This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems.

This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 1

CircuitGTL: An Intelligent Circuit Design

Methodology across Electromagnetic Topologies
with Graph Transfer Learning
Xin Jian, Junyi Zhang, Yaoyao Li∗ , Tianyu Yang, Amr Tolba, Senior Member, IEEE, Osama Alfarraj,
Keping Yu, Senior Member, IEEE, and Mohsen Guizani, Fellow, IEEE

Abstract—Existing deep learning-based circuit design methods These results provide valuable insights for lightweight, and high-
mostly focused on the primary matching of the model itself or precision design of coupled electromagnetic structures.
circuit data, lacking generalizability and ignoring deep repre-
sentation of electromagnetic coupling effects in coupled circuits. Index Terms—Electromagnetic inverse problem, circuit design,
Therefore, it exhibits difficulties to further improve the accuracy graph representation learning, graph transfer learning across
in circuit performance prediction and requires large training electromagnetic topologies.
datasets. To address these challenges, this paper proposes an
intelligent circuit design methodology with graph transfer learn-
ing (CircuitGTL). Specifically, it achieves the weighted graph I. I NTRODUCTION
modeling of complex electromagnetic environment circuits, where
nodes represent components, edges represent the electromagnetic
coupling effect between components, and edge weights signify the
differential strength of electromagnetic coupling. Hereby, a fused
T HE integrated circuit performance highly depends on
the electromagnetic coupling effects between components
which are decided by electromagnetic parameters and layout.
graph representation model, integrating graph isomorphic net- The electromagnetic environment within circuit, especially
work and electromagnetic coupling effect based graph attention coupled circuit, can be very complex and hard-to-model as
network, is proposed to achieve deep representation learning of
graphic circuit data. Then, a model- and data-driven graph trans- the complex coupling relationships between numerous compo-
fer learning mechanism considering joint optimizing of circuit’s nents. Therefore, the design of coupled circuit essentially re-
performance matrix and non-performance indicators is proposed. volves around the optimization of electromagnetic parameters
This is to achieve lightweight cross-electromagnetic topologies and layout which is driven by performance metric, often con-
parameters optimization. Taking Terahertz (THz) resonant fil- stituting NP-complete electromagnetic inverse problems [1].
ters as an example to verify the effectiveness of CircuitGTL,
numerical results on the MITCircuitGNN experimental dataset It usually consists of two processes: a) forward performance
show that: compared with state-of-the-art algorithm CircuitGNN, analyze (FPA) predicts the electromagnetic properties when
CircuitGTL achieves 9.2% improvement in cross-electromagnetic the circuit parameters are given, and b) backward parame-
topologies performance prediction accuracy, 90.9% reduction ters optimization (BPO) tunes the electromagnetic parameters
in model convergence time and 33.6% reduction in the total when specific electromagnetic objectives are known. Tradi-
coverage area of components with only 20% of data require-
ment; additionally, the design of CircuitGTL has lower insertion tional methods rely on electromagnetic simulation software to
loss, steeper skirts, and higher passband intersection-over-union. complete the FPA, and then manually perform extensive BPOs
based on experience or parameter scanning, often leading to
This work was supported in part by National Key Research and Develop- prolonged design cycles and high computational complexity.
ment Program of China (2023YFB3306900), in part by the National Natural With the development of deep learning, it has been widely used
Science Foundation of China under grant U20A20157 and U21A20448, in
part by the Fundamental Research Funds for the Central Universities under in circuit field such as hardware design [2; 3], analog circuit
grant 2024CDJGF-012, in part by the Researchers Supporting Project Number sizing [4; 5], and performance prediction [6–8]. Furthermore,
(RSP2024R102), King Saud University, Riyadh, Saudi Arabia. (Correspond- given that circuit, netlist, and layout are inherently graph
ing author: Yaoyao Li.)
Xin Jian and Junyi Zhang are with the School of Microelectronics and Com- structures, deep learning methods such as Graph Represen-
munication Engineering, Chongqing university, Chongqing 401331, China tation Learning (GRL) [9] can effectively extract complex
(email: [email protected]; [email protected]). graph structural information, offering novel perspectives for
Yaoyao Li is with the Department of Electronic Information Engineering,
Beihang University, Beijing 100083, China (email: [email protected]). the intelligent design of circuits which are exemplified by
Tianyu Yang is with the School of Computer Science and Engineering, Terahertz (THz) resonant filters [10]. Currently, researches
Chongqing University of Science and Technology, Chongqing 401331, China related to GRL on intelligent circuit design primarily focu on
(email: [email protected]).
Amr Tolba is with Computer Science Department, Community Col- FPA. Reference [11] utilizes Graph Neural Network (GNN)
lege, King Saud University, Riyadh 11437, Saudi Arabia (email: with edge attention mechanisms to predict the performance of
[email protected]). operational transconductance amplifier (OTA) under various
Osama Alfarraj is with Computer Science Department, Community Col-
lege, King Saud University, Riyadh 11437, Saudi Arabia (email: oalfar- layouts. Reference [12] employs GNN to predict the node volt-
[email protected]). ages of operational amplifier circuits, integrating them into the
Keping Yu is with the Graduate School of Science and Engineering, Hosei optimization loop algorithm BabNet [13], thereby enhancing
University, Tokyo 184-8584, Japan. (email: [email protected]).
Mohsen Guizani is with the Machine Learning Department, Mohamed Bin computational efficiency in circuit design optimization loops.
Zayed University of Artificial Intelligence, UAE (email: [email protected]). Reference [14] utilizes GNN to predict layout parasitics and

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 2

circuit parameters, achieving over 90% reduction in simulation To summarize, this paper proposes a novel methodology
errors compared to manual design. Reference [15] integrates named CircuitGTL for circuit design through graph transfer
graph message-passing mechanism and graph descriptor to learning, realizing iterative optimization of electromagnetic
identify missing components in circuits and predict connec- parameters across electromagnetic topologies. To be more
tions for new components. Reference [16–19] utilize GNN to specific, the contributions of this paper are:
extract netlist parameter features, using them as input states for 1) For FPA, this paper proposes an electromagnetic cou-
reinforcement learning, thus achieving automated optimization pling effect weighted graph representation method
of chip layouts and parameters. Addressing the researches (EMCoE-AT), realizing a deep graph representation of
related to GRL for BPO on intelligent circuit design. Refer- electromagnetic structures and further improving the for-
ence [20] utilizes Graph Convolutional Network (GCN) [21] ward prediction accuracy, where nodes represent com-
to identify and label different circuit block topologies, facili- ponents, edges represent the electromagnetic coupling
tating layout optimization of OTA and RF circuits. Reference effect between components, and edge weights signify
[22] introduces CircuitGNN, based on graph message-passing the differential strength of electromagnetic coupling.
mechanism, taking the size, orientation, and relative position 2) For BPO, to achieve high-performance filtering and meet
of components as feature inputs, achieving a four-order-of- the compact design and manufacturing requirements
magnitude acceleration in distributed circuit design compared of modern multi-channel THz filter banks, this paper
to manual design. Reference [23] uses GCN as the encoder proposes a joint optimization method for the out-of-
and for the first time, leverages reinforcement learning (RL) band signal rejection capability and total coverage area
to transfer knowledge between different technology nodes and of components based on meta-learning, achieving area
topologies, achieving optimal Figures of Merit in amplifiers minimization of the designed filter while improving
and regulators. The aforementioned researches indicate that performance.
GRL is an effective method for circuit design. 3) Furthermore, this paper proposes a model- and small-
However, for researches on FPA, there is currently a gap data-driven graph transfer learning mechanism to realize
on effective electromagnetic coupling effects modeling for lightweight cross-electromagnetic topology parameters
electromagnetic coupled structure circuits, which resulting in optimization. To the best of our knowledge, we are
low prediction accuracy of FPA. For the limited researches on the first to apply graph transfer learning to the field of
BPO of filter, they mainly consider the passband performance, electromagnetic inverse problems.
ignoring the out-of-band signal rejection capability and non- 4) We validate the effectiveness of CircuitGTL on the
performance indicators (e.g., coverage area of components), in MIT-CircuitGNN experimental dataset [24]: compared
this case, the designed filter has a problem of transition band to CircuitGNN proposed in [22] which is state-of-
broadening, and the coverage area of components does not the-art algorithm in the field of intelligent circuit de-
meet the compact design and manufacturing requirements of sign, CircuitGTL achieves 9.2% improvement in cross-
modern filter banks. Meanwhile, existing researches on BPO electromagnetic topology performance prediction accu-
are mostly limited to the internal representation and learning racy, 90.9% reduction in model convergence time and
within large databases composed of similar types of circuit 33.6% reduction in the total coverage area of compo-
topologies, failing in deep representation and learning across nents with only 20% of data requirement. Compared
electromagnetic topologies and requiring a huge amount of with manual design, CircuitGTL achieves lower in-
training data. The research trending exhibits to further achieve sertion loss, steeper skirts, higher passband IOU and
cross-electromagnetic topologies transfer, rapid automation, stopband rejection ratio, and a four-order-of-magnitude
and lightweight intelligent circuit design. reduction of design cycle.
Addressing the aforementioned challenges, this paper in- The rest of this paper is organized as follows. Section II
novatively proposes a methodology for cross-electromagnetic describes our proposed CircuitGTL. Section III presents the
circuit design with graph transfer learning (CircuitGTL). This numerical results of experiments, while Section IV describes
proposed methodology introduces a graph representation net- the conclusions and future p demonstrates how to generalize
work weighted by electromagnetic coupling effects. It also uti- our methodology to other coupled circuits with different
lizes a joint optimization method for the out-of-band signal re- resonant unit types. Table 1 gives a summary of the symbols
jection capability of THz resonant filter and total coverage area in the formula.
of components based on meta-learning. This achieves iterative
optimization of electromagnetic parameters and coverage area
minimization of the designed filter, improving performance II. P ROPOSED C IRCUIT GTL
to accommodate the compact design requirements of modern Figure 1 presents the framework of CircuitGTL, with
filter banks, thereby making it more suitable for practical man- its algorithmic pseudocode detailed in Section II.D. Firstly,
ufacturing processes of multi-channel THz filter banks. Fur- CircuitGTL introduces the electromagnetic coupling effect
thermore, we propose a model- and small-data-driven graph weighted graph representation method (EMCoE-AT), where
transfer learning mechanism, achieving lightweight design nodes represent resonators, edges represent the electromag-
across electromagnetic topologies. Using THz resonant filters netic coupling effect between resonators, and edge weights
as a case study without loss of generality, the effectiveness of signify the differential strength of electromagnetic coupling.
CircuitGTL is demonstrated. Subsequently, a fused graph representation model (EMGNN)

