Content uploaded by Fanghao Tian
Author content
All content in this area was uploaded by Fanghao Tian on Jul 02, 2024
Content may be subject to copyright.
Embedding-Encoded Artificial Neural Network
Model for MOSFET Preselection: Integrating
Analytic Loss Models with Dynamic Characteristics
from Datasheets
1st Fanghao Tian
Dept. of Electrical Engineering (ESAT)
KU Leuven - EnergyVille
Diepenbeek - Genk, Belgium
fanghao.tian@kuleuven.be
2nd Shirong Li
Dept. of Electrical Engineering (ESAT)
KU Leuven - EnergyVille
Diepenbeek - Genk, Belgium
shironghkr@gmail.com
3rd Xiaobo Ning
Dept. of Electrical Engineering (ESAT)
KU Leuven - EnergyVille
Diepenbeek - Genk, Belgium
xiaobo.ning@kuleuven.be
4th Diego Bernal Cobaleda
Dept. of Electrical Engineering (ESAT)
KU Leuven - EnergyVille
Diepenbeek - Genk, Belgium
diego.bernal@kuleuven.be
5th Wilmar Martinez
Dept. of Electrical Engineering (ESAT)
KU Leuven - EnergyVille
Diepenbeek - Genk, Belgium
wilmar.martinez@kuleuven.be
Abstract—The rapid emergence of wide-bandgap devices, such
as SiC and GaN, has greatly expanded the choices of components.
Efforts have been devoted to developing analytical models for
calculating the power losses. However, the complexity arises
due to the dependence of power losses on several variables
and operating conditions, making the computation intensive and
challenging when comparing hundreds of components. To address
these limitations and expedite power loss estimation for a wide
range of component choices and operating conditions, this paper
proposes the application of an embedding encoding Artificial
Neural Network (ANN) for power MOSFETs loss modelling.
The ANN models the power losses with the dynamic parameters
from datasheets under the required operational conditions. In
the meantime, embedding encoding is implemented to integrate
the discrete category information into the ANN. Furthermore,
the structure of the proposed embedding ANN is optimized to
reach a low training, validating and testing losses to ensure the
regression accuracy. Experimental validation on a SiC device
confirms the efficacy of the proposed ANN model for estimating
power losses on MOSFETs. This research accelerates the pre-
selection of MOSFETs, benefiting the design automation of power
electronics.
Index Terms—Artificial Neural Network, MOSFET, Machine
Learning, Power Losses
I. INTRODUCTION
Power MOSFETs play a pivotal role in power electronics
owing to their extensive range of applications. Nonetheless,
the losses generated during the switching process remain
inevitable and complicated to calculate due to the highly
dynamic properties of these devices. The advent of wide-
bandgap semiconductor devices, such as Silicon Carbide (SiC)
and Gallium Nitride (GaN), alongside advancements in Silicon
(Si) based devices like CoolMOS [1], has introduced varied
switching characteristics, further complicating analytical loss
estimation [2], [3].
There has been much research regarding the analytical
loss models considering the dynamic data and maximizing
datasheet utilization. For instance, a two-step piecewise con-
stant capacitance method is proposed in [4]. In the meantime,
[5] considers the transconductance as the channel current’s
function instead of constant. Moreover, due to the superior
properties of the WBG devices, many improved analytical
models also proposed by considering the physical insights.
Several analytical models for SiC MOSFETs are proposed
in [6], where the short-channel effect (SCE), drain-induced
barrier lowering effect (DIBL) and the temperature-sensitive
effect (TSE) of SiC devices are considered. [7] proposed an
iteration way to estimate the channel current and can predict
the switching losses with only around 10% error compared
with the experiment validation. Additionally, this model is
also feasible for Si devices due to the consideration of the
body diode model. Similarly, researchers have also explored
analytical models for GaN devices, including enhancement-
model [8] and cascade structure [9], considering their distinct
internal structures. One of the key assumptions in the GaN
loss model is the presence of a Miller ramp during the turn-
on transient, while the Miller plateau disappears during the
turn-off transient [10], [11].
