In magnetic memory, the orientation of a magnetic material normally stores binary "0" or "1" information. But the continuous motion of magnetization as it switches direction could, in principle, be used to make a device that emits electrical spikes like a nerve cell. A research team led by Steven Louis built a simple circuit model that connects one NMOS transistor and one magnetic tunnel junction (MTJ) in series, and in a circuit simulator it successfully learned the XOR problem, a basic machine-learning task. Research is advancing on using memory devices as artificial neurons within the "CMOS+X" framework, which combines spintronic devices with existing CMOS circuits.

AD

The "CMOS+X" design: repurposing a memory device as an artificial neuron

Steven Louis and colleagues proposed the "NMOS+MTJ neuron," in which one NMOS transistor and one MTJ are connected in series. The preprint was first posted to arXiv on December 3, 2025 (v1), followed by v2 on December 4, and a revised version (v3) on August 11, 2026. After peer review, it was published in IEEE Transactions on Circuits and Systems I: Regular Papers (Vol. 73, No. 10, pp. 6488–6500, October 2026 issue).

The design does not replace existing CMOS circuits with spintronic devices. It follows the "CMOS+X" approach of combining a new device with conventional voltage-driven circuits. The basic structure is nearly identical to the "one-transistor, one-MTJ (1T-1MTJ)" cell used in MRAM, a non-volatile memory already in commercial use.

The MTJ is used differently from conventional MRAM, however. In typical MRAM, the high and low electrical resistance states of the MTJ record non-volatile binary information. In this design, the continuous rotation of magnetization inside the MTJ is used to generate spike signals equivalent to the firing of a nerve cell.

How the "downward spike" works: from a 5V supply down to 4.85V

The mechanism that produces spikes with just two components lies in the magnetic properties of the MTJ. In the circuit, the input voltage is applied to the gate of the NMOS transistor, and the output voltage is measured at the drain side. Because the NMOS and MTJ are in series, the same current flows through both. This current produces spin-transfer torque (STT), which rotates the magnetization of a magnetic layer inside the MTJ called the free layer. The MTJ's electrical resistance changes with the magnetization angle, and the result is reflected in the drain output voltage.

The simulated waveforms in v3 of the paper show behavior that differs from the 0-and-1 switching of ordinary digital circuits. With a 5.0V supply and a 3.16V bias applied to the gate, about 100µA flows through the circuit and the magnetization settles at an angle of about 80°. In this state the MTJ resistance is about 500Ω and the output voltage stays at 4.95V.

When a short pulse raising the gate voltage to 3.45V is applied at 5ns, the current temporarily increases to about 120µA. Under these bias conditions, the input voltage threshold for generating a spike is about 3.4V; above this, the magnetization crosses the 90° critical angle. The magnetization then completes a 360° precession (rotation) over about 2ns and returns to a stable angle equivalent to the original 80°.

During this single rotation, the MTJ resistance rises from about 500Ω to 1500Ω and then falls back to 500Ω. As resistance increases, the voltage drop across the MTJ grows, so the output voltage against the 5V supply temporarily falls to about 4.85V. In other words, the neuron does not produce an upward pulse or a digital square wave, but a "downward voltage spike" that momentarily dips from 4.95V to 4.85V.

AD

Six neural functions reproduced without additional control circuitry

Using the circuit simulator LTspice, the team confirmed that the NMOS+MTJ neuron reproduces several behaviors resembling biological nerve cells. These are "threshold firing," in which it fires only when a threshold is exceeded; "response delay" between input and firing; a "refractory period," in which the next response is suppressed right after firing; "synaptic integration," which accumulates multiple inputs over time; "inhibition," which suppresses firing; and "frequency modulation," in which firing rate changes with input strength. These behaviors arise from the magnetization dynamics inside the MTJ, so no additional control circuitry is needed on the individual neuron.

