h1

%0 Thesis
%A Schnieders, Kristoffer
%T Statistische Charakterisierung und Modellierung von memristiven Bauelementen
%I Rheinisch-Westfälische Technische Hochschule Aachen
%V Dissertation
%C Aachen
%M RWTH-2025-09470
%P 1 Online-Ressource : Illustrationen
%D 2025
%Z Veröffentlicht auf dem Publikationsserver der RWTH Aachen University
%Z Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2025
%X Recently, large language models (LLMs), such as ChatGPT, have demonstrated significant potential as valuable tools for facilitating a wide range of tasks. Consequently, a restructuring of many professions due to LLMs appears unavoidable. In parallel, concerns regarding the energy consumption and computational complexity of LLM training and inference have emerged. The power consumption of data-heavy computational tasks, such as LLM training, is partially attributed to the limitations of conventional von Neumann architecture. In this architecture, the central processing unit (CPU) and memory unit are spatially separated, leading to increased latency and energy consumption. This so-called memory wall, or von Neumann bottleneck, can be circumvented by implementing brain-inspired computational architectures that integrate computing and memory tasks within a computing-in-memory unit (CMU). Memristive devices hold promise as key components in CMUs, as they are capable of performing both computational and memory tasks. One promising class of memristive devices is valence change mechanism (VCM) devices, which exhibit excellent endurance, stackability, CMOS compatibility, and low energy consumption. However, the intrinsic variability of these devices remains a substantial hurdle. In particular, read noise phenomena threaten the computational accuracy and memory function of VCM devices, as the initially programmed state can vary over time. Therefore, a deep understanding of this phenomenon and strategies for mitigating its effects are essential to align VCM devices with appropriate applications. This thesis adopts a multi-level approach to investigate read noise and its impact on application-level performance. The content spans from verifying correlations between material properties and read noise, derived from theoretical physics, and investigating read noise engineering approaches, to developing applications in which device type and noise characteristics are optimally matched. First, this thesis compares the characteristics of read noise as a function of switching material and switching modes of VCM devices. Next, methods to mitigate read noise are explored, concluding with a discussion on potential applications and statistical modeling. Initially, physical explanations for read noise are validated and their contributions quantified. The influence of material properties on the read noise characteristics is investigated by comparing VCM devices based on different switching materials. The results link the read noise characteristics to the energy gap between the conduction band (CB) and oxygen defect states. For materials with small gaps, electrons cross the Schottky barrier by tunneling from the CB (Type 1), whereas for materials with larger gaps, electrons tunnel from defect states (Type 2). Type 1 devices exhibit low read noise, while Type 2 devices display stochastic, abrupt current jumps of varying amplitude. These jumps are attributed to the dislocation of oxygen vacancies near the Schottky barrier. Building on these insights, the next section examines how switching modes modulate read noise characteristics. For Type 1 materials, few noise events are observed in the filamentary switching mode, while the area-dependent mode consistently exhibits low noise amplitudes. For both switching modes, read noise remains significantly lower than that of Type 2 devices. To address the pronounced read noise in Type 2 materials, the tunability of read noise via device fabrication and read parameter variation is investigated. This includes modifying fabrication processes and altering read voltages. The read noise of thermally oxidized TaO_x devices is measured across a range of read voltages. It is observed that the frequency and amplitude of noise events increase with the duration and temperature of annealing. Furthermore, the read noise amplitude increases with higher read voltages in set polarity, while relative read noise remains nearly constant for all reset polarities. Finally, potential applications of VCM devices are analyzed with an emphasis on read noise characteristics. The state instability of TaO_x-based 1R and 1T1R devices is evaluated. The read noise is found to be the dominant contributor to state instability. The number of practically achievable states is assessed, and instability data is analyzed using statistical methods to ensure reliable interpretation. The application chapter explores how read noise impacts VCM device applications. To this end, a statistical array model is developed to simulate the state instability of VCM-based 1R and 1T1R arrays. This model identifies key properties essential for the reliable operation of large-scale VCM crossbar arrays while significantly reducing simulation runtimes compared to conventional compact models. Additionally, a read noise-based true random number generator is introduced. This generator showcases a method for harnessing the intrinsic randomness of VCM devices, highlighting an additional potential application of VCM arrays integrated within CMUs.
%F PUB:(DE-HGF)11
%9 Dissertation / PhD Thesis
%R 10.18154/RWTH-2025-09470
%U https://publications.rwth-aachen.de/record/1021101