Sim-to-real for Inertial Confinement Fusion: a step toward faster zero-carbon energy

Deep decarbonisation of the electricity grid will need firm, dispatchable low-carbon generation to complement wind and solar, and inertial confinement fusion (ICF) is one of the few candidates in that role. Progress toward a working ICF power plant is limited less by physics than by the cost of each experimental laser shot: any tool that helps an optimiser choose the next shot more informatively directly shortens the R\&D loop and its embodied carbon. Machine-learning surrogates trained on radiation-hydrodynamics simulation are a natural candidate, but simulation alone is not enough: on real shots, a sim-only surrogate has roughly $7\times$ the mean absolute percentage error (MAPE) of any model that has seen even a handful of experimental data. Sim-to-real transfer is therefore mandatory, but how to transfer under extreme experimental scarcity ($N{\leq}290$ shots) is still an open question. We benchmark nine adaptation strategies, from frozen-encoder heads to basin-diverse deep ensembles, scoring them on both accuracy and on the uncertainty calibration that downstream active-learning optimisers actually consume. Once adaptation closes the accuracy gap, eight of the nine methods sit inside a $0.4\%$-MAPE band; calibration then does the work of selecting a winner. A $K{=}10$ ensemble diversified across four axes matches the lowest in-distribution MAPE ($4.48\%$ vs.\ $4.47\%$ for the runner-up) and roughly doubles the Spearman calibration score under sawtooth shift ($\rho{=}0.56$). The recipe is portable: an adapted, well-calibrated surrogate lets the optimiser spend expensive experiments where they actually reduce uncertainty, and the same regime applies to catalyst discovery, plasma control, and geothermal design.