A rare form of vanadium oxide could be the basis for a new generation of computers that mimic the human brain, researchers from the Electronic Materials Engineering (EME) department have found.
Trivanadium pentoxide (V3O5) has properties suited to making devices for neuromorphic computing, which is based on a network of artificial neurons and synapses.
Neuromorphic computers use an analogue architecture that can process complex real-world data more effectively and efficiently than conventional digital computers, said EME PhD student Sujan Das.
“For example, every human has a different face, so it’s hard to train a normal computing system to recognise faces,” Mr Das said.
“The qualities of neuromorphic computers would enable patterns such as faces to be recognised much more easily.”
Neuromorphic processors carry out computations more efficiently than conventional computers, thanks to a combination of parallel processing and distributed memory.
The challenge scientists have been working on is to create a ‘neuristor’, a device that mimics the neurons in our brain by responding in a nonlinear way to the amplitude and timing of input signals.
For their neuristor, the EME team chose V3O5, which transitions from insulating to metallic behaviour when heated. Voltage pulses applied to a V3O5-based device heat the material and cause a rapid change in the device’s resistance.
V3O5 is not the only oxide that exhibits this behaviour – many studies have looked at the more common vanadium dioxide (VO2), which has similar behaviour. However, it’s insulator-metal transition occurs at around 70 degrees Celsius, which is too low for reliable operation in high-density integrated circuits. Another candidate, niobium dioxide (NbO2), has an operating temperature of over 800 degrees, which is too high for energy-efficient operation.
“A phone using VO2 neuristors would not work on a sunny day in summer!” said Dr Sanjoy Nandi, a postdoctoral fellow in EME.
“We realised V3O5 had the goldilocks operating temperature of around 140 degrees Celsius.
“But it’s very challenging to make – you need to have very accurate control over the oxide composition.”
Fortunately the team were able to reach out to a research group in Puerto Rico, led by Professor Armando Ruà, who had recently developed methods to deposit high-quality V3O5 films. Working with this group, the EME team were able to fabricate and test their V3O5 neuristors.
The first task was to thoroughly study the rare oxide’s electrical, thermal and physical properties, said Mr Das.
“Because this form of vanadium oxide is rare, there is limited information on its properties – information that is critical for modelling the device’s behaviour and energy performance.”
“We were pleased that the properties of V3O5 were well suited for use as a neuristor. The devices were found to be extremely reliability and had impressive endurance, which is important for practical applications,” Mr Das said.
The EME team fabricated their neuristors from the V3O5 films by adding suitable metal electrodes, and successfully demonstrated the nonlinear behaviour required of a neuron.
“The simplicity of the device makes it very scalable.” said Dr Nandi.
They then added a second neuristor and explored the dynamics of the coupled devices, behaviour that underpins the operation of a neural network.
“The simplicity and scalability of these oxide-based devices provides a strong foundation for developing a new generation of computers based on neural networks,” Professor Elliman said.
“Such computers can learn in a way that conventional computers cannot and are much more tolerant of errors and incomplete data.”
“They have the potential to revolutionize the way we process and analyse information.”
The research is published in Advanced Materials.