Artificial Intelligence (AI) has gathered significant atten-
tion and extensive applications in various domains of power
electronics, including design, control, and maintenance [12].
The remarkable modelling capabilities of Artificial Neural
Networks (ANNs) stand out particularly, serving as surrogate
models for complex functions that substantially reduce sim-
ulation time without compromising modelling accuracy. This
proficiency of AI has also catalyzed the emergence of design
automation (DA) in power electronics, marking a promising
new domain for exploration [13].
With the proliferation of WBG devices in the market, the
primary objective of design automation in the field of semi-
conductor devices is to enable the data to be machine-readable
[13]. This is crucial as power losses are significantly influ-
enced by the dynamic characteristics provided in datasheets.
Fortunately, a tool proposed in [14] and presented online as
PowerBrain [15] facilitates the aggregation of comprehensive
datasheet data, thereby laying the groundwork for both loss
estimation and comparative analyses of MOSFET devices
from diverse manufacturers.
To expedite the loss estimation process, an Artificial Neural
Network (ANN) is trained to accurately represent power
losses across different operational scenarios. ANNs generally
require continuous inputs. For example, in [16] and [17],
the ANN power loss prediction model was proposed for one
single MOSFET trained by the data from SPICE simulation
in various operating conditions. In order to use one single
ANN for more component loss prediction, it is necessary to
integrate discrete categorical information such as the choice of
MOSFETs. However, the incorporation of discrete input into
ANNs has presented an obstacle [18] and label and one-hot
encoding have emerged as prevalent techniques to address this
issue. Nevertheless, the utilization of label encoding involves
assigning a numerical value to each category, establishing
an ordinal relationship between distinct categories. However,
this approach is unsuitable for addressing the topic in this
research work since different components contain no ordinal
information. In contrast, using one-hot encoding addresses
the concern of ordinality by expressing each category as a
binary vector, in which a value of 1 presents at the position
that corresponds to the category, while all other positions
are presented by 0s. This approach, however, significantly
amplifies the size of the ANN. As a result, the utilization
of embedding encoding serves to decrease dimensions and
avoids the unnecessary ordinary relationship simultaneously,
presenting a more effective approach for managing categorical
data within the ANN framework [19].
The principal contribution of this paper lies in training an
embedding encoding ANN to predict power losses for mul-
tiple MOSFETs under various operation conditions, thereby
facilitating an efficient component selection process within the
domain of multi-objective optimization.
II. ANALYTICAL POWER LOSS MODEL
Due to the superior properties of the WBG devices, the
accuracy of the analytical loss models based on Si devices
has approached the limitation of predicting the losses of the
SiC- and GaN-based devices. As a result, This paper adopts
the datasheet parameters related loss model for SiC and GaN,
respectively.
A. SiC MOSFETs
The power losses on SiC MOSFETs consist mainly of
conduction losses and switching losses. The conduction losses
can be accurately determined by the temperature and drain
current-dependent on-state resistance shown in (1).
Econd =I2
ds(rms)Rds(on)T(1)
where Econd is the conduction loss in one period T,Ids(rms)
is the drain to source current, and the dynamic on-state
resistance, which can be determined from the figures in the
datasheet, is indicated by Rds(on).
However, due to the dynamic transient process, The switch-
ing losses are more complicated to model. The simplified
switching transient waveform is shown as Fig. 1.
Fig. 1: Switching Process of a SiC MOSFET.
For turn-on loss, the following equation is obtained:
ET on =1
2triVds Io+1
2tfvVds (Io−2Ioss) + trsVdsIo+Er r
(2)
Where tri is the time for the drain-source current Ids to
reach the current Ioflows through the MOSFET when the
device is completely turn-on, tfv is the time for the drain-
source voltage Vds reduces to Vds,on,trs is the time for the
Ids to reach the maximum value because of the peak reverse
recovery current, and Err is the loss in channel caused by the
reverse recovery effect. Ioss is the discharging current from
the output capacitance of the device.
