Content uploaded by Fanghao Tian
Author content
All content in this area was uploaded by Fanghao Tian on Dec 23, 2024
Content may be subject to copyright.
Content uploaded by Wilmar H Martinez
Author content
All content in this area was uploaded by Wilmar H Martinez on Sep 11, 2024
Content may be subject to copyright.
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 1
Automated Extraction of Data from MOSFET
Datasheets for Power Converter Design Automation
Fanghao Tian, Student Member, IEEE, Qingcheng Sui, Student Member, IEEE, Diego Bernal Cobaleda, Student
Member, IEEE, and Wilmar Martinez, Senior Member, IEEE
Abstract—Power electronics design automation, implementing
Artificial Intelligence (AI) to optimize the design of power
converters, has emerged as a novel research topic given the
complexity of power converter design, whose key challenges
include power loss modeling across the enormous number of
available components. This paper proposes a novel end-to-end
AI-based tool for extracting nonlinear dynamic properties from
semiconductor datasheets, which can enhance the power loss
estimation model and accelerate the optimal design of power
converters. Firstly, thousands of images from power transistor
datasheets are collected and annotated to construct a training
database. Then, CenterNet, a neural network for image object
detection, is trained for figure segmentation from datasheets and
key element detection from figures. Optical Character Recogni-
tion (OCR) and morphological image processing techniques are
utilized to extract the specific dynamic data. The results illustrate
that the customized tool for power transistor device datasheets
in this paper can accurately extract the data, significantly
reducing the time consumption for transistor data collection
and its characteristic modeling work, promising pathways to
streamline and optimize power electronics design. The tool has
been published online and is actively being updated and improved
via http://www.powerbrain.ai.
Index Terms—Artificial intelligence, image processing, design
automation, power electronics.
I. INTRODUCTION
WITH the development of decarbonized systems such
as the electrification of transportation and renewable
energy generation, the demand for power electronic converters
has expanded substantially due to their extensive applica-
tions [1]. Increasingly strict requirements on efficiency, power
density, and lifetime for the design of power converters are
required in the meantime. However, developing the optimal
power converter for a specific application is challenging,
especially when facing the abundance of available Metal-
Oxide-Semiconductor Field-Effect Transistors (MOSFETs) in
the market [2]. In addition, emerging novel semiconductor
technologies like Silicon Carbide (SiC) and Gallium Nitride
(GaN) improved the power converter performance and further
expanded the choices of power switches. With the increasing
number of component alternatives, it is challenging to compare
F. Tian, Q, Sui, D. Bernal Cobaleda, and W. Martinez are with
the department of electrical engineering, Katholieke Universiteit
Leuven, 3000 Leuven, Belgium and EnergyVille, 3600 Genk, Belgium.
(e-mail: fanghao.tian@kuleuven.be, Qingcheng.sui@kuleuven.be,
diego.bernal@kuleuven.be, wilmar.martinez@kuleuven.be.)
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org.
Fig. 1. Illustration of a classic CNN.
them experimentally in order to define the optimal choice
based on power losses, volume and cost. Consequently, the
preselection is necessary based on their power loss compu-
tational results in various operating conditions. Power losses
generated during the switching process remain inevitable and
intricate to calculate due to the highly dynamic and nonlinear
properties of silicon, SiC MOSFETs, and GaN High-Electron-
Mobility Transistors (HEMTs). An extensive study has been
conducted on analytical loss models that consider dynamic
data, aiming to maximize datasheet utilization. For example,
it has been noted in [3] that the transconductance is a function
of the channel current rather than being constant. The authors
of [4] provide a comprehensive method for estimating power
losses in SiC MOSFETs using datasheet characteristics. This
method employs an iterative approach to estimate the channel
current and successfully predicts switching losses with an
experimental validation demonstrating an error margin of
approximately 10%. The latest research in [5], [6] mentions
that the Threshold Voltage Hysteresis (TVH) effect during
the switch-on and off period should also be considered,
where the relevant information can also be acquired from the
figures of datasheets. The aforementioned models indicate the
possibilities of estimating power losses on MOSFETs using
dynamic characteristics from datasheets, which indicates that a
comprehensive database of dynamic data from a large number
of components could benefit power converter designers for
evaluating losses efficiently and choosing the optimal option
from thousands of choices in the market.
However, the process of manually extracting dynamic data is
time-consuming and may result in generating inaccurate data.
Nevertheless, the incorporation of artificial intelligence (AI)
methodologies has enabled the automation of numerous facets
within the realm of power electronics [7], thereby facilitating
the design automation of power electronics. With their ex-
0000–0000/00$00.00 © 2021 IEEE
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 2
ceptional proficiency in object detection tasks, Convolutional
Neural Network (CNN)-based image processing algorithms
can significantly aid in automating the process of extracting
dynamic data from datasheets [8], [9], such as the example
showed in Fig. 1.
There are publications on general information extraction
from PDF documents. For instance, a tool for extracting
figures and tables is proposed in [10], which can detect various
page elements in scholarly documents and locate the figures
and tables by reasoning about empty regions. F igureS eer
proposed in [11] is a complete system for automatically
finding figures in academic papers, especially those containing
subfigures, and sorting them into result-related figures and
other figures using CNN. All the information contained in
the figures is then extracted based on Optical Character
Recognition (OCR) algorithms, which identify text present
in figures, such as titles and legends. In addition, legends
in various colors or types are used to distinguish multiple
lines in the line chart. However, figures from the datasheet
are often monochromatic, with legends distributed irregularly,
which leads to failure when directly applied. Another universal
model called ChartOCR in [12] gathers the key information
and identifies the chart type at the same time. However, when
taking into account the ability to generalize across a large
number of images in various styles, this method lacks the
precision required to extract dynamic data from numerous
datasheets, resulting in erroneous power losses calculations.
LineEX focused specifically on line charts by adapting vision
transformers and pose detection methods to detect the key
points of a line chart [13]. In this case, the coordination of
the key points of the real data from the figure should be
annotated as the training data, which makes it impossible to
train MOSFETs datasheet figures in the same way.
The aforementioned tools cannot be utilized directly on
MOSFETs datasheets, unfortunately, due to a variety of distin-
guished characteristics of MOSFETs datasheets from general
documents:
1) There are multiple figures in a fixed layout on certain
pages.
