Leveraging Domain Walls and Stochasticity in Ferromagnetic Materials for Unconventional Computing

Future unconventional computing systems will require computers with high energy efficiency, low compute-memory bottlenecks, can be immersed in harsh conditions, are adaptive to sensory inputs, and can perform analog computations on the edge. A leading material class to tackle these extreme conditions are ferromagnetic materials and associated devices based on magnet tunnel junctions (MTJ). Devices made from magnetic materials, for example spin transfer torque magnetic random-access memory (STT MRAM), have relatively low switching energy, are nonvolatile, can be directly integrated with silicon CMOS, and are robust to radiation. They also have dynamical behaviors that can be leveraged for new computing paradigms, such as neuromorphic computing, for real-time training and inference on the edge. They also can sense electromagnetic fields, which could eventually lead to sensor-memory-processing combined analog systems.<br/><br/>We will present our results on leveraging domain walls (DWs) and stochasticity in ferromagnetic materials for these applications. We will show experimental results on how DWs integrated into MTJs (DW-MTJs) can act as both artificial synapses and artificial neurons [1]. We will show our results applied the devices to neural networks, as well as to a nearer-term application of analog content addressable memory (aCAM). We will show how the stochasticity of the DW motion aids in online learning without forgetting; how the stochasticity of the MTJ can be used as a random number generator with many knobs to control its randomness; and how the combination of both stochastic DWs and stochastic MTJs can be used to implement efficient Bayesian neural networks [2]. These results elucidate the wide design space for using ferromagnetic materials for unconventional computing.<br/><br/>References:<br/>[1] T. Leonard, S. Liu, M. Alamdar, H. Jin, C. Cui, O. G. Akinola, L. Xue, T. P. Xiao, J. S. Friedman, M. J. Marinella, C. H. Bennett, and J. A. C. Incorvia. “Shape-Dependent Multi-Weight Magnetic Artificial Synapses for Neuromorphic Computing.”<i> Advanced Electronic Materials, </i>2200563 (2022). &lt;DOI 10.1002/aelm.202200563&gt;.<br/>[2] S. Liu, T. P. Xiao, J. Kwon, B. J. Debusschere, S. Agarwal, J. A. C. Incorvia, and C. H. Bennett. “Bayesian Neural Networks Using Magnetic Tunnel Junction-Based Probabilistic In-Memory Computing<i>.” Frontiers in Nanotechnology </i>4: 1021943 (2022). &lt;DOI 10.3389/fnano.2022.1021943&gt;.