Optimization of Memristors for Information Storage and Neuromorphic Computing

Abstract

[eng] The total amount of global information to be stored and computed is increasing exponentially since the beginning of the 21st century. The semiconductor industry made great efforts to push the silicon-based integrated circuits (ICs) to follow Moore’s Law to achieve better performance and meet the requirements. However, the basic element of the modern ICs, the transistor, is approaching its physical limitation of further scaling down. Current leakage in the vertical direction across the dielectrics and the crosstalk in the horizontal direction and heat between adjacent devices are unavoidable for smaller devices with thinner dielectrics and higher device density. There are two main solutions are proposed to solve these problems. One is “More than Moore” by replacing the transistor with a new electronic device, for example, a memristor. The other one is “More Moore” by introducing two-dimensional (2D) materials into the transistor structure. Compared to transistors, memristors are competitive with simpler device structure, scalability, and three-dimensional stackability. 2D materials have been included in the technology progression plan in the International Roadmap for Device and Systems (IRDS). In this PhD thesis, I combined the advantages of both memristor and 2D materials and carried out a deep study on 2D materials-based memristors, especially using a 2D insulator hexagonal boron nitride (h-BN) as the resistive switching medium. I fabricated h-BN based memristors with van der Waals structure and demonstrated that the device has the capability of stable and highly controllable tristate operation by controlling the current compliance (CC) and reset voltage (VRESET). I integrated the h-BN based memristor with a commercial complementary metal-oxide-semiconductor (CMOS) circuit, and h-BN memristors with a size of 0.053 µm2 have been achieved. The formed one transistor one memristor (1T1M) cell can show high endurance (millions of cycles with every measurement points recorded), multistate switching, coexistence of volatile and non-volatile switching, and synaptic behaviour like spike-time-dependent-plasticity (STDP), which are useful for implementing a spiking neural

network (SNN). The circuit based on two 1T1M cells can also show matrix operations like OR and IMP logic. I also wrote a perspective paper about proposing a roadmap for the future development of integrated circuits based on 2D layered materials. Apart from the 2D materials- based memristor, I also conducted several experiments based on the metal-oxide memristor and liquid phase exfoliated 2D materials for comparison. Our study is an important step toward the 2D materials-based memristor with wafer- scalable fabrication methods, and its successful integration into silicon-based ICs could inspire other scientists to study 2D materials-based devices integrated into a real microchip with functional circuitry.

[spa] La cantidad total de información que se almacena y calcula a nivel global está aumentando exponencialmente desde principios del siglo XXI. La industria de los semiconductores está haciendo grandes esfuerzos para impulsar los circuitos integrados (CI) basados en silicio para cumplir la Ley de Moore logrando un mejor rendimiento y a la vez que cumple con los requisitos. Sin embargo, el elemento básico de los circuitos integrados modernos, el transistor, se está acercando a su limitación física de reducción de tamaño. Se han propuesto dos soluciones principalmente para resolver estos problemas. La primera es la llamada "More than Moore" que se basa en reemplazar el transistor con un nuevo dispositivo electrónico, por ejemplo, un memristor. La segunda es la conocida como "More Moore" mediante la introducción de materiales bidimensionales (2D) en la estructura del transistor. En esta tesis doctoral, combiné las ventajas de los memristores con los materiales 2D y realicé un estudio profundo sobre los memristores basados en materiales 2D. Concretamente, he usado el nitruro de boro hexagonal (h-BN) fabricado mediante Chemical Vapour Deposition (CVD) como medio de conmutación resistiva. Fabriqué memristores basados en h-BN con estructura de van der Waals y demostré que el dispositivo tiene la capacidad de operar en tres estados estables y altamente controlables mediante el control de la corriente límite (CC) y el voltaje de reset (VRESET). Integré este memristor basado en h-BN con un circuito comercial CMOS logrando memristores h-BN con un área de 0,053 µm2. La celda formada por un transistor y un memristor (1T1M) puede mostrar alta durabilidad (millones de ciclos), conmutación multi-estado, coexistencia de conmutación volátil y no volátil y comportamiento sináptico como spike-time-dependent-plasticity (STDP), que son útiles para implementar la red neuronal de impulsos (SNN). Además, los circuitos memristivos basados en dos celdas 1T1M también puede mostrar operaciones matriciales como las operaciones lógicas OR e IMP. Nuestro estudio es un paso importante hacia el memristor basado en materiales 2D, y su integración exitosa en circuitos integrados de silicio. Esto podría inspirar a otros científicos a estudiar dispositivos basados en materiales 2D integrados en un microchip real con circuitos funcionales.