2) While some of the figures are depicted in multiple
colors, most of the datasheet figures are displayed in
grayscale or monochromatic.
3) The figures are often line charts with grids.
4) The lines are usually curved and smooth without ver-
tices.
As a result, a customized tool needs to be designed to ensure
high accuracy in extracting data from MOSFETs datasheets.
Therefore, a customized data extraction technique is built to
focus especially on line charts in datasheets. In summary, this
work focuses on developing an AI-based tool for extracting
information from the figures of datasheets of MOSFETs and
GaN HEMT devices to develop a database for enhancing
the modelling of power losses computation for simulation
in power converter design automation. To further benefit
the design engineers, the tool is designed as an end-to-end
solution, which entails the ability to extract dynamic data
simply by providing a PDF datasheet.
The structure of this paper is as follows: In section II, an
overview of CNN and CenterNet is provided. The comprehen-
sive step-by-step description of the tool is presented in Section
III. The results of CenterNet’s training and data extraction
are examined in Section IV, followed by a case study of
using the collected data for selection under various operating
conditions in Section V. Concluding remarks are included in
Section VI. The initial result of this work was presented in
the 2022 International Power Electronics Conference (IPEC-
Himeji 2022- ECCE Asia) [14].
II. CE NT ER NET BA SE D OBJECT DET EC TI ON
A. CNN
CNN is a type of deep neural network that is specifically
suited for supervised learning tasks with image input. CNNs
may automatically create powerful feature representations
from raw picture pixel values by using convolutional layers
that capture spatial characteristics. As a result, they are widely
used in computer vision applications, such as image classifi-
cation, where the items in the image can be recognized and
categorized, and object detection, where the recognized items
can be located additionally [15].
Convoluting and pooling are the two primary procedures
in CNNs. Various convolutional blocks are employed to re-
duce the image size and increase the number of channels.
Afterwards, the image size decreases more in the subsequent
pooling stage while the channel number remains constant.
Low-resolution images with a high number of channels are
generated by repeatedly applying convolution and pooling op-
erations to represent the features of the original images, based
on which the CNN may perform classification or recognition
tasks through a fully connected neural network in the end.
Fig. 2. Flowchart of two types of object detection algorithms: one-stage
method (up) and two-stage method (down).
Many object detection algorithms have been developed sig-
nificantly, which can be generally categorized into two classes.
On the one hand, two-stage methods, such as RCNN [16] and
Faster RCNN [17], generate a region proposal that detects all
possible object regions before the regions are further classified
and labelled. On the other hand, one-stage methods, such as
YOLO [18] and CornerNet [19], accomplish localization and
classification in a single flow for object detection [20], [21].
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 3
Fig. 3. Architecture of CenterNet [22].
Fig. 2 illustrates and compares the flowcharts of one-stage and
two-stage object detection algorithms.
In addition, based on whether a pre-defined anchor is
used, object detection algorithms can also be categorized
into anchor-based methods and anchor-free methods, where
anchors refer to pre-defined, fixed-size bounding boxes that
serve as reference points for detecting objects within an
image. Pre-defined anchors enhance recognition accuracy yet
compromise the flexibility and adaptability of the algorithm.
On the contrary, anchor-free approaches eliminate the need for
predefined anchor boxes, enabling them to construct bounding
boxes and localize the objects simultaneously within a single
step. Furthermore, novel anchor-free methods directly predict
and localize objects from feature maps, utilizing computational
efficient techniques such as key point estimation or center
point detection. In this way, this approach simplifies the
detection process while maintaining the speed and efficiency.
Compared to it, anchor-based methods require boxes matching
and adjustment during the training, which brings more com-
putational complexity.
B. CenterNet
CenterNet, a novel anchor-free one-stage object detection
algorithm, has an excellent performance on multiple scales
objects [22]. Fig. 3 shows the CenterNet structure using a line
curve from datasheets as an example input.
1) Backbone: The backbone of CenterNet consists of two
main parts. Firstly feature extraction is applied on a 512×512
images with 3 color channels to obtain a 16 ×16 ×2048
low-resolution high-channel feature map. The original images
will be resized to the required dimensions while maintaining
the aspect ratio. If the image is not square, white blocks will
be added to the sides to achieve the required dimensions
Selecting a backbone CNN is essential since the feature
extracted by it serves as the foundation for subsequent ob-
ject recognition and localization procedures. Many CNNs
with particular architectures have been developed for this
function, such as ResNet, a deep CNN containing residual
blocks making the NN demonstrate exceptional performance
on well-known benchmark image datasets [23]. As a result,
a 50-layer ResNet, named ResNet50, is constructed as the
backbone of the CenterNet in this paper. The second part
is the transposed convolution to obtain a 128 ×128 ×64
high-resolution feature map, where the images are segmented
into 128 ×128 sections. The reverse operation of convolution,
named transposed convolution, is implemented to increase the
image size and decrease the channel numbers in order to obtain
high-resolution feature maps.
2) Pooling: Following the high-resolution feature map,
additional pooling operations are carried out, leading to the
creation of three distinct heatmaps: category heatmap, center
heatmap, and width-height heatmap. The category heatmap
contains a number of channels equal to the number of classes
used to anticipate the presence of a specific object in a
particular section of the heatmap. The central heatmap has two
channels showing the central point’s location. Lastly, instead
of using the cascaded corner heatmaps mentioned in [22], a
2-channel heatmap is created to predict the width and length
of the object.
The definition of losses is crucial for training a CNN as
it directly impacts the performance of object detection of the
trained CNN. The CenterNet training losses consist of three
parts. The first part is focal loss-based heatmap loss, defined
as (1).
Lk=−1
NX
xyc ((1 −ˆ
Yxyc)αlog(ˆ
Yxyc)Yxy c = 1
(1 −Yxyc)β(ˆ
Yxyc)αlog(1 −ˆ
Yxyc)otherwise
(1)
where x, y indicates the position on the heatmap while
cindicates the channel that correspondingly refers to one
category. Yxyc shows the spatial distribution of possibilities
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 4
of the center points locations onto a heatmap by a Gaussian
kernel. ˆ
Yxyc is the predicted distribution where the predicted
center point is 1 and other points are 0. In addition, α, β are
coefficients, and Nis the number of key center points.
