Imagine highways that not only track traffic flows on a daily basis, but open new lanes for rush hour traffic. Tiny devices called memristors can control electrical signals flowing through them in the same way based on a ‘memory’ of previous electricity flows—opening new possibilities in the way we build computers.
Most computers today rely on shuttling data (as electrical signals) between two separate electronic units: one for memory, and one for processing. That ‘travel distance’ can make each calculation cost a little extra time and power, which adds up over millions of instances. Memristors both store and process data on a single unit, making them potentially faster, more energy-efficient building blocks for ‘brain-like’ (neuromorphic) supercomputers that train artificial intelligence (AI) systems on seas of data.
But today’s memristors don’t quite live up to their full potential just yet—the random, unpredictable movement of ions within them leads to inconsistent performances. This affects learning accuracy in neural networks and hinders implementation of hardware accelerators.
To address these limitations, a research team led by Kah-Wee Ang and Yong-Wei Zhang, Principal Scientists at A*STAR’s Institute of Materials Research and Engineering (IMRE) and Institute of High Performance Computing (IHPC) respectively, turned to ultrathin, two-dimensional (2D) palladium selenide (PdSe2) as a novel material foundation for next-generation memristors.
“We chose PdSe2 as it’s incredibly stable when exposed to air and allows layer-by-layer oxidation when exposed to ozone, which combine 2D materials and three-dimensional oxides to create ultrathin bilayer heterostructures.” said Ang and Zhang. “These properties give us the upper hand for tackling the scalability and variability issues that plague existing memristors.”
Working with colleagues from the National University of Singapore and the Wuhan University of Technology, China, the team designed and engineered a novel dual-layered memristor using PdSe2 and its oxide form (PdSeOx) through an innovative fabrication process involving ultraviolet light and ozone. The base layer, comprised of PdSeOx, is porous and flexible enough for a free flow of ions, while the second layer, comprised of PdSe2, is more compact to restrict ion movement significantly.
“This unique design allows us to control ion movement more effectively, making a more predictable and reliable memristor with enhanced computing performance,” the duo explained.
In a series of simulations, the team tested different configurations of their memristor in a convolutional image processing application. With a ‘crossbar kernel’ memristor configuration, the application correctly recognised images with about 94 percent accuracy. The memristor also enhanced the stability and accuracy of computing applications with a low variability in set and reset voltage values (4.8 and -3.6 percent, respectively.)
“We firmly believe that our findings—a reliable memristor design and an innovative fabrication approach—will significantly contribute to the evolution of neuromorphic computing hardware systems based on 2D materials,” said Ang and Zhang. “In addition, our memristor design shows promise in high-precision image recognition and convolutional image processing, which could advance non-von Neumann hardware accelerators.”
Looking ahead, the team plans to develop a scalable production method for PdSeOx/PdSe2 heterostructures, which will support the commercial-level fabrication of memristors based on their novel material.
The A*STAR-affiliated researchers contributing to this research are from the Institute of Materials Research and Engineering (IMRE) and the Institute of High Performance Computing (IHPC).