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 3

BPO FPA Deep Graph

Joint Optimization Representation of Graph transfer learning Few-shot
Electromagnetic Target Domain
Equation (18) Topology
Amplitude (dB)

(Data-Driven)
1 2 2 1 （Model-Driven）
EMGNN EMCoE-AT
GIN  31
h1 h3
13
Section III.C
0 Frequency (GHz)  21 12  23
Performance Area
Section III.D
. .
Transfer Function of
. h2
 32
h4
Filter to be Optimized
NO
Optimized YES Iteration NO
Parameters over? Optimize the
Less Source Domain
Electromagnetic error?
Parameters Different Topologies
Target Passband YES
Different Numbers
Amplitude

EM
GAT Solver

0 Frequency Predicted Value Ground Truth

Design Objective Back propagation

Fig. 1: The framework of CircuitGTL

is introduced, combining Graph Isomorphic Network (GIN) TABLE I: Notation Table For the Formulas
[25] and Graph Attention Network (GAT) [26], to achieve Variables Description
cross-domain deep representation of electromagnetic graph
k the index of the layers
data between the source and target domains. Finally, we add Wn the learnable parameter of MLP
area constraints on the basis of optimizing objectives for We the learnable parameter of MLP
jointly regularized transfer functions of stopbands and tran- A the normalized adjacency
∆l the distance between nodes
sition bands. Additionally, we propose a meta-learning-based C the electromagnetic topology matrix
joint optimization method for enhancing the out-of-band signal hi node features
rejection capability of filters and minimizing the total cov- εij edge features
Hkij edge encoding
erage area of components. This method leverages the shared
zki the fused features of nodes and edges
parameters of EMGNN to iteratively optimize electromagnetic W the learnable weight matrix
parameters and compact the coverage area of the target domain ckij the electromagnetic coupling effect between
filters. CircuitGTL deeply integrates the graph representation nodes
capability of GNN [27] with the lightweight capability of αkij the normalized attention coefficients
Lp the loss for the main task
Graph Transfer Learning [28], realizing cross-electromagnetic wBP the learnable weight vector
topology intelligent circuit design, as detailed in Sections λ1 ,λ2 the learnable weight vector
III.C-III.D. Next, this paper employs the open-source THz Q0 the minimum area typically provided by the
manufacturing vendor
resonant filter circuit data, MIT-CircuitGNN [24], as the sub- m the move rang
ject of study. Each circuit consists of several microstrip open- ∇ψp log Lp the gradients of the loss values
loop square resonators [29], with their parameters depicted in cosim the cosine similarity
La the parameters of the weighted model
Figure 2(a). f (cosimi ; Λ) the weight generated by the weighted model
t
LDf ine the training loss function on the target domain
τ, ι the number of EMGNN layers for pre-training
A. Graph Representation Learning of Electromagnetic Topol- and fine-tuning
ogy lr the learning rate
ϑιpre the shared model parameters after pre-training
In order to model the THz resonant filter as graph, where on the source domain
ϑιf ine the parameters of fine-tuned model
nodes represent resonators, edges represent the electromag-
netic coupling effect between resonators, and edge weights sig-
nify the strength of electromagnetic coupling, the electromag-
netic coupling effect weighted graph representation method of the graph data at the current state to explain the learning
EMCoE-AT will model the nodes and edges separately. The process of each part of EMCoE-AT.
node and edge features that we list in Table 2 fully model the 1) Node-level Modeling: Feature Mapping of Resonators
dimensional information and structural characteristics of the As shown in Table 2, node features model the physical infor-
filter. EMCoE-AT completes step ⃝ 1 in Figure 1. Here, we mation of individual resonator, which include the input/output
will detail the principles of EMCoE-AT for deep modeling of ports: [input, output] = [1, 0] or [0, 1] or [0, 0], side lengths
node-level and edge-level electromagnetic graph data. At the of resonators: l, angular position of the slit: θ ∈ [− π4 , π4 ],
end of each section, we will provide the feature dimensions orientation: dir = Binary[a, b, c, d] where a,b,c, and d are

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 4

of dimension dim = 8 ([[input, output], l, θ, dir]), have been

Resonator1 Mixed ( x3 , y3 ) mapped to dimension dim = 800.

.
Resonator3 coupling
l
Input Output MLP LeakyReLU

. 
r .
(x4 , y4 )
(x5 , y5 )
.
( x1 , y1 )
S I /O
.. .

l hi

...
...
w ( x2., y2 )
Resonator4 Magnetic
D Resonator5 coupling dir  [ a, b, c, d ]

Resonator1 Electric resonator 

Resonator2 coupling
Node Features
(a) Physical Parameters of THz Resonant Filter Wn
Node Embeddings hi
Node-level （dim=400）
modeling network

h3 53 Fig. 3: Schematic Diagram of Node-level Modeling Network

 31
35 2) Edge-level Modeling: Graph Representation of Electro-
13 45 magnetic Coupling Effects
h4 54 h5 Electromagnetic coupling effects comprehensively reflect
h1  32  23
12 25 factors influencing the strength of electromagnetic coupling,
including distance, coupling modes (electric coupling, mag-
 21 52 netic coupling and mixed coupling), and the orientation(dir)
and angle of the slit(θ) [29]. The edge-level network con-
h2 siders the impact of each factor separately to characterize
differentiated electromagnetic coupling strengths based on
attention mechanism. These are used as initial relative edge
(b) Graph Representation of THz Resonant Filter weights for aggregating electromagnetic signals and will be
iteratively updated in the GAT layer of EMGNN model. This
Fig. 2: An Example of EMCoE-AT enables graph representation of the complex electromagnetic
environment within THz resonant filter banks, thus enhancing
TABLE II: Node/Edge Features in the THz Resonant Filter the graph representation capability of traditional sum neigh-
Graph Modeling borhood aggregation [22], as shown in Figure 4.
Parameters Feature Name Description  a11 a12 a13 a15  c11k c12k c13k c15k 
a a22 a23 a25  Eq.(5) Eq.(2)  k 
I/O Input/output port Eq.(3) A =  21 c ck ck
C   21 22 23
c25k 
Eq.(4)
 
k
H   (h , h , ε ) × a21
k
21
k k
2
k
1
k
21
c21  
   k 
l side length of the square  a51 a52 a53 a55  k c51 c52 c53
k k
c55 
k

Node Features k
H 23   k (h2k , h3k , ε23
k
) × a23 c23
dir slit orientation of the resonator ×
k
θ angular position of the slit z 1
Wk  21k Σ
h4 z2k h4
k
z z3k Self-
W Eq.(7) Attention
k Eq.(8) ×
∆x = |xi − xj |, ∆y = difference distance 1
h3 h5 z3k h3 h5
h1  23k
|yi − yj |  21k
 23k Wk Eq.(9) h1
 k
 23k
Edge Features Edge-level
21

S gap of resonators h2
modeling networks C 2k
h2
z2k Eq.(10)
D shift of resonators Layer=k

