Neural Network-Based Frequency Optimization for Superconducting Quantum Chips

Optimizing frequency configurations for qubits and gates in superconducting quantum chips presents a complex NP-complete challenge, critical for mitigating decoherence and crosstalk. This paper introduces a neural network-based approach, leveraging the network as a surrogate model to predict frequency errors. The method employs a closed-loop Bayesian optimization framework to iteratively refine configurations, guided by the network’s knowledge of nonlinear error mechanisms. By focusing on localized chip windows, the optimization identifies optimal frequency settings that minimize errors. The approach is validated through randomized and cross-entropy benchmarking, showing improved energy calculations when optimizing frequency configurations for a crosstalk-aware hardware-efficient ansatz in variational quantum eigensolvers on superconducting quantum chips.