Development of an artificial intelligence surrogate and analysis model for PIXE data based on a minimal experimental effort

Artificial intelligence has been applied to ion beam analysis since the beginning of the century. Modern platforms have significantly reduced the barrier to entry in this field, and, with the fast development of machine learning algorithms and the resurgence of interest in artificial intelligence, applications to ion beam analysis have multiplied in recent years. Most works have used simulated data to train the neural networks, taking advantage of the widespread availability of simulation codes for techniques such as Rutherford backscattering. Artificial intelligence models are most useful when they can be trained either with simulated data or with a minimal experimental effort. We developed an autoencoder-based artificial intelligence model for PIXE data based on such a small experimental effort, that is capable of analysing PIXE spectra without background subtraction, providing elemental concentrations, and of generating PIXE spectra from given elemental concentrations. We report on the model and its results.