The simulation also specifies the MTJ's structure and magnetic properties in detail. The modeled MTJ free layer is an elliptical cylinder with a semi-major axis of 30nm, a semi-minor axis of 20nm, and a thickness of 3nm, giving a volume of 5.65487×10⁻²⁴m³. The compact model extends the conventional two-layer ferromagnet model to three layers, computing the equation of motion for magnetization coupled with the equations of the electrical circuit.

Other main conditions include an NMOS transistor from the LTspice standard library, a 5.0V drain supply, a 5.3mT external magnetic field, a Gilbert damping constant of 0.1, and zero anisotropy field. As the team explains, however, this three-layer model is a simplification of the stacked structure of commercial MTJs, which include features such as synthetic antiferromagnetic pinning layers and low-anisotropy free layers.

An analog circuit that runs one XOR training epoch in 40 nanoseconds

The "on-chip machine learning" in the paper's title was verified with an analog multilayer neural network built in LTspice. The feedforward network to be trained consists of two input-layer neurons and one output neuron, each outputting at most one spike per input pattern. In addition, five encoding neurons are placed: two that convert inputs A and B into spike waveforms, two that supply bias signals to the input and output layers, and one that generates the target spike indicating the correct firing time during training.

The synapses connecting the neurons are represented by voltage amplifiers, and multiple input signals are combined by ideal voltage adders. The team suggests that in real circuits these could be replaced by standard CMOS adder circuits or similar. Training uses a supervised loss function, defined as half the squared difference between the actual output spike time and the target spike time. To reduce this error, synaptic weights are updated by gradient descent based on spike timing.

The timing is also designed in nanoseconds. Each XOR input pattern is allotted 10ns, so one epoch, processing the four patterns in sequence, takes 40ns of simulated time. Each pattern's processing includes inference and a weight update; taking advantage of a typical neuron response delay of under 5ns, error evaluation and weight updates are done in the remaining time.

As training progressed, the synaptic weights began to stabilize at around 16 epochs, and the loss dropped sharply toward zero. At epoch 22, about 900ns into the simulation, the output spike times matched the target times almost exactly for all four XOR input conditions. The total simulated training time is about 1µs. By contrast, the figure in the paper of "about 15 seconds on a typical laptop" is the actual computation time LTspice needed to complete the simulation, not a measurement of training time or inference latency on real hardware of the proposed circuit.

In this experiment, the four XOR patterns were learned under the modeled conditions. However, the network as a whole includes idealized voltage amplifiers and adders as well as functional blocks for learning. The model of the single neuron that computes MTJ magnetization dynamics must be distinguished from the circuit model of the entire network, including the learning process.

AD

Real-device verification and power evaluation remain

The team has previously studied an artificial neuron combining a spin valve (SV) and an MTJ. In a paper published in 2024 (arXiv:2409.09268), they reported simulation results of generating spikes of about 1ns without bias current. The present work adds current control by an NMOS transistor and extends the scope of verification to connecting multiple neurons into a trainable analog network.

Neither the proposed NMOS+MTJ neuron nor the learning network, however, has been fabricated as an actual semiconductor chip. The device behavior and network learning results all come from numerical simulation in LTspice. No measurements of spike waveforms or learning behavior from devices fabricated in the lab have been made.

The paper lists experimental verification, optimization of energy efficiency, and scaling to larger networks as future challenges. Because parts of the synapses and learning circuitry in this network are treated as idealized functional blocks, the power consumption and energy efficiency of the whole system have not been evaluated quantitatively. There are also no data for measured energy consumption per spike, or for judging any power advantage over existing GPUs or CMOS neurons.

First author Steven Louis points out that in real devices, manufacturing variation, thermal fluctuations, differences in temperature and resistance, and self-heating may affect spike timing. The present model does not quantitatively evaluate the effects of these factors, and the successful XOR learning is limited to four input patterns. In the next stage, it will be important to confirm whether a prototype can reproduce the roughly 2ns magnetization precession and the spike waveform dropping to 4.85V, and to measure how device variation and thermal fluctuations affect firing timing and learning performance.