Secondly, offset loss accounts for the resolution difference
resulting from downsampling or resizing when comparing the
heatmap with the original images. This is defined as (2).
Lk=1
NX
p
ˆ
O˜p−(p
R−˜p)(2)
where ˆ
O˜pis the predicted offset and (p
R−˜p)shows the real
offset. More specifically, Ris the downsampling ratio between
the original image and heatmap size, and ˜pis the real position
on the heatmap resolution.
Lastly, size loss stands for the width and length loss of the
bounding boxes. For the k−th object detected, the size loss
is defined as (3).
Lsize =1
N
N
X
k=1
ˆ
Spk−sk(3)
where ˆ
Sp indicates the predicted length and width, while s
is the predefined length and width in annotation.
Once the structure of CenterNet and the training losses
are well established, CenterNet is prepared for training. The
trained CenterNet models will be utilized to perform two
object detection tasks: detecting figures from the datasheet
page and detecting key elements from line figures, which will
be further discussed in the following sections. The aforemen-
tioned CenterNet structure and losses definition are identical
for both models.
III. DATA EXTRACT IO N
In order to establish an end-to-end tool with a single
workflow that can directly extract the dynamic data from
the line chart by inputting the original datasheet file in PDF
format, the task is organized as an overview in Fig. 4.
Fig. 4. Overview of data extraction tasks.
Prior to processing individual figures, they are detected and
separated from the original PDF pages by CenterNet model 1.
Then, the individual figures are passed into CenterNet model 2
for object detection. Based on the detection results, the figures
are processed by the classic image processing algorithms to
extract the line in pixel coordination. Combined with the text
recognition result, the final line data in number coordination
are extracted. In the whole process, CenterNet is utilized for
localizing figures in PDF pages as model 1 and detecting
the key elements in figures as model 2, which assists in the
subsequent line data extraction.
A. Dataset Construction
Like other supervised learning algorithms, the dataset plays
a crucial role in the performance of the CenterNet. CenterNet
must be utilized for two tasks: detecting figures from PDF
pages and identifying key elements within those figures.
Therefore, two datasets need to be prepared.
Datasheets from different manufacturers have a variety of
formats available. Some pages contain six figures, while some
include only four. To strengthen the generality of the object
detection method, datasheets from various manufacturers in
different layouts are collected and annotated manually. Firstly,
2000 screenshots of datasheet pages form a dataset where all
line charts are annotated. Secondly, 3800 individual figures
cut out from datasheets form another dataset where six key
elements are annotated: title, legend, coordinate origin corner,
label, x-axis and y-axis tick values, and other information.
Fig. 5 depicts some examples of the annotated pages and
figures, in which all the elements mentioned above are anno-
tated by rectangular boxes in different colors with different
labels.
B. Training
After labelling the dataset and establishing CenterNet, 300
epochs of training are carried out. The training process is
facilitated by utilizing the frozen scheme, a training strategy
that fixes certain neural network parameters. Initially, a pre-
trained ResNet50 model with parameters from the Pascal
Visual Object Classes (VOC) dataset [24], a large benchmark
dataset, is loaded. The ResNet50 remains fixed during the
initial 150 training epochs, allowing the network to stabilize
and effectively learn the pooling layers without altering the
backbone parameters. This technique leverages the pre-trained
model’s ability to extract useful features, which accelerates the
training process and mitigates potential issues arising from the
relatively small size of our dataset.
In the subsequent 150 epochs, the entire CenterNet model,
including the ResNet50 backbone, is unfrozen and trained,
ensuring that both the backbone and the pooling parts are fine-
tuned to our specific dataset. This phased training approach
not only stabilizes the learning process but also enhances the
model’s performance by allowing comprehensive training once
the initial features are well established.
The two datasets are divided randomly into training, vali-
dation, and test datasets, with proportions of 80%, 10%, and
10%, respectively.
C. Data Extraction
Two CenterNet models are trained using two distinct
datasets for figure separation and line data extraction tasks,
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 5
Fig. 5. Some examples of annotated PDF pages (left) and annotated line charts (right).
(a) (b) (c)
Fig. 6. Line extraction process (a) Original image. (b) Binarized image. (c) Line skeleton. [14]
respectively. The initial datasheet is processed by the first
trained CenterNet model to separate line charts, which then
require individual analysis for extracting dynamic line data
information. Object detection is the first step to localize
and identify the key elements of a line chart, such as the
title, legend, coordinate origin corner, label, x-axis and y-
axis tick mark values, and other information, which can be
realized by passing the individual figures into the second
trained CenterNet model. Then, text recognition is necessary
to recognize the text and numbers required to comprehend
the coordination information. Traditional image processing
algorithms, such as edge detection and morphological image
processing techniques, are finally utilized to extract line data
by analyzing the pixels of the lines.
1) Text recognition: The auxiliary information for the line
charts is contained in the text content. OCR, which is used for
digital text recognition, has advanced significantly in recent
years due to the development of machine learning algorithms.
In this paper, the open-source OCR engine Tesseract is utilized
[25]. In this way, texts in titles, labels, and legends are
recognized digitally, as well as the numbers in mark tick
values, which are important for extracting data accurately.
2) Line data extraction: Prior to digitizing data in a line
chart, it is important to specify the tick mark values and
coordinate details. To begin, all recognized mark numbers are
grouped into the x-axis and y-axis based on their positions.
The figure is then converted to binary to detect vertical and
horizontal lines corresponding to data points on the grid in
datasheet figures. By matching the tick numbers with the grid,
the accurate coordination information is determined.
If the figure is polychromatic, an additional clustering pro-
cedure is conducted. In this way the original figure is separated
into several monochromatic figures, each figure contains only
one line in one color. If the original figure is monochromatic,
this step is skipped.
The next stage involves eliminating the background grid
and extracting data from each individual line. Morphological
image processing techniques are employed to eliminate the
grid because the lines are thicker than the grid. Morphological
closing, which involves a dilation operation followed by an
erosion, is used to remove the thin black grid from the
background. The lines’ skeleton is extracted to reduce line
width to one pixel, making data extraction straightforward.
