RuCl3 Memristors for Neuromorphic Computing

The information age seems poised to transition to an era where machine learning and artificial intelligence approaches to information processing pervade almost every field. In this context, there is a notion that new hardware specifically designed to mimic the learning circuits in the human brain will be beneficial. Such hardware can be termed "neuromorphic" [1] and can in principle be realized using silicon integrated circuits as demonstrated in the Intel Lohi chip [2]. However, the complexity and power consumption of the silicon approach is significant and the hope is that new active materials cab be discovered with intrinsic electrical response that is sufficiently nonlinear to enable neuromorphic applications more efficiently.

An electrical device response that is useful in neuromorphic applications is the memristor [3] for which the current-voltage response shows a hysteresis loop (Figure 1). This essentially indicates a history-dependent dynamic resistance in the material that is essential for learning applications. The experimental realization of memristors is comparatively recent [4] and typically involves complex and defective materials incorporated into nanoscale geometries [5].

We have discovered a high voltage memristive response in bulk crystals of RuCl3 [6]. Figure 1 shows the "pinched hysteresis loop" that defines memristive behavior. The size of the loop scales with the voltage sweep rate as qualitatively expected for meristor action. Our understanding of the origin of the effect is still developing but there are a few new ideas that have come about from phenomenological modeling. We can simulate the hysteretic I/V curves as a function of sweep rate and temperature by assuming a Shockley diode equation governing charge injection at the interface synergistically coupled to Joule heating due to the injected current. This is similar to other models [7] but includes an anomalously large diode ideality factor. Such large factors are known to arise in some cases from strong disorder in nanostructures [8] and we hypothesize that some combination of metal-insulator domains and crystallographic stacking faults are the origin of the disorder in our devices.

Importantly, the general topic of charge transport in RuCl3 and related layered quasi-2D Mott insulators is not well developed. We believe that there is room to improve the memristor performance of this material by optimizing materials quality, electrical contact engineering ,and possibly the fabrication of nanoscale thin films. This situation is reminiscent of the early days of the transistor industry where the materials science and device engineering protocols were not yet in place. Perhaps strongly correlated layered halides will be the "silicon" of the neuromorphic computing field? Our group continues to work on this possibility.

References

[1] de Valle et al., J. Appl. Phys. 124, 211101 (2018).

[2] Davies et al., IEEE Micro 38, 82 (2018).

[3] Chua, IEEE Transactions on Circuit Theory 18, 507(1971).

[4] Strukov et al., Nature 453, 80 (2008).

[5] Pickett et al., J. Appl. Phys. 106, 074508 (2009).

[6] Frick et al., Appl. Phys. Lett. 116, 183501 (2020).

[7] Gibson et al., Appl. Phys. Lett. 108, 023505 (2016).

[8] Brotzmann et al., J. Appl. Phys. 106, 063704 (2009).