A multi-scale spatio-temporal network for time-series LIBS spectra and its application in heavy metal quantification

Proposes an end-to-end LIBS framework eliminating manual delay time selection. • Introduces a Physics-Inspired Multi-Scale Spatio-Temporal Network (MS-STN). • Develops ACSNOA algorithm balancing global exploration and local exploitation. • Achieves 43.4 % MAE improvement (2.566 µg/g) for 10 µg/g Cr extrapolation. The rapid detection of heavy metal contamination is crucial for environmental and public health. Laser-induced breakdown spectroscopy (LIBS) has great potential for in situ analysis, but its quantitative accuracy is limited by empirical delay-time selection and insufficient precision at trace levels. This study proposes an end-to-end intelligent framework to overcome these limitations. We introduce a physics-inspired multi-scale spatio-temporal network (MS-STN), which shifts multi-scale analysis from the spectral to the temporal dimension of plasma evolution. This allows automatic feature extraction from full time-series spectra, eliminating tedious pre-selection of a single delay time based on metrics like signal-to-noise or signal-to-background ratios. An adaptive chaotic simulated newtonian optimization algorithm (ACSNOA) is also proposed for stable hyperparameter tuning. Experiments on chromium (Cr)-spiked licorice samples show that MS-STN achieves a mean absolute error of 2.566 µg/g and a prediction root mean square error of 2.889 µg/g for 10 µg/g Cr extrapolation, significantly outperforming benchmark models. Ablation studies validate each core component, and ACSNOA shows superior optimization stability. This work provides a high-precision, automated, and physics-informed paradigm for LIBS-based heavy metal analysis, advancing its real-time, on-site monitoring applications. Download: Download high-res image (335KB) Download: Download full-size image