On the contrary, the turn-off process can be calculated in
following:
ET of f =1
2trv Vo(Io−2Ioss) + 1
2tfi(Vo+VLd)(Io−2Ioss)
(3)
Where trv is the time for drain to source voltage Vgs rising
to Vo,tfi is the time for the channel current ich decrease to
0. Ioss now is the charging current for the output capacitance,
and VLd is the voltage on the parasitic drain inductance Ld.
The current rising time tri and the current falling time tf i
are mainly related to the threshold voltage Vth, miller voltage
Vmil, and transconductance gm, which can be extracted from
the transfer characteristic curve in the datasheet. The tri and
tfi can be expressed in form:
tfi =−ln(Vth +Vg
Vmil +Vg
)(CgsRg +Lsgm)(4)
tri =−ln(1 −Io
gm(Vg−Vth))(Cgs Rg +Lsgm)(5)
On the other hand, the voltage rising trv and falling time
tfv are estimated by the charging and discharging process of
the output capacitance:
trv =Qcoss,discharge
Icoss,discharge
(6)
tfv =−Qcoss,charge
Icoss,charge
(7)
The charge storage in the output capacitance is obtained by
the charge equivalent method based on the dynamic parasitic
capacitance curve in the datasheet:
Qoss =1
VoZVo
0
v·Coss(v)dv (8)
Similarly, the parasitic capacitances are approximated by:
Cx,eq =1
VoZVo
0
Cx(v)dv, x =iss, oss, rss (9)
The recharge current required for the parasitic capacitance
is solved by the following equation:
−2Ls
QcossRg,on
I2
oss +2
gmRg,on
+Cgd
(Cgd +Cds)Ioss
+1
Rg,on
(Vg,on −Vth −Io
gm
)=0 (10)
2Ls
QossRg,of f
I2
oss +2
gmRg,of f
+Cgd
Cgd +Cds Ioss+
1
Rg,of f
(Vg,of f −Vth −Io
gm
)=0 (11)
Since the transconductance gmis dependent on the chan-
nel current, it will be changed by updating the value of
the recharging current Ioss. Therefore, the Ioss needs to be
determined by solving the equation above iteratively.
Additionally, the reverse recovery related parameters are
determined by the body diode characteristic of the opposite
side devices. To fit different operation point, the correlation of
the single operation point parameters needs to be made. Based
on the power diode dynamic model, some diode characteristic
value including drift region time τm, the effective carrier
lifetime τcand the time constant τrr during the recovery phase
can be solved by:
Qrf =Qrr −
I2
rr,data
2dif/dt
τrr =Qrf
Irr
Irr,data =dif
dt ·(τc−τrr )(1 −e−T1,data/τc)
1
τrr
=1
τc
−1
Tm
(12)
After knowing these values, the time of the storage injected
minority charge carriers reached to zero T1can be solved by
combined the equations below:
qm(T1) = dif
dt τc(T0+τc−t−τce−t/τc), t < T1
qe(T1) = 0 = qm(T1) + Tm(Io−dif
dt T1)
(13)
Now the peak reverse current Irr, charge storage in recovery
session Qrf , and reverse conducting time trs under different
operation can be obtained by:
trs =T1−T0
Irr =trs
Io
tri
Qrs =1
2trsIr r
(14)
Finally, the reverse recovery loss in analysis switching
device Err can be calculated by the following equation:
Err =Qrs Vds +IrrVds
τrr
tfv
(tfv −τrr +τrr e−tfv /τr r )(15)
B. GaN HEMTs
GaN HEMTs, on the other hand, have different structures
compared to the Si and SiC devices, as they do not contain
a physical body diode and have lower parasitic capacitance,
resulting in a lower power loss [11]. This paper’s analytical
model for GaN HEMTs is mainly based on the model proposed
in [10], focusing specifically on losses of GaN HEMTs in a
hard-switching b half-bridge circuit.
Compared thee transition process with SiC MOSFETs, one
of the key assumptions in this model is the presence of a
Miller ramp during the turn-on transient, while the Miller
plateau disappears during the turn-off transient. The simplified
switching waveforms for turn-on and turn-off transitions are
illustrated below:
As is shown in Fig. 2, during the turn-on process, the gate-
to-source voltage Vgs increases linearly till the plateau voltage
Vpl while Ids starts increasing when Vgs exceeds Vth till Vo.
