Researchers from University College London, University of Leeds, Tohoku University, Imperial College London and Japan Atomic Energy Agency have demonstrated that spin transfer torques in nearly isotropic CoFeB-based magnets can dynamically stabilize magnetic states that are unstable in equilibrium, realizing a nanoscale spintronic analogue of the Kapitza pendulum.
The team uses MgO∣CoFeB∣W multilayers, where the demagnetizing field favors in-plane magnetization while an interface-induced perpendicular magnetic anisotropy (PMA), tuned by post-growth annealing, counter-balances this tendency. By carefully optimizing the growth-annealing protocol, the shape and interface anisotropies almost cancel, yielding magnets with vanishingly small effective anisotropy that are nearly isotropic on the Bloch sphere. This near-isotropy is crucial because it suppresses conventional auto-oscillations and lowers the critical current needed to drive the system into a strongly nonlinear dynamical regime.
In this geometry, a static magnetic field normally tilts the free-energy landscape so that the magnetization settles in one of two minima aligned close to the field direction. Under equilibrium conditions, a state pointing opposite to the applied field corresponds to a maximum of the free energy and is therefore unstable. The engineered films provide a clean platform where this unstable configuration can instead be stabilized dynamically by current-driven spin torques, allowing direct exploration of far-from-equilibrium spin dynamics in a simple macrospin-like system.
Nonlinear magnetization dynamics are driven by spin transfer torque originating from spin-orbit interaction in the tungsten layer, i.e., a spin-orbit torque generated by the spin Hall effect. A charge current flowing along the W layer is converted into a spin current normal to the interface, with spin polarization transverse to the applied field, which exerts a torque on the CoFeB magnetization. As the dc current increases, the spin torque first compensates magnetic damping, then overwhelms it, destabilizing the usual equilibrium state with magnetization aligned parallel to the external field.
In the optimized nearly isotropic films, this process leads not to standard auto-oscillations but to a steady state in which the magnetization is dynamically stabilized near the free-energy maximum, pointing opposite to the applied field. Two independent electrical measurements show that this inverted configuration becomes the attractor of the dynamics under appropriate current and field conditions, contradicting the common assumption that strong spin torque always drives the system into a limit-cycle auto-oscillation. This behavior constitutes a spintronic analogue of the Kapitza pendulum, where rapid driving reshapes the effective energy landscape so that an upside-down pendulum—or here, an anti-aligned magnetization—becomes dynamically stable.
In an intermediate current regime, before the fully inverted state is stabilized, the spin transfer torque drives large-amplitude fluctuations of the magnetization direction that explore essentially the entire Bloch sphere. Analytical estimates and numerical simulations show that the probability distribution of magnetization over the sphere can be tuned “at will” by adjusting current and magnetic field, providing a controllable continuous stochastic variable rather than a simple binary up–down state.
Experiments confirm that a vanishing effective anisotropy is key to this behavior, as it suppresses coherent auto-oscillations that would otherwise dominate at high drive and instead promotes stochastic excursions across many metastable directions. The resulting electrically driven magnets do not settle into a fixed orientation but continuously wander among many possible directions, with statistics determined by the balance of spin torque, damping, thermal noise and applied field. These stochastic magnetic states are promising building blocks for hardware that implements probabilistic rather than deterministic computing primitives.
Leveraging this behavior, the researchers demonstrate a proof-of-concept simulation of a probabilistic neural network based on restricted Boltzmann machines, where the fluctuating magnetization states emulate stochastic neurons for generating two-dimensional images. In this context, the continuous magnetization direction on the Bloch sphere acts as a rich analogue variable that can encode probability distributions beyond the binary states used in standard magnetic memory elements.
Because the nearly isotropic magnets operate at low critical currents and do not require large static fields to access complex dynamical regimes, they offer a route to compact, low-power spintronic hardware for neuromorphic and probabilistic AI workloads. The work establishes isotropic magnets as a platform for studying far-from-equilibrium spin phenomena, including regimes sometimes described as anti-magnonics, and aligns closely with efforts to develop energy-efficient computing architectures for next-generation AI.