algorithm with
performance of the proposed optimization algorithm. The results
show that the proposed DL-enabled PSO accelerates optimization DL models and CST simulations, the proposed method aims
by over 500 times in an iteration and decreases return loss by to enhance both the efficiency and accuracy of microwave
10 dB across the pass-band of the designed filter. filter design optimization. The DL model helps ensure that
Index Terms—Deep learning, microwave filter design, particle the optimization process is quicker and more efficient. Addi-
swarm optimization (PSO), hairpin filter, CST simulation. tionally, PSO aids in adjusting the convergence direction of
the learning model, thereby enhancing the overall accuracy
I. I NTRODUCTION of the design optimization. Combining the strengths of the
In the field of microwave engineering, simulation is a critical DL model and traditional EDA optimization techniques, this
step in designing and optimizing microwave devices, such as research offers a robust solution to the long-standing trade-off
antennas, filters, and power amplifiers. Traditional electronic between efficiency and accuracy in microwave filter design
design automation (EDA) software, including popular tools, optimization.
e.g., CST Microwave Studio, Ansys HFSS, and Comsol Mul- In the following sections, the paper discusses the proposed
tiphysics, has long been used for achieving accurate simulated DL-enabled rapid optimization approach and demonstrate its
results reference required for factual statements like this one advantages over the EDA-centric benchmark model through a
[1]. However, when electromagnetic (EM) effects become design optimization case study of a hairpin band-pass filter.
complex, these traditional methods could be highly repetitive
II. O PTIMIZATION M ETHODOLOGY
and time-consuming for high-density mesh calculations. This
inefficiency arises from the need for extensive computational A. Optimizer Structure
resources and long processing time to achieve accurate simu- Fig. 1 illustrates the structure of the proposed PSO algo-
lation results [2], [3]. rithm, which integrates three key components: a basic PSO
Recent advancements in deep learning (DL) have introduced algorithm serving as the optimizer, a DL model, specifi-
new possibilities for improving the efficiency of microwave cally a multilayer perceptron (MLP) in this study, and a
filter simulation [2]–[4]. Leveraging the input-output mapping CST simulator. Firstly, the microwave filters’ goals, such as
relations of DL models, microwave engineers can potentially bandwidth and insertion loss, are specified to set the design
reduce the time required for simulations. Despite this promise, targets. An appropriate filter topology is then selected, and its
there remains a significant challenge, viz., to ensure that these geometrical parameters are calculated. The PSO algorithm is
DL-based simulations can achieve the same level of accuracy used to optimize and predict the behavior of the target filter.
as EDA software based on the computation-intensive EM To accelerate the PSO process, a well-trained MLP model is
simulation. utilized to find appropriate initial points of particle position for
The DL-enabled microwave filter design presents additional PSO. Afterward, the PSO process updates particles through
challenges. The primary obstacle to DL design model con- the real-time simulation of CST, which slightly adjusts the
vergence is the non-uniqueness problem [5], [6], where the accuracy of predicted results. This iterative approach ensures

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARA. Downloaded on April 25,2025 at 12:05:49 UTC from IEEE Xplore. Restrictions apply.
979-8-3503-6354-8/24/$31.00 ©2024 IEEE 1287
 
 


ࡸ૝ ࡿ

  

!
 ࡸ૜


  


ࢃ૚ ࡸ૞ 濷濷
ࡸ૚ ࡸ૛
 
 
 



      ࢃ૛ ࢃ૜
(a)




 



 

(b)
   
Fig. 3. Models of seventh-order hairpin band-pass filter: (a) 2D model; (b)



3D model.

 
 

Fig. 1. Flow chart of the proposed DL-enabled rapid optimization for

represents the best position of the ith particle and pgbest
microwave filter design. represents the global best position of the swarm. Then the
particle position is updated by
z Xi (t + 1) = Xi (t) + Vi (t + 1) . (4)
Whether to update the global best position depends on
ܺ௜ାଵ
ܺ௜ the fitness function f (x), which defines the best solution of
optimization and is usually a convex function. To achieve
the minimum of the fitness function, the global best particle
‫݌‬௚௕௘௦௧ ܸ௜ାଵ ሺ‫ݐ‬ሻ position is updated as per
ܸ௜ ሺ‫ݐ‬ሻ
pi,lbest = Xi when f (Xi ) < f (pi,lbest ), (5)
x
y ‡ƒ” Š‹‰ „‘—†ƒ”› and pgbest = pi,lbest when f (pgbest ) < f (pi,lbest ). (6)