Fig. 6 shows the process of transforming the original figure
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 6
into a binarized figure for the purpose of matching tick
mark values and then into a line skeleton for the purpose of
extracting line data. During the binarization process, titles and
labels are removed to eliminate background noise.
Lastly, for the situation where multiple lines appear in
one monochromatic figure, and several lines cross each other,
an additional stage of data correlation is necessary, which
means the trajectories of each individual line can be utilized
to estimate the value of the following point. After the line
skeletons are extracted, every line has a single pixel width.
However, after sampling with a small step on the x-axis, some
dots from different lines may share the same x-axis coordinate
values. Those dots sharing the same xicould be represented as
[(xi, yi1),(xi, yi2),· · · ,(xi, yin)]. As a result, the y-axis value
of the following dot can be estimated by the coordinates of
the previous two dots on the same line as (4):
ˆy(i+1)j=yij −y(i−1)j
xi−xi−1
(xi+1 −xi) + yij (4)
The point nearest to ˆy(i+1)jis considered to be the associ-
ated point for this line j.
Algorithm 1 Line Data Correlation Algorithm
Sampling:
Divide interval on x-axis [a, b]into nparts, obtaining
points xi=a+i·b−a
nfor i= 0,1, . . . , n.
For each xi, search all mpixels, construct the database
{(xi, yi1),(xi, yi2),...,(xi, yim)}, where i∈ {0, . . . , n}.
Initialization:
Combine all data of i= 0,1,2randomly to form a
sequence of data as a line.
Each line has an initial score sc.
Main:
for i= 3 to ndo
for each line do
Predict ˆyby (4).
Find the closest yik among all yi.
if |ˆy−yik |< p then
Link (xi, yik)to the line.
else
sc =sc −1.
if sc ≤0and line length <30 then
Delete the line.
end if
end if
end for
Remove all the dots in yithat have been linked to a line.
if yiis not empty then
Combine the rest of the data of i=iwith points in
i=i+ 1, i + 2 randomly to form new lines.
end if
end for
Delete duplicate lines and short lines.
End.
In addition, to prevent incorrect correlation or noise in-
terference, the distance between the predicted position and
the closest dot is limited to a certain range. The detailed
description of the correlation algorithm is listed in Algorithm.
1.
Algorithm 2 The Complete PowerBrain Algorithm
Input: Datasheet in PDF format
Output: Extracted figure data
Read the PDF file.
for each page i in the PDF file do
Convert page i to a .png file.
end for
for each page i.png file do
Run object detection using CenterNet Model 1 (trained
for figure detection).
for each detected box j labeled as ‘figure’ do
Cut the boxed area and save as page i figure j.png.
end for
end for
for each individual figure file page i figure j.png do
for each box k do
if box k is ‘title’, ‘legend’, ‘label’, or ‘other’ then
Run OCR using Tesseract and save the detected text.
else if box k is ‘tick’ then
Run OCR to recognize the text in numbers.
Save the number and position in all ticks.
end if
end for
end for
fig ←Load the original figure.
fig cut ←Cut areas of ‘title’, ‘legend’, ‘label’, ‘tick’, and
‘other’.
if the figure is in multiple colors then
Cluster the figure into several monochromatic figures.
end if
for each monochromatic figure or the original figure if it is
monochromatic itself do
Morphological Operation to remove the grid and keep
lines.
Extract the skeletons of each line.
Extract sequential line data using Algorithm 1.
end for
Save all lines data in pixel coordination into lines pixels.
Categorize all ticks into x ticks and y ticks.
Detect long horizontal and vertical lines from grids.
Match x ticks with vertical lines.
Match y ticks with horizontal lines.
for each data point in lines pixels do
Project the pixel coordinates to numbers coordinates.
end for
Save the final result in the machine readable format.
The complete working flow of the algorithm is shown in
Algorithm 2. It starts with separating the original PDF file
into pages. The CenterNet model 1 then detects all the figures
and separates them. Each figure will be detected by CenterNet
model 2 for key elements detection. Following that, each figure
is processed by the following steps: clustering if the figure is
in multiple colors, cutting off the parts of the title, legend,
label, and others, removing the background grid, extracting
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 7
the skeleton of lines, and categorizing and correlating each line
by Algorithm 1. Combined with the text recognition results,
the extracted sequential line data are projected to numerical
coordinates as the final results.
IV. RES ULTS
The approach proposed in this paper is an end-to-end
solution for extracting dynamic data from MOSFETs and
GaN HEMTs datasheets. The system consists of two main
parts, a CNN-based object detection algorithm and a line
chart data extraction method based on classic image processing
techniques. The CenterNet object detection method precisely
localizes and categorizes the key elements in the figure, assist-
ing the conventional image processing algorithms in accurately
extracting the data.
In contrast to full AI-based methods such as LineEX [13],
the method proposed in this paper guarantees the precision of
the collected data, which significantly influences power losses
estimation. In addition, the customized annotated dataset from
the original datasheet guarantees the accuracy of the object
detection.
A. Object Detection Result Analysis
Two CenterNet models were trained by the customized
database using the annotated datasheet pages and annotated
line charts from the datasheets. Fig. 7 depicts the training and
validation losses.
Fig. 7. Training and validation losses for two CenterNet models.
As shown above, the losses converge rapidly at the begin-
ning of the training and then become stable. After unfreezing
the backbone, the losses further decrease and steadily reach
a low value, which further illustrates the effectiveness of
the frozen training strategy. More specifically, The training
errors of two CenterNet models converge to 0.5671 and
0.7909, respectively, while the validation errors attain 0.4502
and 0.6533, respectively, which indicates a high accuracy in
recognizing the key elements from the images. Regarding the
test dataset, all tested figures undergo manual verification.
In objection detection algorithms such as CenterNet, the
confidence threshold is a parameter that determines the lowest
level of certainty required for the model to identify the
presence of an object. In order to choose an optimal threshold,
several tests with different thresholds are demonstrated on the
same test dataset. The evaluation criteria are listed below:
•True Positive (TP): the cases when the objects are cor-
rectly detected.
•False Positive (FP): the cases when the non-object areas
are wrongly recognized as objects.
•False Negative (FN): the cases when the object areas are
not recognized.
