Multilayer thin film oxides for resistive switching

Abstract

Resistive switching devices are hotly being pursued for use as the fundamental units in next-generation hardware deep-learning or neuromorphic systems. However, these devices are still tricky both to fabricate and to operate with consistency. We present strategies which guarantees that switching devices are functional post-fabrication, and with switching cycles that are consistent both from cycle-to-cycle and device-to-device. The resistance of all observed high and low resistance states (HRS/LRS) spanned just 0.23 and 0.19 decades on the logarithm scale across all devices, with both states spanning 0.05 within single devices and all SET transitions falling within a 0.3V span in our multilayer FIB-processed device. DFT simulations suggest that Au atoms from the top metal electrode implanted deeper in the device by FIB would serve as bridging atoms for oxygen vacancies filament by promoting the formation of these vacancies.

In addition, multilayer thin oxide films reduce the stochasticity of filament formation and further improves the switching consistency. This strategy for high consistency resistive switching devices was subsequently exploited for a few purposes. Firstly, multi-bit switching was demonstrated to yield approximately 7 distinguishable states with a single blind set, i.e, without having to program current compliances or use iterative schemes. Secondly, further insights into the SET and RESET mechanisms using pulsed measurements could be obtained since switching stochasticity no longer obscures subtle trends in experimental data. The implications of this study goes beyond the demonstration of a single high consistency device. Future understandings in resistive switching devices shall be achieved more easily since causality between processing parameters and device behaviors can now be quickly established under the significant reduction in switching stochasticity.

The new degrees of freedom introduced here in designing resistive switching devices will also hasten the search for an optimal device, bringing forward the realization of large scale resistive RAM arrays for machine learning or hardware neuromorphic computing applications towards a nearer future.