Black-Box Optimization of Superconducting Circuits Using Reduced-Complexity Neural Networks

Abstract

Single-flux quantum (SFQ) logic based high-speedperiodic-threshold flash converter circuits require multiple non-linear and correlated parameters tuned precisely to function optimally. These parameters cannot be pre-determined from simulations and change with the clock frequency, thus requiring manual optimization every time the clock frequency is changed. In this work, we demonstrate automated bias optimization of an 8-bit superconducting SFQ flash ADC for the first time. A closed-loop 4X lower computational complexity hybrid particle-swarm-gradient-descent-optimization is demonstrated that first uses particle swarm optimization (PSO) to coarse tune the ADC biases and then applies gradient descent (GD) optimization for fine tuning. This results in a 12X reduction in the blind calibration time from several days manually, to a few hours using the proposed optimization scheme, with a resultant performance 2 dB better than an optimization done by a highly skilled human.