Oakland University simulates an artificial neuron from a magnet and a transistor
Researchers at Oakland University say a magnetic tunnel junction plus one NMOS switch reproduces spiking behavior, though the machine-learning demonstration ran in simulation
By Letizia Giugliano
3 min read

Engineers at Oakland University in Michigan have designed an electronic neuron built from two parts: an ordinary transistor and a nanoscale magnetic junction. They say neuron-like behavior, including the timing of electrical spikes, comes from the physics of the magnet rather than from added control circuitry. The team published the work in IEEE Transactions on Circuits and Systems.
How a magnetic junction produces a spike
One half of the circuit is a magnetic tunnel junction (MTJ), the nanoscale invention already used in magnetoresistive random access memory, or MRAM. It consists of magnetic layers split by an extremely thin insulating barrier, and one of those layers can rotate when current flows through it. The other half is a single NMOS transistor, the on-off switch found throughout modern chips, which sets the current passing through the magnet. Electrons carry a property called spin, a small magnetic orientation, and as they move they apply a twisting force known as spin-transfer torque to the magnetic layer. When the layer turns, the junction's resistance changes and the voltage in the circuit shifts. Under the right conditions that produces the short pulse, or spike, a neuron uses to signal.
“The human brain uses 20 watts of power to work; the typical AI model needs about 500 times that for inference and another 50 times for each question you ask it.”
— Dr. Steven Louis, lead researcher, Oakland University
“That's too much power for everyday devices. We want AI to work locally, without relying on massive data centers.”
— Dr. Steven Louis, lead researcher, Oakland University
Thresholds and a refractory period without extra circuits
Biological neurons add up incoming signals, fire only once the total crosses a threshold, then pause briefly before they can fire again. The Oakland simulation reproduced threshold spiking, latency, refractory periods, synaptic integration and inhibition. A conventional electronic neuron usually needs many transistors and dedicated control logic to do the same. Here, the team says, those behaviors fall out of the magnetic dynamics of a single junction. Steven Louis, who leads the group, frames the target as computational: the circuit copies the behavior that makes neurons useful, not the biology behind them.
Read also on The Patent: Hamster cells keep algal chromosomes for 155 generations · Auckland researchers test a voice AI coach for type 2 diabetes
Photo: Fabio Sasso / Unsplash
An XOR test that ran in software
Networks of the simulated neurons were given the XOR gate, a standard machine-learning benchmark for artificial intelligence, and learned it by using the timing of spikes instead of conventional digital signals. The result is a small demonstration, and the limit is clear: the work is simulation-based. The paper describes no fabricated chip and no measured power figure for the complete circuit. TechXplore reported the study on November 17, 2026.
20 watts against 500 times as much
The case for the design is energy. Louis puts the human brain at 20 watts, and says a typical AI model needs about 500 times that for inference, plus another 50 times for each question asked. Conventional machines keep memory and processing apart: a processor fetches data, operates on it and writes the result back. GPUs make that flow efficient, but shuttling data between memory and compute still costs power. Neuromorphic hardware tries to remove the trip by holding a state in a physical device that stores information and takes part in the calculation. In this circuit the junction's changing magnetization helps decide when the neuron fires; it is not just being read.
The record names no patent filing
The publication reports a device concept and simulation results. It names no patent application, no granted patent and no licensing agreement with a startup, and the source gives no sign that Oakland University has sought protection on the design. The intellectual property position is therefore unstated. What the inventors present as their innovation is the use of an MTJ's own dynamics, rather than added circuitry, to generate neuron-like spiking. Whether that ends up in a filing is not something the sources say, though the concept could interest fields from robotics to space.
AI-assisted article



