Machine Learning based

device model for AlGaN/GaN

HEMT

UG PROJECT

MEMBERS
Dhruvi Jain - 21095040
Neeraj Saketh Vamsi Pitla - 21095083
Vedansh Arya - 21095126

Under the supervision of:

Dr. Jaya Jha

DEPARTMENT OF ELECTRONICS ENGINEERING

INDIAN INSTITUTE OF TECHNOLOGY (BHU) VARANASI

1
 CERTIFICATE

This is to certify that the UG Project entitled “Machine Learning based device
model for AlGaN/GaN HEMT” submitted by Dhruvi Jain (21095040), Neeraj
Saketh Vamsi (21095083) and Vedansh Arya (21095126), to the Department of
Electronics Engineering, Indian Institute of Technology (Banaras Hindu
University) Varanasi, in partial fulfillment of the requirements for the award of
the degree “Bachelor of Technology” in Electronics Engineering is an authentic
work carried out at Department of Electronics Engineering, Indian Institute of
Technology (Banaras Hindu University) Varanasi under my supervision and
guidance on the concept vide project grant as acknowledged.

Dr. Jaya Jha

Assistant Professor
Department of Electronics Engineering,
Indian Institute of Technology (BHU) Varanasi

2
 DECLARATION

I hereby declare that the work presented in this project titled “Machine
Learning based device model for AlGaN/GaN HEMT” is an authentic record
of our own work carried out at the Department of Electronics Engineering,
Indian Institute of Technology (Banaras Hindu University), Varanasi as
requirement for the award of degree of Bachelors of Technology in Electronics
Engineering, submitted in the Indian Institute of Technology (Banaras Hindu
University) Varanasi under the supervision of Dr. Jaya Jha, Department of
Electronics Engineering, Indian Institute of Technology (Banaras Hindu
University) Varanasi. It does not contain any part of the work, which has been
submitted for the award of any degree either in this Institute or in other
University/Deemed University without proper citation.

Dhruvi Jain Neeraj Saketh Vamsi Pitla

(21095040) (21095083)

Vedansh
Arya
(21095126)

3
 ABSTRACT

In our study, we present a novel approach utilizing machine learning (ML) for
modeling AlGaN/GaN high electron mobility transistors (HEMTs). We
conducted analyses on two distinct devices: one featuring an AL2O3 oxide layer
and the other without such a layer.We also focused on predicting the
S-parameters of devices for different frequency ranges.
For the device with the oxide layer, we focused on predicting the drain current
values at the minimum Gate to Source Length (LGS). Conversely, for the device
lacking the oxide layer, our attention was on forecasting the drain current
values at higher drain and gate voltages.
Our research holds significant promise, particularly in scenarios where
experimental characterization at elevated voltage ranges proves challenging. In
such situations, our ML-based analysis offers a rapid and resource-efficient
means of understanding device behavior. This methodology stands to greatly
benefit both in terms of computational resources and time.
Moreover, our approach introduces a cost-effective alternative to traditional
experimentation. Instead of exhaustive testing across all possible parameter
variations, we can strategically conduct experiments for select values and utilize
ML predictions to infer behavior for other configurations. This streamlined
approach not only saves valuable time but also conserves financial resources.

4
 Contents

Abstract 4

1 Introduction

1.1 Overview of AlGaN/GaN HEMTs and Their Importance 6

1.2 Adoption of Machine Learning in Semiconductor Device Analysis 7

2 Detailed Literature Survey of Existing Technologies

2.1 Compact Models 9

2.2 Recent Researches 10

3 Work Done
3.1 Objectives and Methodological Approach 12

3.2 Comparative Analysis and Practical Implications 13

4 Experiments/Simulations and Results 14

5 Challenges and Conclusions

5.1 Challenges for Machine Learning based Device modeling 21
5.2 Conclusion 22

6 Bibliography 24

5
 Chapter 1

INTRODUCTION

1.1. Overview of AlGaN/GaN HEMTs and Their Importance

AlGaN/GaN high electron mobility transistors (HEMTs) hold immense

importance in the semiconductor industry due to their unique properties and
potential applications. Here's why they are crucial:

1. High Electron Mobility: AlGaN/GaN HEMTs exhibit exceptional

electron mobility, making them well-suited for high-frequency and
high-power applications. This characteristic enables efficient operation at
microwave and radio frequencies, making them invaluable in wireless
communication systems such as cellular networks and radar systems.

