Superparamagnets for Probabilistic and Reservoir Computing

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

This research program aims to optimize superparamagnetic tunnel junctions for low power probabilistic and reservoir-based computation. The results will have impact on low energy sensors and hand-held electronic devices, as well as in high performance data encryption and probabilistic decryption. Superparamagnetic tunnel junctions are devices that spontaneously fluctuate between two resistance states, and have a time-averaged resistance that can be tuned using a smaller voltage than that needed for conventional switching, enabling lower power consumption, for example in a smart phone. Probabilistic computing performs logic operations based on combinations of the time-averaged signals. The randomness of the resistance fluctuations and ability to design superparamagnetic tunnel junctions for high speed are important features for cyber security applications. Reservoir computing is a type of hardware-based accelerator for neural networks, and here interacting superparamagnets will be used to form different kinds of reservoir. A unique feature is that only the input and output will require electrical connections, which could dramatically reduce power consumption. The aim of this component of the research program is to quantify the speed and short-term memory for different geometries of superparamagnet arrays, in order to evaluate them for use in artificial intelligence applications. The impact of magnetic reservoir computing would come from a better understanding of the algorithms that enable high energy efficiency and complex processing. A graduate student will develop extensive nanofabrication, high frequency electronics, and machine learning skills. There will be multiple options for undergraduate research projects, and a teaching module for a nanofabrication laboratory will be developed. There are two interconnected thrusts to this research program, both centered on electrical control of superparamagnets. In the first, non-interacting superparamagnetic tunnel junctions are optimized for high average fluctuation rate and low bias voltage tunability of the time-averaged resistance. Multiple tunnel junctions are interconnected with variable feedback in order to demonstrate probabilistic logic gate behavior. The effect of the feedback amplitude and averaging time on the statistical preference for different logic states will be determined, and the power consumption measured, in order to benchmark superparamagnet-based logic devices. The second thrust involves investigation of assemblies of magnetostatically interacting nanomagnets for reservoir computing. They are controlled by the magnetic fringe field of a superparamagnetic tunnel junction input, and their response in picked up by the fringe field generated at a superparamagnetic tunnel junction output. Magnetostatically coupled patterns have previously been used for logic devices, but applications have been limited by the need for an external magnetic field. Here electronic control and detection will be used, enabling high speed operation and making integration with semiconductor electronics easier. Investigation of magnetostatically driven output coupling could eliminate an important bottleneck in the design of high performance magnetic logic devices imposed by the weak signal from the inverse spin Hall effect. By combining superparamagnetic tunnel junctions, electronic feedback, and magnetostatically coupled patterns, the proposed research program will develop an accelerator for machine learning and a toolkit for exploration of reservoir computing. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.