Presentation Information
[8p-PB2-9]A Cross-Exposure Transfer Rule of Self-Supervised Denoising for AR-HAXPES Depth-Profile Reconstruction
〇Satoshi Toyoda1, Masaki Ando2, Atsushi Ogura2,3, Toyohiko Kinoshita1, Masatake Machida1 (1.Vacuum Products, 2.Meiji University, 3.MREL)
Keywords:
AR-HAXPES,self-supervised denoising,depth profiling
AR-HAXPES at laboratory sources is photon-starved, so guaranteeing the accuracy of depth profiles reconstructed from short-exposure data is challenging. We use self-supervised (Noise2Noise-type) DNN denoising as a front-end to an L1-regularized inverse solver, and study how a denoiser trained at one exposure can be reused at another. On a C/Al2O3/TiO2/Si multilayer measured at three exposures (4.8/24/120 s/frame), we build a 3×3 train-by-inference matrix and identify an empirical directional rule: train S/N ≥ inference S/N. A denoiser trained on a higher-S/N exposure transfers stably to a lower-S/N target, whereas the locally-trained Self-DNN collapses at the most photon-starved condition and propagates into constraint-saturated inverse solutions. Reusing a higher-S/N-trained Cross-DNN preserves accuracy while greatly reducing the effective acquisition time, and is directly applicable to microbeam AR-HAXPES and 4D-XPS.