Reinforcement Learning for real-time context-aware gate calibration

Quantum gates on state-of-the-art devices reach average fidelities that are hard to further push by optimal control. One key limitation of current gate calibrations is that they are often computed in a context-agnostic manner, making it difficult to account for specific correlations related to the quantum circuit within which the gates are executed. Recent efforts have aimed to suppress gate errors from spatially and temporally correlated noise, but they often relate to static components, typical of device characteristics such as crosstalk or frequency collisions. We propose a gate calibration protocol designed to encompass cases where noise is circuit-dependent. By deploying a model-free reinforcement learning method for quantum control to context-aware fidelity estimation and real-time parameter inference, we introduce a new paradigm for gate-based quantum computation where each gate in the circuit carries a pulse calibration tailored for the circuit context it appears in.

We further outline how advanced real-time pulse modulation and classical processing enable such scheme with no additional runtime or controller memory costs, paving the way for promising changes in the gate model paradigm, critical for error correction scenarios dealing with advanced decoding and noise profiling. Our findings illustrate the building of efficient bridges between abstraction layers of the quantum computing stack, pushing the limits of near-term quantum devices before reaching fault-tolerance.