Experimental demonstration of a skyrmion-enhanced strain-mediated physical reservoir computing system

Yiming Sun, Tao Lin, Na Lei, Xing Chen, Wang Kang, Zhiyuan Zhao, Dahai Wei, Chao Chen, Simin Pang, Linglong Hu, Yang Liu, Enxuan Dong, Li Zhao, Lei Liu, Zhe Yuan, Aladin Ullrich, C. H. Back, Jun Zhang, Dong Pan, Jianhua Zhao, Ming Feng, A. Fert, Weisheng ZhaoNature Communications 2023 · DOI

TL;DR

Researchers have developed a novel physical reservoir computing system that leverages the unique properties of magnetic skyrmions and strain-induced electrical changes. This "skyrmion-enhanced" system demonstrates impressive performance in complex tasks like waveform classification and time series prediction, paving the way for ultra-low-power neuromorphic computing.

What Problem Does This Paper Address?

Reservoir computing (RC) offers a computationally efficient approach to processing sequential data, mimicking the recurrent neural networks (RNNs) but with a fixed, complex "reservoir" and only training the output layer. Physical implementations of RC are highly sought after for their potential to achieve high speeds and low power consumption, moving beyond traditional silicon-based computing. Spintronic and strain-mediated systems are particularly promising due to their non-volatility, nonlinear dynamics, and multi-parameter tunability. However, existing spintronic RC systems often rely on electrical currents or magnetic fields as inputs, which can be energy-intensive. This paper addresses the need for more energy-efficient and highly performant physical RC systems by exploring how magnetic skyrmions, combined with strain-mediated electrical control, can create a more powerful and efficient reservoir.

Methods

The core of this work is a strain-mediated physical reservoir built upon a multiferroic heterostructure: Pt/Co/Gd multilayers deposited on a piezoelectric PMN-PT(001) substrate. The key innovation lies in the "skyrmion-enhancement" achieved by simultaneously tuning magnetic skyrmions and electrical resistivity through applied electric fields.

Here's a breakdown of the methodology:

  • Strain-Mediated Control: An electric field (E-field) is applied across the PMN-PT substrate. This induces a piezoelectric strain, which in turn affects the magnetic properties (specifically, the magnetization, $M_z$) and the longitudinal resistivity ($p_{xx}$) of the Pt/Co/Gd layers.
  • Skyrmion Dynamics: Magnetic skyrmions, which are topologically stable magnetic whirls, are present in the Pt/Co/Gd multilayers. The applied strain influences the density, size, and shape of these skyrmions.
  • Anomalous Hall Effect (AHE) as Output: The system's output is measured via the anomalous Hall effect (AHE), which generates a Hall voltage ($V_{xy}$). This voltage is a function of both the applied strain (and thus the E-field) and the magnetic state, specifically incorporating contributions from the magnetization ($M_z$) and resistivity ($p_{xx}$). The relationship is described by: $p_{xy}(E) = R_0H_z + R_s4\pi M_z(E) + \alpha p_{xx}(E)$.
  • Reservoir Computing Framework: The device acts as the physical reservoir. Input signals are applied as voltage sequences to the PMN-PT substrate. The resulting $V_{xy}$ output signals are then processed by a trained linear output layer to perform classification and prediction tasks.
  • Tasks Performed:
    • Sequential Waveform Classification: The system is tested on its ability to recognize current, last, and second-to-last waveforms in a sequence of square and sine waves.
  • Mackey-Glass Time Series Prediction: The system is used to predict future values of a chaotic time series generated by the Mackey-Glass differential equation.

Datasets

The paper does not explicitly mention or introduce new benchmark datasets. Instead, it utilizes:

  • Synthetically generated input signals: Random sequences of square and sine waveforms for the classification task, and a Mackey-Glass chaotic time series for the prediction task.
  • Experimental output data: The anomalous Hall voltage ($V_{xy}$) generated by the physical reservoir in response to the input signals.

