A Machine Learning Approach to Superconducting Circuit Design

Superconducting qubits are electrical circuits comprising inductors, capacitors, and Josephson junctions. Although circuit parameters may in principle assume a wide range of values, existing architectures generally span only a small subset of the available design space, and design variations are often based largely on experience and engineering intuition. However, future circuit architectures – such as radically new qubit designs or many-body couplers for adiabatic quantum computing – will likely extend beyond an intuitive regime. Towards this end, we have developed a numerical search engine that screens the parameter design space to achieve circuits with desired properties. The engine performs Bayesian optimization on a graphical model of arbitrary circuits with several nodes. We present novel circuit designs with targeted spectral properties, robustness to parameter variation, and noise sensitivity. We expect our work to facilitate the design of more complex and robust superconducting circuit architectures.