Memcapacitive-Dominated Reservoir Computing with Self-Organizing Nanowire Networks.

Nanowire networks have emerged as a promising platform for physical reservoir computing. However, networks based on memristive switching exhibit inherent limitations in dynamical richness and short-term memory capacity. Memcapacitors are expected to overcome these dynamical constraints through their topology-dependent dynamics. In this work, inspired by the distributed processing of neural systems, we endow Ag nanowire networks with memcapacitive properties via a temperature-modulated fabrication process, realizing a multi-path channel-dependent reservoir system with richer reservoir dynamics and establishing its circuit model. Benefiting from its memcapacitive nature, the system exhibits channel-specific responses and approximate linear superposition, enabling parallel and spontaneous fusion of multi-input signals with controllable and predictable behavior. This allows input states to expand exponentially with channel count. Leveraging this capability, we designed a multi-threshold encoding scheme that generates 256 highly separable states within four time steps while consuming only 18.98 pJ/pulse. This approach reduces the information loss inherent in conventional binarization and delivers over 94% accuracy on a compressed MNIST recognition task, demonstrating a potent strategy for high-performance neuromorphic edge computing.