Abstract
The burgeoning big data era has been pushing our limits of computing and processing. While hardware miniaturization and Moore’s law have driven innovation in digital electronics for several decades, scaling hardware has recently become challenging due to the high energy cost of computing and the increasingly steep memory wall caused by the segregation of the memory and processing units. One alternative is to utilize spins of electrons instead of charge to store and process information, popularly known as spintronics. Nanomagnetic devices are the building blocks in spintronics and exhibit non-volatility, which is attributed to the energy barrier between different spin states. While nanomagnets are now integrated onto silicon as commercial non-volatile memory elements, size scaling them for compactness reduces their energy barrier and turns them volatile. This dissertation addresses this considerable challenge by exploring different modes of energy barrier engineering in nanomagnetic devices that leverage topological invariants and traditional nanomagnets for unconventional computing. First, we utilize ultrasmall (~ 10 nm) and ultrafast (~ 1000 m/s) topological excitations — magnetic skyrmions with topologically protected barriers for temporal computing. We engineer the topological barrier by tuning the interfacial Dzyaloshinskii-Moriya interaction (DMI) to control the size and stability of skyrmions. We model a racetrack with periodic notches to produce and tune the energy barrier to hold skyrmions in place in the presence of thermal jitter and further employ machine learning to automate the process. Second, while topology (skyrmions) is one way to artificially increase the barrier of a tiny magnet, an alternate approach is to utilize instead the truly random white noise in a low barrier magnet (LBM) to accelerate fast optimization algorithms. This concept of probabilistic or p-bits, in effect, a binary stochastic neuron, relies, however, on magnets of perfect circular symmetry and is hyper-sensitive to process variations. We study the reliability of the computational networks built from LBMs and utilize their inherent stochasticity for inferencing tasks. Additionally, we model more realistic, commercially viable medium barrier magnets (MBMs) actuated by short current pulses for energy-efficient and robust probabilistic computing. Third, we turn to potentially more energy-efficient strain-gated dynamical barrier lowering of a nanomagnet in a heterogeneous piezoelectric/magnet/topological insulator (TI)/magnet stack that naturally encompasses logic and memory in an in-memory computing architecture with a minuscule energy cost and negligible footprint to circumvent the memory wall bottleneck. We solve a coupled stochastic Landau-Lifshitz-Gilbert (LLG) equation for the heterostructure, analyze the phase space of the working device, estimate the device level energy cost, and project it to the architecture level for in-memory logic operations. This dissertation may provide a comprehensive approach to energy barrier engineering in nanomagnetic devices, potentially leading to alternative technologies beyond complementary metal-oxide-semiconductor (CMOS) and facilitating more than Moore era.
