Authors:	Schnitzspan, Leo
Advisor:	Jakob, Gerhard
Title:	Superparamagnetic tunnel junctions - true randomness, electrical coupling and neuromorphic computing
Online publication date:	4-Apr-2024
Year of first publication:	2024
Language:	english
Abstract:	This research deals with superparamagnetic tunnel junctions (SMTJs), encompassing their fabrication, characterization and potential applications in the context of neuromorphic computing and as a random number generator. Magnetic tunnel junctions (MTJs) based on a magnesium oxide barrier between cobalt iron-boron alloys exhibit a significant tunnel magnetoresistance (TMR) effect, typically on the order of 100−200 %. This characteristic high TMR signal has led to their widespread commercial use in sensing or as magnetoresistive random access memory (MRAM). Here, our developed TMR stack with in-plane magnetization exhibits a TMR ratio of over 200 % at a resistance area product of 550 Ωµm2. Within this work at JGU Mainz, the first successful MTJ nanopillar fabrication was accomplished in the group, and numerous optimization steps have been undertaken for the development of superparamagnetic tunnel junctions exhibiting nanosecond fluctuations. In the superparamagnetic regime, the magnetization of the ferromagnetic “free layer” fluctuates solely due to thermal excitation, resulting in a volatile MTJ. This inherent fluctuation, occurring naturally, can serve as an entropy source for random number generators, which makes stochastic MTJs attractive for applications with demanding requirements on random number generation, such as Monte Carlo simulations. In this work, it is demonstrated that a random number generator based on SMTJs exhibits true randomness of nanosecond time scale when combined with logic XOR gates. The quality of true random bit generation is assessed by evaluating all randomness tests of the statistical test suite provided by the National Institute of Standards and Technology (NIST). The fluctuation rate and state probability can be manipulated by external magnetic fields, applied currents or voltages, or by the temperature. Electrons that cross the tunnel barrier transfer a torque to the magnetization of the ferromagnetic free layer due to their spin (referred to as “spin-transfer-torque”) and thus significantly influencing the stochastic behavior of the SMTJ. In addition, the Joule heating affects the fluctuation rate at high current densities. It is demonstrated that both contributions, the Joule heating and the effect of the “spin-transfer-torque”, can be determined from dwell time analysis. Furthermore, coupling in the switching behavior can arise when two or more stochastic MTJs are electrically connected. In this work, the coupling strength between two SMTJs has been analyzed using the cross-correlation of the voltage fluctuation as a function of the applied source voltage. This approach was both simulated and experimentally verified using time series measurements for two stochastic MTJs. A network of multiple SMTJs can also be used to generate a Gaussian probability distribution, which might potentially be beneficial for noise-based neuromorphic computing approaches. At the end of this thesis, a neuromorphic circuit implementation based on SMTJs, diodes and transistors is presented. This implementation allows for an analog computation of a noise based local learning algorithm (node perturbation) in a neuromorphic hardware. The proposed approach represents a hardware-based alternative to the established backpropagation algorithm, since the neural network enables an analog and local calculation of the weight adjustment by a learning rule called “node perturbation”.
DDC:	530 Physik 530 Physics
Institution:	Johannes Gutenberg-Universität Mainz
Department:	MaxPlanck GraduateCenter
Place:	Mainz
ROR:	https://ror.org/023b0x485
DOI:	http://doi.org/10.25358/openscience-10119
URN:	urn:nbn:de:hebis:77-openscience-0fa4418b-f926-4511-855e-126a59cdf7f67
Version:	Original work
Publication type:	Dissertation
License:	CC BY
Information on rights of use:	https://creativecommons.org/licenses/by/4.0/
Extent:	xvi, 166 Seiten ; Illustrationen, Diagramme
Appears in collections:	JGU-Publikationen