DNN-enhanced spatially offset Raman spectroscopy for non-destructive subsurface material characterization

Spatially offset Raman spectroscopy (SORS) enables non-destructive subsurface analysis, but traditional methods are limited by the need to manually set parameters and lack full automation. This paper presents a deep neural network (DNN)-enhanced SORS method that simplifies spectrum reconstruction. The DNN model takes only two inputs (zero and non-zero offsets) to reconstruct the pure Raman spectra of deeper layers, eliminating the influence of surface layers. Experimental results demonstrate that the DNN achieves high reconstruction accuracy (<i>R</i><i>M</i><i>S</i><i>E</i><0.02, <i>P</i><i>C</i><i>C</i>>0.96) for four powders, highlighting its potential for fully automated SORS data analysis.