(xi , yi ), (xj , yj ) center position Fig. 4: Schematic Diagram of Edge-level Modeling Networks
Others w width of the Microstrip line
r length of the slit
The impact of distance is modeled by a normalized adja-
cency matrix. Let aij be the element corresponding to node
i and node j in the normalized adjacency matrix A =
binary elements to represent the four slit directions of left, (aij )N ×N , then aij can be expressed as
up, right, and down respectively,(e.g. dir = [1, 0, 0, 0] means (
the slit direction is left). Subsequently, the node features are 1 ∆l < L
mapped to a high-dimensional space for encoding through the aij = L ∆l > L (2)
2
node-level network, which consists of a Multi-layer Perceptron (relu(∆l − L)) + L
(MLP) and the LeakyReLU activation function, as depicted in  
a11 a12 · · · a1N
Figure 3. The encoded node features hi can be expressed as  a21 a22 · · · a2N 
 aij ∈ (0, 1]
A= . (3)

.. .. ..  a = 1
hki = MLPkWn (LeakyReLUk ([input, output], l, dir, θ) (1)  .. . . .  ii
aN 1 aN 2 · · · aN N
where k represents the index of the layers and the learnable q
2 2
parameter Wn of MLP will be learned and updated in where ∆l = (∆x) + (∆y) denotes the distance between
EMGNN model. At this point, the node features hi , initially nodes, L represents the maximum distance at which sufficient

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 5

electromagnetic coupling effects occur between nodes, as node features represented by hk , the edge features represented
shown in [30], and relu(·) signifies the activation function. by εk , and the edge weights represented by α. At this point,
The expression aij in Equation (1), substituting the binary edge features εij , initially of dimension dim = 4 ([[∆x, ∆y],
adjacency matrix elements from [22], adequately characterizes S, D]), have been mapped to dimension dim = 800.
the property of electromagnetic coupling strength rejection 3) Node and Edge Parameters Learning: Deep Graph Rep-
with distance. resentation Learning
The impact of coupling modes, the orientation and angle Figure 5 presents the schematic diagram of the integrated
of the slits are modeled by an attention-based mechanism to graph representation model EMGNN. The circuit diagram
represent the degree of influence. Let C = (cij )N ×N denotes weighted by the electromagnetic coupling effects provided
the electromagnetic topology matrix, where ckij represents the by EMCoE-AT is input into EMGNN, based on the graph
electromagnetic coupling effect between nodes, C can be structural characteristics generated by EMCoE-AT, we provide
expressed as the rationale for the model architecture design: we consider
that the ability of GNN to capture graph structures is key to
their transferability [32], given that the graph G is isomorphic,
 
C = [ckij ]N ×N = aij Hkij N ×N

a11 Hk11 a12 Hk12 · · · a1N Hk1N
 i.e. G = ((v1 , v2 , v3 , ..., vN ), E), (v1 , v2 , v3 , ..., vN ) ∈ V ,
 a21 Hk21 a22 Hk22 · · · a2N Hk2N and it’s known that GIN has been proven to enhance the
 (4) model’s ability in distinguishing different graph structures and
=
 
.. .. .. .. 
 . . . .  achieve the best expressive powerP for isomorphic graphs [25],
aN 1 HkN 1 aN 2 HkN 2 · · · aN N HkN N i.e. h1i = Υ1 ((1 + ς 1 ) · h0i + j∈N (i) h0j ), where ς is a
learnable parameter and Υk (·) represents the update function.
Hkij = η k (hki , hkj , εkij )
(5) The node feature h0i undergoes one propagation and update
= MLPWe (LeakyReLUk (hki , hkj , εkij ))
k
in the GIN layer to obtain h1i , which is then extracted as
where η k (·) represents an MLP with learnable weight parame- the generalized feature representations. Given that the graph
ter We and LeakyReLU activation. Equation (5) concatenates G has edge weights αij and attention-based message passing
edge features εij with node features of i and j and uses η k (·) cannot be achieved by the GIN network alone. Inspired by
to map these features into a high-dimensional space to obtain the fusion network [33], we consider that the initial layers of
edge encoding Hkij . εij denotes the edge features containing GNN learn generalized feature representations, while deeper
the gap S and shift D between two resonators i and j, while layers learn complex features specific to the target task [34].
k represents the index of the layers. Equation (4) multiplies Therefore, we cascade GAT layers after the GIN layer to
Hkij by the corresponding adjacency matrix elements aij to form the fused graph representation model EMGNN. The GAT
construct the electromagnetic topology matrix. layers process the generalized features from the GIN, serving
Subsequently, the self-attention mechanism [31] is executed as deeper layers specific to downstream task. They describe
on the node i to compute the normalized attention coefficients the electromagnetic coupling strength arising from different
αij , which αij is then used as the edge weights to represent distances, coupling modes,the orientation and angle of the
the degree of electromagnetic coupling differentiation of each slit, thereby enhancing the accuracy of differential passing
neighboring node to the node i. of electromagnetic coupling effects between nodes. In GAT
layers, message passing and node updates are performed based
zki = hki ||εkij (6) on the node’s own features and the weighted electromagnetic
coupling effects of those from the neighborhood, and can be
e(zki , zkj ) = LeakyReLU(δ T · [Wzki ||Wzkj ]) (7) expressed as
k
αij = softmax((e(zki , zkj ))) X
Cki = (ckij αij
k
) (k > 1) (9)
exp(e(zki , zkj ))
=X j∈N (i)
exp(e(zki , zkj )) (8)
k∈N (i) X
exp(LeakyReLU(δ T · [Wzki ||Wzkj ])) hki = Υk (hk−1
i , Cki ) = Υk (hk−1
i , (ckij αij
k
)) (10)
=X j∈N (i)
exp(LeakyReLU(δ T · [Wzki ||Wzkj ]))
k∈N (i)
where Cki
aggregates the total weighted electromagnetic cou-
where zki and zkj denote the fused features of nodes and edges, pling effects received by node i using Equation (4) and
respectively, || is the vector concatenation operation, e(zki , zkj ) Equation (8). Taking Figure 4 as an example, in layer k, the
′
indicates the importance of node j to node i. W ∈ RF ×F total electromagnetic coupling effect received by node 2 using
denotes the learnable weight matrix (F indicating the number Equation (9) can be expressed as
of node features), Ni represents the neighborhood of the node
′
i, δ ∈ R2F denotes the learnable weight vector, and (·)T Ck2 = ck21 α21
k
+ ck23 α23
k
k
indicates the transpose operation. To this end, the combination =η ((h2 , h1 , ε21 )a21 )α21 + η ((h2 , hk3 , εk23 )a23 )α23
k k k k k k k

of equations (3) and (7) yields a graph representation method (11)

weighted by the electromagnetic coupling effect, composed of Through the cascaded network layers of GIN, GAT and
the electromagnetic topology represented by the matrix C, the fully connected network, iterative training of the learnable

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 6

parameters Wn , We , W and δ in Equation (1),(5),(7) is electromagnetic parameters, achieving high-performance filter-

performed to achieve the propagation and learning of elec- ing for the target filters. Addressing the issue of transition band
tromagnetic signals, ultimately resulting in the filter’s transfer widening due to the sole consideration of the passband without
function. The transfer function output by EMGNN reflects fitting the transition band [22], this paper proposes a joint
the electromagnetic characteristics of the filter thus achieves a regularization method for both the stopband and transition
deep representation of electromagnetic graph data. The input band. The improved objective function Lp can be represented
of EMGNN is the features represented in a high-dimensional as
space of the graph (both node and edge encoding dimensions
are 800), and the output dimension through a fully connected X X 2 2
network is the number of discretization points of the transfer Lp = (|yi | − 1)2 + λ1 · |yi | + λ2 · |w∗ |
function, which we set to 6000. i:wi ∈wBP i:wi ∈w
/ BP
(12)
∗
Node and edge parameters w = |w1 − w2 | (13)
Architecture of EMGNN
graph representation learning
EMCoE-AT
 52
where yi represents the predicted transfer function, wBP
h5  25 h2 represents the set of passband frequencies, λ1 and λ2 are
 53
 35  21 learnable weight parameters that balance the transition band
12 2 2 2 2
 45  54 h3
13
 31 Wn width and out-of-band rejection penalty term. w∗ represents the
 41 h1 We
h4 14 3 3 3 3 + transition band width, which is calculated as the difference
aggregation
+ 1 1 1 1 between the passband and stopband widths. The passband w1
1  is the frequency region where the amplitude of the transfer
node: W n Eq.(1)
edge: W Eq.(5) function is less than 3 dB below its maximum amplitude,
e
GIN Layer and the stopband w2 is defined as the frequency region
where the amplitude of the transfer function is less than 10
h2
c21 2 2 2 2 dB below its maximum amplitude. Using Lp as the target
h3
3 3 3 3
 21 attention
for iterative optimization of electromagnetic parameters, it
h1 +
c31  31 aggregation
enhances the rejection in the transition band, leading to the
+ 1 1 1 1