However, due to the very small input and Miller capacitance,
the gate voltage continues rising but with a smaller slope. In
the meantime, the drain current and channel current rise due
to the discharge and charge of the two output capacitances
Fig. 2: Switching Process of the GaN HEMT.
and drop back to Ioafterwards. On the contrary, the channel
current quickly drops to 0 and turns off the device completely
before V ds rises completely during the turn-off process. Thus,
the turn-off loss is only related to the current falling time tcf .
More specifically, the turn-on process is separated into the
current rising period and the voltage falling period. The total
turn-on loss can be expressed by the equation:
ET on =1
2VdsIo(tcr +tvf )(16)
During the current rising period, Ids increases to the Io
while Vgs reaches the miler voltage from the threshold voltage.
This period of time can be estimated by the dynamic charge
as (17).
tcr =Qgs2
Ig,cr
(17)
Where Qgs2is the equivalent charge required to the tran-
sition between the miller plateau voltage and the threshold
voltage, and the Ig,cr is the gate current drop during the turn-
on process.
Qgs2can be estimated by a linear approximation given a
gate charge case by datasheet, as shown in (18).
Qgs2,data =Qgs −Qgs(th)
Qgs2=Qgs2,data ·∆Ids
Ids,data
(18)
The real-time Ig,cr , however, depends on the dynamic Vgs.
An average value is estimated as (19).
IG cr =
Vgon −Vgth +Vpl
2
Rgon
(19)
During the voltage falling period, the Vgs continues to
increase a bit instead of keeping constant as a plateau because
of the much smaller parasitic capacitance and faster switching
speed of the GaN HEMTs. This time can be calculated by:
tvf =2(Qoss1+Qoss2)
gm∆Vgs,vf
(20)
where gfs indicates the transconductance and ∆Vgs,vf , the
rise value of Vgs can be approximated by (21).
∆Vgs,vf =Ig,vf
gm
Coss1(0) + Coss2(Vds)
Crss1(0) (21)
The turn-off process also has two sections, including the
current falling and voltage rising period. However, because
of the much lower gate resistance of the GaN FETs, the
miller plateau is completely avoided, and the current will
pass through the channel quickly even before the Vds fully
increases. Therefore, there are no losses in the channel in
voltage rising time, and the turn-off loss can be calculated
by:
ET of f =1
6tcf Io∆Vds,cf (22)
Where ∆Vds,cf is the increase value of Vds, which can be
approximated by:
∆Vgd,vf =1
2
tcf Io
Coss1(0V) + Coss2(Vds)(23)
In addition, the total loss on a GaN HEMT also includes
gate charge loss and reverse conduction losses. More details
can be found in [10].
III. ARTIFICIAL NEURAL NETWORK (ANN)
Artificial Neural Networks (ANNs) have been successfully
applied to a diverse array of practical problems, largely due to
their robust capabilities in regression analysis and predictive
modelling, which significantly accelerates the calculation and
analysis process. For rapid estimation of power switch losses
under diverse operating conditions or for comparing different
devices under identical operating conditions, ANNs can be uti-
lized to construct a loss prediction model, trained by extensive
power loss-related datasets.
A. ANN STRUCTURE AND EMBEDDING ENCODING
A fundamental ANN is composed of multiple layers of
neural nodes, with each node in one layer interconnected
with nodes in the adjacent layers, collectively forming a
complex network. The specific architecture of an ANN can
vary depending on the characteristics of the dataset, including
its size and data type. The dimensions of the input and output
layers are determined by the requirements of the problem
being addressed, whereas the number and structure of the
hidden layers can be adapted based on the complexity and
volume of the data.