Key Insights and Ideas

  • Synergistic Enhancement: The combination of magnetic skyrmions and strain-induced electrical changes creates a richer, more nonlinear reservoir than either phenomenon alone. The skyrmion state, in particular, shows enhanced nonlinearity compared to a saturated magnetic state.
  • Strain-Tunable Skyrmion Properties: Applied electric fields can non-volatilely create, deform, and influence the density of skyrmions. This strain-mediated control is crucial for the reservoir's dynamic response.
  • Multiferroic Heterostructures as Ideal Reservoirs: The use of multiferroic materials like PMN-PT, which exhibit coupled ferroelectric and magnetic properties, is ideal for achieving multi-parameter fusion and efficient control.
  • Short-Term Memory Demonstrated: The system exhibits a clear short-term memory effect, evidenced by its ability to recognize the "last" waveform in a sequence, with the skyrmion-enhanced reservoir outperforming the saturated state.
  • Ultra-Low Power Potential: The estimated energy dissipation is on the sub-fJ scale, significantly lower than other spintronic reservoirs, highlighting the potential for energy-efficient neuromorphic computing.

Key Results

  • Waveform Classification: The skyrmion-enhanced reservoir achieved a recognition rate of 99.3% for the "last" waveform in a sequential classification task. This is a significant improvement over the saturation state.
  • Mackey-Glass Prediction: The system achieved a normalized root mean square error (NRMSE) of 0.2 for a 20-step prediction of the Mackey-Glass time series. An optimized output reconstruction method further improved accuracy.
  • Nonlinearity: The skyrmion state exhibits a more pronounced nonlinear response to applied E-fields compared to the saturated magnetic state, which is crucial for reservoir computing performance.
  • Energy Efficiency: Estimated energy dissipation is less than 1 fJ per waveform, orders of magnitude lower than other spintronic reservoirs.

Limitations and Future Work

  • Experimental Speed: The current experimental setup has a relatively slow input data acquisition rate (10 data points per second), which limits the overall speed of the reservoir.
  • Output Signal Amplitude: While the strain-mediated resistivity control helps increase the output signal amplitude for the Mackey-Glass task, the anomalous Hall voltage itself is typically in the µV range. The authors suggest using magnetic tunnel junctions (MTJs) as an alternative to Hall bars to magnify the output signal induced by skyrmion deformation.
  • Downscaled Devices: To improve performance for real-world applications, further efforts are needed to develop downscaled devices that can achieve higher operating speeds and computational capabilities.
  • Enhanced Memory and Speed: The authors suggest that involving faster magnetization dynamics could significantly enhance memory and speed capabilities.
  • Potential for Ultrafast Operation: The intrinsic ultrafast switching of ferroelectric and ferromagnetic domains in such heterostructures suggests the potential for gigahertz operation frequencies.

Summary updated

Abstract

Physical reservoirs holding intrinsic nonlinearity, high dimensionality, and memory effects have attracted considerable interest regarding solving complex tasks efficiently. Particularly, spintronic and strain-mediated electronic physical reservoirs are appealing due to their high speed, multi-parameter fusion and low power consumption. Here, we experimentally realize a skyrmion-enhanced strain-mediated physical reservoir in a multiferroic heterostructure of Pt/Co/Gd multilayers on (001)-oriented 0.7PbMg<sub>1/3</sub>Nb<sub>2/3</sub>O<sub>3</sub>-0.3PbTiO<sub>3</sub> (PMN-PT). The enhancement is coming from the fusion of magnetic skyrmions and electro resistivity tuned by strain simultaneously. The functionality of the strain-mediated RC system is successfully achieved via a sequential waveform classification task with the recognition rate of 99.3% for the last waveform, and a Mackey-Glass time series prediction task with normalized root mean square error (NRMSE) of 0.2 for a 20-step prediction. Our work lays the foundations for low-power neuromorphic computing systems with magneto-electro-ferroelastic tunability, representing a further step towards developing future strain-mediated spintronic applications.
BibTeX
@inproceedings{sun2023experimental,
  author = {Yiming Sun and Tao Lin and Na Lei and Xing Chen and Wang Kang and Zhiyuan Zhao and Dahai Wei and Chao Chen and Simin Pang and Linglong Hu and Yang Liu and Enxuan Dong and Li Zhao and Lei Liu and Zhe Yuan and Aladin Ullrich and C. H. Back and Jun Zhang and Dong Pan and Jianhua Zhao and Ming Feng and A. Fert and Weisheng Zhao},
  title = {{Experimental demonstration of a skyrmion-enhanced strain-mediated physical reservoir computing system}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-39207-9},
}
Figure