Engineering relaxation timescales in Ag/SiO x /Pt volatile memristors via a thin Al 2 O 3 layer

Neuromorphic computing, exploiting time as a computational variable, needs the development of devices able to show rich dynamics over various timescales. Volatile electrochemical metallization (ECM) memristive devices are among the most promising candidates for this role thanks to their tunable relaxation dynamics, showing retention times typically in the ns to ms range. This paper investigates the complex relaxation timescales of Ag/SiO x /Pt volatile ECM devices emerging from the addition of a few nm thick Al 2 O 3 layer at the active interface. We show that, upon the insertion of the alumina layer, retention time increases up to several seconds and can be further extended with repetition of programming stimulation through a cumulative effect. The accumulation effect is reversible provided that the device is given enough time to recover. In summary, the overall device relaxation is composed of three different timescales, that can be controlled with programming parameters and alumina layer thickness and that we relate to different physical processes inside the device. Intrinsic retention is related to partial filament disruption in pristine (or recovered) devices; cumulative retention is caused by the incomplete dissolution of the filament inside Al 2 O 3 which causes, cycle after cycle, an increase in the retention time; recovery time is needed to recover intrinsic state in a previously stimulated device, and it is related to the retraction of the filament in the Al 2 O 3 layer. The latter two relaxation timescales, typically not reported in literature, can be exploited to implement neuronal and synaptic primitives thanks to their integrative nature.