h2
C1   31c31   21c21 steep skirts and high out-of-band rejection, thus improving
h3
Attention   ( W, δ)
Eq.(7)
filtering performance, as detailed in Figure 8 and Figure 10.
coefficient:
h1 GAT layer
2) Coverage Area Minimization
. Furthermore, with the modern electronic technology’s de-
. mand for miniaturized device design, constraining the area of
.
.
3 layer GAT individual filter from the perspective of multi-channel THz
Shared weight filter banks’ manufacturing processes enables the integration
of more electronic devices or filter banks with additional
channels on the same chip, such as THz channelizers [35]. This
Input graph
embedding also reduces device power consumption and material costs,
concat
Output  enhancing design reliability and consumer appeal. appeal.
This holds significant engineering applications. Therefore,
Pooling this paper introduces coverage area minimization into the
Fully Connected Network
optimization objectives and further proposes a meta-learning-
Label
Training based joint optimization method for the out-of-band signal
...
rejection capability of filter and the total coverage area of com-
ponents. Specifically, to prevent collisions among filter banks
Yes leading to ineffective designs, we model the total coverage
Less error?
Update parameters
Transfer function
area geometrically based on the boundary area, as shown in
No
predicted value Figure 6. We assume that the minimum area limit imposed
by the manufacturing process is Q0 (typically provided by the
Fig. 5: Schematic Diagram of EMGNN model
manufacturing vendor). Thus, the loss for the auxiliary task
can be represented as the l1 loss between the current area Q
and the minimum area Q0
B. Joint Optimization for Filtering Performance and Coverage
Area
Q0 = max|xi + mi dx| ∗ max|yj + mj dy| i, j ∈ N (14)
1) Filtering Performance Error Minimization
Building upon the deep representation of electromagnetic La = ||Q − Q0 ||1 (15)
graph data, this section introduces the optimization objective
of jointly regularizing the stopband and transition band of the where mi , mj represents the move range, in which node
transfer function. This facilitates the iterative optimization of i and node j move in their respective directions. dx, dy

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 7

Area=X*Y
X
Area 33.6% coverage area and is used as the approximation target. The
X （Table 11）
electromagnetic parameters are automatically updated in the
EMGNN through the gradient optimization algorithm [38]
Area
Y and re-parameterization [22]. This achieves minimization of
ConstraintsY
coverage area on the basis of high-performance filtering, to
4-channel More devices better meet the requirements of process manufacturing.
filter bank

Fig. 6: Schematic Diagram of Area Minimization C. Graph Transfer Learning Across Electromagnetic Topolo-
gies
To achieve lightweight design across electromagnetic
represents the direction vector in horizontal direction and
topologies, this section introduces a dual-driven cross-
vertical direction.
electromagnetic topology high-performance graph transfer
We take the area minimization as an auxiliary optimization
learning method, incorporating 1) a graph representation net-
objective, which is combined with Equation 11 to jointly
work with strong electromagnetic coupling effects representa-
serve as the final optimization objective [18], achieving the
tion power and 2) a small dataset of target electromagnetic
minimization of both the expected filtering performance error
specifications. This method addresses the challenge of ex-
and the coverage area.
pensive and scarce labeled data simulation for target circuit
Achieving the joint optimization with weighted objectives
design, achieving a lightweight design of the target circuit.
for the primary task and auxiliary task is essentially a multi-
Specifically, while ensuring the high-performance filtering
task learning problem [36]. The success of multi-task learning
of the designed circuit, this paper further introduces a cross-
depends on effectively combining auxiliary task with the pri-
electromagnetic topology graph transfer learning mechanism
mary task, i.e., assigning different weights to tasks. Traditional
driven by shared parameters of the EMGNN model and a small
methods manually set these weights through prior intuition and
dataset of target electromagnetic specifications. We define the
hyperparameter tuning, leading to randomness in the weights.
source domain as Ds = {γx1 , γx2 , ..., γxK }, where γxi ,i ∈ K
Inspired by [37], we utilize meta-learning to automatically
represents a filter data composed of x resonators, and the
assign weights to tasks in performance and area minimization
target domain as Dt = {βy1 , βy2 , ..., βyV }, where V << K
tasks, aiming for optimal task balance. Specifically, we firstly
and x ̸= y. This paper employs an improved graph transfer
construct a weighted model f (cosim, Λ) based on an MLP
learning algorithm [39; 40] to transfer the structure and feature
network to outputs the weights, which quantifies the positive
knowledge learned by EMGNN in the source domain Ds to the
impact of area constraints on performance by learning the
target domain Dt filter design while avoiding negative transfer
cosine similarity cosim between the gradients of the auxiliary
[41].
task and the primary task [37]. Λ represents the parameters
of the model. We hope to find that the coverage area can be
ϑτf ine = ϑτpre (τ ≤ k) (19)
iteratively reduced while the filtering performance is improved,
t
therefore, we hope that the auxiliary task i.e. Equation (15), ϑιf ine = ϑιpre − lr∇ϑιpre LD ι
f ine (Ψ(ϑpre , N, E, A)) (τ + ι = k)
is in the same half-space as the gradient of the main task (20)
(Equation (12)). where, Ψ(ϑι , N, E, A) represents the model, N , E, A respec-
tively represent node features, edge features, and adjacency
t
cosima = cos(∇ψp log Lp , ∇ψa log La ) (16) matrix of the input graph, and LD f ine represents the training

where we wrap the task loss within a logarithmic operator, loss function on the target domain. The loss function adopts
∇ψp log Lp and ∇ψa log La represent the gradients of the loss the l1 loss. τ , ι respectively represent the number of EMGNN
values for the primary and auxiliary tasks, cosima is taken layers for pre-training and fine-tuning. lr represents the learn-
as the input to the weighted model. This model is considered ing rate. ϑιpre represents the shared model parameters after pre-
as meta-parameters and optimized during the training process training on the source domain, replacing the process of random
based on the loss from the primary task. initialization, serving as the model’s initial parameters in the
fine-tuning stage. The fine-tuned model with its parameters
Next, we represent the joint optimization objective function
ϑιf ine can then be used for the forward design of filters.
using weights provided by the model as follows
The algorithm selectively freezes and fine-tunes the weights
ρa = f (cosima ; Λ) (17) of each layer of the pre-trained EMGNN. This process en-
hances the electromagnetic representation capability specific
Jjoint = ρp log Lp + ρa log La
(18) to the target domain filters while retaining the general electro-
= f (cosimp ; Λ) log Lp + f (cosima ; Λ) log La
magnetic features learned in the source domain, eliminating
where Jjoint represents the joint objective function, and the need to retrain the model with a large amount of target
f (cosimi ; Λ) represents the weight generated by the weighted domain data.
model. So far, we combine Equation (12) and expand Equation Table 3 provides the optimal freeze-finetune settings for
(18) to get the final joint optimization objective function. sections III.C and III.D. For transfer task 1, we define the
In this paper, the objective function Jjoint represents as source domain consisting of 4 kinds of topology structures and
equation (18) jointly weighted with filtering performance and the target domain consisting of another 6 kinds of topology

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 8

structures. For transfer task 2, we define the source domain as 22: STEP4 Lightweight Design of Circuits:
a circuit consisting of 4 resonators, and the target domain con- 23: 1) Input target transfer function and electromagnetic spec-
sists of circuits composed of 5 and 6 resonators respectively. ifications.
Subsequently, through sequential experiments, we get the best 24: 2) Load the fine-tuned EMGNN model.
transfer settings for the aforementioned transfer algorithm in 25: 3)
these transfer design tasks in section III: for cross-topology 26: for i in range M (M represents number of initial tem-
structure design, the optimal transfer setting involves fine- plates) do
tuning all layers of EMGNN with no layers frozen. For the two 27: Predicted transfer function = EMGNN(h(i) , A).
tasks of cross-unit quantities design, the best settings are fine- 28: Construct f (sim, λ), taking as meta-parameters.
tuning all layers of EMGNN and freezing the first layer (GIN 29: Based on Equation (12),(15), build Lp and La .
layer) of EMGNN while fine-tuning the remaining layers. 30: Based on Equation (16), calculate cosim, taking as
input.
TABLE III: Optimal Freeze-Finetune Settings for Transfer 31: Based on Lp , optimize meta-parameters.
Algorithm under Different Transfer Tasks 32: Based on Equation (18), backward optimize input
Source Domain Target Domain Optimal Setting electromagnetic parameters until Jjoint is minimized.
Task
Data Data 33: end for
Cross-Topology
First Type of Second Type of Finetune all layers of 34: 6) Select i corresponding to the lowest loss value as the
Topology Structure Topology Structure EMGNN
Structure Design final design result.
(4 kinds) (6 kinds)
Finetune all layers of
Filter of 5 units
EMGNN
Cross-Unit Filter of 4 units Freeze the first layer of
Quantities Design Filter of 6 units EMGNN, finetune the III. N UMERICAL R ESULTS
remaining layers

