An all-magnonic neuron with tunable fading memory

Abstract

Magnonics offers nanometer-scale wave propagation and strong nonlinearities, making it attractive for neuromorphic applications such as artificial neurons. Yet, magnonic elements with interconnections solely within the magnonic system remain challenging, preventing the realization of interconnected magnonic neurons to date. Here, we experimentally demonstrate an all-magnonic neuron that reacts to magnon inputs with thresholded, amplified magnon firing and subsequent self-reset, enabling all-magnonic operation and cascading. Our approach is based on micro-antenna excitation on an ultra-low damping garnet with perpendicular magnetic anisotropy (PMA), where we exploit the positive magnon frequency shift to realize nonlinear activation. Using Brillouin light scattering spectroscopy, we uncover a transient neuron response with tunable fading memory: A 25% change in pump power results in a 3-order-of-magnitude tuning in memory time, which we harness, demonstrating temporal integration of up to 50 magnon pulses. Finally, we realize neuron triggering in a cascade of 3 neurons, highlighting its potential for connected magnonic circuits.

Figures (6)

• Figure 1: | Device layout. a Conceptual diagram of the neuron principle: The neuron is triggered when incoming magnon stimuli exceed the threshold given by its nonlinear activation function, leading to the emission of a magnon pulse, before self-resetting back to baseline. b Colorized SEM micrograph of the applied microstructure, consisting of a vertical CPW antenna, here functionally representing the neuron, and two diagonally placed CPW antennas used for input generation. The structures are placed on a Ga:YIG film, which is magnetized in-plane using a bias magnetic field of $\mu_{\mathrm{0}}H_{\mathrm{app}} \approx 86mT$ along the indicated direction. The emitted magnons are measured using a time-resolved BLS microscope, which is schematically depicted on the right.
• Figure 2: | Nonlinear excitation mechanism.a Left axis: Magnon dispersion relation $f_{\mathrm{k}} = \omega_{\mathrm{k}}/2\pi$ for vanishing amplitude $c_{\mathrm{k}} = 0.0$ (blue) and with nonlinear magnonic self- and cross-frequency shift for increased amplitude $|c_{\mathrm{k}}| = 0.34$ (red) of mode $k \approx 8.3rad\per µm$. Right axis: CPW excitation efficiency $\eta(k)$ as a function of the magnon wavevector $k$ (gray). b Nonlinear dependence of the BLS intensity $I_{\mathrm{BLS}}$ of the emitted magnons as a function of up- (blue, continuous) and downsweep (blue, dashed) of the excitation power $P_{\mathrm{N}}$. The results of the numerical model are shown for comparison (black, dashed).
• Figure 3: | Triggered neuron activation and decay.a Experimental schematic: Two RF-pulses of duration $\Delta\tau_{\mathrm{1}} = 15ns$ and $\Delta\tau_{\mathrm{N}} = 1µs$ and frequency $f_{\mathrm{1}} = 2.2GHz$ and $f_{\mathrm{N}} = 1.69GHz$ are generated and applied to input 1 and to the neuron, respectively. Note that $f_{\mathrm{1}}$ was optimized for high pulse quality (low pulse dispersion, high $v_{\mathrm{G}}$), but triggering with other frequencies is generally possible if enough positive nonlinear frequency shift can be induced, see also Fig. \ref{['fig:cascade']}. b Temporal evolution of the neuron output BLS intensity and numerical model predictions for fixed pump power $P_{\mathrm{N}}' = -0.45dB$ and different trigger powers $P_{\mathrm{1}}' = P_{\mathrm{1}} - P_{\mathrm{Trigger}}'$, where $P_{\mathrm{Trigger}}$ is the absolute trigger power yielding 50% activation (with regard to intensity maximum). c Neuron activation function: maximum neuron output BLS intensity vs maximum input BLS intensity by input 1 and numerical model for comparison. d Tunable neuron decay: Variation of the pump power $P_{\mathrm{N}}'$ for a fixed trigger power $P_{\mathrm{1}}' = 1.1dB$. e Neuron decay time extracted from exponential decay fits from panel (d) as a function of the pump power $P_{\mathrm{N}}$ in mW. f Change of threshold pump power $P_{\mathrm{N}}^{\mathrm{th}}$ for triggered activation at a fixed trigger power $P_{\mathrm{1}}' = 1.1dB$ as a function of the pump frequency $f_{\mathrm{N}}$.
• Figure 4: | Multi-input neuron triggering.a BLS intensity of the neuron, triggered by 10 consecutive 15-ns-long input pulses from input 1, each with a pulse spacing of $\Delta t = 80ns$ and frequency $f_{\mathrm{1}} = 2.2GHz$. Shown are different curves for varying pump powers $P_{\mathrm{N}}'$, at the pump frequency $f_{\mathrm{N}} = 1.69GHz$. The peak positions are indicated by a semi-transparent line as a guide to the eye. b BLS intensity of the neuron, triggered by both inputs 1 and 2, each sending an independent 15-ns-short trigger pulse with varying pulse delay between them, with frequencies $f_{\mathrm{1}} = f_{\mathrm{2}} = 2.2GHz$. Measured at the output side of the neuron with a pump power of $P_\mathrm{N}' = -0.7dB$ and pump frequency $f_{\mathrm{N}} = 1.69GHz$. Inset: Maximum neuron output as a function of the pulse delay $\Delta t_{\mathrm{1,2}}$ between the two pulses.
• Figure 5: | Cascaded neuron activation.a Colorized SEM micrograph of the employed nanostructure: Three CPW antennas are placed on top of a 1-$\upmu$m-wide Ga:YIG waveguide. The neurons N2 and N3 are driven at powers in the bistability power regime $P_{\mathrm{N2}} = -2dBm$ (compare Supplementary Fig. S6). BLS measurement is performed on the right side of N3, 'behind' the chain of neurons. b Cascaded neuron experiment, with one trigger pulse by N1 and two pump pulses by N2 and N3, showing the time-resolved BLS intensity for combinations of pump states (ON/OFF) (binarized in legend). The pulse timings are indicated by the shaded areas. Frequencies used: $f_{\mathrm{N1}} = f_{\mathrm{N2}} = f_{\mathrm{N3}} = 2.1GHz$. c Repetition of a similar experiment with a different pump frequency for the center neuron $f_{\mathrm{N2}} = 2.2GHz$, and $f_{\mathrm{N1}} = f_{\mathrm{N3}} = 2.0GHz$ yielding similar results as (b). Shown is the BLS intensity as a function of time and BLS frequency for the (1 1 1) case.