Automated total micro-IBA using Advanced Image Processing and Machine Learning

Abstract

We have developed a Python code that aims at automatizing the analysis of generic micro-IBA data by associating statistical methods and machine learning algorithms. The code is organized in two parts: hyperspectral image analysis and composition prediction. In the first stage, main phases and local anomalies are detected separately using PCA and DWEST methods, respectively. In the prediction stage, we use the model generated by a trained artificial neural network. The network is fed with simulated particle and x-ray spectra generated from the SIMNRA and Gulys software codes. For particle spectra, we paid particular attention to the cross section selection that goes beyond already implemented SIMNRA functionalities. To limit the impact of the simulation time on the overall code performance, we make use of data augmentation. When using simulated data as input, we found that the trained neural network predicts stoichiometries and thicknesses with an excellent agreement, even for complex targets composed of several elements and layers. Regarding realistic experimental data, we still get reasonable predictions but remain dependant of cross section quality. The code can combine data from RBS, NRA, ERDA and PIXE and should pave the way for fully automatized micro-ion beam analysis.