Artificial neuromorphic devices based on magnetic tunnel junctions

<?A3B2 tlb=1.27h?><?A3B2 th=9.5p?>The development of artificial intelligence technology is intrinsically linked to advancements in hardware technology. Traditional Von Neumann architecture, which separates computational units and data storage units, faces significant bottlenecks in terms of processing speed and power consumption. These challenges, often referred to as the “storage wall” and the “power consumption wall”, pose substantial barriers to further progress in artificial intelligence technology. As the complexity and computational demands of artificial intelligence algorithms continue to grow, overcoming these limitations has become an urgent priority. Neuromorphic devices, which aim to mimic the low-power operation of the human brain, present a promising pathway for addressing these challenges. By emulating the fundamental structural and connectivity features of the brain’s neural systems, these devices enable hardware-based neuromorphic computing that can bypass many of the limitations of traditional architectures. Neurons and synapses, the basic building blocks of biological neural networks, serve as the core components simulated by neuromorphic devices. Among the various technologies being explored for this purpose, magnetic tunnel junctions stand out due to their numerous advantages. As core elements of spintronics, magnetic tunnel junctions offer long operational lifetimes, low power consumption, non-volatility, high speed, and multifunctionality, making them highly suitable for simulating artificial neurons and synapses in neuromorphic systems. Their unique ability to exploit the spin properties of electrons provides a versatile platform for memory and computation functions. Magnetic tunnel junctions can perform a variety of functions by manipulating the movement of magnetic domains. This manipulation adjusts the resistance of the device, enabling both memory and computational capabilities within a single memristor, which is crucial for simulating neural synapses. Furthermore, applying electrical current to magnetic tunnel junctions generates spin-polarized currents. These currents produce spin-transfer torques, which induce oscillations in the free-layer magnetic moments, resulting in high-frequency voltage signals. This property allows for the execution of weighted calculations and nonlinear function simulations based on resonance frequencies, facilitating the emulation of synaptic and neuronal functions. <?A3B2 tlb=1.26h?><?A3B2 th=9.5p?>Additionally, the barrier properties of magnetic tunnel junctions can be modulated through electrical and thermal means. This modulation causes the magnetic moments in the free and pinned layers to transition between parallel and antiparallel states. These resistance fluctuations generate random signals, which are particularly useful for true random number generation and probabilistic computations. By tailoring the applied conditions, these random signals can mimic various neuronal behaviors, further enhancing the functionality of neuromorphic systems. <?A3B2 tlb=1.26h?><?A3B2 th=9.5p?>This paper begins by examining the development challenges faced by current artificial intelligence hardware, focusing on the limitations of traditional architectures. It provides a concise overview of the progress in artificial neuromorphic devices and the materials employed in their fabrication. The fundamental structure and operating principles of spintronic devices, particularly magnetic tunnel junctions, are then introduced. The discussion subsequently delves into the specific roles of magnetic tunnel junctions-based neuromorphic devices, including spin memristors, spin oscillators, and spin random number generators, in simulating artificial neural synapses and neurons. Current research efforts in these areas are reviewed, highlighting the advances made as well as the remaining challenges. Finally, the paper explores the prospects for the development of spintronic-based neuromorphic devices, emphasizing their potential to revolutionize artificial intelligence hardware and pave the way for more efficient and powerful computing systems.