2. High Power Density: These HEMTs can handle high power

densities, making them ideal for power amplifiers in various applications,
including satellite communications, military radar systems, and wireless
infrastructure.

3. High Temperature Operation: AlGaN/GaN HEMTs can operate at

elevated temperatures, making them suitable for harsh environments such as
automotive and aerospace applications where reliability under extreme
conditions is paramount.

6
4. Energy Efficiency: Their high efficiency and power-handling
capabilities make them attractive for power management applications,
including power supplies, inverters, and electric vehicle systems, contributing
to energy savings and reduced environmental impact.

5. Miniaturization and Integration: AlGaN/GaN HEMTs offer the

potential for miniaturization and integration, enabling the development of
compact and lightweight devices for various applications, including portable
electronics, wearable devices, and IoT (Internet of Things) sensors.

1.2. Adoption of Machine Learning in Semiconductor

Device Analysis

Our analysis using machine learning (ML) techniques can significantly enhance
the understanding and optimization of AlGaN/GaN HEMTs:

1. Predictive Modeling: ML-based modeling allows for accurate

predictions of device characteristics under different operating conditions,
facilitating the design and optimization of AlGaN/GaN HEMTs for specific
applications without the need for exhaustive experimental testing.

2. Resource Efficiency: By leveraging ML algorithms, our analysis

streamlines the characterization process, reducing the need for extensive
experimental testing and conserving valuable resources such as time, materials,
and equipment.

3. Enhanced Performance: Through advanced data analysis and pattern

recognition, our approach can identify key factors influencing device
performance and optimize device design parameters to maximize
performance metrics such as power efficiency, gain, and linearity.

7
4. Accelerated Development: By providing rapid insights into device
behavior and performance, our analysis expedites the development cycle of
AlGaN/GaN HEMTs, allowing semiconductor manufacturers to bring
innovative products to market more quickly and cost-effectively.

8
 CHAPTER 2
Detailed Literature Survey of Existing Technologies

2.1. Compact Models

Two traditional types of modeling for AlGaN/GaN HEMTs are Physics based
Simulations and Empirical Modeling. While they both have their own merits
, they have some disadvantages too which are :-

Physics-Based Modeling:

● Solving complex equations can be computationally intensive.

● Models may oversimplify device behavior, leading to inaccuracies.
● Models may not easily adapt to different scenarios or device geometries.
● Difficulty Incorporating Nonlinear Effects: Modeling
nonlinear phenomena can be challenging.

Empirical Modeling:

● Models may not generalize well to different conditions or devices.

● Selection of equations can introduce bias.
● Handling Complex Relationships: Capturing complex interactions can
be difficult.

Model Types :-
1. Physical models- all parameters have physical significance.
2. Empirical equivalent circuit models- based on empirical observations.
3. Table based models- based directly on the measured data.

Compact model requirements:

1. The model must be stable and convergent.

2. Circuit simulation speed and accuracy is critical for timely design.
3. The model should work in all simulation modes(transient and
harmonic balance) and must model all the relevant physics affecting
the device behavior

