🔬 applied physics

Magnonic Full Adder Based on 2D Chiral Magnonic Resonators

This paper demonstrates through micromagnetic simulations that machine learning can be applied to arrays of chiral magnonic resonators to construct a robust magnonic full adder, achieving high classification accuracy with minimal physical output signals when utilizing a multilayer perceptron neural network as the output layer.

Original authors: K. G. Fripp, Y. Wang, O. Kyriienko, A. V. Shytov, V. V. Kruglyak

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: K. G. Fripp, Y. Wang, O. Kyriienko, A. V. Shytov, V. V. Kruglyak

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: A "Brain" Made of Spinning Waves

Imagine you are trying to build a super-fast, energy-efficient calculator. Traditional computers (like your laptop) use electricity flowing through silicon wires. This generates heat and takes up a lot of space.

This paper proposes a different way: Magnonics. Instead of electricity, they use spin waves. Think of spin waves like ripples in a pond, but the "water" is a magnetic material, and the ripples are tiny vibrations of magnetic atoms.

The researchers built a Full Adder. In computer language, a full adder is the basic building block that adds three numbers together (like $1 + 1 + 1$ ) to get a sum and a "carry" bit. It's the arithmetic engine inside every calculator.

The Setup: A Magnetic Pool and Floating Disks

To build this calculator, the researchers created a virtual playground using computer simulations:

The Pool (The Waveguide): Imagine a thin, flat sheet of magnetic material called YIG. It's like a calm pond where waves can travel very far without losing energy.
The Sources (The Inputs): They have three "fingers" (Inputs A, B, and C) that can dip into the pond to create ripples.
- If a finger dips lightly, it's a 0.
- If it dips hard, it's a 1.
The Obstacles (The Reservoir): In the middle of the pond, they placed a grid of 8 tiny, floating metal disks (Permalloy). These aren't just passive rocks; they are "smart" obstacles.
- When a small ripple hits them, they bounce it back gently.
- When a big, strong ripple hits them, they get excited and change their shape slightly, bouncing the wave back in a completely different, complex pattern. This is the nonlinearity—the magic ingredient that makes the system "smart."

The Problem: How to Read the Answer?

When the three input fingers create ripples, the waves crash into the 8 disks, bounce around, interfere with each other, and create a chaotic, beautiful mess of patterns across the pond.

The challenge is: How do we look at this messy pattern and instantly know the answer to the math problem?

The researchers tested two different ways to "read" the result.

Approach 1: The "Simple Translator" (Linear Output)

Imagine you have a team of 3 translators standing at the edge of the pond. They are looking at the water at three specific spots.

How it works: They measure the height of the waves at their spots. They don't try to understand the complex physics; they just add up the heights with a simple formula (like $2 \times \text{Spot 1} + 1 \times \text{Spot 2} - 3 \times \text{Spot 3}$ ).
The Result: They found that if you use 3 specific spots, the math works perfectly. The simple formula can tell the difference between a "Sum of 1" and a "Sum of 0."
Pros: It's very simple and cheap to build.
Cons: It's a bit fragile. If the wind blows (noise) or the fingers dip slightly too hard or too soft (input errors), the translators might get confused and give the wrong answer.

Approach 2: The "Super-Intelligent AI" (Nonlinear Output)

Now, imagine you only have one translator standing at a single spot on the edge of the pond. But this translator is connected to a super-computer brain (a Neural Network).

How it works: The translator sees the wave pattern at just one location. To a human, it looks like random noise. But the AI brain has been trained to recognize subtle, complex patterns in that single signal. It's like a master chef who can taste a single spoonful of soup and tell you exactly which 20 spices were used, even if the flavors are mixed up.
The Result: Even with just one reading, the AI gets the answer right almost 100% of the time, even when the wind is blowing hard (high noise).
Pros: Extremely robust. It can handle messy, noisy data and still get the math right.
Cons: It requires a complex electronic brain (the AI) to do the reading, which uses a bit more power than the simple translator.

The "Aha!" Moment: Why This Matters

The paper shows that nature (the magnetic disks) does the hard work of mixing the information. The "reservoir" of disks creates a complex interference pattern that holds all the answers.

The Linear Approach is like trying to solve a puzzle with a simple ruler. It works if the puzzle is clean, but fails if it's messy.
The AI Approach is like using a detective with a magnifying glass. It can solve the puzzle even if it's covered in mud.

The Takeaway

This research proves that we can build logic circuits (like adders) using magnetic waves and tiny disks instead of just silicon chips.

If you want simplicity, you can use a few sensors and simple math.
If you want reliability in a noisy world, you can use a single sensor paired with a smart AI.

This is a step toward "Beyond Von Neumann" computing—computers where memory and processing happen in the same place, using waves instead of electricity, potentially leading to devices that are faster, smaller, and use a fraction of the energy of our current smartphones.

1. Problem Statement

As computational demands increase, conventional CMOS architectures face bottlenecks due to the non-locality of memory and computation, leading to high energy consumption and performance penalties. To address this, "beyond von Neumann" architectures, specifically neuromorphic computing, have been proposed. While artificial neural networks (ANNs) can be simulated on CMOS, they are more efficient on bespoke hardware that integrates memory and computation.

Magnonics (the science of spin waves) offers a promising platform for such hardware due to:

Low energy consumption (magnons are quanta of spin waves).
Low damping in materials like Yttrium-Iron Garnet (YIG).
Inherent strong nonlinearity in resonant structures.

