🔬 physics

Spin Vector Control for Heisenberg-Inspired Probabilistic Computing

This paper demonstrates a scalable spintronic platform using dual ferromagnetic spin injections into monolayer graphene to achieve continuous vector control of spin accumulation, thereby enabling the direct implementation of Heisenberg-inspired probabilistic computing architectures for low-power, non-Boolean applications.

Original authors: Zhihong Chen, Yuanqiu Tan, Rahul Tripathi, Saleh Bunaiyan, Ryan Wagner, Neil Dilley, Kerem Camsari, Joerg Appenzeller

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhihong Chen, Yuanqiu Tan, Rahul Tripathi, Saleh Bunaiyan, Ryan Wagner, Neil Dilley, Kerem Camsari, Joerg Appenzeller

Original paper licensed under CC BY 4.0 (https://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 the world of computers as a bustling city where information travels like cars on a highway. For decades, this city has run on a strict rulebook: every car is either a "1" (go!) or a "0" (stop!). This binary system works well, but as our cities get bigger and our traffic gets heavier, the old roads are clogging up, and the fuel (energy) needed to keep everything moving is running out. Scientists are now looking for a new kind of traffic system, one that doesn't just rely on "on" and "off" switches but uses the natural, wobbly, and probabilistic nature of the universe to solve problems faster and with less energy.

At the heart of this new idea are tiny magnetic particles called "spins." Think of a spin not as a spinning top, but as a tiny compass needle that can point in any direction, not just North or South. In the old binary world, we force these needles to point only Up or Down. But in the new "probabilistic" world, scientists want to let these needles point anywhere—North-East, South-West, or anywhere in between. This allows computers to handle complex puzzles, like training artificial intelligence or solving massive optimization problems, by mimicking how nature naturally finds the most stable, energy-efficient state. The big challenge has been: how do we get these tiny compass needles to point exactly where we want them to, and how do we make them talk to each other to add up their directions?

This paper is about a team of researchers who built a tiny, high-tech playground to answer that question. They created a device using a single layer of graphene (a material as thin as a sheet of paper, made of carbon atoms) as a smooth road for these magnetic compass needles to travel on. Instead of just pushing one needle at a time, they set up two "injectors" (like two water hoses) that shoot streams of spinning compass needles into the graphene road from different angles. One hose shoots needles pointing at a 45-degree angle, and the other shoots them at a 135-degree angle.

The magic happens when they turn the knobs on these two hoses. By adjusting how much current (water pressure) comes out of each hose, they can make the two streams of compass needles mix together right on the graphene road. If they turn up the first hose, the mixed stream points closer to 45 degrees. If they turn up the second, it points closer to 135 degrees. But here is the cool part: if they turn both on, the needles don't just pick one direction or the other; they actually perform a "vector sum." It's like if two people push a shopping cart from different sides; the cart doesn't go straight left or straight right, but moves in a new direction that is the perfect blend of both pushes.

The researchers found that they could control this resulting direction with incredible precision. They showed that by simply changing the ratio of the electrical currents, they could make the spin direction rotate smoothly and continuously between the two angles. They even proved that when the two streams were balanced in a specific way, they could cancel each other out, making the signal disappear, just like two waves meeting and flattening the water. This wasn't just a guess; they measured it with sensitive equipment at very cold temperatures (20 Kelvin) and confirmed it with computer simulations that acted like a digital twin of their experiment.

What does this mean? It means they have built the first experimental proof that we can do "vector math" with spins in real life. Instead of just flipping a switch, we can now steer a magnetic state to point anywhere in a circle. This is a crucial step toward building the next generation of computers that use these "probabilistic bits" (or p-bits) to solve problems that are too hard for today's machines. While this specific experiment was done in a lab at very low temperatures, it lays the groundwork for future devices that could one day run complex AI models or optimization tasks with a fraction of the energy we use today. The paper doesn't claim to have built a full computer yet, but it has successfully demonstrated the fundamental steering mechanism that such a computer would need to work.

