Volatile memristive switching devices for neuromorphic computing

Neuromorphic systems, which aim to mimic the function of the biological brain, are promising candidates to overcome the Von Neumann architecture. Traditional architectures, where memory and processing unit are separated, cannot withstand the exponential growth of computation power and data transfer rate requirements. The attractiveness of biological neural networks is the ability to concurrently work both as computing and storage unit. In this context, memristors are increasingly gaining interest as artificial neurons or synaptic elements thanks to their low power consumption, scalability and CMOS technology compatibility. Memristors are two-terminal devices consisting of two electrodes sandwiching a switching layer. By applying a voltage across the two terminals, the resistance of the device can be changed from low (ON state) to high (OFF state) and vice-versa. Among memristive technologies, there are the electrochemical metallization memories (ECMs). In ECMs, one of the two electrodes is based on metals such as Cu or Ag which can easily diffuse into the switching layer. The transition to ON state (set) relies on the formation of a conducting filament inside a solid electrolyte: when a positive bias is applied, metal cations migrate from anode (active electrode) to cathode (inert electrode). In general, ECMs can be both volatile and non-volatile. In the first case, the reset is achieved by the filament self-dissolution, in the second, disrupting the metallic filament requires the application of a negative voltage on the active electrode. This thesis addresses the study of volatile and non-volatile functionality of ECM cells and how this dual regime can be controlled by engineering the materials stacks and the programming conditions. The devices investigated in this thesis are based on Pt (bottom electrode) / switching layer / Ag (top active electrode) stacks. As switching layers, we first consider single dielectrics as 10 nm SiOx and 10 nm Al2O3. The first device shows volatile behaviour, whereas the second one is non-volatile. To explore the coexistence of both volatile and non-volatile functionality, bi-layer stacks are developed by engineering a thin Al2O3 layer positioned at the top or bottom electrode interfaces: 10 nm SiOx / 1 nm Al2O3 (top), 10 nm SiOx / 2.5 nm Al2O3 (top), 2.5 nm Al2O3 (bottom) / 10 nm SiOx. Devices are electrically characterized by performing quasi-static measurements. Moreover, a subset of devices is programmed through pulses. Data statistical analysis show that the bi-layer dielectric stacks made of SiO2 and Al2O3 in diverse combination differ in forming/threshold voltage and leakage current values. Moreover, by controlling the maximum current flowing into the device during the filament formation (current compliance, CC), it is possible to achieve in the same device the volatile or the non-volatile behaviour. The thicker alumina layer, the higher the forming and threshold voltage, the lower the leakage current and the maximum supported CC for volatile operation. These results illustrate how critical is the role of an interlayer at the metal/insulator interface, between different materials inside the switching layer, on the switching capabilities of the device. This thesis work was carried out in CNR-IMM Unit of Agrate Brianza and in the framework of the EU H2020 MeM-Scales project (grant n. 871371).