Objective Laser-induced breakdown spectroscopy (LIBS) has emerged as a promising technique for rapid, in-situ characterization of rock-core samples owing to minimal sample preparation requirements and simultaneous multi-element detection capability. In geological exploration and resource evaluation, accurate quantification of major elements such as sodium (Na), magnesium (Mg), potassium (K), iron (Fe), calcium (Ca), and aluminum (Al) is essential for assessing mineral composition and economic potential. However, quantitative LIBS analysis of complex geological matrices faces significant challenges: strong pulse-to-pulse intensity fluctuations driven by transient plasma variability, pronounced matrix effects from heterogeneous mineral composition, and microscale surface non-uniformity of pressed powder pellets. Conventional single-shot approaches are susceptible to these disturbances, yielding unstable predictions and limited cross-matrix generalization. The present work aims to develop a practical, hardware-independent approach that exploits repeated-shot statistics to enhance repeatability, accuracy, and dynamic-range coverage of LIBS quantitative models for rock-core analysis.

Methods Forty-three certified national reference rock powder samples spanning soils, sediments, and ores were pressed into pellets and measured using a 1064 nm pulsed Nd:YAG LIBS system with a side-axis fiber-coupled spectrometer covering 200-800 nm (Fig.2). For each sample, 100 individual laser shots were acquired in a fixed-distance array pattern with 5 mm inter-point spacing; five pre-ablation shots were applied before each sequence to remove surface contamination. Key acquisition parameters are summarized in Tab.1. After outlier rejection, baseline correction, normalization, and z-score standardization, partial least squares (PLS) regression models were independently built for six target elements using 35 training-set samples. A multi-shot fusion strategy was implemented at the prediction stage by arithmetically averaging the 100 per-shot predictions for each test sample, without altering the learned spectrum-to-concentration mapping. To address wide dynamic concentration ranges, a binary high-low modeling scheme was introduced: a low-range sub-model trained in logarithmic concentration space for trace-level sensitivity, a high-range sub-model with concentration-proportional sample weighting for elevated-level robustness, and a tunable Sigmoid gating function for smooth blending of both sub-model outputs (Fig.3).

Results and Discussions Single-shot predictions for all six elements exhibited considerable scatter, with large per-sample standard deviations reflecting stochastic plasma fluctuations and local surface non-uniformity (Fig.4(a)). After applying multi-shot fusion, predicted values shifted markedly toward the ideal 1∶1 reference line and residual error bands narrowed substantially across both training and test samples (Fig.4(b)). Quantitatively, test-set root mean square error (RMSE) decreased by 10.5%-66.7% and the coefficient of determination (R²) improved for all six elements (Tab.2). The largest gain was recorded for Fe, whose RMSE dropped from 2.903 wt% to 0.966 wt% (66.7% reduction), consistent with its exceptionally broad concentration span across the sample set; K also benefited markedly, with test-set R² rising from 0.878 to 0.964 and RMSE falling by 45.8% (Tab.2). These improvements are attributed to statistical averaging of pulse-to-pulse random errors and the broader spatial coverage afforded by the 5 mm-pitch array pattern. For elements whose concentrations span multiple orders of magnitude, the binary high-low PLS framework provided further gains. Taking Fe as the representative example, linear-scale and logarithmic-scale scatter plots confirm that the low-range sub-model systematically underestimates at high concentrations, whereas the high-range sub-model loses precision in the dilute region; the dual-model blend eliminates both deficiencies (Fig.5, Fig.6). Quantitatively, the Sigmoid-gated dual model reduced Fe mean relative error (MRE) from 115.12% (high-range sub-model alone) to 20.11%, while simultaneously lowering RMSE from 0.958 wt% to 0.666 wt% (Tab.3). Consistent MRE and RMSE reductions were observed across all six elements under the binary scheme (Tab.3), confirming that the approach generalizes beyond Fe. The complete inference pipeline processed approximately 1674 spectra per second (0.597 ms per spectrum), confirming real-time suitability for on-site geological deployment.

Conclusions A prediction-level multi-shot fusion strategy combined with a binary high-low PLS modeling scheme and a Sigmoid gating mechanism substantially enhances LIBS quantitative stability for complex rock-core matrices. Multi-shot fusion alone reduces RMSE by up to 66.7%, and the binary dual model further suppresses relative errors in the low-concentration region while preserving high-concentration accuracy. Compared with prior approaches based on single-shot spectral normalization or double-pulse hardware enhancement, the proposed framework extracts additional predictive value from repeated-shot data without increasing measurement complexity or hardware cost. Classical PLS regression and statistical aggregation form the sole computational basis, facilitating deployment on portable or industrial LIBS platforms. Millisecond-scale inference satisfies the real-time demands of field geological surveys, resource evaluation, and environmental monitoring. Future research may explore adaptive outlier-aware shot selection, cross-instrument domain adaptation, and more expressive nonlinear regression modules to further improve robustness across extreme matrices and ultra-trace concentration ranges.