Based on the three values, the following metrics are utilized
for evaluation:
•Precision indicates the proportion of correctly detected
objects among the total number of identified regions,
defined as:
Precision =TP
TP +FP (5)
•Recall indicates the proportion of correctly detected ob-
jects among the total number of actual objects, defined
as:
Recall =TP
TP +FN (6)
•The F1-score provides a harmonic mean of precision and
recall, offering a single metric for overall performance
evaluation, defined as:
F1-score = 2 ×Precision ×Recall
Precision +Recall (7)
For the evaluation of model 1, 200 datasheet pages from
various manufacturers are tested, and the results under differ-
ent confidence thresholds are presented in TABLE I.
As shown in TABLE I, a low threshold results in high FP,
and a high threshold value results in reducing the FP cases
but increasing the FN cases. To successfully detect as many
figures in a PDF file with a minimal rate of recognizing the
non-figure area as figures, a threshold value of 0.2 is chosen
for CenetrNet model 1.
For the evaluation of model 2, 384 figures from multiple
datasheets are in the test dataset. The 6 key elements in the
figures are analyzed separately given the threshold of 0.3, 0.4,
and 0.5. The results are listed in TABLE II. Based on the F1-
score, the best threshold value is 0.4 for the title, corner, label,
and tick numbers and 0.3 for the legend and others.
In contrast to model one, the inaccuracies observed in
model 2 performance are more diverse, encompassing several
types of errors. The errors include mixing labels with titles,
erroneously recognizing short symbol labels as tick numbers,
falsely detecting additional text as titles, misidentifying non-
relevant parts as corners, and failing to detect some texts.
Although there are various inaccuracies, not all of them have a
substantial impact on the data extraction process. For instance,
an additional verification process for tick numbers is applied to
filter and eliminate detections that are not numbers. Moreover,
the identification of corners is mostly used as additional data
in cases where figures do not have the grid in backgrounds,
which is a small portion of the total figures examined. The
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 8
TABLE I
TES T RES ULT OF CE NTERNE T1
Confidence
threshold
Number of
detected objects TP cases FP cases FN cases Precision Recall F1-score
0.05 889 854 35 1 0.9606 0.9988 0.9794
0.1 859 853 6 2 0.9930 0.9977 0.9953
0.2 854 853 1 2 0.9988 0.9977 0.9982
0.3 853 852 1 3 0.9988 0.9965 0.9977
0.4 852 851 1 4 0.9988 0.9953 0.9971
0.5 845 844 1 11 0.9988 0.9871 0.9929
TABLE II
TES T RES ULT OF CE NTERNE T2
Key Confidence
threshold
Number of
detected objects TP cases FP cases FN cases Precision Recall F1-score
0.3 393 387 6 1 0.9847 0.9974 0.9910
title 0.4 385 382 3 6 0.9922 0.9845 0.9884
0.5 366 365 1 24 0.9973 0.9383 0.9669
0.3 785 767 18 1 0.9771 0.9987 0.9878
corner 0.4 768 764 4 4 0.9948 0.9948 0.9948
0.5 652 651 1 117 0.9985 0.8477 0.9169
0.3 801 776 25 4 0.9688 0.9949 0.9817
label 0.4 791 774 17 6 0.9785 0.9923 0.9854
0.5 766 761 5 19 0.9935 0.9756 0.9845
0.3 5262 5193 69 0 0.9869 1.0000 0.9934
tick 0.4 5228 5193 35 0 0.9933 1.0000 0.9966
0.5 5119 5111 8 82 0.9984 0.9842 0.9913
0.3 1050 1046 4 0 0.9962 1.0000 0.9981
legend 0.4 1030 1014 0 16 1.0000 0.9845 0.9922
0.5 975 904 0 71 1.0000 0.9272 0.9622
0.3 251 243 8 2 0.9681 0.9918 0.9798
others 0.4 238 236 2 9 0.9916 0.9633 0.9772
0.5 213 212 1 33 0.9953 0.8653 0.9258
line data extraction method in this paper can further reduce
the potential negative impacts of certain errors.
Given the low training and validation losses, the trained
CenterNet is able to identify the figures from a datasheet page
and key elements of a line chart. Fig. 8 shows some examples
of the object detection results on both figure separation and
key elements recognition tasks. On the left, the red border
boxes identify and select all the line charts on screenshots
of the datasheet pages with a number on the top left corner
showing the confidence level between 0−1, which indicates
the possibility that it is recognized as a figure. Similarly, the
four figures on the right show the key elements of the line
charts being recognized and highlighted by bounding boxes
and categorized in accordance with the pre-established classes.
B. Data Extraction Result Analysis
In this section, the results are presented in 2 dimensions.
1) Quantitative Analysis: The results of PowerBrain are
compared with those obtained using the LineEX method. A
total number of 170 figures from datasheets were tested using
both methods to evaluate their performance. Firstly, the ground
truth data was obtained by directly analyzing the decoded PDF
files, since some datasheets encode their figure data in scalable
Vector Graphics (SVG) format within the PDF file. Several
lines in one figure are represented as sequences of continu-
ous data points [(x1, y1),...,(xm, ym)]. Secondly, the data
obtained from PowerBrain and LineEX were interpolated to
obtain values at the exact same xipositions. If the ground truth
set is represented as [(x1, yg1),...,(xm, ygm)], the interpo-
lated datasets are represented as [(x1, yi1),...,(xm, yim)] for
i= 1 or 2. Additionally, to eliminate the influence of different
scales in different figures, the evaluation metric is defined as
(8).
err =1
N
N
X
k=1
1
Mk
Mk
X
i=1
|ˆy−ygi |
max yg−min yg
(8)
Where err represents the average relative error of all the
lines in one figure, Nis the total number of lines, and Mk
indicates the number of points on the kth line. ˆyand ygi
represent the y values from the comparing method and the
ground truth under the same xvalue, respectively. Lastly,
max ygand min ygindicate the maximal and minimal y values
of the ground truth. The results are listed in TABLE III.