Fig. 2. Illustration of the movements of particle swarm by the PSO algorithm. C. DL-Enabled Simulator
In this study, an MLP, a feedforward neural network, is
utilized to generate initial particle positions in the PSO. The
an efficient and adaptive optimization process, leveraging both input data of the MLP model consists of the geometrical
DL and real-time simulation feedback to converge on an parameters of the topology of hairpin band-pass filter, and
optimal microwave filter design. the output data are the predicted S-parameters. Mean absolute
B. PSO Algorithm error (MAE) is used for calculating the fitness function.
PSO minimizes the loss or maximizes the benefits by III. R ESULT C OMPARISON
generating a group of particles that are able to update their
individual states and make decisions on movement according The simulations are conducted on AMD Ryzen 5900HS
to a set of predetermined rules and criteria. In Fig. 2, each with Radeon Graphics and NVIDIA GeForce RTX 3050 Ti
random particle has its position (X) and velocity (V ) and Laptop GPU @ 4GB (128 bits). The MLP model consists
will keep searching over the parameter space within a preset of an input layer, two hidden layers, each containing 128
boundary until it reaches the minimal loss, as known as the
global best position. In each round of training, a particle has neurons, and an output layer. The training employs the Adam
a position as follows: optimizer with a learning rate of 0.001, batch size of 32, and
200 epochs [9]. The optimization algorithm employs the PSO
Xi = (xi1 , xi2 , ...xiD ) , i = 1, 2, ...N, (1)
with acceleration constants c1 of 1.5 and c2 of 1.5.
where dimension D and the number of particles N are pre-
defined as hyperparameters. The ith particle has a velocity: A. Geometry of Hairpin Filter
Vi = (vi1 , vi2 , ...viD ) , i = 1, 2, ...N, (2)
Taking a seventh-order hairpin band-pass filter as an exam-
ple to simulate, Fig. 3(a) shows the geometrical parameters of
which is a velocity towards the best particle expanded as its topology, which are the width W , length L, and spacing S
follows: of the microstrip lines on the patch. These nine parameters of
Vi = ωVi (t) + c1 r1 (t) [pi,lbest (t) − Xi (t)] design components determine the dimension of each particle
(3)
+c2 r2 (t) [pgbest (t) − Xi (t)] , in the optimizer and the number of input layers in the MLP
where c1 and c2 are acceleration constants; ω is inertia weight; model. Fig. 3(b) illustrates the 3D model of the designed
r1 and r2 are random numbers between 0 and 1; pi,lbest hairpin filter in CST.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARA. Downloaded on April 25,2025 at 12:05:49 UTC from IEEE Xplore. Restrictions apply.
1288
 0.25 TABLE I
EDA-centric C OMPARISON OF MINIMUM LOSS AND RUNNING TIME OF DL- ENABLED
0.28
0.2 DL-enabled
AND EDA- CENTRIC APPROACHES .
Loss(MAE)

0.26

Loss(MAE)
0.24 0.15 DL-enabled EDA-centric
0.22 MLP:iter=100 MLP:iter=0
0.1 Particles Loss Runtime(s) Loss Runtime(s)
0.2
5 0.19 63.53
0.18 0.05
0 50 100 0 5 10
CST:iter=0 10 0.18 113.65 N/A
Iteration Iteration 20 0.04 323.70
D E 5 0.07 3402.52 0.12 3187.16
0 0 CST:iter=10 10 0.07 8487.30 0.04 14874.12
20 0.02 28472.07 0.03 15038.31
-10 -10 5 0.05 6211.09 0.06 7619.44
CST:iter=20 10 0.03 16654.89 0.03 24411.54
Magnitude (dB)

-20
Magnitude (dB)

