Privileged Bases for X-Ray Fluorescence Spectra Robust Automatic Classification

EasyChair Preprint 15575

Abstract

X-Ray Fluorescence (XRF) is an analytical technique that furnishes complex elemental spectra for element identification. Its non-invasiveness and portable nature have given the technique a broad application across various fields, each of which shows an idiosyncratic spectra type, resulting in a landscape where full automation of XRF analysis is challenging for Artificial Intelligence (AI) techniques.
In this contribution, we make use of recent results and hypothesis on the performance and interpretability of AI networks (superposition theory) to explore the prospects of using AI techniques to overcome the bottleneck of XRF automation analysis. In particular, we suggest that an autoencoder of XRF spectra whose (monosemantic) latent space dimensions match the number of elemental lines present in the input should have improved performance and interpretability.
In addition, we will discuss some of the implications and difficulties in this process, as well as some very preliminary results in this direction.

Keyphrases: Synthetic Dataset Generation, X-ray fluorescence (XRF), cultural heritage, deep learning, fundamental parameters, interpretability, machine learning

@booklet{EasyChair:15575,
  author    = {Fernando García-Avello Bofías and Alessandro Bombini and Chiara Ruberto and Francesco Taccetti},
  title     = {Privileged Bases for X-Ray Fluorescence Spectra Robust Automatic Classification},
  howpublished = {EasyChair Preprint 15575},
  year      = {EasyChair, 2024}}