Servicenavigation

Project Details

Back

Projekt Print View

Locally active memristive data processing (LAMP)

Subject Area Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
Term from 2015 to 2021
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 273537230
 
Final Report Year 2022

Final Report Abstract

Throughout the LAMP project, as a result of acquiring a large set of experimental data, an accurate physics- based model has been developed, which captures the temperature-activated Frenkel-Poole conduction mechanism underlying the nonlinear dynamics of our NbOx devices. Importantly, the introduction of this model ended a longstanding debate over the physical reason for the threshold switching effect in niobium oxide, proving that threshold switching at low currents is caused merely by temperature-activated Frenkel-Poole conduction, thus ruling out the emergence of a metal-insulator transition in this physical process. An adapted version of the model later introduced to account for the electrodes’ contact resistances and parasitic conduction paths. Furthermore, the impact of the device dynamic behavior on its electrical characterization was thoroughly investigated based on a given measurement setup. This analysis revealed, particularly, the appearance of multiple-NDR effects within the device quasi-static current-voltage characteristics in those scenarios where, the presence of a sufficiently large electrical RC element in the measurement setup, induced the emergence of oscillations during the voltage sweep. Our study uncovered the true origin for the multiple NDR phenomenon, which stems simply from the time-averaging operation, which the Source Measure Unit (SMU) applies to the acquired memristor current data. These results enabled to stop the potential dissemination of misconceptions in the literature. System-theoretic techniques were adopted to gain an in-depth understanding of the mechanisms underlying threshold switching effects in niobium oxide. In particular, the origin for the instability of the Negative Differential Resistance (NDR) branch in the device DC current-voltage characteristic, extracted under voltage stimulation, was identified, and a suitable bias circuit, enabling the stabilization of this branch, was devised. This study guided us toward the design of a robust memristor-based relaxation oscillator, and a good understanding of the role of each model parameter with respect to the characteristics of the oscillator. It also enabled to study static and dynamic pattern formation in resistively-coupled arrays of identical cells comprised of such oscillators. Those patterns emerge for design parameters poising the isolated cell in the locally active and stable domain, also referred to as Edge of Chaos, which is where the origin for complexity may be found. This body of knowledge was utilized to gain a better understanding of how to optimize the layer stack so as to boost the device performance in a given application. The physical realization of multilayer stacks endows the resulting devices with a higher degree of nonlinearity, by increasing the voltage extension of the NDR region while preserving their symmetric switching behavior. In the framework of the project, we established a fabrication process flow for devices with an area below 1 μm². The composition, and the design of these devices were optimized to enhance their performance in the applications of interest. Finally, the chips, diced from the wafer, and hosting a number of devices, were bonded via gold wires to DIL packages. The physics- based model was first fitted to experimental data from these new devices and then endowed with an additional parameter, which allows reproducing the statistical variability in their electrical parameters. This final variant of the model was employed to explore the potential of memristor oscillatory networks to solve a complex combinatorial optimization problem. Namely, the example of graph coloring was investigated, by exploiting the locally-active dynamics of the NbOx devices, which can induce an asymptotic separation between the relative phases of the oscillators. This difference in the phases can be utilized to partition the graph vertices into a set of different color groups, which poses a valid solution to the graph coloring problem. Two strategies, aiming to enhance the capability of the proposed computing engine to solve complex graph coloring problems with multiple valid solutions, were illustrated. Both strategies are based on resolving the impasse situation, that originates when the phase pattern of a given network converges toward a local minimum of a pre- specified optimization goal function. Furthermore, the fundamental role of those two strategies in enabling the proposed bio-inspired memristive network to outperform alternative state-of-the-art software and hardware implementations, was demonstrated.

Publications

 
 

Additional Information

© 2024 DFG Disclaimer / Copyright / Privacy Policy

Textvergrößerung und Kontrastanpassung