A. Experimental Setup
This paper uses state-of-the-art algorithm CircuitGNN [22]
D. Pseudocode of CircuitGTL as the baseline model to validate the superiority of our
Combining Sections A to C, Algorithm 1 presents the CircuitGTL algorithm. The parameter settings of CircuitGNN
pseudocode of CircuitGTL. are the default settings of the original paper. This paper
also uses the THz resonant filter circuit dataset sourced from
Algorithm 1 CircuitGTL: Circuit Intelligent Design
MIT’s open repository [24] as an example, which comprises
Algorithm
approximately 400,000 circuit data samples with various topo-
1: STEP1 EMCoE-AT Graph modeling: Based on Section logical structures consisting of 3-6 square resonators. The
II.A. feature descriptions are as shown in Table 4. During simulation
2: STEP2 EMGNN Pre-training based on Source Do- validation, the sampled data is divided into training, validation,
main: and test sets in an 8:1:1 ratio.
3: 1) Input h(0) , A
4: 2) TABLE IV: Dataset Feature Description
5: for i in range N (N represents number of units) do
Data Feature Value
6: Layer 1 (GIN): Neighborhood aggregation and up-
(1) Number of Units 3-6
date, output hi .
Operating Frequency 200GHz-400GHz
7: for k from 2 to K do Side Length 50µm-90µm
8: Layers 2 to K (GAT): Inter-Unit Gap 0.625µm-45µm
9: Based on Equation (2)-(3), output A. Angular Position [0, 2π]
10: Based on Equations (4)-(5), output C.
11: Based on Equations (6)-(8), compute α for i-th Table 5 provides the optimal parameter settings for Cir-
neighborhood. cuitGTL algorithm. And, we evaluate our method from two
12: Based on C and α, output h(K) . aspects:(1) model accuracy and convergence, which is repre-
13: end for sented as (a-b), and (2) designed filter performance, which is
14: end for represented as three evaluation metrics (c-e): insertion loss,
15: 3) Graph pooling and prediction: passband IOU, and out-of-band rejection ratio to thoroughly
(K) (K)
16: Graph embedding = concat(hI , hO ) evaluate the the signal transmission capability and signal re-
17: Predicted transfer function = f c(Graph embedding). jection ability of the designed filters. These evaluation metrics
18: 4) Compute l1 -Loss between predicted transfer function can be listed as follows:
and Ground truth. a. Transfer function error (Error)(dB): we use transfer
19: 5) Backpropagation. function error to represent model accuracy. It is defined as the
20: 6) Repeat Steps 1)-5) until l1 -Loss is minimized. mean absolute error (MAE) between the predicted values y
21: STEP3 EMGNN Fine-tuning with Small Target Do- and actual values y of transfer function. And it is computed
main Dataset: Based on Table 3. Output fine-tuned N
as error(y, y) = N1
P
EMGNN model. |20log10 (|yi |) − 20log10 (|yi |)|,
i=1

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 9

b. Convergence time(s)/rounds: we use convergence time units. Table 7 shows the forward predictions of CircuitGTL
and rounds it takes from the start of training to the convergence and CircuitGNN on the target domain test set. From Table 7,
of the model to represent convergence [42]. it can be observed that CircuitGTL achieves lower error with
c. Insertion loss(dB): it is defined as the absolute value of only 10,000 data samples compared to CircuitGNN trained on
the maximum magnitude of transfer function. 50,000 target domain data samples, while reducing runtime
d. Passband IOU(%): it is defined as the overlap ratio of by 31.8%. These results indicate: (1) the graph representa-
actual passband W to target passband. And it is computed as tion network with electromagnetic coupling effect weighting
W ∩W
Wexp : | W ∪Wexp
exp
|∗ 100%. introduced by CircuitGTL improves the accuracy and compu-
e. Out-of-band rejection ration (ORR)(%): it is defined as tational efficiency of cross-topology performance prediction
the normalized bandwidth value at -10dB (below -10dB is con- with the same number of units, and (2) the introduced cross-
WL R
−W−10dB
sidered a stopband) and it is computed as | −10dB |, electromagnetic topology graph transfer learning mechanism
W L −W R
where W L − W R represents the total bandwidth of signal reduces the required amount of target domain data by 80%.
observation region. A smaller value of ORR indicates the
TABLE VII: FPA Performance Comparison of CircuitGTL and
stronger out-of-band rejection capabilities.
CircuitGNN during Cross-Topology Transfer
TABLE V: CircuitGTL Algorithm Parameter Settings
No. of Training Error Convergence Convergence
Method
Data (dB) Rounds Time(s)
Stage Model Parameters Description
40,000 3.112 256 2662
Backward Optimizer Adam Optimizer [43] CircuitGNN 50,000 2.409 218 2289
Batch Size 64 60,000 2.381 213 2258
Pre-training Training Rounds 1200 CircuitGTL 10,000 2.406 605 1560
Initial Learning Rate 2 × 10−4
Weight Decay 0.5/200 rounds
Figure 7 shows a visual prediction of the transfer function
Initial Learning Rate 2 × 10−5 for cross-topology structures by CircuitGTL with the same
Fine-tuning
Dropout 0.5 number of units. From Figure 7, it is evident that CircuitGTL
accurately predicts the transfer function curves for cross-
topology structures with much less data, demonstrating the
B. FPA Evaluation on the Source Domain Dataset model’s powerful electromagnetic structure representation ca-
Table 6 compares the prediction accuracy and runtime of pability.
CircuitGTL and CircuitGNN on the source domain dataset
(The source domain dataset mentioned here and the subsequent
target domain dataset are described in sections II.C and
Table 3). We also introduce the classic GIN, GAT, and GCN
models without our proposed EMCoE-AT graph representation
method as baseline models. From Table 6, it can be observed
that compared to CircuitGNN, CircuitGTL has improved
prediction accuracy by 19.4% and reduced the convergence
time by 36%. These results indicate that CircuitGTL can
learn the electromagnetic characteristics of existing filter data
with higher accuracy and efficiency, providing support for
subsequent high-precision transfer learning of cross-topology
structures and unit quantities.

TABLE VI: Comparison of Predictive Performance of Circuit-

GTL and CircuitGNN in Source Domain Fig. 7: CircuitGTL Cross-Topology Visualization of
Prediction Results
No. of Source Error Convergence Convergence
Method
Domain Data (dB) Rounds Time(s)
Figure 8 shows the transfer functions corresponding to
GAT 60,000 4.256 586 8179 filters designed by CircuitGTL, CircuitGNN, and manual
GCN 60,000 5.012 1001 14965
GIN 60,000 4.947 192 3317 design. The design task involves a THz bandpass resonant filter
CircuitGNN 60,000 2.414 262 3354 consisting of 4 resonators with the THz passband from 220
CircuitGTL 60,000 1.945 153 2142 GHz to 250 GHz. Figure 8 shows that all three design methods
accurately concentrate their passbands within the target THz
frequency regime.
C. FPA and BPO Evaluation of Cross-Topology Structure Table 8 shows the BPO average performance of filters
Design obtained by CircuitGTL, CircuitGNN, and manual design
This section evaluates the performance of CircuitGTL in in ten different design tasks with different target passbands.
cross-topology structure transfer with the same number of From Table 8, it can be observed that: (1) filters obtained

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 10

(a) CircuitGTL (b) CircuitGNN (c) Human

TABLE IX: FPA Performance Comparison of CircuitGTL and
0 0 0
CircuitGNN when Transferred to a Filter Composed of 5 Units
-6 -6 -6
No. of Training Error Convergence Convergence
Method
S21 (dB)

S21 (dB)

S21 (dB)
-10 -10 -10 Data (dB) Rounds Time(s)
40,000 3.216 321 4623
CircuitGNN 50,000 2.905 305 4392
-20
170 220 250 300
-20
170 220 300
-20
170 220 300
60,000 2.481 431 6206
250 250
Frequency (GHz) Frequency (GHz) Frequency (GHz)
CircuitGTL 10,000 2.747 285 396
Fig. 8: Transfer function corresponding to the results of
CircuitGTL, CircuitGNN, and manual design, where the TABLE X: FPA Performance Comparison of CircuitGTL and
green area represents the target passband CircuitGNN when Transferred to a Filter Composed of 6 Units
No. of Training Error Convergence Convergence
Method
by CircuitGTL cross-topology transfer design perform as the Data (dB) Rounds Time(s)
best in all three metrics with slightly better insertion loss, and 40,000 3.356 336 4906
CircuitGNN 50,000 3.165 317 4627
the passband IOU increased by 1.5%, the stopband rejection 60,000 2.713 411 6000
ratio improved by 14.5%, and the required data amount was
CircuitGTL 10,000 2.875 402 747
only 20% of that required by CircuitGNN, and (2) the design
process of CircuitGTL is equally rapid as CircuitGNN, taking
less than 1 minute on a single GPU (NVIDIA GeForce RTX
4090), whereas manual design would take several days. These Figure 9 shows a visual prediction of the transfer function
results indicate that filters designed by CircuitGTL exhibit the for different numbers of resonators by CircuitGTL. From
best filtering performance and design efficiency, requiring only Figure 9, it is evident that CircuitGTL accurately predicts
20% of the data compared to CircuitGNN, thereby achieving the transfer function curves for cross-unit quantities with
lightweight electromagnetic parameter optimization and cross- much less data, although there are some minor errors in
topology design. certain frequency ranges away from the passband.However, the
overall results demonstrate the model’s strong electromagnetic
TABLE VIII: BPO Average Performance Comparison of Cir- structural representation capability when facing more complex
cuitGTL, CircuitGNN, and Manual Design transfer tasks.

