Electronegative metal dopants improve switching variability in ${\mathrm{Al}}_2{\mathrm{O}}_3$ resistive switching devices

Resistive random-access memories are promising for nonvolatile memory and brain-inspired computing applications. High variability and low yield of these devices are key drawbacks hindering reliable training of physical neural networks. In this paper, we show that doping an oxide electrolyte, ${\mathrm{Al}}_2{\mathrm{O}}_3$, with electronegative metals makes resistive switching significantly more reproducible, surpassing the reproducibility requirements for obtaining reliable hardware neuromorphic circuits. Based on density functional theory calculations, the underlying mechanism is hypothesized to be the ease of creating oxygen vacancies in the vicinity of electronegative dopants due to the capture of the associated electrons by dopant midgap states and the weakening of Al-O bonds. These oxygen vacancies and vacancy clusters also bind significantly to the dopant, thereby serving as preferential sites and building blocks in the formation of conducting paths. We validate this theory experimentally by implanting different dopants over a range of electronegativities in devices made of multiple alternating layers of ${\mathrm{Al}}_2{\mathrm{O}}_3$ and WN and find superior repeatability and yield with highly electronegative metals, Au, Pt, and Pd. These devices also exhibit a gradual SET transition, enabling multibit switching that is desirable for analog computing.

Figure 1

Electronegative Au dopants and multilayering of the oxide films are two key factors to enable low variability switching. (a) Device cross-section imaged using transmission electron microscopy (TEM). For further elaboration on device design and TEM imaging, see Fig. 2 and Supplemental Material Fig. S7 [58]. Panels (b) and (c) compare the device performance without and with Au dopant implanted into the multilayered switching devices. (b) Multilayer ${\mathrm{Al}}_2{\mathrm{O}}_3$ cannot switch consistently when focused ion beam (FIB) milling is performed to define the device area before Au deposition. (c) Multilayer ${\mathrm{Al}}_2{\mathrm{O}}_3$ switches consistently when FIB milling is performed to define the device area after Au deposition. These devices did not require electroforming. The $I\text{−}V$ plot shows 300 superimposed switching cycles (15 devices ×20 consecutive cycles in each device). (d) The variability improvement using the Au doping strategy can be seen from the short span of resistances in the cumulative distribution functions (CDFs) observed over the 300 switching cycles, covering only 0.10 and 0.34 decades for the low-resistance state (LRS) and high-resistance state (HRS), respectively. Typical spans of other ${\mathrm{Al}}_2{\mathrm{O}}_3$ [60, 61] devices from the literature are shown for comparison. (e) The device-to-device CDF in (d) can be displayed separately for the 15 devices to show cycle-to-cycle variations <0.05, indicating that most of the variation seen in (d) comes from device-to-device differences. (f) CDF plots show the narrowing of the spread in the LRS and HRS resistances from a few orders of magnitude down to 0.10 and 0.34, respectively, with increasing the number of oxide layers from one to three.

Figure 2

Experimental investigation on the effect of focused ion beam (FIB)-processed device perimeter to measure conductance of individual devices at high-resistance state (HRS) and low-resistance state (LRS). (a) Three-dimensional (3D) perspective view of the FIB milling carried out to define the area of a single resistive switching device. (b) Scanning electron microscopy (SEM) image of region outlined with red dashed lines in (a). The darker region shows the exposed Si substrate, while the lighter regions are the top Au film. The red dash line [cross-section shown in (c)] will be affected by Au implantation due to the proximity to the FIB milling. (c) Stack schematic showing the layers deposited above a Si substrate. The FIB milling implants Au atoms sideways into the edge of ${\mathrm{Al}}_2{\mathrm{O}}_3$ layers. An atomistic schematic of the regions outlined in red dashed lines is shown in (d). (e) An alternative fabrication scheme where a 30-nm-thick Au contact pad of variable size was deposited with e-beam deposition through a shadowmask onto the usual triply repeated WN/${\mathrm{Al}}_2{\mathrm{O}}_3$ stack. A line with variable length was then FIB milled into this Au contact pad. (f) Measured device conductance data indicates a linear relation with increasing FIB milled lengths but no dependence on the contact pad area.

