CISE-ANR: FET: Small: Hybrid Stochastic Tunnel Junction Circuits for Optimization and Inference – StochNet

We will develop digitally compatible, low-barrier magnetic tunnel junctions (MTJs), hybrid measurement and control circuits, and architectures to use their natural stochasticity in the context of various biologically inspired probabilistic computational models. This work is motivated by research in the neuroscience community highlighting the inherently stochastic behavior of neural processes. Stochastic MTJs, implemented with small modifications to the commercial back-end-of-line (BEOL) integrated MTJ process, provide a natural substrate for computational models exploiting this randomness. Lowering the barrier close to room temperature puts devices in a superparamagnetic regime, where thermal fluctuations change the device state. We will design and test both rate coded and temporally coded stochastic circuit primitives. Building on experiments, we will design probabilistic computing architectures such as energy-based machine learning models, non-equilibrium oscillatory models, and time-of-arrival based temporal models. Leveraging the cross-stack expertise of our team from material growth and device fabrication to VLSI design and computer architecture, we will bridge the gap between these domains by developing computing abstractions that are informed by the practical challenges of next-generation technology integration while being guided by applications that can make the most of such randomness effectively. Keywords: Magnetic tunnel junctions; Spintronic-CMOS circuits; Markov models; Probabilistic computing; Stochastic devices; Temporal Computing.

Our project follows the grand neuroscientific challenge of reverse-engineering the brain into technological development. We start with the idea that the probabilistic nature of the brain may be fundamental to its computational ability as postulated by researchers, based on the stochastic nature of basic neural processes such as synaptic release. The path from biology to technology involves diverse fields such as mathematics, electrical engineering, computer science, material science, physics, and neuroscience. We will build and test circuit primitives that convert the inherent randomness of superparamagnetic MTJs (SMTJs) into usable encoding schemes. We will then utilize these primitives as building blocks of architectures that are energy-efficient, reliable, and robust to variations by design. The possibility of straightforward integration of our magnetic-tunnel junction-based circuits into commercial fabrication raises the possibility of large scale networks of coupled stochastic devices as a substrate for subsequent investigations into physics, network computation, and neural processes operating on the edge of chaos.

The results of this research will advance knowledge and understanding in the area of emerging technologies for novel computing. The design of digitally interfaced, tunable, and energy-efficient stochastic unit cells will enable the compact physical realization of complex networks exhibiting little-explored non-equilibrium dynamics and unique stochastic temporal behavior. This understanding will be important for the development of low power, distributed intelligence for edge computing where low energy optimization and inference is important. We will disseminate the scientific advances from this research to the public through a variety of avenues such as lab tours for students, parents, government staff, and science journalists. We will conduct seminars at the University of Maryland and CEA (SPINTEC), as well as produce pedagogical YouTube videos in English, French, and Spanish describing our work. Additionally, we will collaborate with institutional summer internship programs that foster a research mindset in high school and undergraduate students to increase participation from the local community. Exposure to diverse scientific topics and international collaborations will be very beneficial to researchers, graduate students, and postdocs hired with requested funds.