Technical Summary: Spin Vector Control for Heisenberg-Inspired Probabilistic Computing

Problem Statement
Conventional computing architectures face bottlenecks as transistor scaling reaches physical limits, particularly for energy-intensive tasks like machine learning and combinatorial optimization. While probabilistic bits (p-bits) based on binary Ising models have shown promise in spintronic hardware, they are limited to discrete states. A critical challenge in advancing probabilistic computing is realizing and manipulating vector spin information. Such capabilities are essential for mapping complex energy-based models, specifically the classical Heisenberg Hamiltonian, which requires continuous spin states rather than binary projections. Previous spintronic implementations, such as spin majority gates, have largely captured only scalar projections of spin polarizations, failing to exploit the full vectorial nature of spin currents required for richer logic and continuous-variable computation. Furthermore, the controlled vector summation of multiple spin polarizations within a single nonmagnetic channel remains an unresolved experimental challenge.

Methodology
The authors present a spintronic platform utilizing monolayer graphene channels to achieve real-space vector summation of spin currents. The experimental setup involves lateral non-local spin valve (NLSV) devices fabricated with the following key features:

Device Architecture: A monolayer graphene channel connects ferromagnetic (FM) permalloy (Py) electrodes. To address conductance mismatch and improve spin injection efficiency, a thin Al2O3 tunneling barrier is formed between the FM contacts and the graphene.
Dual Injection Geometry: The core innovation is a "tilted dual injection" design. Two injector electrodes (Py1 and Py2) are patterned with needle-shaped edges to mitigate demagnetizing fields and are oriented at 45° and 135° relative to the easy axis of a third, fixed detector electrode (Py4).
Electrical Control: Spin currents are injected simultaneously from Py1 and Py2 using synchronized sinusoidal current sources. By independently tuning the magnitudes of the injection currents ( $I_1$ and $I_2$ ), the authors modulate the relative weight of the spin vectors injected from each electrode.
Characterization: Spin transport is measured at 20 K using lock-in techniques to detect non-local resistance ( $\Delta R_{NL}$ ). Magnetic Force Microscopy (MFM) is employed to track the magnetization evolution of the tilted injectors under varying in-plane magnetic fields. Theoretical validation is provided through analytical vector summation models and HSPICE spin-circuit simulations based on a generalized 4×4 spin-circuit approach.

Key Contributions and Results

Experimental Realization of Vector Summation: The study demonstrates the first comprehensive experimental realization of real-space addition of two spin vectors in a nonmagnetic medium. Unlike previous scalar approaches, this device performs true vector addition, where the resulting spin accumulation vector's direction and magnitude are continuously tunable via electrical current ratios.
Continuous Angular Tunability: By varying the ratio of injection currents ( $I_1/I_2$ ) while maintaining a constant total current, the authors achieve continuous control over the spin polarization angle. The resultant spin vector rotates smoothly between the magnetization directions of the two injectors (45° and 135°).
Signal Modulation and Cancellation: The non-local spin signal ( $\Delta R_{NL}$ ) exhibits a cos-1 dependence on the spin vector angle relative to the detector axis. The experiments reveal a regime of geometric signal suppression near 90°, where the x-components of the spin vectors from the two injectors cancel each other out, minimizing the measurable signal without requiring spin dephasing. This confirms the vectorial interplay between non-collinear spin currents.
Validation of Models: The experimental data shows excellent agreement with theoretical predictions derived from vector summation models and HSPICE simulations. The results validate that the spin accumulation follows coherent addition of non-collinear spin polarizations.
Robustness Across Configurations: The vector summation behavior is validated across different injector geometries, including the primary 45°/135° configuration and an alternative 15°/165° setup, confirming the generality of the method.

Significance and Claims
The paper claims to establish a foundational building block for next-generation analog and probabilistic spin logic architectures. By demonstrating that spintronic devices can directly implement vector-based summation, the work provides a pathway to map classical Heisenberg models using stochastic low-barrier magnets (LBMs). This capability is posited as essential for hardware-efficient implementations of energy-based models, such as training modern Hopfield networks, which are expected to consume significantly more energy if implemented digitally compared to spintronic realizations.

The authors assert that this scalable pathway, based on two-dimensional materials, offers new opportunities for low-power, non-Boolean computing architectures. The work does not claim to have fully realized a Heisenberg-based computer but rather provides the critical experimental proof-of-concept for the vector control mechanisms required to build such systems. Future work is identified as focusing on room-temperature operation and the integration of stochastic spin vectors from LBMs to realize functional, reconfigurable spin-vector processors.

More like this