TR-MOFE Spin Dynamics in Magneto-Optical Garnets
Always-on biosignal analysis is limited less by the sensor than by the energy of the digital computation that classifies the signal. Brain-inspired physical reservoir computing offers a way around this: a nonlinear dynamical material, driven by the input, does the computation in its own physics, and only a simple linear readout is trained. Magnetic materials are a natural fit, because spin dynamics supply both fading memory and nonlinearity, and neural-network models have a deep link to spin-glass physics (the Hopfield network maps onto the Sherrington-Kirkpatrick model).
My work is measurement-first. Using time-resolved magneto-optical Faraday-effect (TR-MOFE) pump-probe, a 400 nm pump excites the magnetization of a transparent garnet film and a delayed probe reads the Faraday rotation in transmission. A single time-domain trace yields three parameters: the precession frequency, the ring-down (memory) time, and an effective damping. This reaches an ultrafast regime that MRI and NMR cannot, and it is a characterization tool rather than a biosignal sensor.
Across four garnets, bismuth-substituted YIG and pure YIG show clean coherent precession, while aluminium-substituted and YAG-substrate films do not. The heavy bismuth ion gives strong spin-orbit coupling and a large magneto-optic signal, whereas aluminium lowers the magnetization and destroys coherence. The precession frequency rises roughly linearly with the in-plane field, which a movable two-magnet geometry and a position scan are used to map and calibrate.
The ring-down is the reservoir's memory kernel: the memory time sets how long the system remembers, the frequency sets its timescale, and large-angle precession supplies the nonlinearity. The measured parameters then parameterize a Landau-Lifshitz-Gilbert reservoir model, which is validated on a real biosignal task such as ECG arrhythmia classification rather than left as a pure simulation. Because a reservoir only sees a time series, the same approach extends from electrical signals such as ECG and EEG to magnetic ones such as MCG and MEG.
This is preliminary work from my first six months at the Tabata Laboratory, University of Tokyo.