Reservoir computing is a promising approach for processing time-dependent data, but hardware implementations often struggle because their internal dynamics are locked to the material's natural relaxation time. Researchers at Seoul National University and Yonsei University have now demonstrated a monolithic 3D stack that overcomes this limitation by pairing memristors with thin-film transistors (TFTs) in a vertical arrangement.
The stack uses Ru/HfO2/TiN memristors integrated with In2O3-channel TFTs. The TFTs act as control elements, modulating the memristors' conductance dynamics and allowing the reservoir kernel to be electrically programmed across a wide range of timescales. This means a single device can be reconfigured for different temporal processing tasks rather than being fixed at fabrication time.
The work, published in Nature Communications, highlights how monolithic 3D integration can add programmability to neuromorphic hardware without resorting to complex external circuitry. The authors describe the result as an electrically programmable multimode reservoir computing kernel, pointing toward more flexible and efficient hardware for temporal signal processing.