A MA-SSA-optimized XGBoost-MLP framework using LIBS for rapid classification and quantitative analysis of heavy...

Laser-Induced Breakdown Spectroscopy (LIBS) holds significant value for rapid elemental detection; however, strong spectral interference, matrix effects, and high-dimensional data characteristics pose considerable challenges to accurate quantitative analysis. To enhance the performance of LIBS quantitative analysis, this study proposes a novel machine learning framework that integrates XGBoost and Multilayer Perceptron (MLP), optimized by a Multi-dimensional Adaptive Sparrow Search Algorithm (MA-SSA). The framework employs XGBoost for automated feature selection, eliminating redundant spectral variables while retaining critical information, and utilizes MA-SSA to optimize the hyperparameters of the MLP in regression tasks, significantly improving model stability and prediction accuracy. Experimental results demonstrate that the proposed method achieves 100 % accuracy in multi-class classification, outperforming traditional classifiers such as Random Forest, XGBoost, and standalone MLP. In terms of quantitative detection, the MA-SSA-optimized model achieves an RMSE of 4.43 µg/g, surpassing other hybrid optimization models including XGBoost-SSA-MLP (RMSE=4.62 µg/g), XGBoost-PSO-MLP (RMSE=5.225 µg/g), and XGBoost-GA-MLP (RMSE=5.584 µg/g). XGBoost-based feature selection effectively reduces spectral dimensionality while maintaining predictive performance. The proposed MA-SSA algorithm further enhances convergence efficiency and generalization capability. This study provides a robust, efficient, and scalable solution for LIBS analysis, with broad application potential in the field of real-time quantitative detection.