Source-linked AI summary
Magnetic skyrmion artificial synapse for neuromorphic computing
Kyung Mee Song, Jae-Seung Jeong, Biao Pan, Xichao Zhang, Jing Xia, Sun Kyung Cha, Tae-Eon Park, Kwangsu Kim, Simone Finizio, Joerg Raabe, Joonyeon Chang, Yan Zhou, Weisheng Zhao, Wang Kang, Hyunsu Ju, Seonghoon Woo
TL;DR
A fully electrical skyrmion device could provide a building block for low-power neuromorphic computing, but integrating writing, motion, detection, and deletion remained an open challenge. The paper demonstrates a room-temperature ferrimagnetic skyrmion artificial synapse and evaluates it in pattern-recognition simulations, reaching approximately 89% accuracy.
Problem
A fully electrical skyrmion device integrating creation, motion, detection, and deletion had remained elusive despite separate demonstrations of these operations.
Method
The paper combines current-induced skyrmion operations in a single ferrimagnetic device with STXM imaging, electrical resistance measurements, and neuromorphic simulations.
Results
~89% learning accuracy was achieved in supervised handwritten-pattern recognition simulations using skyrmion synapses.
Takeaways & Limitations
The demonstrated skyrmion artificial synapse provides a fully electrical building block for spintronics-based bio-inspired computing.
Takeaways & Limitations
Resistivity changes slightly deviate from skyrmion-number changes at highly populated states, attributed to statistical skyrmion-diameter variation.
Abstract
from arXiv · showhide
Since the experimental discovery of magnetic skyrmions achieved one decade ago, there have been significant efforts to bring the virtual particles into all-electrical fully functional devices, inspired by their fascinating physical and topological properties suitable for future low-power electronics. Here, we experimentally demonstrate such a device: electrically-operating skyrmion-based artificial synaptic device designed for neuromorphic computing. We present that controlled current-induced creation, motion, detection and deletion of skyrmions in ferrimagnetic multilayers can be harnessed in a single device at room temperature to imitate the behaviors of biological synapses. Using simulations, we demonstrate that such skyrmion-based synapses could be used to perform neuromorphic pattern-recognition computing using handwritten recognition data set, reaching to the accuracy of ~89 percents, comparable to the software-based training accuracy of ~94 percents. Chip-level simulation then highlights the potential of skyrmion synapse compared to existing technologies. Our findings experimentally illustrate the basic concepts of skyrmion-based fully functional electronic devices while providing a new building block in the emerging field of spintronics-based bio-inspired computing.
I. INTRODUCTION
Magnetic skyrmions offer a potential basis for energy-efficient devices, but a fully electrical device integrating their key operations had remained elusive. This work introduces a room-temperature skyrmion artificial synapse for neuromorphic computing.
- Magnetic skyrmions are topologically nontrivial spin textures proposed for energy-efficient data storage, processing, and transmission.
- The work presents the first single device in which skyrmions are written, driven, read, and erased electrically at room temperature.
- The device is designed as an artificial synapse for neuromorphic computing, where analog memory supports synaptic operations.
- Skyrmions’ rigid-body, particle-like behavior allows multiple nanoscale skyrmions to accumulate within a defined area without interacting with topographic defects.
- The work contributes to spintronics-based bio-inspired computing alongside magnetic artificial neural networks and probabilistic skyrmion architectures.
A. STXM imaging on ferrimagnetic heterostructures
The study combines STXM imaging with electrical measurements to characterize ferrimagnetic multilayers supporting room-temperature skyrmions. Hall resistance tracks magnetic-domain configurations and varies linearly with applied magnetic field.
- The Pt/GdFeCo/MgO ferrimagnetic multilayer has perpendicular magnetic anisotropy and stabilizes approximately 200 nm skyrmions at room temperature.
- 0.98 mJ m-2 is the reported DMI magnitude for the asymmetric ferrimagnetic multilayer stack.
- The experiment alternates electrical pulse application with resistance measurement while recording magnetic-domain images after each pulse using STXM.
- Hall resistance and magnetic-domain configurations are measured as functions of out-of-plane magnetic field Bz.
- Both normalized Hall resistance and normalized out-of-plane Fe magnetization vary linearly with applied magnetic field.
- Topological Hall contributions can be suppressed by destructive contributions from the antiferromagnetically coupled ferrimagnetic sublattices.
B. Electrical operation of magnetic skyrmion artificial synapse
The device uses current-controlled skyrmion accumulation and dissipation to modulate synaptic weights, producing analog potentiation and depression through electrical operation. Measurements show multiple resistive states, while deviations from ideal linearity emerge at highly populated states.
- Synaptic weights are proportional to the number of skyrmions accumulated or dissipated within the active device area.This current-controlled population change imitates potentiation and depression.
- 24 distinct resistivity states were measured during potentiation as current pulses generated and moved skyrmions across the Hall cross.The pulses used 1.69 × 10^10 to 4.24 × 10^10 A m^-2 current densities and 100 ns duration.
