AI-Assisted Quantum Control

AI-assisted quantum control uses learned models or policies to choose pulses, measurements or feedback actions. The research question is when learning improves achievable performance or reduces calibration cost relative to a strong conventional controller. A high reward in a simulator is evidence about that simulator and reward, not automatic evidence of robust physical control.

Required background. Optimal control develops control objectives and constraints; pulse-level control connects them to hardware.

The objective must represent the experiment

A state-transfer fidelity can be combined with penalties for leakage, fluence and sensitivity to uncertain parameters. For example,

$$J(u)=1-F(u)+\lambda\int_0^T |u(t)|^2\,dt,$$

subject to amplitude and bandwidth limits. The penalty weight has units and changes the problem; it is not a harmless training parameter. A gate objective should test a channel or a complete input basis rather than a single state.

Model-based gradient methods exploit a differentiable simulator. Reinforcement learning can use sparse rewards and histories. Bayesian or adaptive experimental design can reduce measurement cost. Compare each method with access to equivalent prior information and the same number of physical evaluations.

Evidence and open limitations

Bukov and colleagues compared reinforcement learning with optimal-control methods in simulated interacting-qubit problems and analyzed their control landscapes. Established: learning methods can find useful control policies in specified models. Active: robustness to drift, partial observability, non-Markovian environments and sim-to-device mismatch. Conjectural: that a larger neural policy automatically outperforms conventional control once training and measurement costs are counted. Primary study.

A learned policy can exploit discretization errors, unrealistic pulse edges or omitted leakage. Reward hacking here is a scientific model failure: the optimized number no longer measures the intended operation.

Open questions

Can a controller identify its own model mismatch? How much prior structure should be encoded rather than learned? Can measurement-efficient policies remain robust across calibration episodes? What information from a pulse family transfers between devices without assuming identical Hamiltonians?

Graduate project: robustness under a fixed experiment budget

Compare a learned pulse policy with a gradient-based optimizer on a driven two-level system extended by one leakage level. Impose the same amplitude, bandwidth and duration bounds. Train on a declared detuning distribution and evaluate on held-out drift and noise. A laptop simulator is sufficient for the first stage.

Deliver infidelity distributions, worst-case leakage, total simulator/measurement calls and reproducible pulse files. Failure tests refine the integration step, distort the transfer function and evaluate parameter values beyond the training range. The baseline receives an equal tuning budget. Research begins when a policy improves performance on an independently calibrated device or a justified uncertainty model and retains its gain after training overhead is included.

Entry points and scope

Read the stable control page before adopting learning terminology. Literature checked on 11 October 2026; selected methodological evidence: the 2018 study. A measured robustness or sample-efficiency improvement is an update trigger; simulation reward alone is not.

References

• M. Bukov, A. G. R. Day, D. Sels, P. Weinberg, A. Polkovnikov and P. Mehta, “Reinforcement Learning in Different Phases of Quantum Control,” Physical Review X 8, 031086 (2018). Preprint.
• G. Wilson et al., “Best Practices for Scientific Computing,” PLOS Biology 12, e1001745 (2014). Article — independent implementation and testing of the control simulator.