Machine learning-driven SERS quantification of phosphates through MB-Mo-AgNPs spectral transduction

An indirect surface-enhanced Raman spectroscopy (SERS) approach based on an MB-Mo-AgNPs probe system was developed to quantify aqueous KH₂PO₄ standard solutions. Silver nanoparticles (AgNPs), methylene blue (MB), and ammonium molybdate served as the SERS substrate, Raman reporter, and phosphate-responsive reagent, respectively. All concentrations were expressed as KH₂PO₄ mass concentrations. Across 0.0625-10 mg L^-1 KH₂PO₄, the MB bands near 1400, 1470, and 1626 cm^-1 decreased with increasing concentration. UV-Vis spectroscopy, qualitative dried-state transmission electron microscopy, dynamic light scattering, and zeta-potential measurements supported changes in the probe's colloidal state, particle association, and interfacial environment, consistent with a possible indirect signal-modulation pathway. These measurements did not establish a specific molecular mechanism or demonstrate MB desorption. The dataset comprised 2040 spectra from 51 independently prepared physical samples at 17 concentrations. In physical-sample-level grouped training-validation-test rotations, the 40 replicate spectra per sample remained within one subset, preventing sample overlap. Single-peak linear regression was compared with full-spectrum partial least squares regression, support vector regression (SVR), a conventional one-dimensional convolutional neural network (1D-CNN), and an attention-enhanced 1D-CNN. The full-spectrum models yielded lower errors than single-peak linear calibration. SVR achieved the lowest quantitative error among the compared models, with an RMSE of 0.0072 ± 0.0010 mg L^-1 across three grouped rotations. In the prespecified grouped analysis, the attention-enhanced 1D-CNN yielded lower quantitative error than the conventional 1D-CNN. The results apply only to aqueous KH₂PO₄ standard solutions and do not validate total-phosphorus determination, real-wastewater analysis, or performance in external matrices.