- 8.1 ± 1.7 nΩ cm was the average resistivity decrease per added skyrmion during potentiation.
- 41 accumulated skyrmions produced a final resistivity change of Δρxy,f = -3.1 × 10^2 ± 4.1 nΩ cm.
- 16 different states were generated during depression, with an average resistivity change of |Δρxy,sk-d| = 8.9 ± 0.1 nΩ cm per annihilated skyrmion.Reversed pulses systematically annihilated skyrmions and increased resistance.
- At highly populated states, resistivity changes deviate slightly from skyrmion-number changes because skyrmion diameters vary with local defects and mutual interactions.The authors suggest material optimization could improve linearity.
C. Neuromorphic pattern-recognition simulation using skyrmion synapse
The study evaluates skyrmion synapses in multilayer neural-network and chip-level simulations for handwritten pattern recognition. The simulations use experimentally measured device characteristics and model circuit architectures for neuromorphic computing.
- The MNIST simulation uses a three-layer network with 784 input neurons, 100 hidden neurons, and 10 output neurons.
- The chip-level simulation models skyrmion synapse arrays in a 2T1R configuration within tiles, processing elements, buffers, accumulation units, activation units, and pooling units.The architecture represents a multi-level hardware hierarchy.
- ~10 nm-size skyrmions and magnetic-tunnel-junction reading schemes are identified as routes to improve scalability and electrical characteristics.
A. Potential of magnetic skyrmion artificial synapse and neural network
Chip-level simulations indicate that skyrmion-synapse ANNs can offer energy-efficiency and device-operation advantages relative to RRAM, while retaining comparable circuit-level performance. Their potential is balanced by lower on/off ratio and larger cell area.
- Chip-level comparison: Comparable circuit-level performance and improved energy efficiency arise from the skyrmion ANN’s reduced operation voltage relative to RRAM.Skyrmion synapses use few-millivolt, sub-ns weight-update pulses, whereas RRAM generally requires over 1 V for several nanoseconds.
- Device characteristics: The skyrmion synapse has a reduced on/off ratio and larger bit-cell and array area than the RRAM counterpart.The reduced on/off ratio increases dynamic-reading energy, while multiple skyrmions and the 2T1R structure increase area.
- Device characteristics: Skyrmion synapses may provide good linear weight distributions because synaptic weight directly depends on the number of accumulated skyrmions.The cited passage contrasts this with linearity challenges in RRAM and PCM switching mechanisms.
- Device characteristics: Skyrmion-synapse weight variation can be controlled through skyrmion number, unlike RRAM/PCM variations associated with temporal resistance formation.
IV. CONCLUSIONS
The study demonstrates a fully functional, all-electrical skyrmion artificial synapse and evaluates its use for supervised neuromorphic pattern recognition. The device reaches approximately 89% learning accuracy in simulation.
- IV. CONCLUSIONS: Current-induced skyrmion creation, motion, detection and deletion are integrated into a single ferrimagnetic artificial synapse.
- IV. CONCLUSIONS: Synaptic weights are modulated by the number of skyrmions in the active device area.
- IV. CONCLUSIONS: ~89% learning accuracy is reached in supervised pattern-recognition simulation using the skyrmion synapse.
- IV. CONCLUSIONS: The authors identify potential applications beyond multi-level synaptic memory, including racetrack memory, logic, magnonic and radio-frequency devices.
V. METHODS
The methods combine ferrimagnetic multilayer fabrication, simultaneous electrical and STXM measurements, micromagnetic modeling, and ANN and circuit simulations.
- Material growth and device fabrication: Pt/GdFeCo/MgO films were fabricated as Hall bars with electrical contacts for current-pulse experiments.The films were deposited by DC magnetron sputtering, and the Hall bar had a 4.5 µm stripe width.
- STXM measurements: STXM with XMCD at the Fe L3 absorption edge imaged out-of-plane Fe magnetization during simultaneous electrical measurements.Dark and bright contrasts represented upward and downward Fe magnetization, respectively.
- MNIST pattern recognition simulation: Experimentally measured resistance levels and error ranges supplied conductance means and deviations for ANN weights using selected potentiation and depression states.
- Skyrmion synapse simulation: OOMMF micromagnetic simulations used the LLG equation augmented by antidamping-like spin-transfer torque from the spin Hall effect.The model describes time-dependent spin dynamics in the skyrmion synapse.
- Skyrmion synapse simulation: Potentiation simulations applied positive current pulses to nucleate and drive skyrmions into the chamber, while depression reversed the current polarity to remove them.A localized out-of-plane field was used numerically to promote nucleation.
- ANN and circuit simulation: NeuroSim+ simulations modeled MNIST recognition with 1024 input neurons, 100 hidden neurons and 10 output neurons, alongside 32 × 32 circuit subarrays.