9
2.2. Recent Researches

A variety of research papers have been published in recent years on ML-based

device modelling AlGaN/GaN HEMT. They focused on analyzing process
dependencies on threshold voltage in recessed gate AlGaN/GaN
MIS-HEMTs, Improving semiconductor device modeling for electronic
design automation using machine learning techniques, etc.
● One of the research papers investigates the use of machine learning
(ML) techniques to model the behavior of AlGaN/GaN
high-electron-mobility transistors (HEMTs). They face challenges due to
material variations, process variations, and device geometry. The study
explores two primary ML approaches: Tikhonov, which solves an
inverse problem by adding a regularization term to the objective function
to mitigate overfitting and provide a stable solution, and Artificial
Neural Networks (ANNs), specifically feedforward neural networks that
are trained using labeled data to predict device characteristics based on
input features such as gate length, gate-recessed depth, and dielectric
thickness.
Accurate data collection during device fabrication is crucial for this
modeling framework. Key parameters include gate-recessed depth, gate
length, dielectric thickness, and other process-related variables. The ML
model is trained using a dataset that includes both simulated and
experimental data. Using Tikhonov regularization, the ML-based model
analyzes the impact of fabrication processes on the threshold voltage
(V_th) of AlGaN/GaN HEMTs. By varying process parameters, the
model predicts changes in V_th and identifies critical steps affecting
device performance.
Experimental validation of the model’s predictions against actual
measurements demonstrates good agreement between predicted and
measured V_th values, underscoring the effectiveness of the ML-based
approach. Gate-recessed depth and dielectric deposition play crucial roles
during device fabrication, and optimal values for these parameters are
determined based on the ML model’s recommendations.

10
 Flowchart of this work using ML-based TR approach to model the VTH.

● Another research paper discusses the integration of machine learning

(ML) techniques into electronic design automation (EDA), specifically
for semiconductor device modeling, and highlights the crucial role of
quality training datasets in enhancing ML model performance. It
introduces a self-augmentation strategy using variational autoencoder
(VAE) techniques designed to improve ML-based device modeling.
This innovative strategy reduces the reliance on extensive experimental
data points and eliminates the dependency on technology
computer-aided design (TCAD) tools.
The application of this approach is demonstrated through its use in
predicting the ohmic resistance value in gallium nitride (GaN) devices,
which achieves a remarkable 70% reduction in mean absolute error
(MAE) for the predictions.Additionally, the flexibility of the proposed
method allows for its adaptation to various tasks within the
semiconductor industry, underscoring its broad applicability and potential
to enhance next-generation EDA simulations and modeling. This
suggests promising avenues for future research and implementation in the
field, aiming to leverage ML advancements to streamline and improve
the accuracy of semiconductor device design and manufacturing
processes.

11
 CHAPTER 3
WORK DONE
3.1. Objectives and Methodological Approach

Objectives:
Our primary objective is to develop a robust predictive model for AlGaN/GaN
HEMTs using machine learning techniques. We aim to leverage Python
libraries such as Pandas and NumPy for data handling, Seaborn and Matplotlib
for exploratory data analysis (EDA), and CatBoost,ARIMA method for
predictive modeling. Our goal is to provide accurate and reliable predictions to
enhance device characterization.

Methodological Approach:

1. Data Collection and Preparation:

We sourced data from MATLAB text files and transformed them into Pandas
data frames for analysis. This preprocessing step ensures compatibility with
our Python-based workflow.

2. Exploratory Data Analysis (EDA):

Utilizing Seaborn and Matplotlib, we conducted EDA to understand the data's
distribution, identify patterns, and detect outliers. This step informs our
subsequent data processing and modeling decisions.

3. Data Filtering and Feature Selection:

We systematically filtered out devices exhibiting aberrant characteristics,
focusing our analysis on those displaying consistent patterns. This
selective approach improves the quality of our model by excluding
erroneous data.

12
4. Model Training with CatBoost, XGBoost and ARIMA:
Employing the CatBoost algorithm, we trained our predictive model on the
refined dataset. During training, we observed that negative current values
negatively impacted prediction accuracy. Consequently, we implemented a
filtering mechanism to remove such anomalies, enhancing model performance.
For the s- parameters, we used the ARIMA approach with XGBoost algorithm
for predicting the s-parameter values at higher frequency values by using lower
frequency values. We used normalized data for training the model.

5. Batch Training of Multiple Devices:

Our approach allows for simultaneous training of various devices, facilitating
efficient model development and comparison across different device
configurations.

3.2. Comparative Analysis and Practical Implications

Comparative Analysis:
Our comparison between predicted and experimental data revealed a relatively
low margin of error. While not perfect, this suggests that our method has some
potential for practical application in industry. The modest level of error
indicates that our predictive model captures key device characteristics with
reasonable accuracy.