This paper proposes an ANN model for predicting the
losses for a wide array of power MOSFET components, each
operating under diverse conditions. A significant challenge
arises from the mixed datatype inputs required by the ANN:
these inputs comprise both discrete variables, which represent
the categorical information of device selection, and continuous
variables indicative of operational conditions such as voltage,
current, temperature, frequency, duty cycle and gate resis-
tances. Traditional ANNs struggle with such heterogeneous
data because they are designed to handle only one type of
input at a time, either categorical or continuous, which limits
their predictive power in complex scenarios like this.
To overcome this limitation, the application of embedding
encoding within an ANN is adapted as shown in Fig. 3. This
technique involves the addition of a specialized embedding
layer to the network. The primary function of this layer is to
transform discrete categorical data, meaning the device types,
into a numerical format that the ANN can process concurrently
with continuous data. By converting categories into a series of
numbers, the embedding layer creates a ’code’ that the ANN
interprets, effectively translating qualitative information into a
quantitative form that augments the ANN’s predictive accuracy
and processing efficiency. This encoding not only preserves the
categorical information but also enables the ANN to establish
nuanced relationships between device types and operational
conditions, significantly enhancing the network’s performance
in predicting power losses across different MOSFET compo-
nents.
Fig. 3: Embedding ANN for MOSFET Loss.
B. DATASET PREPARATION
By implementing a comprehensive data collection strategy,
we can dynamically monitor and record the losses experi-
enced by each component under a spectrum of operational
conditions. These conditions comprise key variables such as
voltage, current, junction temperatures, frequency and duty
cycle, which the power losses strongly depend on. Those data
are then fed into the previously discussed analytical calculation
model, where the power losses assigned to all the operation
conditions are obtained. Since conduction losses and switching
losses are dependent on different parameters, two different
groups of data are collected for ANN training. For conduction
loss, considering 5 operational parameters, which are voltage,
current, temperature, frequency and duty cycle, 2000 data for
each component is collected.Similarly, for switching losses, 5
operational parameters are also considered, which are voltage,
current, temperature, external gate on resistance and external
gate off resistance, 2500 data for each component are col-
lected. The range of all operational variables are shown in
TABLE I.
TABLE I: Range of Variables
Vds/Vds max 0.1-0.9
Ids/Ids max 0.1-0.9
Junction Temperature (°C) 25-100
Frequency (kHz) 10-500
Duty Cycle 0.1-0.9
Rgon 0-28
Rgoff 0-28
The losses from 90 components are collected, forming a
database with 180,000 sets of data for conduction loss and
225,000 sets of data for switching losses. Those data are
randomly divided into 3 categories: training dataset (60%),
validation dataset (20%) and testing dataset (20%).
It is worth mentioning that the power loss analytical model,
as previously described, necessitates the integration of dy-
namic characteristics, which are derived from the tables and
figures presented in component datasheets. Consequently, an
extensive database is formulated, consisting of a spectrum of
both static and dynamic characteristics of the components.
This database includes critical parameters like the dynamic
on-state resistance which exhibits dependency on drain current
and temperature variations, drain-to-source voltage-influenced
parasitic capacitances and transfer characteristics. Establishing
this comprehensive database is the preparatory work essential
for conducting the power loss analytical model and providing
reliable data for ANN model.
C. HYPERPARAMETERS OPTIMIZATION
Hyperparameters in ANN comprise the architecture data
such as number of hidden layers and nodes, and the training-
specific variables [20]. These parameters have a high impact
on the ANN structure definition and its regression accuracy.
The ANN in this paper is optimized by Optuna, which is an
advanced hyperparameter optimization framework proposed in
[21]. An optimal structure for both ANNs are listed in TABLE
II.
TABLE II: Optimal ANN Structure
Conduction ANN Switching ANN
Embedding Layer Neuros 13 17
Other Layers Number 3 4
Other Layers Neuros [141,140,121] [164,59,174,142]
D. Training Result
1000 epochs of training is conducted in this case. In order
to facilitate the training process, a normalization process is
implemented on both inputs and outputs data. The MSE loss
on training and validation dataset during the training process
is as shown in Fig. 5 and Fig. 6.
Fig. 4: Comparison of the ANN Output and the Experimental Result.
