Publication date: 9th January 2023
While non-volatile memristive devices find their best application as memory elements, volatile ones offer a promising solution for the development of bio-inspired neural networks that make use of dynamic elements with long time constants for sensory and learning systems.
This is due to their ability to mimic decay mechanisms with time constants that can reach behavioral time scales of minutes or more.
Here, we present SrTiO3 valence-change-mechanism (VCM) devices which can exhibit either non-volatile or volatile behaviour [1]. The relaxation of the volatile devices consists of the combination of an electronic and ionic process that modulates the Schottky barrier in the metal-SrTiO3 interface.
We demonstrate that, by adjusting the parameters of their programming pulses, both fast (<100 ms) and slow (seconds to minutes) relaxation dynamics can be implemented.
We describe how this device is fabricated, present its electrical characterization, and propose a model of its relaxation behavior. We furthermore show how these devices can be interfaced to mixed-signal CMOS circuits for neuromorphic computing applications.
Such volatile devices can potentially complement the conventional CMOS technology to build more compact trace circuits on synapses that implement temporal correlation-based synaptic behaviours, e.g., spike-timing dependent plasticity (STDP) or long-lasting eligibility traces [2] for solving the distal reward problem [3]. Similarly, the range of time constants that such devices demonstrate can be exploited by event-based sensing circuits for detecting spatio-temporal features across multiple sensory modalities.