TABLE III
QUAN TITATI VE RE SULT COMPARISON
Method Average
Error
cases
(<1%)
cases
(<2%)
CPU time
per figure
GPU time
per figure
LineEX 5.57% 50 85 83.39s 27.93s
Our Method 1.61% 149 157 14.93s 12.64s
In general, our method demonstrates significantly better
accuracy, as indicated by the majority of results (87.6%)
having a relative error of less than 1% and 92.3% having a
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 9
Fig. 8. Some examples of CenterNet outputs on pages (left) and line charts (right).
Fig. 9. Comparisons of Line Extraction Results of PowerBrain and LineEX .
relative error of less than 2%. Among the failure cases, the
primary issue was incorrect results from the OCR, which led to
errors in converting pixel coordinates to numerical coordinates.
As for the running time, the test was conducted on a HP
ZBook 15v G5 laptop with CPU i7-9750H 2.60GHz and GPU
NVIDIA Quadro P600. The results show that PowerBrain
is more time-efficient. Compared to LineEX which relies
more on GPU accelerator, our method is easier and faster to
be implemented on CPU, and relatively lower computational
power is needed.
2) Qualitative Analysis: The qualitative analyses are con-
ducted by visually comparing the results with those from the
LineEX extraction method.
Fig. 9 illustrates four examples of the comparisons of two
methods. The first column shows an example of a monochro-
matic curve, where the result by LineEX shows some devia-
tions on several points, which indicates potential influence by
the grid of the original datasheet chart during LineEX ’s key-
point extraction inference process. However, it still follows the
overall trend and might be considered reasonably accurate. The
second column shows an example of polychromatic figure with
two close lines overlapping. Our method recognized both lines
while LineEX missed the one below. The third column shows
an example of crossing lines. It can be seen that LineEX
results are not as good as PowerBrain, especially near the
crossing region. In addition to the three good examples, the
last column shows one failure case. In this case, our method
failed since number 20 was wrongly recognized as 2, which
caused error when the line data in pixel coordination being
transferred into number coordination. Among 170 cases, this
failure happened 3 times. Generally, it indicates PowerBrain
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 10
can extract the data with a high accuracy and very low failure
rate.
It should be noted that LineEX cannot function properly
when dealing with a monochromatic figure that contains
multiple lines, as it relies on color detection to determine the
number of lines. However, the majority of figures in datasheets
are monochromatic, which further illustrates the necessity of
developing a specialized tool for processing semiconductor
datasheets.
V. CA SE ST UDY: USING POW ER BRA IN DATA FOR
MOSFET SEL EC TI ON
The selection of the MOSFET component depends on many
factors, including the packaging, size, price, and power losses
under the given operating conditions. Among them, the power
losses of the device is the most complex one that requires
more modeling and calculation.
The losses in semiconductor devices are primarily catego-
rized into conduction losses and switching losses. Conduction
losses occur when the device conducts current while ON state.
Switching losses, on the other hand, arise during the transition
between the ON and OFF states. These losses are attributed to
the dynamic behavior of the device’s channel as it transitions
to the conducting state. During this process, losses associated
with switching frequency and stored energy are significant.
The following sections provide a detailed model for calculating
power losses based on datasheet specifications.
A. Conduction Loss
To estimate the conduction losses Pcond, the ON-Resistance
Rds(on), the current flowing through the device Ids(rms), and
the temperature dependant are needed.
Pcond =I2
ds(rms)Rds(on)(9)
Since Rds(on)depends on the drain current and temperature,
the accurate parameters for this calculation should be extracted
from the plots: On-resistance vs. Drain Current (For various
temperatures) and On-resistance vs. Temperature. A case study
is conducted on C2M0040120D SiC MOSFET. After data
extraction and interpolation, Fig. 10 shows a look-up map for
On-resistance under various operating conditions.
Furthermore, the conduction losses calculated by (9) are
distributed as Fig. 11.
By establishing the look-up table as illustrated in Fig. 10
and Fig. 11, design engineers could fast determine the accurate
conduction losses in the specific design situation.
B. Switching Losses
To estimate switching losses, complete transitions are con-
sidered, including the ON and OFF process as shown in Fig.
12.
The modeling of switching losses is complex due to the
fact that many parameters are dependent on the dynamic
characteristics of the device, which can be acquired from the
figures inside the datasheet. A simplified switching loss model
Fig. 10. On resistance distribution depending on drain to source current and
junction temperature.
Fig. 11. Power losses distribution based on drain to source current and
junction temperature.
Fig. 12. Switching process of a MOSFET: turn-on (left) and turn-off (right)
[26].
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 11
based on the datasheet information is described as follows, and
more details can be found in [4].
The switching-on losses are calculated by (10).
ET on =1
2triVds Io+1
2tfvVds (Io−2Ioss) + trsVdsIo+Err
(10)
tri represents the time required for the drain-source current
Ids to reach the current Io, which flows through the MOSFET
when the device is fully on. tfv denotes the time it takes
for the drain-source voltage Vds to decrease to Vds,on.trs is
the time for Ids to reach its maximum value due to the peak
reverse recovery current, and Err represents the energy loss
in the channel caused by the reverse recovery effect. Lastly,
Ioss is the current discharged from the output capacitance of
the device.
In the meantime, the turn-off process is calculated by (11).
ET of f =1
2trv Vo(Io−2Ioss) + 1
2tfi(Vo+VLd)(Io−2Ioss)
(11)
In this context, trv is the time it takes for Vds to rise to Vo,
and tfi is the time for the channel current ich to decrease to
zero. Here, Ioss represents the charging current for the output
capacitance, and VLd is the voltage across the parasitic drain
inductance Ld.
However, the parameters in (10) and (11) depend on the
characteristics and operating conditions of the device, which
require further calculations. For example, tri,tf i,tr v, and tf v
are determined by (12)-(15).
tfi =−ln(Vth +Vg
Vmil +Vg
)(CgsRg +Lsgm)(12)
tri =−ln(1 −Io
gm(Vg−Vth))(Cgs Rg +Lsgm)(13)
trv =Qoss,discharge
Ioss,discharge
(14)
tfv =−Qoss,charge
Ioss,charge
(15)
It can be seen that tri and tf i are primarily influenced by
the threshold voltage Vth, the Miller voltage Vmil, and the
transconductance gm. Meanwhile, trv and tf v are calculated
based on the charge in the output capacitance.
