This presentation focuses on the development of an automated procedure for detecting inhomogeneous regions in nuclear microprobe elemental maps. Nuclear microprobe analysis uses a focused MeV light-ion beam, scanned across the sample surface, to generate 2D maps from RBS and PIXE spectra recorded with OMDAQ software. These maps provide information on the surface and depth elemental composition of the sample and help identify defects that may influence its structural, optical, or electrical properties. The proposed approach treats the 2D elemental maps as images and applies a U-Net convolutional neural network to automatically identify regions of interest associated with material inhomogeneities. This work represents a step toward autonomous nuclear microprobe experiments, reducing manual intervention and improving the efficiency of defect detection and analysis.