Denoising strategies for three-dimensional visualization of multilayer thin-film interfaces via angle-resolved hard x-ray photoelectron spectroscopy

Nondestructive three-dimensional visualization of buried interfaces in multilayer thin films is central to semiconductor device characterization. Angle-resolved hard x-ray photoelectron spectroscopy (AR-HAXPES) enables such depth-resolved analysis, but short-exposure measurements suffer from photon shot noise that degrades the reconstructed depth profiles—a problem that becomes acute as the technique is pushed toward spatiotemporal “4D-XPS” (spatially and temporally resolved XPS) at focused-beam instruments. Here, we benchmark two denoising strategies—statistical bin-pool (BP) resampling and self-supervised deep neural network (DNN) denoising—applied to AR-HAXPES data from a C/Al2O3/TiO2/Si multilayer film acquired with a laboratory Ga Kα source (9.25 keV) under two per-frame noise levels (24 s/frame × 2500 frames; 120 s/frame × 400 frames). Depth profiles are reconstructed per pixel via L1-regularized inversion and scored against a pseudoground truth using the depth-profile root mean square error (Depth RMSE), film-thickness fidelity Δthk, and the first-order Wasserstein distance W1. The two strategies turn out to be complementary rather than competing: under high-noise conditions, the self-supervised DNN dominates with up to a 27-fold equivalent-exposure gain, whereas under low-noise conditions, BP matches or slightly surpasses the DNN at roughly a six- to eightfold gain and retains a clear advantage in reconstructed layer-thickness fidelity. BP—training-free and computationally trivial—is, therefore, the preferred choice whenever the per-frame statistics are sufficient for the inversion to resolve interfaces, while the self-supervised DNN remains justified specifically when photon starvation is severe.