High-accuracy prediction for the NARMA task achieved by eight detectors in nonlinearly interfered spin waves-based physical reservoir computing

Abstract Physical reservoir computing enables highly efficient artificial intelligence hardware by replacing neural reservoirs with physical systems exhibiting nonlinearity, short-term memory, and high dimensionality. Spin-wave interference has been predicted to achieve high computational performance through multiple detectors, but performance beyond two channels has not been examined experimentally. In this study, we evaluate the nonlinear autoregressive moving average (NARMA) task performance of a spin-wave reservoir with up to eight detectors and quantify its short-term memory capacity (MC). The normalized mean square error by variance for the NARMA10 task decreased from 0.155 with two channels to 0.130 with eight detectors. Additionally, the MC increased monotonically as the number of detectors was expanded from 2 to 8, demonstrating the effectiveness of spatial multiplexing. These findings confirm that increasing the number of detectors effectively enhances computational capability and provides a promising route toward advanced spin-wave reservoir computing hardware.