Insertion Passband Convergence

Method ORR
Loss(dB) IOU Time
CircuitGTL 3.80 0.833 0.375 33s
CircuitGNN 3.81 0.821 0.438 35s
Manual Design 3.92 0.786 0.413 7 days

D. FPA and BPO Evaluation of Cross-Unit Quantities Design

This section evaluates the performance of CircuitGTL in
cross-unit quantities transfer with different numbers of res-
onators. Tables 9 and 10 show the forward predictions of Cir-
cuitGTL and CircuitGNN on different target domain testsets.
From Tables 9 and 10, it can be observed that: (1) for filters
with 5 resonators, CircuitGTL achieves a 5.4% improvement
in prediction accuracy with only 10,000 data samples, com-
pared to CircuitGNN trained on 50,000 target domain data
samples. The design efficiency also improves by 90.9%, and
(2) for filters with 6 resonators, CircuitGTL achieves a 9.2%
improvement in prediction accuracy with only 10,000 data
samples, compared to CircuitGNN trained on 50,000 target
domain data samples. The design efficiency also completed a
83.9% improvement. These results further demonstrate that
when facing transfer across more complex topologies (i.e.,
different unit quantities): (1) the graph representation network
with electromagnetic coupling effect weighting introduced by
CircuitGTL improves the accuracy and computational effi-
ciency of performance prediction and (2) the introduced cross-
electromagnetic topology graph transfer learning mechanism Fig. 9: CircuitGTL Cross-Unit Quantities Visualization of
reduces the required amount of target domain data by 80%. Prediction Results

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 11

Figure 10 shows the transfer functions corresponding to assuming the convergence time of the inverse design is the
filters designed by CircuitGTL, CircuitGNN, and manual same, the total computational overhead for the entire design
design. The design task involves a THz bandpass resonant process of CircuitGTL would be 20% of the data generation
filter consisting of 6 resonators with the THz passband from overhead and 9.1% of the model training overhead of Cir-
220 GHz to 250 GHz. Figure 10 shows that all three design cuitGNN. We believe this represents a significant saving in
methods accurately concentrate their passbands within the computational overhead. We can also see that when expanded
target THz frequency regime. to a circuit with 5 resonant units, the prediction performance
error of CircuitGTL differs by 0.266dB from the result trained
0
(a) CircuitGTL
0
(b) CircuitGNN
0
(c) Human with 60,000 data by CircuitGNN. When expanded to a circuit
with 6 resonant units, the error difference is 0.162dB, reducing
-6 -6 -6 the error gap. This demonstrates that CircuitGTL performs
S21 (dB)

S21 (dB)

S21 (dB)
-10 -10 -10 better when scaled to larger circuits.

-20 -20 -20

E. Area Minimization Verification
170 220 250 300 170 220 250 300 170 220 250 300
Frequency (GHz) Frequency (GHz) Frequency (GHz)
The BPO capacity of CircuitGTL includes not only the
Fig. 10: Transfer function curves corresponding to the results above filter performance optimization, but also the optimiza-
of CircuitGTL, CircuitGNN, and manual design, where the tion of non-performance indicators (coverage area of compo-
yellow area represents the target passband nents). This section validates the area constraint capability of
CircuitGTL taking a filter consisting of 4 resonators as an
Table 11 shows the BPO average performance of filters example. Hereby, we randomly generate several initial filter
obtained by CircuitGTL, CircuitGNN, and manual design templates and calculate the average overall filter areas and
in ten different design tasks with different target passbands. units side lengths after optimization for 50 identical design
From Table 11, it can be observed that: (1) filters obtained requirements, as shown in Table 12. It is worth noting that the
by CircuitGTL cross-unit quantities transfer design achievs units side lengths of these templates are randomly sampled
the lowest insertion loss, improving it by 5.9% compared from 50µm to 90µm, resulting in a certain range of initial
to CircuitGNN, while passband IOU improved by 5.7% and areas.
stopband rejection ratio improved by 21.2% and (2) the design
TABLE XII: Comparison of Area Constraint Capability of
process of CircuitGTL is equally rapid as CircuitGNN, taking
CircuitGTL and CircuitGNN
less than 1 minute on a single GPU (NVIDIA GeForce RTX
4090), whereas manual design would take several days. These Insertion Passband Total Unit Side
Method ORR
results indicate that when facing more complex topologies Loss(dB) IOU Area(µm²) Length(µm)
(i.e., different unit quantities), CircuitGTL requires only 20% CircuitGTL 3.98 0.816 0.372 21015 61.5
CircuitGNN 4.02 0.823 0.477 31672 71.4
of the data compared to CircuitGNN, achieving lightweight
electromagnetic parameter optimization and cross-unit quan-
tities design with the best performance. We believe that by From Table 12, it can be observed that, while achieving
further increasing the size of the fine-tuning dataset, the similar insertion loss, better passband overlap, and out-of-band
transfer performance of CircuitGTL can be further improved. rejection ratio, the overall filter area of CircuitGTL-designed
filters reduced by 33.6%, and the average unit side length
TABLE XI: BPO Average Performance Comparison of Cir- decreased by 13.9%. These results indicate that CircuitGTL
cuitGTL, CircuitGNN, and Manual Design effectively reduces the design area of filters without compro-
mising filtering performance, making the designed filters more
Insertion Passband Convergence practical for engineering applications and better suited for the
Method ORR
Loss(dB) IOU Time
compact manufacturing requirements of modern multi-channel
CircuitGTL 4.17 0.842 0.338 52s THz filter banks.
CircuitGNN 4.43 0.792 0.429 53s
Manual Design 4.59 0.706 0.412 7 days
IV. C ONCLUSION
CircuitGTL is analyzed in terms of computational overhead, Addressing the existing deep learning-based circuit design
particularly when scaled to larger circuits. From Table 9 and methods failing to achieve lightweight deep representation and
Table 10, we can see that when scaled to a circuit with learning of coupled circuits across-electromagnetic topologies,
6 resonant units, CircuitGTL reduces the data requirement this paper proposes an intelligent circuit design methodology
by 80% compared to CircuitGNN, while achieving better called CircuitGTL. The latter is based on the core idea of
performance. This means CircuitGTL can save 80% of the representing circuits as graph structures weighted by electro-
computational overhead during the initial data preparation magnetic coupling effects. It achieves iterative optimization
using electromagnetic simulation software. Additionally, the of electromagnetic parameters and lightweight design of cross-
time for model training to convergence is reduced from 4627s electromagnetic topologies through graph neural networks and
to 747s, significantly decreasing the computational overhead graph transfer learning mechanisms. Using the THz resonant
during model training. According to Table 9 and Table 10, filter as a case study, the effectiveness of CircuitGTL is verified

