Predicting the future with magnons

Zhiwei Xiong, Christopher Heins, T. Devolder, Fabian Kammerbauer, Mathias Kläui, J. Faßbender, Helmut Schultheiß, Katrin SchultheißarXiv 2025 · DOI · arXiv

TL;DR

Researchers have developed a "magnon-scattering reservoir" (MSR) that leverages the natural nonlinear physics of spin waves in magnetic microdisks to forecast chaotic systems. By transforming simple microwave inputs into high-dimensional spectral outputs, this system achieves high-accuracy predictions of complex time series, outperforming previous magnetic-based computing architectures.

What Problem Does This Paper Address?

Predicting chaotic systems—from weather patterns to secure communications—is notoriously difficult because small changes in initial conditions lead to wildly different outcomes. Traditional digital computers struggle with the energy-intensive task of simulating these nonlinear dynamics in real-time.

Reservoir Computing (RC) offers a shortcut: instead of simulating complexity, we use a physical system (the "reservoir") that is inherently complex and nonlinear. We feed an input into this reservoir, and it projects the data into a high-dimensional "state space." From there, we only need to train a simple, energy-efficient linear layer to interpret the results. The challenge lies in finding a physical medium that is scalable, fast, and possesses the right balance of nonlinearity and "fading memory."

Methods

The authors utilize magnons—collective excitations of electron spins in a magnetic material—as their reservoir.

  1. The Hardware: They fabricated a 5-µm-wide Ni$_{81}$Fe$_{19}$ (Permalloy) microdisk in a magnetic vortex state. In this state, magnetic moments curl in-plane around a central out-of-plane core, providing a stable environment for specific magnon modes.
  2. The Physics: The core mechanism is three-magnon splitting. When a "pump" magnon is excited at a high frequency, it can nonlinearly decay into two lower-frequency magnons. This creates a rich, high-dimensional spectral response from a single-frequency input.
  3. Input Encoding: The chaotic input signal is encoded into a microwave current using frequency modulation (specifically, continuous-phase frequency-shift keying).
  4. Readout: To "see" the reservoir's state, the team used time-resolved Brillouin light scattering (TR-µBLS) microscopy. This optical technique measures the intensity of different magnon frequencies, which serve as the "features" for the prediction model.
  5. Training: A simple linear regression was used to map the measured spectral intensities to the predicted future values of the time series.

Datasets

The system was evaluated using the Mackey-Glass (MG) benchmark. This is a standard chaotic time series generated by a delay-differential equation. It is widely used in the machine learning community because it is deterministic yet appears aperiodic and "random," making it a rigorous test for forecasting models.

Key Insights and Ideas

  • Nonlinearity as a Feature Generator: The intrinsic "scattering" between different magnon modes acts as a natural computer. It takes a 1D input and automatically maps it into a 182-dimensional spectral output, providing the "richness" required for complex tasks.
  • Fading Memory: For a reservoir to predict the future, it must "remember" the recent past. Because magnons have a finite lifetime (nanoseconds), the system naturally possesses a "fading memory" where past inputs continue to influence the current state for a short duration.
  • The Binning Trade-off: The authors discovered that "spectral resolution" is a critical design parameter. If you group frequencies into bins that are too large, you lose the nonlinear detail needed for accuracy. If bins are too small, the model becomes too complex and prone to overfitting. They found an optimal "Goldilocks" zone at roughly twice the natural resolution of their measurement tool.
  • Geometric Diversity: Perhaps the most striking insight is that combining different shapes—disks of different sizes, rings, and squares—improves performance. Each geometry has unique magnon modes, and combining them acts like an "ensemble" of different reservoirs, significantly boosting prediction stability.

Key Results

  • High Accuracy: The MSR achieved a Normalized Root Mean Squared Error (NRMSE) of ~0.031 for a short-term prediction horizon ($t' = 16$).
  • Long-Range Forecasting: The system successfully predicted up to 300 steps into the future (roughly six full oscillation cycles of the chaotic signal). This significantly exceeds the performance of previous magnetic reservoirs, such as those based on artificial spin-ice, which typically only manage 1-2 cycles.
  • Ensemble Boost: By combining the outputs of a 3-µm disk, a 5-µm disk, and a 5-µm square, the researchers reduced the average error across all prediction horizons, achieving an NRMSE of 0.056 even when predicting 300 steps ahead from a single "snapshot" of the reservoir.

Limitations and Future Work

  • Optical Readout: Currently, the system relies on Brillouin light scattering (a bulky optical setup) to read the reservoir state. For this to become a practical "chip," the authors acknowledge the need for electrical readouts, potentially using the inverse spin Hall effect or magnetoresistive elements.
  • Optimization: The specific combination of geometries used in this study was not fully optimized; the authors suggest that "evolutionary algorithms" could be used in the future to design the perfect multi-device reservoir.
  • Integration: While the materials are CMOS-compatible, the next step is full on-chip integration to move from a laboratory demonstration to a scalable hardware platform.

Summary updated

Abstract

Forecasting complex, chaotic signals is a central challenge across science and technology, with implications ranging from secure communications to climate modeling. Here we demonstrate that magnons - the collective spin excitations in magnetically ordered materials - can serve as an efficient physical reservoir for predicting such dynamics. Using a magnetic microdisk in the vortex state as a magnon-scattering reservoir, we show that intrinsic nonlinear interactions transform a simple microwave input into a high-dimensional spectral output suitable for reservoir computing, in particular, for time series predictions. Trained on the Mackey-Glass benchmark, which generates a cyclic yet aperiodic time series widely used to test machine-learning models, the system achieves accurate and reliable predictions that rival state-of-the-art physical reservoirs. We further identify key design principles: spectral resolution governs the trade-off between dimensionality and accuracy, while combining multiple device geometries systematically improves performance. These results establish magnonics as a promising platform for unconventional computing, offering a path toward scalable and CMOS-compatible hardware for real-time prediction tasks.
BibTeX
@misc{xiong2025predicting,
  author = {Zhiwei Xiong and Christopher Heins and T. Devolder and Fabian Kammerbauer and Mathias Kläui and J. Faßbender and Helmut Schultheiß and Katrin Schultheiß},
  title = {{Predicting the future with magnons}},
  year = {2025},
  eprint = {2510.06382},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.06382},
  howpublished = {arXiv},
}

Loading related content...

Figure