Monolithic three-dimensionally integrated memristor-thin-film transistor for electrically programmable multimode reservoir computing

This is the authors' abstract. We don't add key points for this paper.

Reservoir computing offers an efficient framework for processing temporal information, but most hardware reservoirs exhibit fixed or narrowly tunable dynamics determined by intrinsic material relaxation time. A monolithically integrated memristor-thin-film transistor reservoir kernel of this study enables electrically programmable temporal dynamics across a wide timescale. By vertically and monolithically integrating Ru/HfO2/TiN memristors with In2O3 channel thin-film transistors (TFTs), the memristive conductance dynamics are modulated by the TFTs. The temporal response can be actively tuned over three orders of magnitude experimentally and extended to over eight orders of magnitude via circuit configuration, demonstrating temporally unconstrained hardware with general applicability for spatiotemporal data processing. Implemented as an active reservoir array compatible with back-end-of-line integration, the system exhibits reliable spatiotemporal responses. The system demonstrates spatiotemporal computing capability through dynamic motion recognition in vision sensors and ultrasound signal classification, achieving high accuracy with a simple linear readout layer. Current reservoir computing hardware is efficient but limited by fixed dynamics. Shim et al. report a monolithically integrated memristor-transistor reservoir with electrically programmable temporal dynamics, supporting multiple temporal processing modes for motion recognition and biomedical signal classification.