Magnon Reservoir Computing – MARIN

Whether it is on the surface of a lake, in colour patterns of halos and coronas in the atmosphere, or as fringes in optical or gravitational waves, wave interference is a phenomenon which plays an important role in our everyday life and allows us to address the most complex questions. But what about computation? Imagine a bucket of water into which a series of pebbles of different weights are dropped. By observing the resulting interference patterns, which contain information on past and present events, can we deduce the original sequence of these pebbles? The answer is yes! It is known that interference of water waves can be an efficient medium for a liquid state machine, an example of the neuro-inspired paradigm of reservoir computing. Here, the pebbles represent a complex temporal waveform, much like a voice signal; recognising such signals requires not only knowing the constituent frequencies, but also the order in which they arrive. Here, we propose that spin waves, which are elementary excitations of magnetic systems, can offer an efficient implementation of reservoir computing at the submicron scale. Because of their inherent nonlinearities and capacity to couple to transport phenomena, we envisage advanced pattern recognition tasks in magnonic devices at GHz frequencies.
With the MARIN project, we aim to realise experimentally such a magnonic reservoir computer. The MARIN project will investigate experimentally and theoretically the capacity of SWs in micro- and nanostructured thin films to satisfy the three basic requirements of reservoir computing, namely: (i) approximation – whether similar inputs result in similar outputs; (ii) separability – whether distinct input classes result in distinct output classes; and (iii) fading memory – how quickly inputs are forgotten over time. The basic control mechanism is the nonlinear coupling between SWs, which allows orthogonal eigenmodes of the equilibrium state to interact with each other as their amplitudes increase. Because such coupling also involves thresholding events, like for spiking neurons, we can achieve computational tasks with a cognitive nature like classification. This will be applied to demonstrate advanced signal recognition, e.g. on time-series, as a first step towards a proof-of-concept of efficient analogue non-Boolean operations.
Two media will be used: the well-established epitaxial YIG films and the epitaxial Heusler thin which both exhibit very low intrinsic magnetic damping mandatory to be able to efficiently excite non-linear spin-waves. The energy transfer from one non-linear spin-wave to another in a deterministic manner at the hearth of the reservoir computing scheme envisioned will be studied theoretically by micromagnetics simulations in the real and reciprocal space and experimentally by inductive spin wave spectroscopy, magnetic resonance force microscopy and microfocused Brillouin light scattering. Two different realisations of such time-series analysis are proposed as proof of concept devices: (i) a spectral analyser which performs an on-chip spin-wave Fourier transformation and (ii) a classifier of waves which as a first goal will be designed to sort sine from square waves in the GHz regime. This will allow for a new hardware implementation of reservoir computing that relies on the liquid state machine concept at GHz frequencies, which could be useful for processing telecommunications signals.