A Memristor-Based Saturation-Constrained Information Capacity Framework for Temporal Information Processing.

Physical reservoir computing (PRC) has emerged as a highly energy-efficient paradigm for processing high-dimensional spatiotemporal data in edge-intelligent systems. However, the deployment of PRC is often limited by the quantization mismatch between continuous input signals and the finite dynamic range of hardware substrates. Conventional binary time-multiplexing ensures high noise immunity but necessitates protracted pulse sequences that induce high latency and premature physical saturation. Conversely, multi-valued analog encoding enhances information density but frequently induces state overlap, operating devoid of physical quantization boundaries. To resolve this physical-information trade-off, we propose a saturation-constrained information capacity (SCIC) framework. This framework considers task-specific encoding variables and substrate-intrinsic constraints, enabling device-informed optimization of input mapping. We experimentally demonstrate this framework on a highly uniform, selector-free 32 × 32 crossbar array, utilizing the volatile dynamics of memristors. Under the experimentally selected encoding conditions, the SCIC-guided RC achieves exceptional performance in temporal prediction (NRMSE of 0.019), speech recognition (free-spoken digit dataset (FSDD) subset accuracy of ∼94.5%), and real-time obstacle avoidance. Our approach links encoding constraints to solid-state device physics, providing a device-informed method for PRC parameterization at the edge.