Optoelectronic Control of Redox Dynamics in POM Memristors for Noise‐Resilient Speech and Hardware‐Level Motion Recognition

ABSTRACT Edge artificial‐intelligence systems require compact hardware that can sense, process and store temporal information with minimal energy and latency. Here we exploit optoelectronic control of redox‐active polyoxometalate (POM) clusters dispersed in polymer matrices to realize hybrid memristors with mode‐switchable volatile and non‐volatile operation. Light‐driven, reversible POM reduction produces fading‐memory conductance dynamics that implement a physical reservoir and perform in‐sensor preprocessing of optical inputs, whereas electrically programmed redox states provide stable, multilevel synaptic weights in crossbar arrays. Using the volatile mode, an optoelectronic reservoir encodes lip‐motion trajectories and, when fused with noisy audio features, boosts keyword recognition accuracy from 83.3% to over 96.5% and preserves performance under strong acoustic corruption. In the non‐volatile regime, a hardware readout array classifies human motion trajectories with 90.5% accuracy. This unified redox platform connects molecular design with circuit‐level functionality, enabling multimodal neuromorphic hardware for directly processing complex spatiotemporal signals at the sensor interface.