In recent efforts to develop compact and low-power resistive memory devices, self-rectifying memristors (SRMs) based on electronic bipolar resistive switching (e-BRS) have emerged as promising candidates. Unlike conventional memory structures, SRMs op...
In recent efforts to develop compact and low-power resistive memory devices, self-rectifying memristors (SRMs) based on electronic bipolar resistive switching (e-BRS) have emerged as promising candidates. Unlike conventional memory structures, SRMs operate without electroforming and exhibit built-in rectifying behavior, which helps suppress sneak paths in array configurations. These features enable selector-less architectures and facilitate greater vertical integration. This dissertation studies vertically stacked SRM arrays and their potential for in-memory computing applications.
First, a heterogeneous reservoir computing model is implemented using Pt/Ta2O5/HfO2/TiN (PTHT) devices, which operate in either self-rectifying or leaky modes depending on compliance current settings. This dual-mode configuration improved classification performance for time-dependent inputs, such as handwritten digits.
Second, a four-layer 32 × 32 vertical PTHT memristor array containing over 4,000 devices was fabricated and tested. The structure builds on previous work using smaller arrays and confirms scalability across layers using a multiprobe system and switching matrix. Importantly, it was also demonstrated that the vertical PTHT array can function as a heterogeneous reservoir, similar to planar configurations, enabling reliable time-series classification within a fully integrated structure. The array was used to perform multi-task inference on visual datasets.
Lastly, a wire-bonded PCB testing platform was developed to automate the measurement of the electrical characterization of high-density vertical memristor arrays. By integrating shift registers, multiplexers with ultra-low leakage current, and a multi-channel switching matrix, the setup enabled precise per-cell selection and voltage application across thousands of devices. One of the key advantages of this platform is its ability to access individual layers in a stacked array without relying on bulky or specialized probe stations, which are often impractical for repeated or scalable measurements. This design significantly reduced the total testing time from several weeks per layer to just one day, while improving measurement reproducibility and system stability.
Altogether, the results suggest that vertical SRM arrays are a feasible component for neuromorphic computing and memory-focused systems, where integration scale and energy efficiency are key considerations.