Physics-Informed Machine Learning for Predicting the Robustness of Rydberg-Atom Controlled-Z Gate Pulses

Abstract Neutral-atom quantum computers entangle qubits with laser pulses that must stay accurate despite shot-to-shot laser and motional errors. Choosing among candidate pulses therefore requires a robustness score: the expected gate infidelity under a distribution of errors. Computing this score exactly takes many simulations per pulse. This paper asks a deliberately narrow question: for a single-pulse Rydberg-blockade controlled-Z (CZ) gate, when does machine learning (ML) predict this score better than a cheap physics baseline, and when does it not? Using a simulator validated against a published time-optimal pulse (infidelity 1.6 × 10⁻⁵ without decay and 1.17 × 10⁻³ with decay, against 1.13 × 10⁻³ reported), we label 3,000 training and 2,000 test pulses with an exact quadrature of the expected infidelity and compare four predictors: A second-order perturbative baseline (M0). Gradient-boosted trees on raw pulse parameters (M1). Trees on raw plus physics features (M2). The perturbative baseline plus a learned residual (M3). All ML predictors are given split-conformal 90% intervals. The results are mostly negative for ML at realistic parameters and positive only in a regime map. At the measured error strengths, the perturbative baseline already predicts log₁₀ R with an RMSE of about 10⁻⁴, whereas raw ML is about 2,500 times worse. ML helps once quasi-static errors reach several percent (RMSE of M0 grows to 0.117 at σ = 0.15, while M2/M3 stay near 0.013), or when noise is correlated in time. Physics features are essential, since raw ML only overtakes the baseline at the largest error strength we tested. Designs optimised at infinite blockade lose rank correlation (Spearman 0.99 → 0.50) as the blockade strength drops to 8 Ω. A shift test shows that the conformal intervals lose their coverage on pulses from other generators, and fair search baselines show that ML screening offers no advantage over ranking by nominal infidelity at realistic noise. We state limitations plainly: simulation only, one gate and one five-parameter pulse family, no experimental data, and no pulse that beats the published one. The code is short and fully reproducible.