A Comparative Study of Machine Learning Algorithms Trained With Monte Carlo Simulations For X-Ray Fluorescence Analysis

Abstract

X-ray fluorescence (XRF) spectrometry is widely used in quantitative analysis because it is fast, reliable, and nondestructive. However, challenges such as matrix effects, spectral overlap, background interference, and the need for suitable standards can make quantitative XRF analysis difficult, motivating the use of machine learning models that learn composition spectrum relationships from representative datasets. Normalized characteristic peak areas extracted from XRF spectra were used as input to predict the elemental compositions of stainless steel samples. Artificial neural network (ANN), random forest (RF), and support vector regression (SVR) algorithms were trained with datasets generated from Monte Carlo simulations. All trained models were validated using experimental XRF spectra. All models demonstrated compatibility with experimental results, with the ANN emerging as the most effective model, particularly for trace elements. These results demonstrate the feasibility of using simulation-generated datasets to train machine learning models and reduce reliance on extensive experimental data.