Reinforcement learning for optimization of fluxonium two-qubit gates*

Among superconducting qubits, the fluxonium is a promising alternative to the transmon for gate-based quantum information processing. Here we will discuss our recent demonstration (arXiv:2304.06087) of an architecture for fluxonium-fluxonium two-qubit gates mediated by transmon couplers (FTF, for fluxonium-transmon-fluxonium). Relative to architectures that exclusively rely on a direct coupling between fluxonium qubits, FTF enables stronger couplings for gates using non-computational states while simultaneously suppressing the static controlled-phase entangling rate (ZZ) down to kHz levels, all without requiring strict parameter matching. We implemented FTF with a flux-tunable transmon coupler and demonstrated a microwave-activated controlled-Z (CZ) gate whose operation frequency could be tuned over a 2 GHz range, adding frequency allocation freedom for FTF in larger systems. To optimize this gate, we implemented model-free reinforcement learning of the pulse parameters to boost the mean gate fidelity up to 99.922±0.009%, averaged over roughly an hour between scheduled training runs. In this talk, we will discuss the details of this calibration procedure and opportunities for improvements.