Artificial intelligence (AI) applications are creating a huge buzz not only in the scientific community, but also in the general public. What is often forgotten, however, is the immense cost of resources to implement such applications in our conventional computing architecture, which is not optimised for this kind of computation. In addition, the progressing digitalisation of our world leads to exploding electricity demands with estimations reaching up to 30-50% of worldwide electricity being due to Information and Communications Technologies by 2025. It will require new forms of computer memory to mitigate these challenges and one possible memory technology od the future is resistive switching, where information is stored in the resistance of a memory cell rather than the charge on it as in conventional memory. This bears the promise of more energy-efficient computing and most importantly for AI applications, which emulate functionalities of the brain, individual memory cells can store multiple states instead of the conventional binary memory. However, resistive switching is suffering from challenges of reproducibility and uniformity.

In this paper, we present the design of a new class of materials to enable uniform and reliable resistive switching. We achieved vertical orientation in amorphous hafnium oxide thin films deposited at industry-friendly temperatures and the film orientation leads to very uniform and reproducible performance. This realisation of amorphous nanocomposite films is a great accomplishment, because vertical orientation in thin films typically requires high deposition temperatures, which are not industry-friendly. We achieved this by using pulsed laser deposition with a composite target consisting of the electronics industry standard material hafnium oxide and barium hafnate. During deposition, the kinetic energy available for forming the thin film is just high enough to enable the self-assembled phase separation of the two materials into the aforementioned vertical orientation. The resulting columns act as preferential conduction paths for uniform and reliable resistive switching. A patent based on this approach is at PCT stage and talks about commercialisation are under way.

Figure caption: Left: Cross-sectional transmission electron micrograph which reveals the vertical orientation in amorphous films. Right: Learning (increase of ΔG) and forgetting (decrease of ΔG) of a single device depending on the input training signal.