Full-spectrum XGBoost framework for interpretable classification of femtosecond LIBS spectra

Reliable classification of laser-induced breakdown spectroscopy (LIBS) spectra remains challenging when emission signals are weak and strongly fluctuating under low-pulse-energy femtosecond (fs) laser ablation. This study evaluates a full-spectrum machine-learning approach based on eXtreme Gradient Boosting (XGBoost) using an Al/Si thin-film ablation dataset as a representative binary spectral-classification case. The full raw LIBS spectrum was directly used as the model input, and XGBoost was compared with support vector machine, multilayer perceptron, k-nearest neighbors, and Gaussian Naive Bayes classifiers. Model development and evaluation were performed using processing-point-level data separation, with hyperparameter optimization conducted exclusively on the training dataset using repeated five-fold cross-validation. XGBoost achieved a mean cross-validation accuracy of 99.55 ± 0.26%, an F1-score of 99.48 ± 0.31%, and an AUC of 0.99994 ± 0.00008. On the independent test dataset comprising 612 spectra from 68 processing points, XGBoost achieved 100% accuracy, precision, recall, and F1-score, with an AUC of 1.00000. The classification results remained unchanged at probability thresholds of 0.3, 0.5, and 0.7. The measured single-spectrum inference time was approximately 0.105 ms, substantially shorter than the 10 ms interval between successive laser pulses at 100 Hz. SHAP analysis showed that the most influential wavelength channels were concentrated within the 288–289 nm spectral region and were distributed across multiple neighboring channels rather than being confined to the nominal Si I 288.15 nm wavelength. These results demonstrate accurate full-spectrum classification of the investigated low-signal fs-LIBS dataset while retaining wavelength-level information for model interpretation.