A neural network-based Monte Carlo X-ray (MC X-ray) approach for quantitative analysis of homogeneous multi-element bulk samples at 20 kV.

Abstract

Accurate quantitative analysis of homogeneous bulk samples is critical in materials science and mineral industry applications, yet conventional standards-based Energy Dispersive Spectroscopy (EDS) quantitative analysis remains reliable and widely used. However, for some laboratories or sample types, obtaining, maintaining, or matching appropriate physical standards may be inconvenient or impractical. In this study, we introduce a dynamic framework for elemental quantification of homogeneous multi-element bulk samples investigated in an SEM at an accelerating voltage of 20 kV by integrating Monte Carlo X-ray (MC X-ray), a Monte Carlo simulation program, with artificial neural networks (ANNs). Synthetic spectra and corresponding weight fraction pairs were generated using MC X-ray under the same 20 kV measurement condition, and a multilayer perceptron (MLP) was trained to map spectral features to elemental weight fractions. Validation across 11 certified standard samples yielded root mean square error (RMSE) values from 0.0056 to 0.0179 in weight fraction and coefficients of determination (R2) from 0.9873 to 0.9981, even in cases with overlapping peaks. For elements with certified weight fractions of at least 0.05, the mean absolute percentage error was 4.41%. The model, tailored to specific sample and measurement conditions, enables quantitative analysis in under 15 min on consumer-grade hardware. By reducing the dependence on physical standards and calibration procedures, this approach provides a complementary simulation-based method for quantitative microanalysis when suitable measured standards or stored standard libraries are not readily available for the analysis of homogeneous bulk materials.