What data augmentation contributes to quantitative LIBS: Deep neural networks and partial least squares regressions on the LIBS 2022 steel benchmark

Quantitative laser-induced breakdown spectroscopy is limited less by the regression model than by the calibration data: certified targets are few, unevenly distributed, and described by tens of thousands of correlated wavelengths. On the LIBS 2022 steel benchmark we measure what a data-centric intervention contributes. Synthetic spectra are generated by multiplying each measured training spectrum by fixed factors, leaving line ratios and the reference concentration unchanged; because their amplitude exceeds the measured shot-to-shot dispersion of the integrated intensity, 1.7%, by an order of magnitude, they impose invariance to the overall emission level rather than modelling noise. Across nineteen deep-learning configurations, each compared with and without augmentation at otherwise fixed settings, augmentation lowers the per-element error by close to an order of magnitude on validation and by a factor of three on the blind test. A partial-least-squares reference puts these figures on a chemometric scale: without augmentation the linear and the deep models are indistinguishable, 1.90 against 1.87 wt%, whereas augmentation improves the former by 17% and the latter by 66%. In a controlled comparison, architecture accounts for about a tenth of that effect. On the blind targets the analyte-specific pipeline is the only entry among the best on all four elements, and it nearly halves the error of the best contest entry on molybdenum, the most imbalanced analyte. The benefit is thus a property of this calibration problem rather than of the deep-learning family, and, on the invariance reading, should be expected wherever the overall emission level dominates the nuisance variability.