Energy-efficient Tuning of Spintronic Neurons
The human brain executes highly sophisticated tasks, such as image and speech recognition, with an exceptionally low energy budget and much more efficiently than any computer can. The development of energy-efficient and tunable artificial neurons capable of emulating brain-inspired processes has therefore been a major research goal for decades. Researchers at University of Gothenburg and Tohoku University now jointly report an important experimental advance in this direction, demonstrating a novel voltage-controlled spintronic microwave oscillator capable of closely imitating the non-linear oscillatory neural networks of the human brain.
Researchers at the University of Gothenburg and Tohoku University have developed a voltage-controlled spintronic microwave nano-oscillator, where both the microwave frequency and the threshold current can be strongly tuned, with negligible energy consumption, using a voltage-induced electric field. Says Himanshu Fulara, first author of the study: “This is an important breakthrough as these so-called spin Hall nano-oscillators (SHNOs) can act as interacting oscillator-based neurons but have previously lacked an energy-efficient tuning mechanism – an essential prerequisite to train oscillator-based neural networks for cognitive neuromorphic tasks. We believe that this novel mechanism will also allow for tuning of the synaptic interactions between each pair of spintronic neurons in larger and more complex oscillatory neural network.”
Mimics neuron interactions in our brains
Earlier this year, the Johan Åkerman group at University of Gothenburg announced the first successful demonstration of 2D mutually synchronized nano-oscillator arrays accommodating as many as 100 SHNOs while occupying an area of less than a square micron. These networks can mimic neuron interactions in our brain and carry out cognitive tasks, such as vowel recognition, with a network of four SHNOs capable of distinguishing and categorizing 20 different synchronization states at GHz frequencies, which is many orders of magnitude faster than our brain. However, a major bottleneck in training such artificial neurons, to produce different responses to different inputs, has been the lack of individual oscillator control inside such networks.
This time, the Åkerman group has teamed up with the Hideo Ohno and Shunsuke Fukami group at Tohoku University to develop a seemingly similar bow tie-shaped spin Hall nano-oscillator, but made from a different ultrathin W/CoFeB/MgO material stack and with the added functionality of a voltage controlled gate over the oscillating region. Using an effect called voltage-controlled magnetic anisotropy (VCMA), the magnetic and magnetodynamic properties of a few atomic layers thick CoFeB ferromagnet can be directly controlled to modify the microwave frequency, amplitude, damping, and hence the threshold current, of the SHNO. The researchers also found that the bow-tied geometry strongly enhances the effect of the VCMA such that the SHNO damping can be varied up to 42% using voltages from -3 to +1 V. The demonstrated energy-efficient approach is therefore capable of independently turning individual oscillators on/off within a large synchronized oscillatory network driven by a single global drive current. The findings should also attract fundamental scientific interest as a new mechanism of energy relaxation in patterned magnetic nanostructures has been revealed.
Hoping to scale up to larger network sizes
With readily available energy-efficient independent control of the dynamical state of individual spintronic neurons, the researchers hope to efficiently train large SHNO networks to carry out complex neuromorphic tasks and scale up oscillator-based neuromorphic computing schemes to much larger network sizes.
"These exciting results highlight the productive collaboration that we have established in neuromorphic spintronics between University of Gothenburg and Tohoku University", says Johan Åkerman, "a collaboration that will now intensify as Tohoku University has recently joined the Sweden-Japan collaborative network MIRAI 2.0, coordinated by University of Gothenburg and Nagoya University".
For further details, please see “Giant voltage-controlled modulation of spin Hall nano-oscillator damping”, H. Fulara, M. Zahedinejad, R. Khymyn, M. Dvornik, S. Fukami, S. Kanai, H. Ohno, and J. Åkerman, Nature Communications 11, 4006 (2020).
The earlier work on two-dimensional mutual synchronization was published by Zahedinejad et al, Nature Nanotechnology 15, 47–52 (2020).
Top: A scanning electron microscope (SEM) image of the gated spin Hall nano-oscillator device of width 150 nm.
Bottom: Schematic of the material stack displaying layer order along the gated electrode.