Next-Gen AI Hardware Uses Spin Waves to Unlock Unmatched Computing Power

A revolutionary device uses ion-gating to manipulate spin waves, offering unprecedented performance for time-series predictions and paving the way for energy-efficient AI applications.

Research: Iono–Magnonic Reservoir Computing With Chaotic Spin Wave Interference Manipulated by Ion-Gating. Image Credit: Treecha / Shutterstock

A research team from NIMS and the Japan Fine Ceramics Center (JFCC) has developed a next-generation AI device, a hardware component for AI systems, that incorporates an iono-magnonic reservoir. This reservoir controls spin waves (collective excitations of electron spins in magnetic materials), ion dynamics through protonation-induced redox reactions, and their interactions. This innovative approach significantly enhances computational efficiency by leveraging in situ manipulation of spin wave properties. The technology demonstrated significantly higher information processing performance than conventional physical reservoir computing devices, underscoring its potential to transform AI technologies. This research was published in the journal Advanced Science.

As AI devices become increasingly sophisticated, demand for energy-efficient, high-performance solutions continues to grow. The newly developed device generates spin waves using antennas integrated with yttrium iron garnet (YIG) magnets, a critical material for its operation. The device utilizes an ion-gating technique that introduces ions into the YIG magnets, modulating their saturation magnetization and magnetic anisotropy, which in turn affects the spin wave propagation. The interference patterns of these spin waves can be fine-tuned by applying voltage to the magnets and adjusting the number of ions introduced into them. The device can perform computations by leveraging these dynamic interference patterns through an ion reservoir. This approach achieved a normalized mean square error (NMSE) of 6.41 × 10⁻⁵ and NMSEvar of 9.53 × 10⁻³, representing a performance improvement of 76.6% and 52.7% over conventional devices, respectively.

This device demonstrated exceptional performance in time-series predictions, achieving error rates that were less than one-tenth of previous physical reservoir systems. Its prediction accuracy was evaluated using a standard testing method based on the Mackey–Glass equations, commonly used to model complex variations in biological systems. The device's performance was also benchmarked against traditional neural networks, showcasing competitive results without requiring complex hierarchical adjustments.

Magnetic property manipulation through electronic carrier tuning by ion-gating. a) Iono–magnonic with Y₃Fe₅O₁₂ (YIG) single crystal and Nafion and its experimental configuration. b) Optical microscope image of antennas deposited on the single crystal. c) Scanning transmission electron microscope image of the single crystal and its fast Fourier transform image. d) Spin wave spectroscopy of de-embedded transmission signal, measured at gate voltage V_G of 0.0 V. e) Saturated magnetization M_S and anisotropic field H_a as a function of V_G. f) Electron energy-loss spectra variation of Fe–L₃ of “Unbias” and “Bias”. g) Energy-loss variation at various depths from the YIG/Nafion interface. h) Schematic illustration of spin configuration manipulated by ion-gating.

This technology can be implemented in magnetic thin films and single crystals and miniaturized without performance degradation, making it suitable for various industrial applications. By integrating this system with different sensors, it can enable real-time processing for Internet of Things (IoT) applications and edge computing, offering energy-efficient solutions with high precision. When integrated with different sensors, it can potentially enable energy-efficient, high-precision AI devices for a wide range of purposes.

This project was carried out by a research team consisting of Takashi Tsuchiya (Group Leader, Neuromorphic Devices Group (NDG), Research Center for Materials Nanoarchitectonics (MANA), NIMS), Wataru Namiki (Postdoctoral Researcher, NDG, MANA, NIMS at the time of this project; currently Researcher, MANA, NIMS), Daiki Nishioka (Trainee, NDG, MANA, NIMS at the time of this project; currently Research Fellow, International Center for Young Scientists, NIMS), Kazuya Terabe (Group Leader, Ionic Devices Group, MANA, NIMS), Yuki Nomura (Senior Researcher, Nanostructures Research Laboratory (NRL), JFCC) and Kazuo Yamamoto (Group Leader, NRL, JFCC).

This work was supported by funding from the National Security Technology Research Promotion Fund of the Acquisition, Technology & Logistics Agency of the Ministry of Defense.