betaFeedback

WorksEngineering

Convolutions with Radio-Frequency Spin-Diodes

Abstract

The classification of radio-frequency (RF) signals is crucial for applications in robotics, traffic control, and medical devices. Spintronic devices, which respond to RF signals via ferromagnetic resonance, offer a promising solution. Recent studies have shown that a neural network of nanoscale magnetic tunnel junctions can classify RF signals without digitization. However, the complexity of these junctions poses challenges for rapid scaling. In this work, we demonstrate that simple spintronic devices, known as metallic spin-diodes, can effectively perform RF classification. These devices consist of NiFe/Pt bilayers and can implement weighted sums of RF inputs. We experimentally show that chains of four spin-diodes can execute 2x2 pixel filters, achieving high-quality convolutions on the Fashion-MNIST dataset. Integrating the hardware spin-diodes in a software network, we achieve a top-1 accuracy of 88 % on the first 100 images, compared to 88.4 % for full software with noise, and 90 % without noise.

Introduction

Fully spintronic neuronal networks using nanoscale magnetic tunnel junctions (MTJs) can classify non-linearly separable radio-frequency (RF) inputs without pre-processing or digitization [1]. Such a system is predicted to significantly improve energy efficiency, consuming about 10 fJ per synapse and 100 fJ per neuron with MTJs miniaturized to 20 nm—over 100 times more efficient than GPUs. The frequency multiplexing strategy used for neuron-to-synapse communication supports high connectivity, up to 500 synapses per neuron, and is free of sneak paths. However, current hardware implementations are limited to two synapses per neuron due to the complexity of MTJ stacks. To address these challenges, simpler materials can be used, maintaining the principles of RF interconnects and frequency multiplexing. In this work, we demonstrate that two-layer metallic spin-diodes [2] can perform high-quality weighted sums, enabling convolutional filters on the Fashion-MNIST dataset [3].

Metallic spin-diode synapses

The samples studied, shown in Figure 1(a), consist of a 5nm nickel iron bilayer capped with 5nm platinum on a high-resistance silicon wafer. Figure 1(b) and (c) display a typical device contacted with gold electrodes, measuring 5 by 10 micrometers. When an RF current is injected into these stripes, they generate a rectified DC voltage VDCV_{DC} due to the mixing of the input with the anisotropic magneto-resistance (AMR) response, which can be calculated analytically [4].

Figure 1. Spin diode stack (a) and geometry (b)(c).

Figure 2(a) shows the analytically calculated variations of VDCV_{DC} as a function of the applied magnetic field amplitude HH along the z-axis for a fixed frequency RF current. The applied magnetic field monotonically varies the device’s resonant frequency, fres=fres(H)f_{res}=f_{res}(H), from f1f_{1} to 4 GHz here. The voltage is positive when fres(H)<fRFf_{res}(H)<f_{RF} and negative when fres(H)>fRFf_{res}(H)>f_{RF}. For small differences between fres(H)f_{res}(H) and fRFf_{RF}, the voltage simplifies to: VDC∝(fres(H)−fRF)sin⁡(2Φ0)PRF∝WPRFV_{DC} \propto(f_{res}(H)-f_{RF})\sin(2\Phi_{0})P_{RF} \propto WP_{RF}, with PRFP_{RF} the power of the injected current.

Figure 2. (a) Theoretical spin diode voltage while varying the magnetic field, (b) synaptic behavior.

In our design (Figure 1(c)), we set the angle Φ0\Phi_{0} between the stripe and the applied magnetic field to 45° to maximize VDCV_{DC}. Figure 2(b) shows the calculated VDCV_{DC} as a function of PRFP_{RF} for different magnetic field values. The RF power of the injected current (PRFP_{RF}) is the synapse input, and VDCV_{DC} is the output. The weight WW, corresponding to the slope in Figure 2(b), is tuned by the applied magnetic field.

To perform the weighted sum, we use three methods illustrated in Figure 3. First, we electrically connect several devices to sum their DC voltages. Second, we send RF inputs (PRFiP_{RFi}) at different frequencies: IRF(t)=∑iIRFiI_{RF}(t)=\sum_i I_{RFi}. Third, we apply different magnetic fields to each device, giving each a unique resonant frequency fres(Hi)f_{res}(H_i) close to fRFif_{RFi}. This frequency multiplexing technique implements a weighted sum, with, in the linear regime, VDC∝∑i(fres(Hi)−fRFi)sin⁡(2Φ0)PRF∝∑iWiPRFiV_{DC} \propto\sum_i(f_{res}(H_i)-f_{RFi})\sin(2\Phi_{0})P_{RF} \propto\sum_i W_iP_{RFi}.