Furthermore, the output capacitance charge current, which
influences all (12)-(15), needs to be calculated by solving (16)
in turn-on process and (17) for the turn-off process.
−2Ls
QcossRg,on
I2
oss +2
gmRg,on
+Cgd
(Cgd +Cds)Ioss
+1
Rg,on
(Vg,on −Vth −Io
gm
)=0 (16)
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 (17)
Since transconductance gmis influenced by the channel
current, any alteration in the recharging current (Ioss) will
affect it. Determining the value of Ioss requires iteratively
solving the equation above.
During the calculation of switching losses, the dynamic
characteristics contained in the datasheet figures play an
essential role, as listed below.
1) Parasitic capacitance: The charge in the output capaci-
tance Qoss is involved in (14)-(17), and determined by (18).
Qoss =ZVo
0
v·Coss(v)dv =VoCoss,eq (18)
where Coss,eq stands for the equivalent output capacitance.
Since the parasitic capacitances are dynamically changing
according to the voltage, the equivalent capacitance must be
acquired by integration from the parasitic capacitance curve
in the datasheet, as is shown in (19).
Cx,eq =1
VoZVo
0
Cx(v)dv, x =iss, oss, rss (19)
2) Threshold voltage: The threshold voltage changes ac-
cording to the junction temperature. Accurately defining it
based on the figure in datasheet has an impact on the calcu-
lation of switching losses and the following transfer function.
3) Transfer function: The last important parameter is the
transconductance gm, which is a function of the channel
current ich as shown in (20) [4].
gm(ich) = x
sk1ix
ch
ich −k2
(20)
Where k1,k2, and xare constant coefficients. After rewrit-
ing the function as (21), the coefficients can be obtained by
curve fitting on the transfer characteristics figure (Drain to
source current vs gate to source voltage).
ich =k1(vgs −Vth)x+k2(21)
In summary, the dynamic data extracted from the datasheet
benefits a detailed analytical power loss calculation model. In
this way, many components can be evaluated given the applica-
tion conditions without experimentally testing and comparing
them individually.
C. A MOSFET Selection Case Study
In this section, an example of fast comparisons on multiple
devices based on the data collected from the datasheet is
demonstrated.
Firstly, suppose an operating condition of Vds = 50Vand
Ids = 5Ain a half-bridge with a duty cycle of 0.5and
frequency of 50kHz. A comparison of power losses in 43
devices is shown in Fig. 13.
Under the given operating condition, some devices have
relatively low losses while some devices have much higher
conduction losses due to a high on-resistance.
Secondly, suppose every condition stays the same with only
voltage sweeping from 50V to 900V, Fig. 14 illustrates the
changes of the power losses.
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 12
Fig. 13. Power losses comparison among 43 devices.
Fig. 14. Power losses vs. drain to source voltage on multiple devices.
As shown in Fig. 14, the power losses of various devices
are different in the same operating condition. Moreover, some
devices increase faster than others while the voltage increases.
Finally, a comparison when the current sweeps from 5A to
80A is conducted, and the results are shown in Fig. 15.
Fig. 15. Power losses vs. drain to source current on multiple devices.
Similarly, it can be observed that power losses of various
devices differ in the same operating condition. Moreover, if
the line for a device interrupts, it means the maximal current
is reached for that device.
In summary, the power losses of semiconductor devices
are very different and vary dramatically depending on oper-
ating conditions. Thus, a database obtained from datasheets
of multiple semiconductor devices can greatly benefit the
evaluation and comparisons of multiple MOSFETs and help
power electronics designers select the proper device based on
their design requirements.
However, the method proposed in this paper is not yet
capable of intelligently matching the lines with the legend due
to instances of monochrome lines and legends being randomly
distributed in some datasheets. Additionally, traditional image
processing techniques do not effectively address multiple lines
overlapping cases of monochrome lines. Those limitations
ought to be broken through in the future.
VI. CONCLUSIONS
Datasheets are crucial resources for power electronics de-
signers, providing precise and updated information regarding
semiconductor components. However, extracting the data from
datasheet figures is laborious, especially when numerous com-
ponents with comparable qualities need to be analyzed and
compared. Moreover, high accuracy of the data is usually
required for analyzing the power loss model. In the meantime,
the semiconductor components market is expanding daily,
leading to rapid growth in the component library. Ultimately, a
dedicated automatic data extraction tool for MOSFETs, GaN
HEMTs, IGBT, and other semiconductor components will aid
power electronics designers and streamline the automation of
power electronics design.
This study introduces a semiconductor datasheet data ex-
traction tool that converts information from PDF files into
a structured, machine-readable format, facilitating the au-
tomation of power electronics design processes. Firstly, two
datasets for training purposes are constructed, which are 2000
screenshots of datasheets with figures annotated and 3800
line charts with key elements annotated. Secondly, CenterNet
is trained to perform two object detection tasks: separating
figures from the original datasheets and recognizing key
components in line charts. Finally, using key element posi-
tion information, dynamic data is collected from line charts
through classic image processing techniques. In this way,
the average relative error of the extracted data on the test
cases is 1.61%. The results demonstrate that the CenterNet-
based tool can efficiently identify essential parts within a
figure and automatically extract dynamic data from datasheets.
Furthermore, using dynamic data can improve the accuracy of
the power loss calculation model, leading to enhancements in
the optimization of power converter designs. This tool enables
the fast construction of a comprehensive dynamic database for
the design and automation of power electronics.
Nevertheless, datasheets from different manufacturers fre-
quently contain discrepancies that might make it challenging
to create a universal, reliable, and accurate database. Matching
JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS 13
legends and lines remains an issue since manufacturers utilize
varied indication styles. The customized algorithms are imple-
mented for various manufacturers in this tool to address the
above challenges. However, standardizing datasheets across
industries is becoming important for incorporating AI into
power electronics designs and advancing design automation
for a more promising future.
REFERENCES
[1] Z. Tang, Y. Yang and F. Blaabjerg, ”Power electronics: The enabling
technology for renewable energy integration,” CSEE Journal of Power
and Energy Systems, vol. 8, no. 1, pp. 39-52, Jan. 2022.
[2] Mouser Electronics. Accessed on: October 10, 2021. [Online]. Available:
https://eu.mouser.com/c/semiconductors/discrete-semiconductors.
