Research Projects

The project aims to engineer the microwave properties of skyrmions stabilized in Co-based magnetic nanodot structures ( [Co/Pt]n and [Co/Pt/Ru]n/Co/Pt) for application in reservoir computing. Magnetic skyrmions are topologically protected magnetic states with a whirling magnetization texture. These magnetic states are non-volatile and can be controlled by the application of spin currents, magnetic and electric fields using ultralow energies. We plan to utilize skyrmions stabilized in nanodot structures as elements of a physical reservoir network to provide non-linear input-to-output mappings. These reservoir neural networks are extremely useful in classification/prediction tasks and work by projecting non-linear complex inputs into linearly separable outputs in higher dimensions. Since the non-linear part of the computation is delegated to the energy-efficient skyrmionic device and only a final linear weight layer is trained, these networks are expected to consume significantly lower energy for training compared to conventional software-based networks. The proposed multilayer systems will be grown using magnetron sputtering, and their electrical characterization in microwave regime will be studied through spin-torque ferromagnetic resonance (STFMR) set-up, which will be developed as a part of this project. The setup will be used to experimentally identify the skyrmion resonance modes in multilayer magnetic heterostructures and nanodots by injecting a microwave current of varying frequencies and recording the DC voltage response of the system. Simultaneously, numerical simulations will be performed to study the resonant modes of such constrained skyrmionic systems, and the effect of gate voltage on these modes will be investigated. The microwave properties of skyrmion nanodots tuned via experiments and simulations will be used to implement neural networks for reservoir computing. The focus will be on using the characteristic features of the spin torque diode effect (non-linearity near resonant modes) to propose different protocols for encoding input signals (direct injection or via frequency modulation of a carrier signal close to resonance frequency). These neural networks will be evaluated for their performance based on metrics such as short-term memory and parity check capacity and compared with pure software-based models implemented in PyTorch. We will also simulate deep reservoir networks composed of several interconnected layers of individual skyrmion nanodot reservoirs, each optimized individually to operate on a particular range of injected signal frequencies using gate voltages. In such a network, each layer will be trainable, allowing us to selectively extract information from a particular frequency band in the input signal.