Extended Multiplicative Signal Correction-Assisted ResNet1D for Laser-Induced Breakdown Spectroscopy (LIBS)...

Under rough-surface conditions, laser-induced breakdown spectroscopy (LIBS) spectra of uranium polymetallic ores are affected by both complex matrix effects and morphology-induced nonchemical variability, which increases the difficulty of accurate rare earth quantification. In this study, rough and unpolished pelletized uranium polymetallic ore samples were used to investigate the effects of surface roughness on LIBS spectral quality and the quantitative modeling performance for neodymium (Nd), gadolinium (Gd), and samarium (Sm). A within-sample median-based extended multiplicative signal correction (EMSC) strategy was proposed to reduce roughness-induced nonchemical variability. A one-dimensional residual network (ResNet1D) was then constructed for quantitative regression and compared with partial least squares regression (PLSR) and Al 308.215 nm peak-area normalization within a leave-one-sample-out cross-validation (LOSO-CV) framework. After EMSC_median correction, ResNet1D outperformed RAW-ResNet1D for all three elements, with R² increasing from 0.9474, 0.9130, and 0.8810 to 0.9500, 0.9180, and 0.9287, respectively, while RMSE decreased from 2.0179, 2.6967, and 2.9220 to 1.9686, 2.6177, and 2.2615. Attribution analysis further showed that ResNet1D preferentially focuses on information-rich spectral regions associated with the target rare earth elements. These results demonstrate the effectiveness of the proposed method under the current sample system and experimental conditions and provide a methodological reference for in situ rapid LIBS analysis in uranium polymetallic ore scenarios.