Fig. 5: Training Loss and Validation Loss for Conduction Loss
ANN.
Fig. 6: Training Loss and Validation Loss for Switching Loss
ANN.
Finally, for the conduction loss ANN, the training loss
and validation loss reach 1.48 ×10−8and 1.64 ×10−8,
respectively. The result on test dataset shows an MSE error
of 1.65 ×10−8. Moreover, the predicted results and original
test data are reversed to its original data from normalized data
for comparison. The results shows an average relative error of
0.21%. As for the ANN of switching losses, the final training,
validating, and testing losses are 3.04 ×10−6,8.10 ×10−6
and 7.58 ×10−6, respectively. The relative error is 2.65%.
IV. EXPERIMENTAL VALIDATION
A SiC MOSFET C3M0120065 with TO-247 package is
tested on the basis of the EVAL-1ED3491MX12M board
from Infineon. A double-pulse test is conducted for testing
its switching losses and a buck converter is utilized for the
conduction loss in steady operation condition. The experi-
mental setup is as shown in Fig. 7. As is shown above, the
bi-directional power supply works as both power source and
load allowing high current tests. For validating the model, an
isolated voltage probe and CWT Rogowski current transducer
are utilized for sensing the voltage and current, respectively.
The test result of losses in different conditions are illustrated
in Fig. 4. The results show an error of around 10%, which
complies with the conclusion of the analytical loss calculation
model. It proves that the proposed embedding ANN model
can give a relatively accurate power loss result based on the
analytical model based on dynamic datasheet data.
Fig. 7: Experiment Setup.
V. CONCLUSION
This paper has introduced an innovative approach for a fast
modelling of power losses on MOSFETs and GaN HEMTs
through the implementation of an embedding ANN. Firstly, the
power losses of various devices are calculated under various
operating conditions based on dynamic data from datasheets.
With the integration of an embedding encoding, discrete
variables representing device categories were incorporated into
the ANN model, avoiding many separated ANNs for different
components. Two ANNs for condition losses and switching
losses are trained by data from 90 components. The results
show the relative losses of 0.21% and 2.65%. This approach
enables the expeditious calculation of power losses in power
switches under various operational conditions, which paves a
way for efficient component selection, benefiting the design
optimization of power converters.
REFERENCES
[1] B. Zojer, “Application note CoolMOS™ high-voltage super
junction MOSFET - gate drive and switching dynamics,”
[Online]. Available: https://www.infineon.com/dgdl/Infineon-
MOSFET CoolMOS gate drivce switching dynamicsApplication
Notesv01 00EN.pdf?fileId=5546d4626f229553016fb392d5e7749f.
(Accessed: July 15, 2023).
[2] B. Nguyen, X. Zhang, A. Ferencz, T. Takken, R. Senger and P. Coteus,
”Analytic model for power MOSFET turn-off switching loss under the
effect of significant current diversion at fast switching events,” 2018
IEEE Applied Power Electronics Conference and Exposition (APEC),
San Antonio, TX, USA, 2018, pp. 287-291.
[3] A. S. Bahman, F. Blaabjerg, A. Dutta and A. Mantooth, ”Electrical
parasitics and thermal modeling for optimized layout design of high
power SiC modules,” 2016 IEEE Applied Power Electronics Conference
and Exposition (APEC), Long Beach, CA, USA, 2016.
[4] S. Eskandari, K. Peng, B. Tian and E. Santi, ”Accurate Analytical
Switching Loss Model for High Voltage SiC MOSFETs Includes Par-
asitics and Body Diode Reverse Recovery Effects,” 2018 IEEE Energy
Conversion Congress and Exposition (ECCE), Portland, OR, USA, 2018,
pp. 1867-1874.
[5] X. Huang, Q. Li, Z. Liu and F. C. Lee, ”Analytical Loss Model of High
Voltage GaN HEMT in Cascode Configuration,” in IEEE Transactions
on Power Electronics, vol. 29, no. 5, pp. 2208-2219, May 2014.