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 12

through two different transfer tasks: a) cross-topology transfer the 2018 2nd International Conference on Computer
with the same number of units, and b) cross-unit quantities Science and Artificial Intelligence, ser. CSAI ’18.
transfer. New York, NY, USA: Association for Computing
The EMCoE-AT proposed in this study can adequately char- Machinery, 2018, p. 571–575. [Online]. Available:
acterize the complex electromagnetic environment between https://doi.org/10.1145/3297156.3297160
components. Moreover, the joint optimization objective and [6] A. Al-Hyari, H. Szentimrey, A. Shamli, T. Martin,
the graph transfer learning mechanism are not dependent on G. Gréwal, and S. Areibi, “A deep learning
specific component types. Thus, our proposed method is gen- framework to predict routability for fpga circuit
erally applicable and extendable, which means it can be used placement,” ACM Trans. Reconfigurable Technol. Syst.,
to larger circuits and various resonant unit types of coupled vol. 14, no. 3, aug 2021. [Online]. Available:
circuits, such as split-ring, U-shaped, C-shaped, and E-shaped, https://doi.org/10.1145/3465373
with different electromagnetic characteristics, diverse topolo- [7] L. Xing, Y. Wen, S. Xiao, D. Zhang, and J. Zhang, “A
gies, and multiple coupling types of isomorphic component deep learning approach for series dc arc fault diagnosing
circuits. Figure 11 uses split-ring resonators (SRR) [44] as an and real-time circuit behavior predicting,” IEEE Trans-
example to show how we can extend our method. Our study actions on Electromagnetic Compatibility, vol. 64, no. 2,
offers valuable insights into the intelligent, lightweight, and pp. 569–579, 2022.
high-precision design of coupled electromagnetic structures. [8] N. Dimitriou, L. Leontaris, T. Vafeiadis, D. Ioannidis,
T. Wotherspoon, G. Tinker, and D. Tzovaras, “A deep
Electromagnetic structures Figure out the key learning framework for simulation and defect prediction
with SRR resonators factors affecting CircuitGTL
a
electromagnetic Data simulation
Target
applied in microelectronics,” Simulation Modelling Prac-
coupling effect Source
Domain
b Domain
Data
Small
Data
tice and Theory, vol. 100, p. 102063, 2020.
Extract parameters
c
d
of single resonator Node
Embeddings
[9] W. L. Hamilton, R. Ying, and J. Leskovec, “Representa-
Parameter 1: a
Parameter 2: b tion learning on graphs: Methods and applications,” arXiv
Parameter 3: c EMCoE-
Parameter 4: d GAT
Electromagnetic
preprint arXiv:1709.05584, 2017.
Coupling Effect
w
Weighted Edges EMGNN [10] A. M. Melo, M. A. Kornberg, P. Kaufmann, M. H.
after
l Extract parameters
between resonators
Transferring Piazzetta, E. C. Bortolucci, M. B. Zakia, O. H. Bauer,
Parameter 1: l
Parameter 2: w A. Poglitsch, and A. M. A. da Silva, “Metal mesh
Different Topologies
resonant filters for terahertz frequencies,” Applied Optics,
Joint
Optimization vol. 47, no. 32, pp. 6064–6069, 2008.
Lightweight design with Iterative optimization of Design

graph transfer learning electromagnetic parameters

Objective [11] Y. Li, Y. Lin, M. Madhusudan, A. Sharma, W. Xu,
S. S. Sapatnekar, R. Harjani, and J. Hu, “A customized
Fig. 11: Schematic Diagram of the Extensibility of graph neural network model for guiding analog ic place-
CircuitGTL while Taking SRR as an Example ment,” in 2020 IEEE/ACM International Conference on
Computer-Aided Design (ICCAD). IEEE, 2020, pp. 1–9.
[12] K. Hakhamaneshi, M. Nassar, M. Phielipp, P. Abbeel,
R EFERENCES and V. Stojanovic, “Pretraining graph neural networks
[1] D. Sánchez, L. Servadei, G. N. Kiprit, R. Wille, and for few-shot analog circuit modeling and design,” IEEE
W. Ecker, “A comprehensive survey on electronic de- Transactions on Computer-Aided Design of Integrated
sign automation and graph neural networks: Theory and Circuits and Systems, 2022.
applications,” ACM Transactions on Design Automation [13] K. Hakhamaneshi, N. Werblun, P. Abbeel, and V. Sto-
of Electronic Systems, vol. 28, no. 2, pp. 1–27, 2023. janović, “Bagnet: Berkeley analog generator with layout
[2] M. Kawaguchi, N. Ishii, and M. Umeno, “Analog neural optimizer boosted with deep neural networks,” in 2019
circuit and hardware design of deep learning model,” IEEE/ACM International Conference on Computer-Aided
Procedia Computer Science, vol. 60, pp. 976–985, 2015, Design (ICCAD). IEEE, 2019, pp. 1–8.
knowledge-Based and Intelligent Information Engineer- [14] H. Ren, G. F. Kokai, W. J. Turner, and T.-S. Ku, “Para-
ing Systems 19th Annual Conference, KES-2015, Singa- graph: Layout parasitics and device parameter prediction
pore, September 2015 Proceedings. using graph neural networks,” in 2020 57th ACM/IEEE
[3] K. Settaluri, A. Haj-Ali, Q. Huang, K. Hakhamaneshi, Design Automation Conference (DAC). IEEE, 2020, pp.
and B. Nikolic, “Autockt: Deep reinforcement learning 1–6.
of analog circuit designs,” in 2020 Design, Automation [15] A. Said, M. Shabbir, B. Broll, W. Abbas, P. Völgyesi, and
Test in Europe Conference Exhibition (DATE), 2020, pp. X. Koutsoukos, “Circuit design completion using graph
490–495. neural networks,” Neural Computing and Applications,
[4] A. Lberni, M. A. Marktani, A. Ahaitouf, and A. Ahaitouf, pp. 1–13, 2023.
“Analog circuit sizing based on evolutionary algorithms [16] A. Agnesina, S. Pentapati, and S. K. Lim, “A general
and deep learning,” Expert Systems with Applications, framework for vlsi tool parameter optimization with deep
vol. 237, p. 121480, 2024. reinforcement learning,” in NeurIPS 2020 Workshop on
[5] Z. Wang, X. Luo, and Z. Gong, “Application of deep Machine Learning for Systems, 2020.
learning in analog circuit sizing,” in Proceedings of [17] A. Agnesina, K. Chang, and S. K. Lim, “Vlsi placement

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 13

parameter optimization using deep reinforcement learn- “Transfer learning of graph neural networks with ego-
ing,” in 2020 IEEE/ACM International Conference on graph information maximization,” Advances in Neural
Computer-Aided Design (ICCAD), 2020, pp. 1–9. Information Processing Systems, vol. 34, pp. 1766–1779,
[18] R. Cheng and J. Yan, “On joint learning for solving 2021.
placement and routing in chip design,” Advances in [33] A. Zadeh, M. Chen, S. Poria, E. Cambria, and L.-P.
Neural Information Processing Systems, vol. 34, pp. Morency, “Tensor fusion network for multimodal senti-
16 508–16 519, 2021. ment analysis,” arXiv preprint arXiv:1707.07250, 2017.
[19] A. Mirhoseini, A. Goldie, M. Yazgan, J. W. Jiang, [34] A. Kolluru, N. Shoghi, M. Shuaibi, S. Goyal, A. Das,
E. Songhori, S. Wang, Y.-J. Lee, E. Johnson, O. Pathak, C. L. Zitnick, and Z. Ulissi, “Transfer learning using
A. Nazi et al., “A graph placement methodology for fast attentions across atomic systems with graph neural net-
chip design,” Nature, vol. 594, no. 7862, pp. 207–212, works (taag),” The Journal of Chemical Physics, vol. 156,
2021. no. 18, 2022.
[20] K. Kunal, T. Dhar, M. Madhusudan, J. Poojary, [35] C. Han, Y. Wang, Y. Li, Y. Chen, N. A. Abbasi,
A. Sharma, W. Xu, S. M. Burns, J. Hu, R. Harjani, and T. Kürner, and A. F. Molisch, “Terahertz wireless chan-
S. S. Sapatnekar, “Gana: Graph convolutional network nels: A holistic survey on measurement, modeling, and
based automated netlist annotation for analog circuits,” analysis,” IEEE Communications Surveys & Tutorials,
in 2020 Design, Automation & Test in Europe Conference vol. 24, no. 3, pp. 1670–1707, 2022.
& Exhibition (DATE). IEEE, 2020, pp. 55–60. [36] R. Caruana, “Multitask learning,” Machine learning,
[21] T. N. Kipf and M. Welling, “Semi-supervised classifica- vol. 28, pp. 41–75, 1997.
tion with graph convolutional networks,” arXiv preprint [37] X. Han, Z. Huang, B. An, and J. Bai, “Adaptive transfer
arXiv:1609.02907, 2016. learning on graph neural networks,” in Proceedings of the
[22] G. Zhang, H. He, and D. Katabi, “Circuit-gnn: Graph 27th ACM SIGKDD Conference on Knowledge Discovery
neural networks for distributed circuit design,” in Inter- & Data Mining, 2021, pp. 565–574.
national conference on machine learning. PMLR, 2019, [38] S. Ruder, “An overview of gradient descent optimization
pp. 7364–7373. algorithms,” arXiv preprint arXiv:1609.04747, 2016.
[23] H. Wang, K. Wang, J. Yang, L. Shen, N. Sun, H.-S. Lee, [39] Z. Wu and I. Savidis, “Transfer of performance models
and S. Han, “Gcn-rl circuit designer: Transferable transis- across analog circuit topologies with graph neural net-
tor sizing with graph neural networks and reinforcement works,” in Proceedings of the 2022 ACM/IEEE Workshop
learning,” in 2020 57th ACM/IEEE Design Automation on Machine Learning for CAD, 2022, pp. 159–165.
Conference (DAC). IEEE, 2020, pp. 1–6. [40] X. Wang, K. Guan, D. He, A. Hrovat, R. Liu, Z. Zhong,
[24] H. He. Circuit-GNN Dataset. (2019). [Online]. Available: A. Al-Dulaimi, and K. Yu, “Graph neural network en-
https://github.com/hehaodele/circuit-gnn. abled propagation graph method for channel modeling,”
[25] K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How IEEE Transactions on Vehicular Technology, 2024.
powerful are graph neural networks?” arXiv preprint [41] W. Hu, B. Liu, J. Gomes, M. Zitnik, P. Liang, V. Pande,
arXiv:1810.00826, 2018. and J. Leskovec, “Strategies for pre-training graph neural
[26] P. Veličković, G. Cucurull, A. Casanova, A. Romero, networks,” arXiv preprint arXiv:1905.12265, 2019.
P. Lio, and Y. Bengio, “Graph attention networks,” arXiv [42] Q. He, L. Zhang, H. Fang, X. Wang, L. Ma, K. Yu, and
preprint arXiv:1710.10903, 2017. J. Zhang, “Multistage competitive opinion maximization
[27] J. Yang, Z. Guo, J. Luo, Y. Shen, and K. Yu, “Cloud- with q-learning-based method in social networks,” IEEE
edge-end collaborative caching based on graph learning Transactions on Neural Networks and Learning Systems,
for cyber-physical virtual reality,” IEEE Systems Journal, 2024.
vol. 17, no. 4, pp. 5097–5108, 2023. [43] K. D. B. J. Adam et al., “A method for stochastic op-
[28] A. Gritsenko, K. Shayestehfard, Y. Guo, A. Moharrer, timization,” arXiv preprint arXiv:1412.6980, vol. 1412,
J. Dy, and S. Ioannidis, “Graph transfer learning,” Knowl- 2014.
edge and Information Systems, vol. 65, no. 4, pp. 1627– [44] M. Lu, W. Li, and E. R. Brown, “Second-order bandpass
1656, 2023. terahertz filter achieved by multilayer complementary
[29] J.-S. Hong and M. J. Lancaster, “Couplings of microstrip metamaterial structures,” Optics letters, vol. 36, no. 7,
square open-loop resonators for cross-coupled planar pp. 1071–1073, 2011.
microwave filters,” IEEE Transactions on Microwave
theory and Techniques, vol. 44, no. 11, pp. 2099–2109,
1996.
[30] J.-S. G. Hong and M. J. Lancaster, Microstrip filters for
RF/microwave applications. John Wiley & Sons, 2004.
[31] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit,
L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin,
“Attention is all you need,” Advances in neural informa-
tion processing systems, vol. 30, 2017.
[32] Q. Zhu, C. Yang, Y. Xu, H. Wang, C. Zhang, and J. Han,

