Forging Global Frontiers through the Integration of AI Technology in Millimeter-Wave GaN HEMT Development

Kenya Nishiguchi

Deep Learning based Model for Millimeter Wave GaN HEMTs

Kenya Nishiguchi
Transmission Devices Laboratory

The Sumitomo Electric Group has spearheaded the global development of gallium nitride high-electron-mobility transistors, known as GaN HEMTs, as a cutting-edge next-generation transistor. These devices serve as critical components that support our daily lives in advanced technologies such as high-speed smartphone communications and automotive collision avoidance systems. In their research and development, where efficiency and speed are paramount, simulation models have been used to replicate device behavior. However, GaN HEMTs have been considered challenging to fully model due to their complex dynamics and characteristics.

Recently, we have made a significant breakthrough by harnessing the power of deep learning, commonly used in artificial intelligence (AI), to build artificial neural networks (ANNs). This approach enables us to meticulously reproduce the behavior of millimeter-wave GaN HEMTs, which has been notoriously difficult to model. Consequently, we can now create models that perfectly replicate experimental data without relying on the intuition and experience of engineers.

Millimeter-Wave GaN HEMTs, a Next-Generation Technology Expected in the Future Information Society

The HEMT is a semiconductor device that controls the flow of electricity. Leveraging its extensive transistor expertise, the Sumitomo Electric Group has spearheaded the development and commercialization of HEMTs, using gallium nitride with remarkable material properties, leading the market with pioneering innovations.

Now, the times demand even greater capacity in wireless communications, and the "millimeter-wave GaN HEMT" is the answer. Millimeter wave refers to radio waves in the frequency range of 30 GHz (gigahertz) to 300 GHz, providing a high-speed, high-capacity transmission band capable of meeting the demands of diverse fields such as 5G communication, autonomous driving, and space exploration. While millimeter-wave GaN HEMTs hold promise for advanced signal processing and high-speed data communication, their complex circuit design and analysis pose challenges. Therefore, our study delves into creating ANN models by incorporating deep learning through the application of AI.


Replicating Millimeter-Wave Behavior with Deep Learning

In a neural network, which mimics the nervous system of the human brain, a large number of neurons (nerve cells) are interconnected and transmit signals. ANN models incorporating these neurons self-learn from large data sets to predict optimal output for a given task.

However, the application of ANN models to the millimeter-wave band had remained unexplored due to the extraordinary complexity of millimeter waves, characterized by intricate waveforms. The challenge lies in the overfitting problem, where deep learning becomes ensnared in individual data points and fails to capture the precise trends of the entire data set. Thus, our ingenuity led to the conception of a technique that imposes constraints on predictive models. By applying the ANN exclusively to the current sources of electronic circuits, we successfully evaded overfitting and achieved a groundbreaking model that impeccably replicates the behavior of GaN HEMTs in the millimeter-wave band—a pioneering accomplishment in the world.

Furthermore, utilizing the developed ANN model, we prototyped millimeter-wave GaN amplifiers and meticulously verified their computational accuracy. The results demonstrated the precise predictive capabilities of our ANN model for millimeter-wave GaN HEMTs. This breakthrough enables quantitative analysis and sophisticated circuit design of high-frequency amplifiers, potentially contributing to their performance advancements in the future.

Kenya Nishiguchi

Highly Commended Research Achievement at International Conferences

This research was the world's first case of an ANN model in the ultra-high frequency band of millimeter waves, and the achievement has attracted widespread attention both in Japan and abroad. We were honored to have the opportunity to present at the International Microwave Symposium (IMS), a crucial platform for scholarly exploration of technology and trends, where numerous companies and experts in the telecommunications industry participate. The novelty and significance of our research and development were acknowledged at this prestigious stage.

Deep learning holds vast potential as a transformative technology, offering diverse applications limited only by the scope of ideas. Currently, I am dedicating efforts to AI-based automatic design of GaN HEMTs. My background lies in semiconductor research, and my journey into AI and programming involved a lot of trial and error, gathering information from books and the internet. It was through the incessant unraveling of matrix computations, integral to expressing deep learning, that I eventually grasped the true meaning of "AI learning." GaN HEMTs encompass many physical phenomena that have yet to be elucidated, often requiring research and development to rely on the intuition and experience of engineers. In such intricate and specialized domains, I firmly believe that enhancing the quality and quantity of experimental data for AI learning enables more objective analyses and precise designs. In the future, we intend to make our developed model available to customers to demonstrate the allure of our GaN HEMTs and help gain market share on a global scale.

Kenya Nishiguchi

Related links

・[Conference presentation] Neural Network based GaN HEMT Modelling for Millimeter Wave Power Amplifiers, IEEE/MTT-S International Microwave Symposium-IMS 2022

・[Conference presentation] 14th Topical Workshop on Heterostructure Microelectronics

・[Journal article] Journal of Applied Physics 132.17 (2022): 175302


A journal of technical papers explaining the Sumitomo Electric’s technologies. Technical papers are available in PDF format.

Learn more


Learn more