🔬 applied physics

Non-Volatile Vortex MTJs with Opto-Electrical and Spin-Diode Nonlinearities as Multifunctional Neuromorphic Platforms

This paper demonstrates a multifunctional non-volatile spintronic platform based on vortex magnetic tunnel junctions that unifies zero-power weight storage with optoelectrically driven cubic nonlinearities and spin-diode effects to enable compact, high-accuracy neuromorphic computing.

Original authors: Felix Oberbauer, Tristan Joachim Winkel, Clara C Wanjura, Maksim Steblii, Jakob Walowski, Tim Böhnert, Ricardo Ferreira, Markus Münzenberg, Tahereh Sadat Parvini

Published 2026-03-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Felix Oberbauer, Tristan Joachim Winkel, Clara C Wanjura, Maksim Steblii, Jakob Walowski, Tim Böhnert, Ricardo Ferreira, Markus Münzenberg, Tahereh Sadat Parvini

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

Imagine you are trying to build a super-smart computer that works like the human brain. The brain is amazing because it doesn't just store memories in one place and do math in another; it does both at the exact same time, right where the memory is. This makes it incredibly energy-efficient.

Current computers, however, are like a librarian who has to run back and forth between the library shelves (memory) and the desk (processor) to get a book and read it. This "running back and forth" wastes a lot of energy and time.

This paper introduces a tiny, revolutionary new device—a Magnetic Vortex MTJ—that acts like a "smart librarian" who can read, write, and calculate all in one spot. Here is how it works, broken down into simple concepts:

1. The Device: A Tiny Magnetic Whirlpool

Think of the device as a microscopic pillar made of special magnetic layers. Inside this pillar, the magnetic atoms swirl around like a miniature hurricane or a whirlpool. This swirling center is called a "vortex."

The Memory (The Weight): The authors added a special "storage layer" to this whirlpool. You can heat this layer up with a tiny electrical pulse to "reset" the direction of the swirl. Once it cools down, it stays in that new direction forever, even if you turn off the power. This is like writing a number on a sticky note that never fades. This number acts as a synaptic weight (how strong a connection is in a neural network).
The Non-Volatile Feature: Unlike standard computer memory that needs constant power to remember things, this device remembers its "weight" without any electricity. It's like a light switch that stays in the "on" position even when you unplug the house.

2. The Two Superpowers: Light and Spin

The magic happens because this single tiny pillar can do two different types of "math" using two different inputs:

Power A: The "Thermal Calculator" (Opto-Electrical)

Imagine shining a laser pointer at the device.

The Input: The brightness of the laser represents the data (like a pixel in a photo).
The Weight: The electrical current flowing through the device represents the "importance" of that data.
The Magic: When the laser heats the device, it creates a voltage (electricity) based on how hot it gets. But here's the cool part: the relationship isn't a straight line. If you double the heat, the electricity doesn't just double; it changes in a complex, curved way (a cubic nonlinearity).
The Analogy: Think of this like a volume knob on a guitar amp. Turning the knob slightly changes the sound a little, but cranking it up creates a distortion that changes the entire character of the sound. This "distortion" is exactly what a brain needs to recognize complex patterns. It acts as a built-in "activation function" (the brain's way of deciding if a signal is important enough to pass on).

Power B: The "Spin Radio" (Spin-Diode)

Now, imagine sending a radio signal (RF) into the device instead of a laser.

The Input: The frequency (pitch) of the radio wave represents the data.
The Weight: The power (loudness) of the radio wave represents the importance.
The Magic: The device acts like a radio tuner. If the radio frequency matches the natural "hum" of the magnetic whirlpool, the device jumps into action and creates a strong electrical signal. If the frequency is off, nothing happens.
The Analogy: This is like a tuning fork. If you hum the exact right note, the fork vibrates loudly. If you hum a different note, it stays silent. This allows the device to filter and process information instantly.

3. The "Four-Level" Secret

Usually, a memory device is like a light switch: it's either ON (1) or OFF (0).
This device is smarter. Because it can be read in two different ways (by measuring the heat-induced voltage and the electrical resistance), it can actually distinguish four different states (like a dimmer switch with four settings: Off, Low, Medium, High). This means one tiny device can hold more information than a standard binary switch.

4. The Results: Smarter and Faster

The researchers tested these devices by simulating a neural network to recognize handwritten numbers (like the digits 0–9).

The Winner: The smaller devices (600 nanometers wide) performed the best, achieving 95.4% accuracy.
Why? Smaller devices have a sharper, more dramatic "nonlinear" response. It's like a smaller, tighter spring that snaps back more dramatically than a loose, big one. This sharp snap is perfect for making quick decisions in a brain-like computer.
Comparison: Their performance was almost as good as a standard digital computer running complex software, but with the potential to be millions of times more energy-efficient.

5. The Future: The "Super-Device"

The most exciting part is that this single device can do both the "Thermal Calculator" and the "Spin Radio" tricks at the same time.

The Vision: In the future, one single tiny pillar could take in a laser signal, a radio signal, and an electrical signal all at once. It would mix them together to perform complex math (like multiplying two numbers and adding a third) in a single instant.
The Impact: This could lead to computers that are not only brain-like but also capable of processing massive amounts of data with almost zero energy waste.

Summary

The authors have built a tiny, magnetic whirlpool that acts as a non-volatile memory (it remembers without power) and a nonlinear calculator (it processes data like a brain). By using lasers and radio waves, they showed that this single device can perform the complex math needed for Artificial Intelligence, potentially paving the way for computers that are as efficient as the human brain.