Figure 3. Implementation of a spintronic weighted sum performed on frequency multiplexed inputs

Figure 4(a) shows the experimental implementation of a spatial field gradient along the z direction with four permanent magnets. This gradient links the synaptic weight of each device to its position along the z-axis. Figure 5 shows the field components HxH_x, HyH_y, and HzH_z obtained by spatially scanning a 3D Hall probe in the field gradient with a motorized stage, achieving large variations in HzH_z over +/- 10 mT along the z-axis while keeping small variations in other directions.

Figure 4. Experimental measurement setup

Measurement of the magnetic field created with the gradient magnets

Figure 5. Measurement of the magnetic field created with the gradient magnets

Figure 6(a) displays the experimentally measured VDCV_{DC} response of a single device as a function of the z position in the field gradient. Figure 6(b) shows the VDCV_{DC} vs. PRFP_{RF} curves for different z positions, qualitatively aligned with the analytical calculations in Figure 2. The large background signal outside the linear operation region (highlighted in yellow) is problematic for devices in a chain due to overlaps.

Experimental spin diode voltage versus z position

Figure 6. (a) Experimental spin diode voltage versus the zz position in the field gradient, (b) synaptic behavior.

We design differential RF synapses to cancel these backgrounds, leaving a strong response in the desired region only. As shown in Figure 7, such synapse comprises two NiFe/Pt stripes connected in parallel at +45° and −45° to the magnetic field. This angle difference produces opposite VDCV_{DC} signals. Figure 8(b) shows that when fabricated at the same z position, the overall VDCV_{DC} cancels out over a wide range of RF inputs. When slightly displaced by δz\delta_z (40 micrometers in Figure 8(a) and (c)), the signal response is strong within a narrow region, demonstrating frequency selectivity. The peak amplitude is proportional to δz\delta_z and PRFP_{RF}, making the synaptic weight controllable by the shift between devices (Figure 8(d)).

Double diode synapse electric scheme, geometry, and experimental realisation

Figure 7. Double diode synapse electric scheme (a), geometry for different δz\delta z (b) and experimental realisation (c).

Double diode synapse spin diode voltage versus frequency

Figure 8. Double diode synapse spin diode voltage versus frequency for different δz\delta z (a)(b)(c) and associated synaptic behavior (d).

To predict weight values before lithography, we perform a calibration step. We record and average the response of three identical 45° stripes for various fRFf_{RF} and zz positions, obtaining a reference Vref(f,z)V_{ref}(f,z) (Figure 9(a)). The differential pair’s synaptic response is predicted by Vsynapse(f,z)=0.5×(Vref(f,z−δz/2)+Vref(f,z+δz/2))V_{synapse}(f,z) = 0.5 \times(V_{ref}(f,z-\delta_z/2) + V_{ref}(f,z+\delta_z/2)). This response is shown in Figure 9(b) for fixed PRFP_{RF} and δz\delta_z, with peak values decreasing with input frequency. This effect is accounted for by rescaling the injected power: PRF(cal)=PRF×(−40×3.1(GHz)+150)/(−40×fRF(GHz)+150)P_{RF}(\mathrm{cal}) = P_{RF} \times(-40 \times3.1(\mathrm{GHz}) + 150)/(-40 \times f_{RF}(\mathrm{GHz}) + 150).

Calibration diode and spin diode voltage amplitude

Figure 9. (a) Calibration diode. (b) Spin diode voltage amplitude versus frequency, (c) linear fit of this dependance.

For weighted sum operations, the device chains must have an impedance matched to 50 Ω\Omega RF sources for effective signal propagation. Various architectures are possible: series (Figure 10(a)), parallel (Figure 10(b)), or hybrid (Figure 10(c)). We choose the hybrid configuration because the total chain resistance does not depend on the number of devices it contains, which will enable the implementation of larger kernels (of size >9×9> 9 \times9) in the future. Examples of fabricated devices with a target resistance of 50 Ω\Omega are shown in Figure 11(a) and (b).

Different synapse chaining

Figure 10. Different synapse chaining, in parallel (a), series (b) and mixed (c).

Chain of 4 synapses after fabrication, with a zoom on a differential pair of diodes

Figure 11. Chain of 4 synapses after fabrication (a), and zoom on a synapse (b).

Spin-diode convolutions

We use three chains of four differential synapses to implement three 2×22 \times2 convolutional filters on the Fashion MNIST dataset, as illustrated in Figure 12. The twelve weights are precalculated by training a software neural network, as shown in Figure 13. The network consists of a convolutional layer with three 2×22 \times2 filters (stride and padding of 1), followed by max pooling (kernel size and stride of 2), ReLU neurons, a fully connected layer, and a softmax function for classification. During training, Gaussian noise is added to the outputs of the convolutional layer to ensure noise-resistant weights. The test accuracy reaches 87.63% without noise and 86.34% with 50% noise, indicating a small accuracy loss of 1.3%.

Implementation of a convolutional layer with 3 chains of spintronic synapses

Figure 12. Implementation of a convolutional layer with 3 chains of spintronic synapses.

