DISSERTATION

Domain wall synapses for neuromorphic computing

Abstract

The demand for artificial intelligence has reached unprecedented levels in recent years. To address the growing demand, research efforts have expanded into designing artificial intelligence with ever-increasing complexity. However, such complexity comes at a cost of high-power consumption that will be unsustainable with current computing architecture. The recent proposal of a new brain-inspired architecture, known as Neuromorphic Computing, aims to overcome this challenge. The new architecture consists of interconnected systems of artificial neurons and synapses that promote efficient information communication which potentially reduces power consumption by several orders of magnitude. In particular, research in artificial synapses reveal efficient power consumption in domain wall devices containing a single ferromagnetic layer. The domain walls are typically driven by spin orbit torque, which utilises spin angular momentum to manipulate magnetisation. Its potential to reduce power consumption has already been well-established in literature. However, the stochastic nature of domain wall motion presents a challenge in designing domain-wall-based artificial synapses. In this thesis, we utilise the antiferromagnetic exchange coupling arising from a synthetic antiferromagnet to control spin orbit torque-driven domain wall motion of a single ferromagnetic layer. In part one of three of this thesis, we investigate a synthetic antiferromagnet-ferromagnet junction fabricated from a synthetic antiferromagnetic stack consisting of two ferromagnetic layers made of [Co/Pt] repeated multilayers coupled antiferromagnetically by a Ru layer. We apply spin orbit torque driven domain wall motion to demonstrate domain wall pinning at the region containing synthetic antiferromagnets. We reveal that the depinning current can be tuned depending on the antiferromagnetic exchange coupling strength adjusted by varying the thickness of the Ru layer. Micromagnetic simulations of the synthetic antiferromagnet-ferromagnet junction with various interlayer exchange coupling corroborates with the experimental findings. Additionally, we extended the number of pinning sites in both experiment and simulation to demonstrate the multi-resistive states of such an artificial synapse. We also conducted micromagnetic simulation on a second type of synthetic antiferromagnet-based synapse. Here, the single-layer ferromagnetic layer is replaced by a dual-layer ferromagnet with positive interlayer exchange coupling. We show that the domain wall may be trapped within both the ferromagnet and synthetic antiferromagnetic region. Elsewhere, theoretical studies have identified a potentially more efficient means of manipulating magnetisation than spin orbit torque. They revealed the existence of an even fundamental torque, known as orbital hall torque, which utilises the transfer of orbital angular momentum to influence magnetisation. Recent experimental research in orbital hall torque has been encouraged by the discovery of large theoretical orbital hall conductivity in 3d transition metals. In particular, Cr has been the popular choice of non-magnetic orbital source due to its orbital hall conductivity that is an order of magnitude larger than the largest spin hall conductivity in 5d transition metals. In an analogy to the known spin rashba effect, other studies have also detected enhanced efficiency associated with rashba-type interactions of orbital origin including those reported in surface oxidised Cu and AlOx. However, the possibility of rashba type interaction in Cr has yet to be established. Furthermore, unlike spin hall torque, orbital hall torque is sensitive to the non-magnetic/ferromagnetic interface. To the author’s best knowledge, any effect of sample preparation at the non-magnetic/ferromagnetic interface on the orbital-associated torque has yet been recorded. In part two of three of this thesis, we reveal a short-range enhancement in damping-like efficiency in surface oxidised Cr. Unlike spin angular momentum, orbital angular momentum cannot directly influence magnetisation. A Pt intermediate layer of sizable spin orbit coupling was inserted between the non-magnetic layer and ferromagnet to convert orbital angular momentum to spin angular momentum. Cr thickness dependent spin orbit torque characterisation unveils a sharp enhancement in efficiency that is limited to lower Cr thickness regime. At larger Cr thickness regime, we uncover a steady but monotonic increase in efficiency. Repeated spin orbit torque characterisation with unoxidised Cr reveal only the steady monotonic component across all Cr thicknesses. This steady increase in efficiency was previously attributed to orbital hall torque arising from bulk Cr layer. Furthermore, the similar efficiencies presented in both oxidised and unoxidised Cr stack at larger Cr thicknesses indicates that the surface oxidised layer continue to retain some bulk orbital hall torque response. This finding is unexpected because of the sensitivity of the orbital hall torque to crystallinity, which is this context, is further complicated by an exposed Cr. A short literature review on the natural oxidation of Cr uncovers an unusual oxidation mechanism in which its surface is characterised by both Cr and CrOx islands rather than a single surface oxide layer. We speculate that the retention of bulk orbital hall torque and crystallinity may be afforded by such an unusual oxidative mechanism. In the final part of this thesis, we examine the effect of low-voltage sputter cleaning on the damping-like efficiency. Here, a negative 10V of sputter cleaning was conducted on the heavy metal/non-magnetic interface in Pt/Cr and Pt/(surface oxidised Cr) (same stacks as part two but treated with additional sputter cleaning). Repeated Cr thickness dependent spin orbit torque characterisation unveils further enhancement in damping-like efficiency as compared to the stacks that have not undergone sputter cleaning (same stacks as part two). This result is a first in tuning orbital torque associated efficiency by sample preparation technique and presents a new perspective in optimising orbital hall torque. Additionally, we show that in larger complementary spin sources such as Pt/FM/W/(surface oxidised Cr), the additional sputter cleaning becomes a determining factor in further enhancing Pt/FM/W systems with Cr orbital layer. Likewise, when a Cr underlayer was inserted in Cr/Pt/FM/W/(surface oxidised Cr), improvement in efficiency was more significant when sputter cleaning was conducted on the Cr/Pt interface. Overall, the insertion of dual Cr layers supplemented by sputter cleaning of both heavy metal/non-magnetic interfaces reveals a 62 percent increase in damping-like efficiency as compared to Pt/FM/W structure. While it is impossible to qualitatively assess the orbital contributions from spin sources due to the complementary spin properties from W and Cr, Cr thickness dependent spin orbit torque characterisation of Pt/FM/W/(surface oxidised Cr) indicate a similar short-range enhancement in efficiency as those of Pt/(surface oxidised Cr). Finally, the improved damping-like response achieved via sputter cleaning are supported by reduced switching current in magnetisation switching experiments performed on Pt/FM/W/(surface oxidised Cr). Moreover, magnetisation switching conducted on Pt/(surface oxidised Cr) series (from part two) reveals reduced switching current from both the short-ranged enhanced efficiency at low Cr thickness regime and the steady monotonic response arising from bulk orbital hall torque.

Keywords:
Neuromorphic engineering Domain (mathematical analysis) Computer science Computer architecture Neuroscience Artificial intelligence Artificial neural network Psychology Mathematics

Metrics

0
Cited By
0.00
FWCI (Field Weighted Citation Impact)
0
Refs
Citation Normalized Percentile
Is in top 1%
Is in top 10%

Topics

Advanced Memory and Neural Computing
Physical Sciences →  Engineering →  Electrical and Electronic Engineering
Neural Networks and Reservoir Computing
Physical Sciences →  Computer Science →  Artificial Intelligence
Neural Networks and Applications
Physical Sciences →  Computer Science →  Artificial Intelligence

Related Documents

© 2026 ScienceGate Book Chapters — All rights reserved.