Superparamagnetic tunnel junctions have emerged as promising stochastic neurons for low-energy cognitive computing. We report, for the first time, nanosecond-scale mean waiting times between magnetization reversals in perpendicularly magnetized tunnel junctions, enabled by a strong entropic contribution to the thermally activated switching dynamics.
Top: Schematic of the energy profile along the reaction coordinate for the thermally induced magnetization reversal in an SMTJ, where ΔE and ΔS respectively correspond to the change in internal energy and entropy. Bottom: Corresponding random telegraph noise in the voltage-time trace.
Magnetic tunnel junctions (MTJs) consist of a free and a fixed magnetic layers separated by a tunnel barrier. They exhibit two metastable states depending on the relative orientation of the magnetic layers, namely parallel (P) and antiparallel (AP). At finite temperature, the energy barrier between these states can be overcome by thermal fluctuations, causing the magnetization of the free layer to reverse. We focus on the mean waiting times—or mean dwell times—between such reversals.
MTJs are well-established building blocks for non-volatile magnetic memory, where information retention at the scale of years is required. More recently, their low-retention-time (below milliseconds) counterparts, called superparamagnetic tunnel junctions (SMTJs), have emerged as highly appealing for numerous low-energy, unconventional, and bio-inspired computing schemes. In this context, minimizing the mean dwell times improves computational speed and energy efficiency.
We experimentally measure the mean dwell times in perpendicularly magnetized, 50 nm diameter SMTJs, and observe values as low as a few nanoseconds at negligible bias voltage—a timescale previously reported only in in-plane magnetized tunnel junctions. We theoretically explain this observation with a large entropic contribution to the free energy barrier. In perpendicularly magnetized MTJs, magnetization reverses via the nucleation and propagation of a domain wall. The large entropic contribution stems from the existence of low-energy modes that exist in domain walls, but are not present in collinear magnetic states. The transition state with a domain wall is thus favored by the entropy.
Team: Artificial Intelligence, Theory, Microwave Devices, MRAM
Collaboration: NIST-Gaithersburg, University of Maryland (Maryland, USA)
Funding: ANR (StochNet ANR-21-CE94-0002-01, PRIME SPOT ANR P-22-03813), NSF (CCF-CISE-ANR-FET-2121957), the French RENATECH network, the University of Liège (Special Funds for Research, IPD-STEMA Programme).
Further reading: Entropy-assisted nanosecond stochastic operation in perpendicular superparamagnetic tunnel junctions, Lucile Soumah*, Louise Desplat*, Nhat-Tan Phan, Ahmed Sidi El Valli, Advait Madhavan, Florian Disdier, Stéphane Auffret, Ricardo C. Sousa, Ursula Ebels, Mark D. Stiles, and Philippe Talatchian, Phys. Rev. Appl. 24, L011002 (2025). Open access: hal-05149179
Contact: Philippe Talachian, Louise Desplat
