Mechanistic enhancement of mapping in layered-semiconductor-based reservoir computing devices for neuromorphic computing

We investigate Bi2Se3-based reservoir computing devices fabricated via a rubbing-induced site-selective deposition process, focusing on the role of grid density in governing dynamical behavior and computational capability. Devices with identical footprints but different network structures—including a grid-free 1 × 3 channel and cross-linked 3 × 3, 6 × 6, and 9 × 9 grids—are systematically studied. Step-forward dynamical modeling confirms that all devices exhibit stable, state-dependent responses characteristic of physical reservoirs. Dynamic prediction results show that increasing grid density significantly improves prediction accuracy for Mackey–Glass dynamics, demonstrating enhanced computational capability. Additional analysis reveals that increasing the number of output terminals further improves prediction performance, indicating that both structural complexity and readout dimensionality contribute to reservoir effectiveness. These results establish that grid-density-controlled structural connectivity governs reservoir performance and provide practical design guidelines for layered-semiconductor neuromorphic hardware.