All-ferroelectric implementation of reservoir computing.

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Information Processing

Background:

• Reservoir computing (RC) offers efficient temporal information processing with low training costs.
• All-ferroelectric RC is appealing but undemonstrated due to challenges in developing ferroelectric memristors with distinct switching characteristics.

Purpose of the Study:

• To experimentally demonstrate an all-ferroelectric RC system using volatile and nonvolatile ferroelectric diodes (FDs).
• To fulfill the functional requirements for reservoir and readout networks using FDs derived from the same material structure.

Main Methods:

• Utilized Pt/BiFeO3/SrRuO3 structures to create volatile and nonvolatile FDs by manipulating imprint field (Eimp).
• Characterized FD properties, including short-term memory, nonlinearity, and long-term potentiation/depression.

Main Results:

• Volatile FDs with Eimp showed short-term memory and nonlinearity for the reservoir network.
• Nonvolatile FDs with negligible Eimp exhibited long-term potentiation/depression for the readout network.
• Achieved ultralow normalized root mean square error of 0.017 in Hénon map time-series prediction.

Conclusions:

• The all-ferroelectric RC system is competent for various temporal tasks, demonstrating high performance in time-series prediction.
• Both volatile and nonvolatile FDs exhibit long-term stability, high endurance, and low power consumption.
• The developed system shows promise as reliable and low-power neuromorphic hardware for temporal information processing.