[3] X. Huang, Q. Li, Z. Liu and F. C. Lee, ”Analytical Loss Model of High
Voltage GaN HEMT in Cascode Configuration,” IEEE Transactions on
Power Electronics, vol. 29, no. 5, pp. 2208-2219, May 2014.
[4] D. Christen and J. Biela, ”Analytical Switching Loss Modeling Based on
Datasheet Parameters for mosfets in a Half-Bridge,” IEEE Transactions
on Power Electronics, vol. 34, no. 4, pp. 3700-3710, April 2019.
[5] C. Qian, Z. Wang, G. Xin and X. Shi, ”Datasheet Driven Switching
Loss, Turn-ON/OFF Overvoltage, di/dt, and dv/dt Prediction Method
for SiC MOSFET,” IEEE Transactions on Power Electronics, vol. 37,
no. 8, pp. 9551-9570, Aug. 2022.
[6] N. Wang, J. Zhang and F. Deng, ”Improved SiC MOSFET Model
Considering Channel Dynamics of Transfer Characteristics,” IEEE
Transactions on Power Electronics, vol. 38, no. 1, pp. 460-471, Jan.
2023.
[7] S. Zhao, F. Blaabjerg and H. Wang, ”An Overview of Artificial Intelli-
gence Applications for Power Electronics,” IEEE Transactions on Power
Electronics, vol. 36, no. 4, pp. 4633-4658, April 2021.
[8] Z. -Q. Zhao, P. Zheng, S. -T. Xu and X. Wu, ”Object Detection With
Deep Learning: A Review,” IEEE Transactions on Neural Networks and
Learning Systems, vol. 30, no. 11, pp. 3212-3232, Nov. 2019.
[9] L. Aziz, M. S. B. Haji Salam, U. U. Sheikh and S. Ayub, ”Exploring
Deep Learning-Based Architecture, Strategies, Applications and Current
Trends in Generic Object Detection: A Comprehensive Review,” IEEE
Access, vol. 8, pp. 170461-170495, 2020.
[10] C. Clark and S. Divvala, ”PDFFigures 2.0: Mining figures from research
papers,” in 2016 IEEE/ACM Joint Conference on Digital Libraries
(JCDL), 2016, pp. 143-152.
[11] N. Siegel, Z. Horvitz, R. Levin, S. Divvala, and A. Farhadi, “Figureseer:
Parsing result-figures in research papers,” Lect. Notes Comput. Sci.
(including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics),
vol. 9911 LNCS, pp. 664–680, 2016.
[12] J. Luo, Z. Li, J. Wang and C. -Y. Lin, ”ChartOCR: Data Extraction
from Charts Images via a Deep Hybrid Framework,” in 2021 IEEE
Winter Conference on Applications of Computer Vision (WACV), 2021,
pp. 1916-1924.
[13] S. V. P, M. Yusuf Hassan and M. Singh, ”LineEX: Data Extraction
from Scientific Line Charts,” in 2023 IEEE/CVF Winter Conference on
Applications of Computer Vision (WACV), Waikoloa, HI, USA, 2023,
pp. 6202-6210.
[14] F. Tian, D. B. Cobaleda and W. Martinez, ”Automatic Data Extraction
Based on Semiconductor Datasheet for Design Automation of Power
Converters,” in 2022 International Power Electronics Conference (IPEC-
Himeji 2022- ECCE Asia), 2022, pp. 922-927.
[15] L. Alzubaidi et al., “Review of deep learning: concepts, CNN architec-
tures, challenges, applications, future directions,” J Big Data, vol. 8, p.
53, 2021.
[16] R. Girshick, J. Donahue, T. Darrell and J. Malik, ”Rich Feature Hier-
archies for Accurate Object Detection and Semantic Segmentation,” in
2014 IEEE Conference on Computer Vision and Pattern Recognition,
2014, pp. 580-587.
[17] S. Ren, K. He, R. Girshick and J. Sun, ”Faster R-CNN: Towards
Real-Time Object Detection with Region Proposal Networks,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no.
6, pp. 1137-1149, 1 June 2017.
[18] J. Redmon, S. Divvala, R. Girshick and A. Farhadi, ”You Only Look
Once: Unified, Real-Time Object Detection,” in 2016 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779-
788.
[19] H. Law and J. Deng, “CornerNet: Detecting Objects as Paired Key-
points,” Int. J. Comput. Vis., vol. 128, no. 3, pp. 642–656, Mar. 2020.
[20] Z. Zou, K. Chen, Z. Shi, Y. Guo and J. Ye, ”Object Detection in 20
Years: A Survey,” Proceedings of the IEEE, vol. 111, no. 3, pp. 257-
276, March 2023.
[21] Z. -Q. Zhao, P. Zheng, S. -T. Xu and X. Wu, ”Object Detection With
Deep Learning: A Review,” IEEE Transactions on Neural Networks and
Learning Systems, vol. 30, no. 11, pp. 3212-3232, Nov. 2019.
[22] K. Duan, S. Bai, L. Xie, H. Qi, Q. Huang and Q. Tian, ”CenterNet: Key-
point Triplets for Object Detection,” in 2019 IEEE/CVF International
Conference on Computer Vision (ICCV), 2019, pp. 6568-6577.
[23] K. He, X. Zhang, S. Ren and J. Sun, ”Deep Residual Learning for Image
Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2016, pp. 770-778.
[24] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J.
Winn, and A. Zisserman, ”The Pascal Visual Object Classes Challenge:
A Retrospective,” International Journal of Computer Vision, vol. 111,
no. 1, pp. 98-136, Jan. 2015.
[25] R. Smith, ”An Overview of the Tesseract OCR Engine,” in Ninth In-
ternational Conference on Document Analysis and Recognition (ICDAR
2007), pp. 629-633, 2007.
[26] F. Tian, S. Li, X. Ning, D. B. Cobaleda and W. Martinez, ”Embedding-
Encoded Artificial Neural Network Model for MOSFET Preselection:
Integrating Analytic Loss Models with Dynamic Characteristics from
Datasheets,” in 2024 IEEE Applied Power Electronics Conference and
Exposition (APEC), Long Beach, CA, USA, 2024, pp. 1574-1580.
