Deterministic control of the probabilistic phase dynamics in injection-locked spin-torque nano-oscillators

Technical Summary: Deterministic Control of Probabilistic Phase Dynamics in Injection-Locked Spin-Torque Nano-Oscillators

Problem Statement
Spin-torque nano-oscillators (STNOs) are promising candidates for unconventional computing, particularly for Ising machines and probabilistic bit (p-bit) implementations, due to their tunability, low power consumption, and nanoscale dimensions. However, STNOs inherently exhibit thermally driven phase fluctuations, rendering their dynamics stochastic. While injection-locking is commonly used to suppress phase noise, the authors note that at room temperature, coherence remains susceptible to thermal fluctuations, leading to stochastic "phase jumps" between degenerate states.

Existing approaches to probabilistic computing often rely on superparamagnetic tunnel junctions where a DC current modulates energy barriers to tune the probability of a binary state. A critical gap exists in the experimental demonstration of a single nanoscale spintronic device where the output probability can be continuously programmed and deterministically controlled without suppressing the underlying stochastic fluctuations. Specifically, while unbiased random bitstreams have been generated using STNOs, fine control over the bias of these bitstreams—essential for encoding local fields and constraints in Ising machines and stochastic neural networks—has not been experimentally realized in a single STNO device.

Methodology
The authors experimentally investigated a vortex-based STNO operating in the second-harmonic injection-locking (SHIL) regime. The experimental setup involved:

1. Device Operation: A DC current ($I_{DC}$) and a static magnetic field ($H_{DC}$) induced auto-oscillations at a free-running frequency $f_0 \approx 243.76$ MHz.
2. Injection Locking: An RF drive at $f_{IL} \approx 2f_0$ locked the oscillator frequency to half the drive frequency, creating two degenerate phase attractors separated by $\pi$ (defined as 0 and $\pi$ states).
3. Bias Excitation: A weak RF bias excitation at $f_{bias} \approx f_0$ was applied via a field-line antenna. This perturbation was much smaller in amplitude than the SHIL drive.
4. Control Parameters: The system was controlled by two parameters: the bias phase ($\alpha$) relative to the SHIL drive, and the bias power ($P_{bias}$).
5. Measurement: An analog phase-readout circuit, synchronized to a common 10 MHz reference, allowed for real-time acquisition of the stochastic phase dynamics. The authors extracted mean dwell times ($\tau_0, \tau_\pi$) and occupation probabilities ($P_0, P_\pi$) from voltage-time traces.
6. Modeling: The results were supported by numerical simulations using the finite-temperature Thiele equation and a phenomenological toy model based on an effective phase quasipotential derived from a modified Adler-like equation.

Key Contributions and Results
The study demonstrates that the intrinsic randomness of an STNO can be harnessed to create a programmable stochastic element:

• Deterministic Probability Control: By applying the weak RF bias, the authors successfully broke the symmetry of the phase-jump rates between the 0 and $\pi$ states. This allowed for the continuous tuning of the occupation probability ($P_0$) from the unbiased limit ($P_0 \approx 0.5$) to values approaching 0 or 1.
• Dual Control Knobs:
  • The bias phase ($\alpha$) acts as a selector, determining which of the two metastable states (0 or $\pi$) is statistically favored.
  • The bias power ($P_{bias}$) acts as a strength controller, determining the magnitude of the probability imbalance.
• Quantitative Results:
  • At a fixed bias power of $P_{bias} = -50$ dBm, varying $\alpha$ from 0 to $2\pi$ resulted in sinusoidal-like variations in the mean dwell times and probabilities.
  • At a fixed phase $\alpha = \pi/9$, sweeping $P_{bias}$ from $-100$ dBm to $-40$ dBm increased the contrast between dwell times. At $P_{bias} = -40$ dBm, the probability $P_0$ reached approximately 0.95 (favoring the 0 state) or 0.05 (favoring the $\pi$ state, depending on $\alpha$).
  • The mean dwell times varied significantly relative to the unbiased case (e.g., +280% and -72% at high bias power), demonstrating substantial control over the switching timescale.
• Theoretical Validation: A toy model based on a tilted washboard quasipotential, where the RF bias adds a phase-dependent term, quantitatively accounted for the experimental and numerical observations. The model confirmed that the bias phase shifts the potential minima while the bias power alters the barrier heights, thereby modulating the thermally activated escape rates according to Arrhenius-type laws.

Significance and Claims
The paper establishes injection-locked STNOs as programmable stochastic elements capable of functioning as hardware primitives for probabilistic computing. The authors claim that this work provides a route to implementing tunable p-bits and stochastic binary neurons within a single nanoscale device.

Key implications highlighted by the authors include:

• Ising Machines: The ability to deterministically tune the probability of a phase state allows each STNO to encode local fields and problem constraints in Ising machine architectures, where the oscillator phase represents a binary spin variable.
• Brain-Inspired Computing: The system natively implements a tunable probabilistic activation function, suitable for stochastic neural networks and Boltzmann machines.
• Speed and Integration: Unlike superparamagnetic tunnel junctions where fluctuation rates are limited by thermal activation over magnetization barriers, STNO-based p-bits rely on oscillator frequency and relaxation rates, potentially offering access to faster (gigahertz range) stochastic dynamics.
• Scalability: While the current implementation uses external RF sources and analog readout, the authors note that integrating RF generation and phase detection into CMOS-compatible circuitry is a necessary step for scaling to networks of coupled oscillators.

The authors remain modest regarding the current limitations, acknowledging that the integration of RF routing and phase readout remains a challenge and that the regime where the bias excitation competes with the SHIL drive (rather than merely perturbing it) requires further systematic investigation. However, the work successfully bridges the gap between the physics of noisy spintronic oscillators and the requirements for tunable stochastic elements in unconventional computing.