Event Start
Event End
Location
Abstract
The brain's highly energy-efficient computational capabilities have led to a significant shift in the hardware implementation of computing systems. The conventional CMOS-based architectures for deep neural networks face limitations due to the high energy costs of CMOS scaling and the Von Neumann bottleneck. This thesis addresses these challenges by utilizing spintronic physics to develop innovative spintronic device structures for neuromorphic computing applications. Recently, spintronic devices have garnered considerable interest in memory and computing technologies. Additionally, advanced device fabrication technologies have introduced thermally stable and topologically protected spin textures like magnetic domain walls (DW) and skyrmions. Magnetic skyrmion-based devices appeal due to their topological protection, compact size, and low energy dissipation. The fundamental spintronic device is the magnetic tunnel junction (MTJ). This thesis proposes new spintronic device structures that replicate the characteristics of biological synapses and neurons. A thorough methodology involving spin transport modeling, micromagnetic simulations, device fabrication, and characterization is employed to achieve the mentioned neural properties. To realize these neuromorphic devices, we harness the sophisticated physics governing the statics and dynamics of DW and Skyrmion, controlled by magnetic fields, currents, and voltages.
Initially, we introduce spin-orbit torque (SOT) and voltage-controlled DW-MTJ synaptic devices exhibiting features like spike time-dependent plasticity and improved weight update linearity. Following this, we present the magnetic field and current-controlled multilayer ferromagnetic device structure-based spintronic synapse. Upon integrating these synaptic devices into neuromorphic computing architectures such as artificial neural networks, the systems achieve pattern recognition accuracies exceeding 90%. Moreover, we showcase skyrmion-MTJ synaptic devices regulated by SOT and voltage, providing both short-term and long-term synaptic plasticity, with recognition accuracies also above 90% on the CIFAR-10 dataset. Additionally, we introduce confined voltage-controlled skyrmion-MTJ device structures that imitate the leaky-integrate and fire neuron properties for realistic spiking neural networks. We further propose field-free SOT-driven MTJ spiking neuron devices featuring auto-reset properties.
The thesis progresses by presenting the spintronic memory transistor leaky-integrate and fire neuron device based on domain walls and skyrmions. Enhanced neuron spiking characteristics are demonstrated through magnetic field gating and current-controlled schemes. The scaled version of the device dissipates around 12 fJ of energy and achieves classification accuracies above 97% on the MNIST dataset. Lastly, we introduce spintronic devices for stochastic Boltzmann neural networks neuromorphic computing applications. The thesis concludes by discussing our current progress with MTJ crossbar arrays for neuromorphic computing, current challenges with neuromorphic spintronic devices, and future technology prospects.
Brief Biography
Aijaz H. Lone is a PhD candidate in the Integrated Intelligent Systems (I2S) group at KAUST, working under Professor Gianluca Setti. With a background in electrical and computer engineering from IIT Mandi, India, his research focuses on developing spintronic devices and circuits for neuromorphic computing applications.