1. Problem Statement

Modern artificial intelligence faces significant energy and data-movement costs due to the von Neumann bottleneck, where memory and processing are physically separated. Neuromorphic computing aims to solve this by co-locating memory and computation in adaptive physical elements. However, current hardware implementations face three major challenges:

Lack of Non-Volatility: Many neuromorphic devices require standby power to retain state.
Limited Programmability: Devices often lack the ability to continuously tune synaptic weights.
Separation of Functions: State storage and nonlinear computation are typically handled by separate components, increasing complexity and power consumption.

The authors seek a single nanoscale device that unifies non-volatile weight storage, tunable nonlinear computation, and multilevel readout to overcome these limitations.

2. Methodology

The research focuses on CoFeB/MgO/CoFeB vortex Magnetic Tunnel Junction (MTJ) nanopillars integrated with a specific IrMn-based storage layer.

Device Architecture:
- Stack Structure: Includes a Synthetic Antiferromagnet (SAF) reference layer, a free layer hosting a magnetic vortex, and a reconfigurable storage layer (NiFe/IrMn) separated by a Cu spacer.
- Fabrication: Nanopillars with diameters of 600 nm, 800 nm, and 1000 nm were fabricated via electron-beam lithography and ion-beam milling.
- Thermal Programming: A selective thermal window is created using two antiferromagnetic layers with different Néel temperatures ( $T_N \approx 230^\circ\text{C}$ for the reference, $T_N \approx 180^\circ\text{C}$ for the storage layer). Voltage pulses heat the storage layer above its $T_N$ , allowing its magnetization to be reset by an external field, creating a non-volatile, tunable stray field that shifts the vortex resonance frequency.
Measurement & Operation Modes:
- Opto-Electrical Excitation: A 638 nm diode laser (modulated at 1 kHz) provides thermal gradients across the MgO barrier.
- Bias Enhancement: A DC bias current is applied simultaneously with the laser.
- Readout Channels:
  1. TMS (Tunnel Magneto-Seebeck): AC thermovoltage generated by the laser-induced temperature gradient (linear response).
  2. bTMS (Bias-Enhanced TMS): DC thermovoltage generated by the coupling of thermal and electrical drives (nonlinear response).
  3. Spin-Diode Effect: Rectification of RF signals near the vortex gyrotropic resonance into DC voltage.
Simulation:
- The authors constructed a crossbar-array simulation using measured device response maps as interpolation functions.
- They evaluated two operating modes on the same hardware: bTMS mode (optical input, DC bias weights) and Spin-Diode mode (RF frequency input, RF power weights).
- Performance was tested on a handwritten digit recognition task (MNIST) using a single-layer neural network.

3. Key Contributions

Unified Multifunctional Platform: Demonstrated that a single vortex MTJ nanopillar can simultaneously support non-volatile magnetic reconfiguration, opto-electrical thermoelectric transport, and spin-diode rectification.
Discovery of bTMS Nonlinearity: Identified a pronounced cubic nonlinearity in the thermoelectric response when laser heating and DC bias are combined. This serves as a hardware-native "activation function" for analog computation without extra circuitry.
Four-Level Readout Space: Showed that while electrical resistance and thermovoltage switch at the same magnetic coercive fields, they exhibit distinct amplitudes. This allows a single two-state (Parallel/Antiparallel) junction to encode four distinct signal levels, enhancing information density.
Event-Driven Signaling: Observed sharp "Barkhausen-like" spikes in the thermovoltage during magnetization reversal, suggesting a pathway for spike-based, event-driven neuromorphic signaling.
Combined Operating Regime Proposal: Proposed a future mode where bTMS and spin-diode effects act simultaneously, enabling quadratic cross-terms (nonlinear multi-input interactions) within a single nanoscale element.

4. Key Results

Device Performance:
- Non-Volatile Tuning: The storage layer enables continuous, non-volatile tuning of the vortex gyrotropic resonance over a ~15 MHz range with zero standby power.
- Nonlinearity: The bTMS regime exhibits a strong cubic-like response ( $V \propto I^3$ ), ideal for neural activation.
- Size Dependence: 600 nm devices consistently outperformed 800 nm and 1000 nm devices. Smaller pillars confine heat more effectively, creating larger vertical temperature gradients and stronger nonlinear responses.
Neuromorphic Classification Accuracy:
- bTMS Mode: Achieved 95.4% accuracy on the MNIST task.
- Spin-Diode Mode: Achieved 94.9% accuracy.
- Comparison: Both results are comparable to a digital single-layer network with sigmoid activations, validating the hardware's computational capability.
Processing Density Estimates:
- Current footprint (~25 $\mu\text{m}^2$ ): ~0.12 TFLOPS mm $^{-2}$ .
- Industry-realistic footprint (~1 $\mu\text{m}^2$ ): ~3 TFLOPS mm $^{-2}$ .
- This represents a 2–3 order of magnitude improvement over state-of-the-art silicon digital processors (10–50 GFLOPS mm $^{-2}$ ).

5. Significance

This work establishes vortex MTJs as a unified, scalable platform for next-generation neuromorphic hardware. By integrating programmable magnetism (memory), spin caloritronics (opto-electrical computation), and gyrotropic dynamics (rectification) into a single device, the authors address the critical need for non-volatile, low-power, and highly nonlinear computing elements.

The demonstration of hardware-native nonlinearities (cubic bTMS and spin-diode rectification) eliminates the need for external activation circuits, significantly reducing energy overhead. Furthermore, the proposal to combine these effects suggests a path toward quadratic feature interactions (e.g., attention mechanisms) directly in hardware, potentially solving bottlenecks in modern machine learning algorithms that are computationally expensive on digital CPUs/GPUs. The results position spintronic vortex devices as a leading candidate for energy-efficient, brain-inspired computing architectures.