Hybrid Magnonic Reservoir Computing

TL;DR

This work explores magnonic physical reservoir computing as a fast, energy-efficient alternative to traditional neural approaches. It extends the magnonic auto-oscillation ring (AOR) with an inline amplification neural network and introduces a parallel input scattering model (PSM) that uses spin-wave scattering to fuse multiple input channels. Across real-world and synthetic datasets, the AOR benefits from amplitude encoding and the inclusion of the ANN, achieving higher accuracy than baseline readouts, while the PSM demonstrates robust feature mixing and meaningful dimensional reduction, occasionally rivaling or surpassing small neural networks. The findings suggest magnonic RC designs can approach or exceed dense networks on certain tasks, with potential for scalable, low-power computing in future hardware implementations.

Abstract

Magnonic systems have been a major area of research interest due to their potential benefits in speed and lower power consumption compared to traditional computing. One particular area that they may be of advantage is as Physical Reservoir Computers in machine learning models. In this work, we build on an established design for using an Auto-Oscillation Ring as a reservoir computer by introducing a simple neural network midstream and introduce an additional design using a spin wave guide with a scattering regime for processing data with different types of inputs. We simulate these designs on the new micro magnetic simulation software, Magnum.np, and show that the designs are capable of performing on various real world data sets comparably or better than traditional dense neural networks.

Figures (7)

• Figure 1: Layout of the AOR design used. The data is encoded into the input signal which is then used to excite magnetostatic surface waves which propagate down the wave guide. The output antenna converts the spin wave back into an electronic signal which is then sent to 3 different channels: the readout, the feedback line, and the amplification neural network. The feedback line adds the output signal to the incoming signal at the input antenna. The amplification neural network takes the output from the previous interval and uses it to determine some additional amplification to the next intervals input signal.
• Figure 2: a) The result of the preprocessing on a sample of the stock data. b) A comparison of the output behavior to the input values. Output values normalized to be comparable to the input.
• Figure 3: Layout of the PSM design used. The input region is divided into 2 input channels. Spin waves are excited in the input channels propagating down to the scattering region where the two waves mix. The output region was divided into 1, 2, or 3 read out channels (2 readout channels shown here).
• Figure 4: a) Iris classes plotted by petal length and width. The Setosa and Veriscolor classes make up the easily separable set. The Veriscolor and Virginica classes make up the mostly separable set. b) The scaled readout from the 1 output channel simulation for two samples within the easily separable set.
• Figure 5: Benchmark results taken from IRIS with PSM results added for comparison. LM - Linear Mapping result (95.0%). NN - Neural Network result (98.3%).