Fig. 8: Bayesian optimisation comparison.

a–c Log-scale runtime comparison between standard surrogate model fits (dashed red) and simulations (dashed orange), and deep kernel surrogate model fits (solid blue) and simulations (solid green). The difference in line length is due to the lack of surrogate model training during the eight initialisation simulations. d–f Symmetric log-scale simple regret progression for each matched pair, computed relative to the better of the two observed maxima in that pair. Results are shown for the beam splitter (a, d), total energy reflected (b, e), and peak intensity (c, f) objective functions.