A multi-component heavy metal detection method using UV-Vis superimposed spectrum and deep learning

In the field of environmental monitoring, traditional spectroscopic techniques, despite their high sensitivity in heavy metal (HM) detection, are plagued by complex equipment and high costs. The emergence of artificial intelligence (AI) has driven the digitalization and intelligence of spectral analysis. However, the complexity of environmental samples and spectral overlap remain a major challenge. This study proposes a multi-component HM detection method that integrates ultraviolet-visible (UV-Vis) superimposed spectra with deep learning to address the issue of spectral overlap. By using combinatorial chemical probes to enhance the specificity of colorimetric reactions and conducting high-throughput experiments to collect spectra, a Transformer model is trained to extract qualitative and quantitative information in an end-to-end manner. During the method development phase, the model was trained using five representative HMs (Sb, Fe, Ni, Cd, Cu) to verify its recognition ability under complex conditions (average R² = 0.936, RMSE = 0.132 mg L^-1, MAE = 0.054 mg L^-1). The method was then extended to detect ten HMs in real-world samples, demonstrating its scalability and robustness (average R² = 0.681, RMSE = 0.189 mg L^-1). These results show that the method can simultaneously analyze multiple HMs and transition from spectral fingerprints to quantified ecological risk profiles. Additionally, the AI analysis software developed in this study automates the entire process from spectral input to multi-index detection and risk assessment, providing environmental regulatory authorities with minute-level response capabilities and facilitating proactive early warning for HM pollution control.