Temperature-dependent reservoir computing with granular vortex magnetic tunnel junctions

Technical Summary: Temperature-Dependent Reservoir Computing with Granular Vortex Magnetic Tunnel Junctions

Problem Statement
Spintronic devices, particularly magnetic tunnel junctions (MTJs) hosting magnetic vortex states, are promising candidates for neuromorphic and unconventional computing due to their non-volatility, nonlinear dynamics, and compatibility with nanoscale integration. However, conventional device design typically treats material granularity and microstructural disorder as undesirable sources of variability that degrade reproducibility and reliability. In contrast, the reservoir computing (RC) paradigm suggests that such disorder and nonlinear dynamics can serve as intrinsic computational resources. While previous simulation-based studies have proposed that granular vortex-based MTJs (GV-MTJs) can function as physical reservoirs, a critical gap remains regarding their operational robustness under thermal stress. Temperature variations can alter vortex dynamics, relaxation times, oscillation frequencies, and the effective influence of grain-induced pinning. It is currently unresolved whether the computational capabilities of GV-MTJs remain stable across practical operating temperatures or if thermal stress degrades the dynamical features necessary for reliable computation.

Methodology
This work employs a combined experimental and simulation-driven approach to investigate the temperature-dependent performance of individual GV-MTJs with diameters of 600 nm and 800 nm.

• Device Architecture: The devices consist of a multilayer MTJ stack featuring a synthetic antiferromagnet, a reference layer, an MgO tunnel barrier, and a granular free layer (FL) hosting a magnetic vortex. An integrated field line (FiL) allows for local magnetic excitation of the free layer.
• Micromagnetic Simulations: Using MuMax3, the authors simulated temperature-dependent vortex dynamics by manually adjusting material parameters (saturation magnetization, damping, and anisotropy) based on experimental data from Kang et al. The simulations incorporated a stochastic thermal field and modeled the granular free layer using Voronoi tessellation to represent grain-induced pinning landscapes. This allowed for the isolation of material parameter effects on relaxation time and oscillation frequency.
• Experimental Characterization: Extensive electrical measurements were conducted on fabricated devices across ambient temperatures ranging from 20 °C to 90 °C, with extended testing up to 160 °C. Key metrics included:
  • Relaxation Dynamics: Relaxation time and oscillation frequency were extracted from the decay of the voltage response following a magnetic field pulse.
  • Effective Granularity: The number of "effective pinning sites" was quantified by counting abrupt resistance changes (Barkhausen-like jumps) during magnetic field sweeps at various temperatures.
• Reservoir Computing Benchmark: The computational capability was evaluated using the Free Spoken Digit Dataset (FSDD). Input waveforms (time-compressed speech signals) were applied via the field line. The resulting MTJ voltage responses were processed using Mel-scaled filterbanks and classified using compact convolutional neural networks (CNNs) with two lightweight architectures (Network 1 and Network 2).

Key Results

• Temperature-Dependent Dynamics: Both simulations and experiments confirmed that increasing temperature reduces the vortex relaxation time and oscillation frequency. Specifically, relaxation times decreased by approximately 15% for 1000 nm devices and 10% for 600 nm devices between 20 °C and 90 °C.
• Reduction in Effective Granularity: As temperature increased, the number of effective grain-mediated pinning sites decreased (from ~21 at 20 °C to ~18 at 80 °C for a 600 nm device). This indicates that thermal activation allows the vortex core to overcome weaker local energy barriers more easily, resulting in smoother trajectories and reduced dynamical diversity.
• Classification Performance:
  • Stability (20 °C – 90 °C): The GV-MTJ reservoirs maintained stable spoken-digit classification performance across this range. Accuracy degradation was minimal, remaining below 3 percentage points. For 800 nm devices, the average accuracy drop was less than 1 percentage point.
  • Degradation (160 °C): A sharp decline in classification performance was observed at 160 °C. Control experiments using the field-line voltage directly confirmed that this degradation was not due to reduced input signal amplitude (caused by increased field-line resistance) but rather to the deterioration of the reservoir's intrinsic functionality.
• Signal Analysis: At 160 °C, the MTJ response exhibited fewer abrupt transitions and oscillatory features, consistent with the reduced number of effective pinning sites and shortened relaxation times.

Significance and Claims
The authors claim that this work demonstrates that free-layer granularity, traditionally viewed as a fabrication imperfection, can be harnessed as a physical computing resource for reservoir computing. The study establishes that GV-MTJ reservoirs possess operational robustness over a practical temperature range (20 °C to 90–100 °C), making them viable candidates for neuromorphic edge-computing systems.

However, the paper also identifies the thermal limits of this approach. The sharp performance degradation at 160 °C is attributed to a combination of shortened vortex relaxation times, reduced effective grain-mediated pinning, smoother vortex-core trajectories, and weakened dynamical diversity. These findings provide design guidelines for future granular spintronic reservoirs, highlighting the necessity of optimized device geometry, material stability, and thermal management for deployment in variable thermal environments. The work bridges the gap between device-level magnetic dynamics and application-level reservoir computing performance, offering a measurement-driven evaluation of GV-MTJs under thermal stress.