|Thesis abstract: |
Oxide based resistive switching memories represent a strong candidate for next generation solid state non-volatile memory technology. In particular, the extreme ease of fabrication, the relative low operating voltages and high write/erase speed suggest that this technology may became a valid alternative to Flash memories.
The working principle of RRAM is based on the capability of the material to switch reversibly between two different resistance values. The build-up of a thin conductive filament is responsible of the low resistance that characterizes the so-called set state. There are several experimental evidences that support this hypothesis, but still the actual nature of the conductive filament is unknown.
By the use of particular program algorithm we carefully studied both set and reset states, showing in particular that there is a continuous transition between the set state, that is characterized by a metallic conduction, to the reset state that is, on the contrary, semiconductor-like with a certain activation energy for conduction. This change in resistance can be related to a modulation of defect doping and a consequent shift in the Fermi level. Thanks to this characterization we were able to extract an important microscopic parameter, namely the effective diameter of the conductive filament.
Starting from electrical characterization on NiO (unipolar) RRAM we developed physical models for the set (formation) and the reset (rupture) of conductive filaments, and their dependence on filament size/resistance. A Joule heating thermal model for reset accounts for the reset voltage and current as a function of filament size, taking into account size-dependent heat conduction mechanisms and size-dependent diffusion/oxidation effects. Set transition is modeled by threshold switching while the resistance-dependent set voltage and current are well reproduced by a change in doping concentration and activation energy.
The unipolar reset switching model was extended to the case of HfO (bipolar) RRAM by considering field driven ion migration. The physical comprehension of these switching mechanisms lead us to evidence a universal unipolar/bipolar reset model with negligible dependence on materials.
Physical-based numerical models for reset and retention of unipolar/bipolar RRAM have been presented. These models have been validated on electrical characterization results and can be used for numerical simulations of programming, reliability and scaling predictions.
For the first time we demonstrated complementary switching in single stack nonpolar-RRAM devices, based on simulations and DC/pulsed experiments in symmetric HfO-based RRAM. The introduction of this new operation mode, which intrinsically solves the sneak-path problem in cross-bar arrays, seems to be very promising for the development of future high-density memory. Finally, complementary switching allowed us to explain the coexistence of unipolar and bipolar through vertical ion displacement by field-driven migration and electromigration/diffusion.
Reset current reduction has been accessed by the control of the dimension of the conductive filament. This can be done limiting the current drained by the cell during its growth, by mean of an integrated MOSFET. We characterized 1T1R structures, showing that there is a linear relationship between the current compliance imposed by the MOSFET during set, and the reset current. By the use of CAFM techniques we were able to direct manipulate single conductive filaments showing reset currents below 1 ?A. These results are in line with the expected dependence on the size of the conductive filament thus supporting area scaling as a powerful method for program power reduction which is crucial for device scaling itself. In particular, since unipolar-RRAM may be utilized in cross-point array with dedicated selector diode, reset scaling is necessary to assure that the integrated diode may carry sufficient current.
We statistically studied data retention issues, showing that the time to failure for the set state (the reset state is on the contrary stable) depends on temperature with an Arrhenius-like behavior, and on the initial resistance, i.e. the diameter of the conductive filament and the activation energy. As can be qualitatively expected the failure time decreases for increasing resistances. This fact in turn yield to an intrinsic trade-off between retention time and reset current, since the lower is the resistance, the higher is the current drained by the cell during reset.
Power reduction was addressed in in unipolar RRAMs by the use of pulsed program operation. Minimum reset current requires controlled set to avoid overshoot effects. Set-reset instability and set due to threshold switching in semiconductor-like conductive filament affect both DC and pulsed reset and appear as main issues in preventing low-current, stable and fast reset. This issue may be suppressed by careful set algorithms or material engineering aimed at obtaining metallic-like filaments with sufficiently high resistance.
To better assess reliability issues, we showed data and discussed RTN affecting the reset state. We developed a simple model based on trapping/detrapping of a single localized state that is able to reproduce data of relative resistance fluctuation as a function of the resistance. RTN however may hardly cause read mistakes or other issues.
Synthesis and characterization of a novel device, based on the self-assembly of Ni-NiO core-shell nanowires has been addressed. The assembly of complex architectures, starting from simple nanometric building blocks, possesses strong potentials since it may lead to the production of large arrays without lithographic limits. Here we demonstrated the working principle of a NiO RRAM constituted by the two NWs in a cross-bar structure. The contact point of the two NWs has in fact a Ni/NiO/Ni stack. For the first time we showed resistive switching at the crosspoint of two nanowires. Set and reset parameters were shown to be compatible with typical values for planar structures. The missing encapsulation of these devices are probably responsible for their low endurance, that was evidenced during the characterization. Eventually we showed that better endurance can be achieved with hybrid devices made by a single core-shell NW, crossed by an EBL deposited metal strip.