Figure 3

(a) Total density of states (DOS) of (i) Au-doped ${\mathrm{Al}}_2{\mathrm{O}}_3$ and (ii) Au-doped ${\mathrm{Al}}_2{\mathrm{O}}_3$ with one ${\mathrm{V}}_{\mathrm{O}}$ at nearest neighbor (NN) site to Au. (b) Total DOS of (i) undoped ${\mathrm{Al}}_2{\mathrm{O}}_3$ with a 4-${\mathrm{V}}_{\mathrm{O}}$ cluster and (ii) Au-doped ${\mathrm{Al}}_2{\mathrm{O}}_3$ with a 4-${\mathrm{V}}_{\mathrm{O}}$ cluster. In all DOS plots, the valance band maximum is at 0 eV, and both spin-up and spin-down states are plotted. The dotted line marks the position of the Fermi level. (c) Band decomposed charge density profile of Au $s$ orbital in relaxed Au-doped ${\mathrm{Al}}_2{\mathrm{O}}_3$ with a single ${\mathrm{V}}_{\mathrm{O}}$ (isosurface: $0.003\mathrm{eV}/{\AA }^3$), showing charge transfer to Au. (d) Initial structure of the ${\mathrm{V}}_{\mathrm{O}}$ cluster in Au-doped ${\mathrm{Al}}_2{\mathrm{O}}_3$ (positions of the NN ${\mathrm{V}}_{\mathrm{O}}$ are marked by black circles). (e) Band decomposed charge density profile of electronic states within the bandgap for the relaxed Au-doped ${\mathrm{Al}}_2{\mathrm{O}}_3$ with a 4-${\mathrm{V}}_{\mathrm{O}}$ cluster around Au (isosurface: $0.01\mathrm{eV}/{\AA }^3$).

Figure 4

Dopant metals from a range of electronegativities (a) Au, (c) Pt, (e) Pd, (g) Cu, (i) Ti, and (k) Al were tested to observe their effects on switching variability of ${\mathrm{Al}}_2{\mathrm{O}}_3$. Cumulative distribution functions (CDFs) for the low-resistance state (LRS) and high-resistance state (HRS) are plotted in (b), (d), (f), (h), (j), and (l). Dopants with higher electronegativities (Au, Pt, and Pd) have CDF plots with narrower widths, indicating that ${\mathrm{Al}}_2{\mathrm{O}}_3$ layers doped with these metals have low variability switching, consistent with predictions of easier ${\mathrm{V}}_{\mathrm{O}}$ and ${\mathrm{V}}_{\mathrm{O}}$ cluster formation shown in Table 1 and Fig. 3.

Figure 5

The gradual SET process is convenient for demonstrating multibit switching in ${\mathrm{Al}}_2{\mathrm{O}}_3$. (a) Multiple switching cycles were performed with variable terminating voltages between −0.5 to −4 V and a fixed maximum voltage at 3.5 V. Each switching cycle is color coded according to the most negative terminating voltage used. For example, a voltage sweep from −4 to 3.5 V is shown in blue, has the largest hysteresis that indicates the largest extent of switching, and the sharpest RESET onset at ∼2.8 V. On the other hand, a voltage sweep from −1 to 3.5 V is shown in green and has barely any hysteresis or switching. The simple use of a chosen terminating voltage can put the device into a predictable state, as characterized by (b) the switchable current and (c) the device conductance. The switchable current is defined as the difference in the device current as set and the lowest current measured in the OFF (highest resistance) state. Green, red, and blue overlays in (b) show the nature of one $I\text{−}V$ sweep when the device is set by −2, −2.8, and −3.9 V as terminating voltages, respectively. No current compliance was needed to be programmed in the sourcemeter used for these measurements.