Practical Implications:
The findings of our analysis hint at practical benefits for the semiconductor
industry. While not groundbreaking, the ability to predict device behavior with
some degree of accuracy could offer modest cost and time savings compared to
traditional experimental testing methods. Our approach may streamline certain
aspects of device development and testing processes, albeit with some
limitations. However, further refinement and validation would be necessary
before widespread adoption in industrial settings. Nonetheless, these
preliminary results suggest a promising direction for future research and
development efforts in semiconductor characterization and optimization

13
 CHAPTER 4

Experiments/Simulations and Results

Device Model 1:
For Device Model 1, our dataset comprised data from 40 individual devices.

The characteristics of all 40 devices provided valuable insights, aiding in

the identification and removal of erroneous data points.

For IDS and VDS analysis For IDS and VGS analysis

14
IDS and VGS Analysis:

Training for a Single Device

15
Training for Multiple Devices

IDS and VDS Analysis :

Training for a Single Device

Training for Multiple Devices

16
 Device Model 2:

For Device Model 2, our dataset comprised data from 4 individual devices.

Prediction of IDS vs VGS for LGS value 0.5 (VDS = 2 V)

17
 S-Parameter predictions

For S-parameters, we had the s-parameter values over the range of

40 GHz. The values at higher frequency were very noisy so we
have limited our analysis to 20-30 GHz. For our model, we have
used the ARIMA method and XGBoost approach to make our
predictions. These models factor for sequential data which our
frequency data resembles. For S12 values, we have used
frequencies up to 12 GHz for training and predicted the values
from 12GHz up to 20GHz. For S21,we have used frequencies up to
21 GHz for training and predicted the values from 21 GHz up to
30GHz. For both of these parameters, we have plotted the
comparison in lg(dB) scale.

As we can clearly see in that plot above, the values above 20 GHz
have a significant effect of noise in them. For this reason values till
20 GHz have been considered. Following are the plots showing
comparison between the predicted vs actual values for the training
frequency range.

18
The R2 score for the predictions of S12 and S21 were 0.84 and 0.82
respectively.

19
Similarly, for S11 and S22, we trained the model up to frequency values
of 12 GHz for both and predictions were made for frequency values up
to 20 GHz. For these two parameters, we visualized the predictions by
plotting smith charts.

The R2 score for these two predictions was found to be 0.86 and 0.84.

20
 CHAPTER 5
Challenges and Conclusion

5.1Challenges for Machine Learning based Device

modeling:-
1. Data Quality and Quantity: Obtaining high-quality and sufficient data
for training machine learning models can be challenging, especially for
complex semiconductor devices like AlGaN/GaN HEMTs. Data may be
limited in quantity or contain noise, outliers, or missing values, which
can adversely affect model performance and generalization.
2. Feature Engineering: Identifying and extracting relevant features from
raw data to effectively capture device behavior poses a significant
challenge. AlGaN/GaN HEMTs exhibit complex and nonlinear
characteristics, making it challenging to determine which features are
most informative for modeling. Additionally, feature engineering
requires domain expertise and may involve manual trial and error,
leading to
time-consuming and resource-intensive processes.
3. Model Complexity and Interpretability: Developing machine learning
models that strike a balance between complexity and interpretability is
crucial for effective device modeling. Complex models, such as deep
neural networks, may achieve high accuracy but lack interpretability,
making it challenging to understand the underlying physics of device
operation. Conversely, simpler models may sacrifice accuracy for
interpretability, limiting their ability to capture complex device
behaviors accurately.
4. Generalization to New Conditions: Ensuring that machine learning
models generalize well to new operating conditions, device geometries,
and material compositions is essential for their practical applicability.
AlGaN/GaN HEMTs exhibit variability in performance due to factors
such as temperature, biasing conditions, and fabrication processes.
5. Model Validation and Uncertainty Estimation: Evaluating the
performance of machine learning models and estimating their

