Gelatin-Assisted Laser-Induced Breakdown Spectroscopy (LIBS) Coupled with Principal Component Analysis-Particle Swarm Optimization-Support Vector Machine (PCA-PSO-SVM) for Classification of Tobacco Flavors in Complex Organic Matrices

Liquid tobacco flavorings are critical additives in cigarette manufacturing, directly determining product quality and consistency. However, direct analysis via Laser-Induced Breakdown Spectroscopy (LIBS) faces challenges posed by their high viscosity, high sugar content, and complex organic matrices, which lead to severe signal fluctuations due to droplet splashing and plasma quenching. To overcome these challenges, this study proposes a hybrid enhancement framework that combines gelatin-mediated sample stabilization with a principal component analysis-particle swarm optimization-support vector machine (PCA-PSO-SVM) model for classification in complex organic matrices. A 2.50% gelatin solution was used for mediated solidification, effectively mitigating interference from the liquid matrix. Experimental results indicate that samples solidified with 2.50% gelatin yielded a 5.49-fold enhancement in the signal-to-noise ratio (SNR) for the Mg II 279.46 nm line compared with untreated liquid samples. Furthermore, the Relative Standard Deviation (RSD) was reduced by 56.37%, demonstrating significantly improved spectral stability. Building on this, PCA was employed to mitigate spectral dimensionality and eliminate redundant information. Combined with a particle swarm optimization (PSO) algorithm for adaptive optimization of SVM parameters, comparative analysis reveals that the PCA-PSO-SVM model outperforms traditional SVM, Random Forest (RF), and Extreme Learning Machine (ELM) models. Specifically, it achieved classification accuracies of 95.08% for Class L (flavoring-base liquids) and 96.03% for Class B (surface-applied tobacco flavorings; top-dressing flavorings), with F1 scores of 95.09% and 96.04%, respectively. These metrics represent an improvement of over 12 percentage points compared with the ELM model, indicating the effectiveness of combining physical sample stabilization with adaptive hyperparameter optimization via PSO for complex fluid analysis. These results suggest that the integrated strategy of physical solidification combined with adaptive machine learning algorithms can support the classification of complex tobacco flavor samples and may provide a useful analytical approach for quality control in the tobacco industry.