Resistive switching phenomenon and resistive switching devices could form the basis for the next generation non-volatile solid state memories. Particularly promising are structures based on dielectric oxides such as TaO_x and HfO_x which have been demonstrated to large OFF/ON ratio, low energy dissipation per switching cycle, long endurance, and sub-nanosecond switching times. It is widely accepted that the first switching cycle, referred to as electroforming, results in formation of a conducting small diameter filament within the highly resistive matrix. This is due to the electric field which induces the transfer of oxygen ions across the interfaces and into the metal electrodes producing locally oxygen-deficient conducting oxide. Most of the structural changes induced during electroforming and switching have been reported in devices that experienced high current levels during switching cycle.

The devices used in this study were 500×500 nm² crossbar structures with 20 nm thick TiN electrodes and 100 nm of reactively sputtered amorphous TaO_x. The devices were electroformed and electroformed/switched with the current compliance of 10-20 μA. The data have been gathered in several imaging modes of electron microscopy including bright field, High Angle Annular Dark Field (HAADF), Energy Dispersive X-ray Spectroscopy, and low loss and core regions of Electron Energy Loss Spectroscopy (EELS). The entire volume of the functional layer remained amorphous with no signs of crystallization. It contained only one area of bright contrast in HAADF images corresponding to the conductive filament with the diameter of approximately 50 nm. The maps produced by all techniques clearly indicate an increase of [Ta]/[O] ratio within the filament. In addition to overall change of the composition, the core of the filament shows a fine structure of the composition on the scale of 5-10 nm. We interpret the changes in distribution as due to thermodiffusion in the temperature gradient induced by current constriction when the device enters the negative differential resistance region of its I-V. The temperature distribution during this process has been estimated using a self-consistent finite element model.