Physical reservoir computing (PRC), which utilizes physical systems satisfying three characteristics (i.e., “nonlinearity”, “high dimensionality” and “short-term memory”) required for information processing, is a promising method to realize neuromorphic computing. Recently, nonlinear interfered spin waves were shown numerically and experimentally to be feasible for high performance PRC [1,2]. Moreover, numerical validation had shown multi-detection is a promising method to improve the computational performance because of the diverse and complex dynamics achieved by the method [3]. However, while experimental validations have reported the computational performance with 2 detectors, they haven’t been reported with more detectors [2,4]. Hence, in this study, we experimentally developed multi-detection device up to 8 detectors to improve the computational performance.

We created 10 coplanar waveguides on the Y₃Fe₅O₁₂ (YIG) single crystal with various distances between each waveguide. Among 10 waveguides, 2 waveguides are used to excite spin waves, and the remaining 8 waveguides are used to detect the excited spin waves. Using 2 waveguides as exciter makes nonlinear interference of spin waves. We carried out the prediction task of the chaotic time series generated by Mackey-Glass equation to evaluate prediction accuracy of this system. In the PRC utilizing spin wave interference, computational performance is changed by the applied magnetic field and the interval of pulse input because the nonlinear interference in the YIG is generated differently. So, we applied the magnetic field perpendicularly from 164 mT to 204 mT in steps of 2 mT and set the pulse interval as 2, 5, 7, 10, 12, 15 ns, respectively, to search for the condition that has the lowest error. We used root mean square error (RMSE) as the evaluation index. In the result, we achieved the lowest RMSE, 1.63 × 10^-2, in this system when the magnetic field is 184 mT and the pulse interval is 15 ns with 8 detectors. This RMSE is the top-level performance of PRC (e.g., skyrmion [5] is 6.08 × 10^-2 , iono-magnonics [6] is 3.28 × 10^-2 and ion-gating reservoir [7] is 1.63 × 10^-2.). This result proves that in experimental systems multi-detection is an effective method for improving the performance through enhancing expressive power based on the three aforementioned characteristics of the reservoir. We will also discuss the reason for this enhancement by increasing the number of detectors in more detail at the conference.