Calibrating sim-to-real surrogates for Inertial Confinement Fusion

Simulations of inertial confinement fusion (ICF) provide abundant training data but remain imperfect proxies for reality: a surrogate trained solely on simulation yields $31\%$ mean absolute percentage error (MAPE) on experimental shots, roughly $7\times$ worse than models adapted with even a small amount of real data. Sim-to-real transfer is therefore essential, yet the question of how to adapt under extreme data scarcity ($N{\leq}290$ shots) remains largely unresolved. We benchmark nine adaptation strategies, ranging from frozen-encoder transfer to SNGP and deep ensembles, evaluating both predictive accuracy and uncertainty calibration under structured input shifts. We find that predictive accuracy saturates quickly: at $N{=}33$, eight of nine methods fall within a narrow $0.4\%$ MAPE band. Uncertainty calibration, however, reveals substantial differences. A basin-diverse $K{=}10$ ensemble is the only method that simultaneously ties for the best in-distribution MAPE ($4.48\%$) while achieving the strongest Spearman calibration both in-distribution and under sawtooth shifts ($\rho=0.52$ and $\rho=0.56$, respectively). Ablation studies further show that no single source of diversity, including warm starts, bootstrapping, or architectural variation, dominates across all shift regimes. Instead, combining multiple diversity mechanisms emerges as a practical recipe for robust, uncertainty-aware sim-to-real surrogates in data-scarce scientific domains.