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. This is the author's version which has not been fully edit
content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2024.3472275

IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. XX, NO. X, APRIL 2024 14

Xin Jian received the B.E. and Ph.D. degrees in Osama Alfarraj received the master’s and
communication engineering from Chongqing Uni- Ph.D.degrees in information and communication
versity, Chongqing, China, in 2009 and 2014, respec- technology from Griffith University, in 2008 and
tively, where he is currently an Associate Professor 2013, respectively. He is currently a Professor
with the College of Micro-Electronics and Commu- of computer sciences with King Saud University,
nication Engineering. Riyadh, Saudi Arabia.
His current research interests include graph rep- His current research interests include eSystems
resentation learning for electromagnetic topology, (eGov, eHealth, and ecommerce), cloud computing,
radio frequency machine learning. and big data. For two years, he has served as a
Consultant and a member for the Saudi National
Team for Measuring E-Government, Saudi Arabia.

Junyi Zhang received the B.E. degrees in elec- Keping Yu (Senior Member, IEEE) received the
tronic information engineering from Chongqing Uni- M.E. and Ph.D. degrees from the Graduate School
versity, Chongqing, China, in 2022, where he is of Global Information and Telecommunication Stud-
currently pursuing the master’s degree in electronic ies, Waseda University, Japan, in 2012 and 2016,
information at the College of Microelectronics and respectively. He was a Research Associate, Junior
Communication Engineering, Chongqing University, Researcher, and Researcher with the Global Infor-
Chongqing, China. mation and Telecommunication Institute, Waseda
His current research interests include intelligent University, from 2015 to 2019, 2019 to 2020, and
circuit design, graph representation learning for elec- 2020 to 2022, respectively. He is currently an As-
tromagnetic topology, graph transfer learning. sociate Professor, the Vice Director of the Institute
of Integrated Science and Technology at Hosei Uni-
versity, Japan. Dr. Yu has been a Highly Cited Researcher identified by
Clarivate™ (2023) and the World’s Top 2% Scientists identified by Stanford
University (2022, 2023). He received the 5th Japan Research Front Award
from Clarivate Analytics Japan, the Best Symposium Award from IWCMC
2023, the IEEE Outstanding Leadership Award from IEEE BigDataSE 2021,
Yaoyao Li received the Ph.D. degree in Electro- the Best Paper Award from IEEE Consumer Electronics Magazine Award 2022
magnetic Compatibility and Electromagnetic Envi- (1st Place Winner), IEEE ICFTIC 2021, ITU Kaleidoscope 2020, the Student
ronment in 2017 from Beihang University, Beijing, Presentation Award from JSST 2014. He has hosted and participated in more
China. Currently, he is a professor of electronic than 10 projects, is involved in many standardization activities organized by
science and technology at Beihang University. ITU-T and ICNRG of IRTF, and has contributed to ITU-T Standards Y.3071
His current research interests include intelligent and Supplement 35. He has authored more than 200 peer-reviewed research
circuit design, modeling of complex electromagnetic papers and books, including over 100 IEEE/ACM Transactions papers. He
environments, system level electromagnetic compat- has been a Guest Editor for IEEE Transactions on Computational Social
ibility and electromagnetic protection. Systems, IEEE Journal of Biomedical and Health Informatics, and Renewable
Sustainable Energy Reviews. He served as general co-chair and publicity
co-chair of the IEEE VTC2020-Spring 1st EBTSRA workshop, general co-
chair of IEEE ICCC2020 2nd EBTSRA workshop, general co-chair of IEEE
TrustCom2021 3rd EBTSRA workshop, session chair of IEEE ICCC2020,
ITU Kaleidoscope 2016.
Tianyu Yang received the B.E. degrees in electronic
information engineering from Chongqing University
of Technology, Chongqing, China, in 2023, where he
is currently pursuing the master’s degree in artificial
intelligence at the School of Computer Science and Mohsen Guizani (Fellow, IEEE) received the BS
Engineering, Chongqing University of Science and (with distinction), MS and PhD degrees in Electrical
Technology, Chongqing, China. and Computer engineering from Syracuse Univer-
His current research interests include deep learn- sity, Syracuse, NY, USA in 1985, 1987 and 1990,
ing and graph transfer learning. respectively. He is currently a Professor of Ma-
chine Learning and the Associate Provost at Mo-
hamed Bin Zayed University of Artificial Intelli-
gence (MBZUAI), Abu Dhabi, UAE. Previously, he
worked in different institutions in the USA. His
research interests include applied machine learning
and artificial intelligence, Internet of Things (IoT),
Amr Tolba (Senior Member, IEEE) received the intelligent autonomous systems, smart city, and cybersecurity. He was elevated
M.Sc. and Ph.D. degrees from the Mathematics and to the IEEE Fellow in 2009 and was listed as a Clarivate Analytics Highly
Computer Science Department, Faculty of Science, Cited Researcher in Computer Science in 2019, 2020 and 2021. Dr. Guizani
Menoufia University, Egypt, in 2002 and 2006, has won several research awards including the “2015 IEEE Communications
respectively. He is currently a Full Professor in Society Best Survey Paper Award”, the Best ComSoc Journal Paper Award
computer science with King Saud University (KSU), in 2021 as well five Best Paper Awards from ICC and Globecom Confer-
Saudi Arabia. He has authored/coauthored over 180 ences. He is the author of ten books and more than 800 publications. He
scientific articles in top-ranked (ISI) international is also the recipient of the 2017 IEEE Communications Society Wireless
journals, such as IEEE INTERNET OF THINGS Technical Committee (WTC) Recognition Award, the 2018 AdHoc Technical
JOURNAL, ACM TOIT, IEEE SYSTEMS JOUR- Committee Recognition Award, and the 2019 IEEE Communications and
NAL, etc. He served as a TPC Member at several Information Security Technical Recognition (CISTC) Award. He served as
conferences, such as DSIT 2022, CICA2022, EAI MobiHealth 2021, DSS the EditorinChief of IEEE Network and is currently serving on the Editorial
2021, etc. He has been included in the list of the top 2% of influential Boards of many IEEE Transactions and Magazines. He was the Chair of the
researchers globally (prepared by scientists from Stanford University, USA) IEEE Communications Society Wireless Technical Committee and the Chair
during the calendar years 2020, 2021, 2022, and 2023, respectively. His main of the TAOS Technical Committee. He served as the IEEE Computer Society
research interests include artificial intelligence (AI), the Internet of Things Distinguished Speaker and is currently the IEEE ComSoc Distinguished
(IoT), data science, and cloud computing. Lecturer.

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 24,2024 at 12:50:57 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.