Multi-task fusion network with matrix awareness for LIBS-based heavy metal detection in heterogeneous water matrices

Monitoring heavy metals in water is critical due to their environmental and health hazards. Laser-induced breakdown spectroscopy (LIBS) offers rapid, in situ detection but suffers from matrix effects that reduce quantitative accuracy. To address this, we propose a multi-task fusion network (MTFNet) that separates the quantification of elements in heterogeneous matrices into matrix classification and element regression tasks, enabling joint objective optimization to mitigate matrix effects. The model was evaluated across four water matrices: Yangtze River water, reservoir water, groundwater, and deionized water. For chromium (Cr), the coefficient of determination (R²) improved from 0.749 to 0.994, the root mean square error of cross-validation (RMSECV) decreased from 0.200 to 0.026 mg/L, and the average relative error of cross-validation (ARECV) dropped from 46.73% to 6.94%. Comparable improvements were observed for manganese (Mn) and cadmium (Cd). These results demonstrate that MTFNet effectively mitigates matrix effects and enhances LIBS-based quantitative performance, offering a promising solution for environmental analysis. 

© 2025 Published by Elsevier Ltd.