The specific challenge addressed in this paper is the realization of a magnonic full adder (a fundamental logic circuit for arithmetic operations) using a reservoir computing approach. The goal is to demonstrate how a nonlinear physical system (a "reservoir") can process spin-wave inputs to produce logical outputs (Sum and Carry-out) with high robustness against noise, minimizing the complexity of the required readout layer.

2. Methodology

The authors utilized micromagnetic simulations (using the MuMax3 software) to model a magnonic device. The methodology involved three main components:

A. Device Architecture

Substrate: A 50 nm thick YIG waveguide (3.2 μm width, 10.24 μm length) biased by a 12 mT in-plane magnetic field.
Reservoir: A $2 \times 4$ array of 2D chiral magnonic resonators situated 40 nm above the waveguide. Each resonator is a 60 nm diameter, 10 nm thick Permalloy (NiFe) disk.
Inputs: Three spin-wave sources (A, B, C) representing the two bits and the carry-in bit. These are excited by local, circularly polarized harmonic fields with Gaussian spatial profiles at 1.8 GHz.
Global Bias: A global microwave field (0.09 mT, 1.8 GHz) is applied to enhance performance by coupling directly to the disks and waveguide edges.
Nonlinearity: The Permalloy disks exhibit a nonlinear edge mode resonance. As the excitation amplitude increases, the resonance frequency shifts, causing nonlinear scattering of the propagating spin waves. This creates complex, input-dependent interference patterns.

B. Training and Output Layers

The study compared two approaches for the "output layer" (the mechanism converting physical signals to logical outputs):

Linear Output Layer:
- The dynamic stray magnetic field ( $B_z$ ) is sampled at $N$ physical locations.
- Signals are convolved with a Gaussian kernel to mimic antenna size.
- A linear weighting vector ( $W$ ) is optimized to map the sampled signals to the target truth table (Sum and Carry-out) using a least-squares minimization approach.
Nonlinear Output Layer (MLP):
- A single physical sampling point is used.
- The complex signal (Real and Imaginary parts) is fed into a Multilayer Perceptron (MLP) neural network.
- The MLP has two hidden layers (64 and 32 neurons, ReLU activation) and an output layer of 16 neurons (8 for classification, 8 for error correction).

C. Evaluation Metrics

Accuracy: Ability to satisfy the full adder truth table.
Robustness: Performance degradation under input noise (random variations in field amplitude) and output noise (added to the signal).
Sampling Efficiency: The minimum number of physical sampling points required for valid operation.

3. Key Contributions

Demonstration of a Magnonic Full Adder: The first proposal of a 1-bit full adder using a 2D array of chiral magnonic resonators as a nonlinear reservoir.
Comparison of Readout Strategies: A rigorous comparison between a simple linear weighting scheme and a complex neural network readout, quantifying the trade-off between hardware complexity and noise resilience.
Optimization of Sampling Points: Determining that three physical sampling points are sufficient for a linear readout, while one point is sufficient when paired with an MLP.
Noise Resilience Analysis: Quantifying the device's ability to function under realistic noise conditions (thermal fluctuations, fabrication imperfections).

4. Results

Linear Output Layer Performance

Sampling Requirement: A minimum of three sampling points per logical output (Sum and Carry-out) is required to achieve valid logic levels.
Accuracy: With 3 points, the device achieves correct logic outputs. Increasing the number of points reduces error further, but 3 is the practical minimum.
Noise Robustness: The linear configuration is robust to input amplitude perturbations of approximately 10%. However, when output signal noise is introduced, the classification accuracy drops significantly (to ~60%) as the noise parameter increases.

Nonlinear (MLP) Output Layer Performance

Sampling Requirement: The MLP approach successfully classifies the logic state using data from a single physical sampling point.
Accuracy: The MLP achieves nearly perfect classification accuracy (95–100%).
Noise Robustness: The MLP maintains high accuracy even with substantial noise (noise parameter up to 5.0). In contrast, the linear model's accuracy degrades rapidly under similar noise conditions.
Mechanism: The MLP effectively extracts the complex, nonlinear information embedded in the single signal's phase and amplitude, which a simple linear weighted sum cannot resolve under noisy conditions.

Physical Observations

The interference patterns of the scattered spin waves are highly non-uniform in amplitude and phase, even at low excitation levels.
The "000" input state produces a baseline interference pattern, while "111" shifts the interference maxima due to nonlinear frequency shifts in the resonators.
The device operates at low power levels (excitation amplitudes $\approx$ 0.45 mT), suggesting high energy efficiency.

5. Significance and Conclusion

This work demonstrates that magnonic reservoir computing is a viable path toward low-power, high-speed neuromorphic logic circuits.

Trade-off: The study highlights a critical trade-off:
- Linear Readout: Simpler hardware (passive antennas or simple envelope detectors) but requires more sampling points and offers lower noise resilience.
- Nonlinear Readout (MLP): Requires a small external electronic circuit (CMOS-based MLP) but allows for minimal sampling points (1) and exceptional noise resilience, making it more robust for practical implementation where fabrication imperfections and thermal noise are unavoidable.
Future Outlook: The results suggest that while the magnonic system provides the heavy lifting of nonlinear computation, a hybrid approach (magnonic reservoir + electronic neural network readout) may be the most practical route for realizing complex arithmetic logic units (ALUs) in future computing architectures. The methodology can be extended to more complex multi-input/multi-output devices.

In summary, the paper successfully bridges the gap between theoretical magnonic physics and practical logic circuit design, proving that machine learning techniques can optimize the readout of physical spin-wave systems to perform complex arithmetic tasks with high efficiency and robustness.