Creating energy-efficient electronics using machine learning
USC Viterbi researchers have developed a neural network that can model a high-performing new material using machine learning (ML) techniques.
Materials researchers at the USC Viterbi School of Engineering have developed a new machine learning framework to study at an unprecedented scale how light can control materials. Typical simulations to understand light control of materials can usually simulate only a few hundred atoms, even with state-of-the-art computational resources, which seriously limits their applications. By harnessing the power of machine learning, USC Viterbi researchers were able perform simulations of light control of materials with over a billion atoms — 10 million times greater than conventional methods.
The research team used their machine learning model to perform large-scale simulation of light control of lead titanate, a special class of material, called a ferroelectric material, that has an inherent electronic polarization. The polarization can be thought of as a pattern of arrows in the material that can be controlled by stretching, heating and electricity, which makes the material ideal for use in sensors, energy storage, and memory. The research, from Senior Lecturer of Chemical Engineering and Materials Science Ken-ichi Nomura, Professor in Chemical Engineering and Materials Science Priya Vashishta, Professor of Computer Science Aiichiro Nakano, Professor of Physics Rajiv Kalia, and Ph.D. student Thomas Linker and their co-authors was recently published in Science Advances.
The researchers examined how the electronic polarization of the material lead titanate would respond to light. Recently this material has been gaining popularity because it allows researchers to create complex vortex-like patterns in its electronic polarization. When we think of a vortex, we may imagine a chaotic, swirling mass of matter or energy; however, these types of structures have been found to be very stable and efficient in these materials, which is why they are currently being investigated for next-generation energy storage and memory applications. The USC Viterbi researchers wanted to understand if these complex patterns could be controlled with light.
“We wanted to view these large-scale structures with highly accurate simulation methods that use things like quantum mechanics,” Linker said. “But that’s really difficult and very expensive, so we developed a multi-scaled framework where we train a machine learning model to learn a simpler representation of the light matter interaction. Thus, we can simulate much faster.”
“Without machine learning, it would have been impossible to design this kind of simulation,” Nomura said. “By training the machine learning model to learn how the material behaves in response to a strong laser, we can perform our simulation on supercomputers.”
With their framework, the researchers found a new type of phase that was induced by light matter interaction in lead titanate. “If we shine laser (light), we can create a string pattern in the polarization that is topologically different than the original vortex pattern,” said Nomura.
The research team said their machine learning framework offers an exciting new avenue for exploring light control of materials that was not previously possible.
– Edited from a USC Viterbi press release by CFE Media and Technology.