modeling
(i) Investigation into the dynamics of Ti/HfOx synapses via modeling - I developed a compact model based on Trap-Assisted Tunneling and Ohmic transport principles to understand the working principle of adaptive oxide memristors. The model replicated experimental analog resistance change and temperature-driven current-voltage relationships up to ~200°C. This model also enabled simulations of analog pulse impacts on rupture thickness and extraction of barrier-height parameters at the oxide/electrode interface. We observed barrier height impacts the current responses. Consequently, it is anticipated that the active layer dopant will affect the barrier height and, thereby, the resistance change. [Applied Physics Letters]
(ii) Exploration of the impact of active layer dopant - I explored the impact of titanium dopant on the Atomic Layer Deposited HfOx active layer by varying Hf to Ti ratios. This exploration unveiled that increased titanium doping leads to higher vacancy concentration, signified by a decrease in bonded oxygen. The results of this work have demonstrated clear links between dopant concentration and resistive switching performance. Interestingly, the lower bandgap of titanium relative to HfO2 induces an increased tunneling current and decreased forming voltage while also enhancing linearity under negative analog pulses. However, attaining linearity under positive analog pulses remains a challenging pursuit. [Journal of Materials Chemistry C], [Journal of Applied Physics]
(ii) Exploration of the impact of active layer dopant - I explored the impact of titanium dopant on the Atomic Layer Deposited HfOx active layer by varying Hf to Ti ratios. This exploration unveiled that increased titanium doping leads to higher vacancy concentration, signified by a decrease in bonded oxygen. The results of this work have demonstrated clear links between dopant concentration and resistive switching performance. Interestingly, the lower bandgap of titanium relative to HfO2 induces an increased tunneling current and decreased forming voltage while also enhancing linearity under negative analog pulses. However, attaining linearity under positive analog pulses remains a challenging pursuit. [Journal of Materials Chemistry C], [Journal of Applied Physics]
(i) Addition of barrier layer in HfOx devices - Oxygen ion dynamics influence resistance changes. Integrating a SiOx (< 1 nm) barrier layer near the reset anode interface can control this, leading to gradual resistance changes during positive analog pulses. While the HfOx/SiOx devices outperform standard ones, there’s a trade-off between linearity and switching window because of a small oxide formation in the filament. The barrier layer devices, despite better linearity, have high reset current densities, necessitating further low-power synapse investigation. [ACS Applied Electronic Materials]
(ii) Off-current reduction using ultrathin multi-layered low thermally conductive materials - I posited that electrodes with low thermal conductivity could localize heat in the filament, thereby enlarging the rupture oxide area and decreasing current density. While most such materials are unsuitable as electrodes, the MAX phase uniquely combines low thermal conductivity with high electrical conductivity. Because of the ultrathin multi-layered structure, the MAX phase can confine heat and act as a better oxygen reservoir for the memristor. I fabricated memristors with MAX phase electrodes, verifying the fabrication via methods such as Raman Spectroscopy, Transmission electron microscopy, and in-situ XRD. These devices displayed ultra-low reset current density, high on-off ratio, and superior endurance, emphasizing the promise of MAX phase materials in energy-efficient, high-density memory systems. [Advanced Functional Materials]
(ii) Off-current reduction using ultrathin multi-layered low thermally conductive materials - I posited that electrodes with low thermal conductivity could localize heat in the filament, thereby enlarging the rupture oxide area and decreasing current density. While most such materials are unsuitable as electrodes, the MAX phase uniquely combines low thermal conductivity with high electrical conductivity. Because of the ultrathin multi-layered structure, the MAX phase can confine heat and act as a better oxygen reservoir for the memristor. I fabricated memristors with MAX phase electrodes, verifying the fabrication via methods such as Raman Spectroscopy, Transmission electron microscopy, and in-situ XRD. These devices displayed ultra-low reset current density, high on-off ratio, and superior endurance, emphasizing the promise of MAX phase materials in energy-efficient, high-density memory systems. [Advanced Functional Materials]
(i) Recovery of Deep Learning Performance of a degraded array - To examine the practical applications of adaptive oxide-based memristors in deep learning, I collaborated with IBM T. J. Watson Research Center. We demonstrate ReSta—a novel biasing technique. Designed to rejuvenate the filaments of fatigued memristive arrays after repeated training, ReSta significantly boosts long-term neural network accuracy on analog AI hardware. Next, we leveraged open loop training without write verification on this hardware with 98% accuracy. [IEEE Transactions on Electron Devices]
(ii) Demonstration of Transfer Learning in a 14 nm Technology chip - We also demonstrated efficient deep neural network transfer learning utilizing hardware and algorithm co-optimization in an analog HfOx based ReRAM array. We demonstrate that in deep neural network (DNN) transfer learning for image classification tasks, convergence rates can be accelerated by approximately 3.5 times by utilizing co-optimized analog ReRAM hardware and hardware-aware Tiki-Taka v2 algorithm. [Frontiers in Electronics]
(ii) Demonstration of Transfer Learning in a 14 nm Technology chip - We also demonstrated efficient deep neural network transfer learning utilizing hardware and algorithm co-optimization in an analog HfOx based ReRAM array. We demonstrate that in deep neural network (DNN) transfer learning for image classification tasks, convergence rates can be accelerated by approximately 3.5 times by utilizing co-optimized analog ReRAM hardware and hardware-aware Tiki-Taka v2 algorithm. [Frontiers in Electronics]
