Diffusion processes in HfO₂-based memristors investigated by in situ STEM and atom probe tomography

Abstract text (incl. figure legends and references)

Ionic movement is an essential part of Resistive Random Access Memory[1]. During resistive switching (RS) operations, the memory type determines the diffusing species: in the case of electrochemical metallization memory (ECM), the electrode material migrates and for the valence change memory (VCM) type, the oxygen anions and vacancies[2]. Moreover, the interfacial interactions in the metal-insulator-metal (MIM) RRAM cell can further impact RS[3].

To give insight into the diffusion processes for both ECM and VCM type HfO₂-based memristors we have employed in situ scanning transmission electron microscopy (STEM) in combination with atom probe tomography (APT). The goal is to reveal the role of the microstructure of the insulator and the impact of the metal electrode on the ionic migration process.

The investigated memristive devices consist of highly textured TiN (M) and HfO2 (I) thin films subsequently grown on c-cut sapphire by using reactive molecular beam epitaxy (RMBE), complemented by sputter deposition of Cu and Pt (M). These result in an ECM and VCM devices, respectively. From the VCM device, focused ion beam (FIB) lamellae for STEM and APT tips were prepared for a pristine device and a device that has been operated for 100 bipolar resistive switching (BRS) cycles. From the ECM device, FIB lamellae were transferred onto Micro Electrical Mechanical System (MEMS) based heating chips (DENSsolutions) for in situ STEM.

For the BRS cycled VCM device, Figure 1 shows the Hf+O density variation retrieved from one APT data set overlaid on a medium-angle annular dark-field (MAADF) STEM plan view image of TiN. The highest variation occurs along a distinct line which, based on the feature size, is related to a TiN grain boundary (GBs). This indicates that the RS operation could be confined to the TiN GBs. For the ECM device, a low-loss electron energy loss spectroscopy (EELS) spectrum image (SI) map is overlaid on a cross-sectional high-angle annular dark field (HAADF) STEM image after in situ heating, shown in Figure 2. Cu diffusion started at ~350°C and the Cu feature in the EELS map at ~19.6 eV becomes present at a HfO₂ GB. This demonstrates that Cu diffusion could primarily occur along GBs.

We have shown that GBs influence and locally confine diffusion processes. For filament-type RS like in metal-oxide based memristor, this could lead to an improvement in variability and reliability by providing pre-defined diffusion pathways and thus, avoiding random diffusion.

References:

[1] R. Waser et al., Advanced Materials 21, 2632 (2009).

[2] I. Valov, Semicond. Sci. Technol. 32, 093006 (2017).

[3] S. R. Bradley et al., Microelectronic Engineering 109, 346 (2013).

Figure 1: The Hf+O₂ density variation retrieved from one APT data set for a VCM HfO₂-based memristor cycled for 100 BRS operations shows the highest variation along a distinct line. As overlaid here with a plan-view MAADF STEM image from TiN it becomes clear, that the feature shown in the Hf+O density variation must be related to a TiN grain boundary based on the feature size.

Figure 2: The EELS SI map overlaid on the cross-sectional AADFSTEM image of an ECM type HfO₂-based memristor after in situ heating reveals, that the Cu feature located at ~19.6 eV becomes present at the grain boundary (GB). This showcases, that Cu should primarily diffuse along GBs.

The authors acknowledge funding from the ERC "Horizon 2020" under Grant No. 805359-FOXON and Grant No. 957521-STARE.