Plasmonic-upconverting nanosensor for automated colorimetric and ratiometric sensing of Hg²⁺ via digital readout and neural networks

Abstract The integration of photonic nanomaterials with artificial intelligence is driving the development of advanced sensors capable of converting molecular interactions into light-based signals, offering a powerful approach for portable detection of toxic pollutants. This work reports a novel strategy for the spectral modulation of lanthanide-doped upconversion nanoparticles (UCNPs) via plasmonic interactions with gold nanoparticles (AuNPs), enabling sensitive and selective detection of mercury ions (Hg 2+ ). Upon interaction, UCNP–AuNP heterostructures exhibit efficient quenching of green emissions due to plasmonic overlap with the 2 H 11/2 → 4 I 15/2 and 4 S 3/2 → 4 I 15/2 transitions of Er 3+ . This interaction enables a ratiometric luminescence response modulated by Hg 2+ -induced alterations in the AuNP surface plasmon resonance. To address the slow kinetics and weak affinity of citrate-stabilized AuNPs toward Hg 2+ , diethylenetriamine (DETA) was incorporated as a molecular mediator. DETA significantly enhances the AuNP–Hg 2+ interaction, as validated by density functional theory calculations, leading to pronounced red shifts and intensity amplification of the plasmonic band. As a result, the green emission from the UCNPs was progressively suppressed in a concentration-dependent manner, accompanied by a corresponding increase in the red-to-green ( R / G ) emission ratio. Additionally, chromaticity analysis in both the CIE 1931 and RGB color spaces revealed distinct perceptual color changes, enabling quantitative evaluation based on photographic images. To further exploit this optical response, convolutional neural network (CNN) was trained to classify Hg 2+ molecular species and distinguish them from potential interferents based on their distinct scattering and emission patterns, while a separate regression CNN was developed to accurately estimate Hg 2+ concentrations. This AI-assisted strategy achieved a classification accuracy of 100% and a determination coefficient ( R 2 ) of 0.989 for concentration prediction, with a detection limit of 0.25 ppm. Overall, this multimodal plasmon-enhanced nanosensor represents a promising platform for Hg 2+ detection, underscoring the potential of UCNP–AuNP hybrid systems and AI integration for portable, image-based diagnostic applications.