KR-20260066682-A - Spin-Wave-Based Neuromorphic Device Using Ternary Weights and Integrated Neuromorphic Computing Array

Abstract

The present invention relates to a neuromorphic device that physically implements triadic weights (+1, 0, -1) of an artificial neural network by utilizing the phase interference phenomenon of spin waves (Magnone) within a magnetic thin film, and a neuromorphic computing array integrating the same. The spin wave modulator includes a Mach-Zehnder type magnetic thin film interferometer structure and controls the phase difference into three stable states: 0 degrees (+1, constructive interference), 180 degrees (-1, destructive interference), and 90 degrees (0, intermediate amplitude). A physical mechanism is provided in which the 90-degree phase spontaneously stabilizes under competitive magnetic anisotropy conditions, and various embodiments are included that support current, voltage, and optical input methods. Multiple modulation cells form a crossbar array to perform vector-matrix multiplication in parallel within a single spin wave propagation cycle, thereby simultaneously improving inference accuracy, energy efficiency, and noise immunity compared to a binary spin wave device.

Inventors

• 안범주

Assignees

• 안범주

Dates

Publication Date

20260512

Application Date

20260415

Claims (1)

1. In a neuromorphic device that computes a signal using a spin wave (Magnnon), which is a collective excitation phenomenon of spins within a magnetic medium, A spin wave generator that generates spin waves having a specific frequency and phase; A spin wave modulation unit comprising a magnetic thin film pattern and modulating the phase of a passing spin wave by controlling the local magnetization direction or effective magnetic field of the magnetic thin film; and A spin wave detection unit that detects the amplitude or phase of a spin wave that passes through the modulation unit and causes mutual interference; is included, The above modulation unit adjusts the phase difference of the passing spin wave, A first state (+1) exhibiting maximum amplitude due to constructive interference with a phase difference of 0 degrees; A second state (-1) exhibiting minimum amplitude due to destructive interference with a phase difference of 180 degrees; and A neuromorphic device characterized by expressing the triadic weights of an artificial neural network through three stable states: a third state (0) in which the phase difference is superimposed at 90 degrees and represents an intermediate amplitude of the maximum and minimum amplitudes.

Description

Spin-Wave-Based Neuromorphic Device Using Ternary Weights and Integrated Neuromorphic Computing Array The present invention relates to a neuromorphic device that physically performs neural network operations using spin waves (SW) or magnons generated by the collective excitation phenomenon of spins within a magnetic medium. More specifically, the invention relates to a neuromorphic device that directly represents ternary weights +1, 0, and -1 in an artificial neural network at the hardware level by controlling the phase of the spin waves into three stable states of 0 degrees, 90 degrees, and 180 degrees, and to a neuromorphic computing array that integrates the same into a crossbar array structure. Artificial Neural Networks (ANNs) demonstrate excellent performance in various artificial intelligence applications, such as pattern recognition, natural language processing, and image classification, and their implementation requires both large-scale parallel computing and low power consumption. Traditional von Neumann architecture-based CMOS processors have a fundamental limitation in that memory and computing units are physically separated, leading to memory bandwidth bottlenecks (memory walls) and excessive energy consumption associated with data movement during neural network inference. As an alternative to overcome these limitations, the neuromorphic computing paradigm, inspired by the operating principles of the biological brain, is attracting attention. Neuromorphic computing adopts an in-memory computing approach that locally stores synaptic weights in non-volatile devices and performs computation and memory at the same location, thereby fundamentally eliminating energy waste associated with data transfer. While various non-volatile devices, such as Resistive Random Access Memory (ReRAM), Phase Change Memory (PCM), and Ferroelectric Random Access Memory (FeRAM), have been studied as neuromorphic synapses, these devices harbor unresolved physical limitations in terms of cycle endurance, inter-device variability, and analog linearity. In the field of spintronics, it is widely recognized that utilizing not only the electric charge but also the spin degrees of freedom for information processing enables the realization of devices capable of simultaneously satisfying the three requirements of non-volatility, high speed, and low power consumption. In particular, spin waves, which are a phenomenon in which the collective precession of spins within a magnetic medium propagates as waves, can operate across a broadband frequency range from GHz to THz. Furthermore, since they transfer energy without charge transfer, they offer the advantage of not generating Joule heating in principle. Additionally, spin waves simultaneously possess phase and amplitude information, which are inherent properties of waves; consequently, constructive or destructive interference occurs when multiple spin waves overlap at a specific point, and this interference characteristic provides natural analog computation capabilities. Previous studies on spin-wave-based logic and computational devices commonly employ a Mach-Zehnder interferometer (MZI) structure in which spin waves are branched from a pair of waveguides composed of magnetic thin films, phase-modulated, and then rejoined. These studies have focused on implementing binary logic gates by controlling the phase difference of spin waves to only two states: 0 degrees (constructive interference, maximum amplitude) and 180 degrees (destructive interference, minimum amplitude). Research implementing weights of artificial neural networks using spin-wave systems has also been limited to this binary framework, so weights could only have two values: +1 or -1. However, binary weights present a problem in that they severely limit the representational capacity of neural networks. Neural networks composed of binary weights exhibit significantly lower inference accuracy compared to floating-point weight networks, and they suffer from the disadvantage of inevitably high switching energy because weight updates always involve a complete magnetization inversion from +1 to -1. Furthermore, phase noise in a binary environment directly leads to sign-flip errors, causing component-level noise to induce critical errors in the neural network output. Meanwhile, Ternary Neural Networks (TNNs), which have been studied in the software field, have been verified to maintain inference accuracy close to that of floating-point weighted neural networks while limiting weights to three values: +1, 0, and -1. In particular, the presence of zero weights can physically disable the corresponding synaptic path during computation, which critically contributes to energy savings and naturally induces a sparse distribution of weights in deep neural networks. However, implementing these ternary weights in digital CMOS hardware still results in additional digital circuit overhead, which limits energy efficiency. Therefore, there is