-20 20 0.01 50449.77 0.03 27557.47

-30 -30
optimization. The return loss S(1,1) has an average value of
-40 -40
-23.55 dB over the pass-band.
DL-enabled S(2,1)
-50 -50 DL-enabled S(1,1)
MLP S(2,1) EDA-centric S(2,1) IV. C ONCLUSION
MLP S(1,1) EDA-centric S(1,1)
-60 -60 This paper introduces a DL-enabled approach to enhance the
2 4 6 8 2 4 6 8
Frequency (GHz) Frequency (GHz) efficiency of microwave filter design optimization compared
F G to the traditional EDA-centric EM process. The methods
incorporate CST real-time simulation to further refine the
Fig. 4. Loss of optimization and performance of designed hairpin band-pass microwave filter. A hairpin band-pass filter serves as a case
filter: (a) Loss of MLP-based PSO; (b) Loss of CST-based PSO; (c) S(2,1)
and S(1,1) of designed hairpin band-pass filter from MLP-based PSO; (d) study to validate the DL-based PSO algorithm with CST real-
S(2,1) and S(1,1) of designed hairpin band-pass filter from CST-based PSO. time optimization. The experimental results have verified the
effectiveness and efficiency of DL-based PSO in microwave
filter simulation and design. These results also demonstrate
B. Performance of the Proposed PSO Algorithm that the DL-enabled approach is able to considerably accel-
With five particles set by PSO, Fig. 4(a) and (b) illustrate erate the PSO process for filter design, while the real-time
that the MLP simulator rapidly finds the best global solution CST simulation enhances the filter performance in terms of
after 100 iterations, achieving a loss of 0.19. Using the initial bandwidth and insertion loss.
points returned by the MLP model, the DL-enabled PSO
with the CST simulator achieves a loss of 0.07 after an R EFERENCES
additional 10 iterations. By generating random initial points, [1] A. Srivastava, R. K. Chaudhary, A. Biswas, and M. Jaleel Akhtar, “Dual-
the EDA-centric PSO with the CST simulator reaches a loss band l-shaped siw slot antenna,” in Proc. IEEE International Conference
on Microwave and Photonics (ICMAP), Dhanbad, India, Dec 2013, pp.
of 0.12 after 10 iterations. Table I shows that the MLP model 1–3.
significantly surpasses the CST simulator in computational [2] J. Ma et al., “Transfer learning for the behavior prediction of microwave
efficiency, achieving a runtime performance improvement of structures,” IEEE Microw. Wireless Tech. Lett., vol. 33, no. 2, pp. 126–
129, Feb 2023.
over than 500 times per iteration. This dramatic speed increase [3] S. Luo, J. Ma, C. Zhang, S. Dang, and R. Shubair, “Federated learning
enables the optimizer to converge much faster compared to for microwave filter behavior prediction,” IEEE Microw. Wireless Tech.
solely using the traditional CST simulator. As the number Lett., vol. 34, no. 3, pp. 255–258, Mar. 2024.
[4] G. Pan, Y. Wu, M. Yu, L. Fu, and H. Li, “Inverse modeling for filters using
of particles increases, optimization improves at the cost of a regularized deep neural network approach,” IEEE Microw. Wireless
computation resource and sometimes converges to a similar Tech. Lett., vol. 30, no. 5, pp. 457–460, May 2020.
solution. [5] J. Ma et al., “Revolutionizing microwave circuit design with machine
learning: Challenges and opportunities,” IEEE Commun., vol. 62, no. 8,
pp. 100–106, Aug. 2024.
C. Design Optimization of Hairpin Filter [6] F. Feng, W. Na, J. Jin, W. Zhang, and Q.-J. Zhang, “ANNs for fast
DL-enabled optimization approach further improves the parameterized EM modeling: The state of the art in machine learning for
design automation of passive microwave structures,” IEEE Microw. Mag.,
performance of the hairpin filter through the EM simulation vol. 22, no. 10, pp. 37–50, Oct. 2021.
results from the EDA software. Fig. 4(c) shows the predicted [7] J. Ma et al., “A learning-based methodology for microwave passive
behavior of the filter designed by MLP-based PSO with five component design,” IEEE Trans. Microw. Theory Tech., vol. 71, no. 7,
pp. 3037–3050, July 2023.
particles, where the ripple on the pass-band is over -5 dB at [8] X. Chen and N. Hu, “Effectiveness evaluation for UAV air-to-ground
the frequency of 3.85 GHz and the average value of return loss attack based on PSO-BP neural network,” in Proc. IEEE Chinese Control
S(1,1) is -13.25 dB in the range of 3 dB bandwidth. In Fig. Conference (CCC), Dalian, China, July 2017, pp. 3864–3869.
[9] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
4(d), MLP-based PSO uses the CST simulator for fine-tuning arXiv preprint arXiv:1412.6980, Dec. 2014.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARA. Downloaded on April 25,2025 at 12:05:49 UTC from IEEE Xplore. Restrictions apply.
1289