Superconducting Electronics Neural Networks for Wideband RF Signal Processing
Abstract
APPROVED FOR PUBLIC RELEASEWe propose to apply the concept of physical neural networks (PNNs), developed in the PI s group (Wright e,t al. 2021), to implement a machine-learning system ultimately capable of processing extremely wideband (> 50 GHz) radiofrequency (R,F) signals in real time. The key idea of our proposed project is to harness the ultra-low-dissipation dynamics of analog superconduc,to a superconducting circuit without digitization. Two potential use cases include radar signal processing and communications signal, processing.Background: Simulations of superconducting electronics used to perform ~100-Gbps channel equalization (reversal of disto,rtions in a communications system) in the machine-learning paradigm of reservoir computing have shown impressive performance (Rowlan,ds et al. 2021). However, an experimental demonstration has not yet been performed. One of the major difficulties is that reservoir,computing requires multiplication of output signals (interpreted as a vector) by a programmable matrix. This matrix-vector multiplic,ation is difficult to achieve in superconducting electronics, and if instead one aims to read out the signals from the superconducti,ng electronics and perform the matrix-vector multiplication at room temperature in standard digital logic, one is forced to slow dow,n the entire system (obviating the point of using superconducting electronics) or lose information by subsampling (degrading perform,ance). The PNN approach has a number of differences and benefits over reservoir computing, but a crucial practical distinction is th,at a PNN does not involve a matrix-vector multiplication at its output and the entire PNN processing can occur in the analog superco,nducting-electronics domain, avoiding the difficulty that one faces when performing reservoir computing with superconducting electro,nics.Why is this a good idea to do now? There has been a resurgence of interest in analog methods for performing neural-network proc,nt of a number of new methods for training unconventional, analog systems to act as neural networks. We have recently developed a me,thod, Physics-Aware Training, that enabled us to demonstrate canonical image- and speech-processing tasks on hardware that one would, ordinarily suspect incapable of performing any sophisticated task: a sheet of metal that is shaken by an audio speaker; a nonlinear, RLC oscillator comprising just a single resistor, inductor, capacitor, and transistor; and optical pulse propagation through a nonl,inear crystal.In short: we now have the methods to apply the techniques and lessons of the past decade of successes in deep learning, to arbitrary physical hardware, which we did not have until very recently.Why is it important that this work gets done at all? Real,-time RF signal processing is crucial in many areas, including communications and radar, but signal bandwidths continue to increase,, making real-time processing ever more difficult. Merely digitizing signals with 50-GHz bandwidth is challenging, let alone performi,ng sophisticated machine-learning processing on the resulting data streams in real time. Our proposal to construct ,onducting electronics will investigate the ability of analog superconducting circuits to perform machine-learning processing on RF s,ignals directly, without digitization, in real-time. If our research agenda is successful, we will usher in an entirely new capabili,ty in real-time wideband signal processing, enabling new functionality and improved performance in communications and radar systems,, and full spectral awareness for Navy platforms more generally.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 08, 2022
- Source ID
- N000142212352
Entities
People
- Peter L. McMahon
Organizations
- Cornell University
- Office of Naval Research
- United States Navy