Multimodal Unsupervised Learning Framework Integrating External PIXE Mapping and Automated Vision Systems

Ion Beam Analysis (IBA) techniques, particularly Particle-Induced X-ray Emission (PIXE) mapping, have become indispensable for the non-destructive physicochemical characterization of complex materials. This work utilizes the unique external beamline setup at the Laboratory for Material Analysis with Ion Beams of the University of São Paulo (LAMFI-USP). The experimental setup couples a broad MeV-proton probe to a high-precision, large-range XYZ robotic stage. Crucially, to handle surface irregularities on large or complex samples without introducing imaging artifacts, the beamline incorporates an automatic machine-vision positioning correction system with an auto-focus camera. This system ensures rigorous control over the sample-to-detector distance, allowing for simultaneous, high-spatial-resolution topographical adjustments and optical tracking during X-ray data acquisition.

While modern data acquisition transforms wide-field PIXE into a high-throughput source of big data, extracting spatial-chemical correlations from the resulting hyperspectral data cubes remains a challenge. Traditional unimodal workflows apply unsupervised machine learning algorithms, such as Non-Negative Matrix Factorization (NMF) and K-Means clustering, solely to the X-ray spectra to segment different elemental phases. However, this unimodal pipeline fails to achieve complete segmentation accuracy when samples exhibit visual or optical anomalies under irradiation that do not present strong elemental contrasts, leaving critical phases unresolved.

To overcome this limitation, this work introduces a multimodal framework that directly fuses the hyperspectral PIXE data with the contextual RGB optical information captured by the line's integrated auto-focus camera. This methodology was validated on two distinct scenarios: a tourmaline crystal displaying intense localized luminescent features and a complex, large-scale fish fossil from the Araripe Basin (Ceará, Brazil). By combining chemical spectroscopy with automated optical vision, the framework successfully resolved material ambiguities and suppressed artifact propagation. This data fusion ultimately resulted in final analysis images of exceptionally high quality, proving that the integration of optical context significantly enhances the accuracy and visual definition of complex chemical phase segmentation.