Figure 4
(a)–(c) The mean absolute error of the thickness, roughness and SLD increases when fewer data points are available for fitting. The CNN and the growth model co-refinement can analyze sparsely sampled data with a 6–11× lower number of measured reflectivity values compared with the individual fit. Panels (d)–(f) illustrate the effect of simulating increasing amounts of photon shot noise for the
R(q,t) data, which would occur for shorter integration times or weaker X-ray sources. The growth model co-refinement is accurate for noise levels more than two orders of magnitude higher than for the independent fit of individual XRR curves. The CNN can predict correct parameters with even higher noise levels than the co-refinement approach. |