.png){kind=icon}
.png){kind=icon}
Return
Improving quantification of trace-level copper in geological samples using LIBS and Machine Learning techniques
.png){kind=icon}
PRE
AI
A
A
L
M
R
M
DOI：10.1016/j.sab.2026.107519.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Trace-level copper (Cu) quantification in geological materials by Laser-Induced Breakdown Spectroscopy (LIBS) is challenged by strong shot-to-shot spectral variability and matrix effects, especially under low-power and purge-free conditions. This study presents a structured workflow separating within-pellet spectral heterogeneity from inter-pellet transfer to improve Cu quantification in certified geological reference materials. Single-shot LIBS spectra were acquired in ambient air using a single pulse of similar to 9 mJ with a compact, low-energy LIBS setup, yielding 4000 spectra from four certified OREAS (R) reference materials. Spectra were detrended using a Savitzky-Golay approach and normalized by Standard Normal Variate and Min-Max. Latent structure was explored using Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) followed by clustering-based segmentation (K-means, Hierarchical Clustering, and HDBSCAN). Cu regression models were built using Partial Least Squares (PLS) and Support Vector Regression (SVR) with wavelength selection based on Variable Importance in Projection (VIP) for PLS and mutual information (MI) for SVR. Performance was evaluated at two complementary levels. Spectral-level cross-validation was used as an exploratory within-pellet analysis to examine shot-to-shot variability and the effect of segmentation on predictive behavior. Pellet-level leave-one-sample-out (LOSO) validation was implemented to assess transfer across unobserved pellets. Combined with nonlinear regression and informed wavelength selection, within-pellets segmentation identified subsets exhibiting improved predictive performance with the best spectral-level configuration reaching R-2 > 0.997 and RMSE = 0.0184 wt.%. However, pellet-level LOSO showed limited inter-matrix transfer across the four available certified materials. Additional analysis demonstrated that intensity-only T-i screening partially recovered the selected subset but did not match its predictive performance, suggesting that clustering leverages additional spectral-structural information beyond intensity.
Keywords:
Laser-Induced Breakdown Spectroscopy (LIBS)
Trace element quantification
Machine Learning
Spectral segmentation
Cu quantification
Journal
IF:
3.8
Papers:
5.1K
Citations:
8.8K
Organization
Cited Papers