To power homes and other buildings across large land areas, electrical energy is carried on high-voltage transmission lines as alternating current (AC), in which electrons periodically switch direction. A transformer steps down the high voltage in power lines to allow electricity to be safely used in people’s homes. The resulting power from home outlets is AC, but most electronics use direct current (DC) electric power, in which current flows in one direction. This requires using a converter, a device or component within electronic equipment that converts AC to DC. In contrast, an inverter converts DC back to AC.
Researchers at Princeton University have been closely looking at power converters, inverters and transformers, focusing especially on the magnetic components within these devices. These components are often the largest and least efficient parts of a power electronics system, since electrical energy is lost to heat in the magnetic core (“core loss”). Developing the next generation of efficient power magnetics is paramount because of their central importance in a wide range of applications and fields, such as renewable energy, computer chips, automobiles, medical equipment, telecommunication, electro-manufacturing and beyond, according to Minjie Chen, assistant professor of electrical and computer engineering and the Andlinger Center for Energy and the Environment, who is part of a group of researchers studying power magnetics.
Designing efficient power magnetics is difficult because of the numerous electrical, mechanical and thermal factors that govern performance. Also, measuring the core loss within these parts is complex because the loss is typically non-linear. To address this multi-pronged challenge, Chen and his fellow researchers at Princeton and Dartmouth College have developed MagNet, “a large-scale open-source database designed to enable researchers modeling power magnetic material characteristics using machine learning to accelerate the design process of power electronics.”
MagNet has a vast array of data on various magnetic components and their material characteristics. Researchers can use this data and deep machine learning networks to design more efficient magnetic components and to model magnetic materials more accurately than existing computational methods.
This project was supported by Princeton’s Schmidt DataX Fund, which aims to spread and deepen the use of artificial intelligence and machine learning across campus with the aim of accelerating discovery. MagNet won DataX funding in 2021, along with nine other projects. Besides Chen, the principal investigators are Professor Niraj Jha and Assistant Professor Yuxin Chen in the Department of Electrical and Computer Engineering.
In addition to the researchers, Vineet Bansal, a senior research software engineer, who is jointly appointed to the Center for Statistics and Machine Learning (CSML) and the Princeton Institute for Computational Science and Engineering, was a key member of MagNet’s development. According to team members, he was instrumental in deploying MagNet as a web application and its setup on Princeton servers.
MagNet was released to the public in December 2022. The researchers are also launching an international power magnetics modeling competition, the MagNet Challenge, in 2023, based on the MagNet database.
“This is an exciting and fitting project for DataX because it is interdisciplinary in approach and tackles a problem that is difficult to model using older, conventional methods,” said Peter Ramadge, director of CSML, which oversees parts of the DataX Fund. “It utilizes machine learning and neural networks to efficiently and accurately approximate complex calculations and better understand magnetic components for future applications.”
Chen said deciphering the behavior of power magnetics is important because they are the central piece of advanced power electronics.
“The future world needs smarter and more efficient power electronics as many policymakers push for zero carbon emissions. Advanced power electronics will enable the modernization of the grid, upgrade transportation systems, and transform renewable energy, microelectronics, medical devices, robotics, and telecom and data communication systems,” said Chen.
Haoran Li, a doctoral student in electrical and computer engineering who kickstarted the project, said this is the first machine learning-driven large-scale dataset for power magnetics. Li likened MagNet to ImageNet, a well-known massive computer vision research database initiated at Princeton.
“The problem in power magnetics research is that everyone has their own datasets, and they only use small amounts of data to validate their own models. There wasn’t such a thing as a large, wide open-source dataset, and MagNet addresses this need,” said Li.
Research from MagNet has already yielded several research and conference papers, including interest from industry and researchers beyond Princeton’s campus.
“It’s a very useful program because, as designers, we want an accurate prediction of what’s happening within the magnetic material,” said Shukai Wang, a doctoral student in the electrical and computer engineering department and one of the contributors to MagNet. “It becomes critical to model the behavior accurately, especially if you put these power magnetics systems in space-constrained laptops, airplanes or cars, where size, weight and efficiency are all important.”
To access MagNet, click on this link.