[6] Z. Ma, Y. Pei, L. Wang, Q. Yang, Z. Qi and G. Zeng, ”An Accurate
Analytical Model of SiC MOSFETs for Switching Speed and Switching
Loss Calculation in High-Voltage Pulsed Power Supplies,” in IEEE
Transactions on Power Electronics, vol. 38, no. 3, pp. 3281-3297, March
2023.
[7] D. Christen and J. Biela, ”Analytical Switching Loss Modeling Based
on Datasheet Parameters for mosfets in a Half-Bridge,” in IEEE Trans-
actions on Power Electronics, vol. 34, no. 4, pp. 3700-3710, April 2019.
[8] Z. Qi et al., ”An Accurate Datasheet-Based Full-Characteristics Ana-
lytical Model of GaN HEMTs for Deadtime Optimization,” in IEEE
Transactions on Power Electronics, vol. 36, no. 7, pp. 7942-7955, July
2021.
[9] J. Wang, H. S. -h. Chung and R. T. -h. Li, ”Characterization and
Experimental Assessment of the Effects of Parasitic Elements on the
MOSFET Switching Performance,” in IEEE Transactions on Power
Electronics, vol. 28, no. 1, pp. 573-590, Jan. 2013.
[10] A. Lidow, M. De Rooij, J. Strydom, D. Reusch, and J. Glaser. GaN
transistors for efficient power conversion. John Wiley & Sons, 2019.
[11] J. Gareau, R. Hou and A. Emadi, ”Review of Loss Distribution, Analysis,
and Measurement Techniques for GaN HEMTs,” in IEEE Transactions
on Power Electronics, vol. 35, no. 7, pp. 7405-7418, July 2020.
[12] S. Zhao, F. Blaabjerg and H. Wang, ”An Overview of Artificial Intel-
ligence Applications for Power Electronics,” in IEEE Transactions on
Power Electronics, vol. 36, no. 4, pp. 4633-4658, April 2021.
[13] K. Hermanns, Y. Peng and A. Mantooth, ”The Increasing Role of Design
Automation in Power Electronics: Gathering What Is Needed,” in IEEE
Power Electronics Magazine, vol. 7, no. 1, pp. 46-50, March 2020.
[14] F. Tian, D. B. Cobaleda and W. Martinez, ”Automatic Data Extraction
Based on Semiconductor Datasheet for Design Automation of Power
Converters,” 2022 International Power Electronics Conference (IPEC-
Himeji 2022- ECCE Asia), Himeji, Japan, 2022, pp. 922-927.
[15] KU Leuven - EnergyVille, ”PowerBrain.ai”. Available:
https://powerbrain.ai/. Accessed: July 15, 2023.
[16] F. Tian, D. B. Cobaleda and W. Martinez, ”Artificial-Intelligence based
DC-DC Converter Efficiency Modelling and Parameters Optimization,”
2022 24th European Conference on Power Electronics and Applications
(EPE’22 ECCE Europe), Hanover, Germany, 2022, pp. 1-7.
[17] F. Tian, D. B. Cobaleda, H. Wouters and W. Martinez, ”Parameter
Design Optimization for DC-DC Power Converters with Deep Re-
inforcement Learning,” 2022 IEEE Energy Conversion Congress and
Exposition (ECCE), Detroit, MI, USA, 2022, pp. 1-7.
[18] Hancock, J.T., Khoshgoftaar, T.M. Survey on categorical data for neural
networks. J Big Data 7, 28 (2020).
[19] M. K. Dahouda and I. Joe, ”A Deep-Learned Embedding Technique for
Categorical Features Encoding,” in IEEE Access, vol. 9, pp. 114381-
114391, 2021.
[20] T. Yu and H. Zhu, ”Hyper-parameter optimization: A review of algo-
rithms and applications,“ 2020, arXiv:2003.05689.
[21] T. Akiba, S. Sano, T. Yanase, T. Ohta and M. Koyama, ”Optuna: A next-
generation hyperparameter optimization framework”, Proceedings of the
25th ACM SIGKDD International Conference on Knowledge Discovery
Data Mining, pp. 2623-2631, 2019.