21
 uncertainty is critical for assessing their reliability and making informed
decisions. Traditional validation techniques, such as cross-validation and
holdout validation, may not adequately capture the uncertainty inherent
in device modeling. Additionally, quantifying uncertainty in predictions
is challenging, especially in complex semiconductor systems where
variability and nonlinearity are prevalent.
6. Computational Resources and Scalability: Training and deploying
machine learning models for device modeling can be computationally
demanding, requiring substantial computational resources and time.
Scaling machine learning algorithms to handle large datasets and
complex models may pose challenges, particularly for researchers with
limited access to high-performance computing infrastructure.
7. Integration with Physics-Based Models: Integrating machine learning
with physics-based models to leverage the strengths of both approaches
presents technical challenges. Combining empirical data-driven insights
with physics-based principles requires careful calibration, validation,
and interpretation to ensure consistency and accuracy in device
modeling.

5.2Conclusion:-
In conclusion, our investigation showcases the transformative potential of
machine learning (ML) in modeling AlGaN/GaN HEMTs, but significant
challenges remain. Foremost among these is the constrained availability of
comprehensive and diverse datasets, limiting the scalability and robustness of
ML models. Additionally, the intricate interplay between semiconductor
physics, fabrication processes, and device performance presents complexities
that demand sophisticated modeling approaches and computational
resources.

Moreover, the rapid evolution of semiconductor technology introduces a

dynamic landscape, requiring continuous adaptation and refinement of ML
models to remain relevant and effective. Collaborative efforts among
researchers, semiconductor manufacturers, and data providers are crucial in
addressing these challenges. Establishing data-sharing initiatives, standardizing
characterization protocols, and advancing computational methodologies are
essential steps toward overcoming these hurdles.

22
Despite these obstacles, our work underscores the transformative potential of
ML in accelerating device optimization and informing future advancements in
the semiconductor industry. By surmounting these challenges through
collective endeavors, we can harness the full power of ML to drive innovation,
enhance efficiency, and unlock new opportunities in semiconductor research
and development.

23
 CHAPTER 6
Bibliography

1. Tian-Li Wu , Member, IEEE and Sayeem Bin Kutub , Student

Member, IEEE - “Machine Learning-Based Statistical Approach to
Analyze Process on Threshold Voltage in Recessed Gate AlGaN/GaN
MIS-HEMTs” -
https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9247437.
2. Zeheng Wang , Member, IEEE, Liang Li, Ross C. C. Leon, Jinlin Yang,
Junjie Shi, Timothy van der Laan, and Muhammad Usman -
“Improving Semiconductor Device Modeling for Electronic by
Machine Learning Techniques”.-
https://ieeexplore.ieee.org/document/10235876
3. T. Imada, M. Kanamura, and T. Kikkawa, Enhancement-mode GaN
MIS-HEMTs for power supplies,” in Proc. Int. Power Electron.
Conf. (ECCE ASIA), Sapporo, Japan, 2010, pp. 1027–1033, doi:
10.1109/IPEC.2010.5542039.
4. K. J. Chen et al., “GaN-on-Si power technology: Devices and
applications,” IEEE Trans. Electron Devices, vol. 64, no. 3,
pp. 779–795,Mar. 2017, doi: 10.1109/TED.2017.2657579.
5. B. De Jaeger et al., “Au-free CMOS-compatible AlGaN/GaN HEMT
processing on 200 mm Si substrates,” in Proc. 24th Int. Symp. Power
Semiconductor Devices ICs, Bruges, Belgium, Jun. 2012, pp. 49–52,
doi: 10.1109/ISPSD.2012.6229020.
6. T.-L. Wu et al., “The impact of the gate dielectric quality in developing
au-free D-mode and E-mode recessed gate AlGaN/GaN transistors on
a 200 mm Si substrate,” in Proc. IEEE 27th Int. Symp. Power
Semiconductor Devices IC’s (ISPSD), Hong Kong, May 2015, pp.
225–228, doi:10.1109/ISPSD.2015.7123430.
7. P. J. Ballester and J. B. O. Mitchell, “A machine learning to predicting
protein–ligand binding affinity with applications to molecular
docking,” Bioinformatics, vol. 26, no. 9, pp. 1169–1175, May 2010,
doi:10.1093/bioinformatics/btq112.

24