Figure 13. Convolutional network for FashionMNIST classification.

After training, we implement the learned weights of the first convolutional layer in hardware through lithography. The input frequencies are set to 1.75, 2.25, 2.75, and 3.25 GHz. Based on the interpolation of a single diode’s resonance frequency vs. its position along the zz-axis, synapses are spaced with 175, 216, and 245 micrometers to center their frequency peaks at these values. Figure 14 shows the targeted weights, corresponding δz\delta_z values, and measured weights after fabrication. Discrepancies arise from resolution limitations in photolithography leading to unwanted variations in width and position of the diodes. Figure 15 compares the predicted profile (red) from reference diode interpolation and the experimental profile (blue) of fabricated samples. Figure 16 compares the experimental versus predicted operations for each chain.

Theoretical and experimental weights for the 3 chains

Figure 14. Theoretical weights, associated δz\delta z and obtained experimental weight values for the 3 chains.

Theoretical and experimental spin diode voltages versus frequency for the 3 chains

Figure 15. Theoretical (red) and experimental (blue) spin diode voltages versus frequency for the 3 chains.

Figure 16. Comparison between experimental and theoretical weighted sum, voltage values have been converted to pixel values.

To perform the convolution, each 2×22 \times2 region is flattened into a 4-pixel vector. Each pixel is converted to an RF signal, with its frequency encoding the pixel location and power encoding the pixel value (0 to 255). The four RF inputs are combined and sent to three chains of four synapses, each implementing a different convolutional filter. The output voltages are converted back to pixel values to form the three output images. To scan the entire input image, the kernels are slid with a stride of 2.

The experimental network reached an accuracy of 88% on the first 100 images, close to the 88.4% accuracy of the noisy software model, and slightly lower than the 90% accuracy of the model without noise. The accuracy of the noisy software model was determined by passing each of the 100 input images through the network 1000 times with varying random noise.

The confusion matrices for the noisy software model and the experimental model (Figure 17(a) and (b)) show similar misclassifications of pullovers, shirts, and coats, highlighting the comparable accuracy of both models for these relatively similar types of clothes. Figure 18(a) and (b) show an example image after convolution in (a) software and (b) hardware. The difference (c) is close to zero, demonstrating the high quality of the hardware convolutions with RF metallic spin-diodes.

Figure 17. Confusion matrices for the pure software network and the hybrid hardware-software network on the 100 first images of the dataset.

Figure 18. Shoe image convolved by the software network (a) and by the hardware-software network (b), (c) displays the difference between the two images.

Conclusion

This work demonstrated the use of NiFe/Pt bilayer spin-diodes for high-quality RF signal classification. By employing chains of differential RF synapses, we implemented convolutional filters on the Fashion MNIST dataset, achieving a top-1 accuracy of 88% on the first 100 images. This performance closely matches the 88.4% accuracy of the noisy software model and 90% without noise.

This study underscores the promise of spintronic hardware for next-generation neural networks, offering significant energy efficiency and scalability advantages. Future work will focus on non-volatile frequency control of diodes through magneto-ionic effects [5], removing the need of field gradients, and unlocking on-chip training. Miniaturization of synapses down to 10 nm, and their integration in large networks for more complex applications, will pave the way for efficient and scalable RF signal classification systems.

Acknowledgment

This work received funding by the European Research Council advanced grant GrenaDyn (reference: 101020684).

References

  1. [1]A. Ross, N. Leroux, A. De Riz, D. Marković, D. Sanz-Hernández, J. Trastoy, P. Bortolotti, D. Querlioz, L. Martins, L. Benetti et al., “Multilayer spintronic neural networks with radiofrequency connections,” Nature Nanotechnology, vol. 18, no. 11, pp. 1273–1280, 2023.DOI
  2. [2]G. Finocchio, R. Tomasello, B. Fang, A. Giordano, V. Puliafito, M. Carpentieri, and Z. Zeng, “Perspectives on spintronic diodes,” Applied Physics Letters, vol. 118, no. 16, 2021.DOI
  3. [3]H. Xiao, K. Rasul, and R. Wolfgram, “Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms,” 2017.arxiv.org/abs/1708.07747
  4. [4]A. Azevedo, L. Vilela-Leão, R. Rodríguez-Suárez, A. Lacerda Santos, and S. Rezende, “Spin pumping and anisotropic magnetoresistance voltages in magnetic bilayers: Theory and experiment,” Physical Review B—Condensed Matter and Materials Physics, vol. 83, no. 14, p. 144402, 2011.DOI
  5. [5]M. Khademi, A. Kumar, M. Rajabali, S. P. Dash, and J. Åkerman, “Large non-volatile frequency tuning of spin hall nano-oscillators using circular memristive nano-gates,” IEEE Electron Device Letters, 2023.DOI

Paper details

vlnc:W_yzscuywc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents