Andre Slavin: MTJ-based artificial neuron suitable for fully analog machine learning

#machine-learning #MTJs #magnetization #dynamics #neuromorphic #computing

Magnetic tunnel junctions (MTJs) are nanoscale spintronic devices composed of two ferromagnetic layers separated by an insulating barrier. MTJs are CMOS-compatible and widely used commercially in magnetic memory and sensor technologies, with over 108 devices manufactured annually. In this work, we demonstrate a practical analog artificial neuron by pairing a spin-valve/magnetic tunnel junction (SV/MTJ) device with an NMOS transistor. This SV/MTJ-nMOS neuron supports fully analog machine learning and is compatible with conventional electronic systems and neural network architectures.

In a recent theoretical study [1], it was shown that an SV/MTJ device can function as an artificial neuromorphic device replicating fundamental behaviors observed in biological neurons, such as short spike generation, response latency, synaptic integration, refraction, inhibition, adaptation, and spike-train generation. By integrating this SV/MTJ device with an NMOS transistor, both input and output signals become voltage-based. This configuration allows straightforward implementation of synaptic weights through voltage-controlled amplifiers, thus enabling seamless integration into fully analog neural networks.

To validate system-level performance, we constructed a three-neuron SV/MTJ-nMOS neural network in LTspice program, complete with supporting circuitry [2]. A feedforward backpropagation algorithm was employed, performing all learning and synaptic weight adjustments entirely in the analog domain using spike timing (time-encoded signals). An example of realization of an XOR logic gate, a standard neural network benchmark, was used to assess learning effectiveness. After approximately 30 training epochs, the network reliably converged, performing XOR classification successfully. The complete training cycle required less than 2 μs of simulation time. These results confirm the potential of the SV/MTJ-nMOS artificial neuron circuit for fast, low-power, fully analog machine learning, compatible with standard electronic design and established neural network frameworks.

References

[1] S. Louis, H. Bradley, C. Trevillian, A. Slavin and V. Tyberkevych, IEEE Magn. Lett., Vol. 15, p. 1–5 (2024)

[2] S. Louis, H. Bradley, A. Litvinenko and V. Tyberkevych, arXiv preprint arXiv:2503.20813 (2025)

• Add Event to Calendar
  iCal
  Google Calendar

• 1420 Austin Bluffs Pkwy
• Osborne Center A204
• Colorado Springs, Colorado
• United States 80918
• Building: OCSE
• Room Number: A204

---

• Contact Event Host
• Co-sponsored by UCCS Department of Physics and Energy Science