A deep learning strategy for discrimination and detection of multi-sulfonamides residues in aquatic environments using gold nanoparticles-decorated violet phosphorene SERS substrates

Surface-enhanced Raman spectroscopy (SERS) is an emerging technique for rapid and highly-sensitive detection of analytes, but the substrate dependence of enhancement performance and low throughput of spectral analysis limit its widespread application. Herein, gold nanoparticles (AuNPs) decorated violet phosphorene (VP) as SERS substrate was prepared by an in-situ seed-mediated growth method, which exhibited excellent repeatability, high reproducibility, favorable storage stability, an enhancement factor of 1.66 x 106 and a low detection limit of 4.7 ng/mL for sulfamethazine. A deep learning strategy based on a one-dimensional convolutional neural network (1-D CNN) was introduced to solve the problem of differentiating three structurally similar antibiotics (sulfa-methazine, sulfadiazine, and sulfamethoxazole) at 0.005-10.00 mu g/mL with similar characteristic peaks. The model achieved 100% accuracy over traditional machine learning such as principal component analysis and t -distributed stochastic neighbor embedding. The quantitative analysis model built with a 1-D CNN was also successfully used for the quantitative analysis of three sulfonamides as well, with output parameters of Rp2 >= 0.9786 and RPD >= 6.35. This work will provide a new reference for the preparation of metal nanoparticles decorated with two-dimensional nanomaterials as SERS substrates and the discrimination and detection of multi-analytes with similar structures.

Automated summary

This study prepared gold nanoparticles (AuNPs) decorated violet phosphorene (VP) as a surface-enhanced Raman spectroscopy (SERS) substrate using an in-situ seed-mediated growth method. The substrate exhibited excellent repeatability, high reproducibility, favorable storage stability, an enhancement factor of 1.66 x 106, and a low detection limit of 4.7 ng/mL for sulfamethazine. Furthermore, a deep learning strategy based on a one-dimensional convolutional neural network (1-D CNN) was introduced to successfully differentiate three structurally similar antibiotics at 0.005-10.00 μg/mL.