The problem the researchers set out to solve is a familiar one. Seafood is densely packed with nutrients, and even when it is chilled or frozen, endogenous enzymes and microorganisms gradually break down proteins and other compounds, releasing volatile amines such as ammonia and biogenic amines. These gases are the key contributors to total volatile basic nitrogen, or TVB-N, the gold-standard chemical indicator of seafood freshness. Conventional laboratory methods for measuring TVB-N, including chemical-microbiological assays, chromatography, mass spectrometry, and spectroscopy, are accurate but slow, destructive to the sample, and dependent on expensive instrumentation and skilled technicians. They cannot tell a shopper whether the shrimp in the display case is fresh, and they cannot track freshness continuously inside a sealed package.

Fluorescent sensing tags have emerged as a promising alternative because they can be placed inside packaging and photographed with a smartphone. Early versions, however, relied on a single fluorescence signal that switched on or off, which made them vulnerable to environmental interference and difficult to judge by eye. The field responded with ratiometric probes, which carry two well-resolved emission bands so that one signal serves as an internal reference for the other. But many dual-emission probes built from dyes or quantum dots suffer from aggregation-caused quenching, in which the fluorescent signal collapses when the molecules cluster together on a solid surface. Aggregation-induced emission, or AIE, flips that behavior, glowing brighter when aggregated, and it has become a favored design principle for solid-state sensors.

The team, led by Wenyang Zhang with co-authors You Tian, Yanwu Chen, and Min Wei, built their tag around a relatively unexplored class of AIE materials: bimetallic nanoclusters. Using a microwave-assisted synthesis run at 90 degrees Celsius for just 150 seconds, they prepared silver-copper nanoclusters capped with D-penicillamine, abbreviated DPA-AgCuNCs. Transmission electron microscopy showed dense aggregates of particles averaging about 3 nanometers in diameter, and energy-dispersive X-ray mapping confirmed a homogeneous distribution of silver, copper, sulfur, carbon, and nitrogen. Spectroscopic characterization revealed a strong orange emission at 585 nanometers when excited at 360 nanometers, a large Stokes shift of 225 nanometers, and a long photoluminescence lifetime of about 4 microseconds, all hallmarks of a ligand-to-metal charge transfer from the sulfur atom of the capping ligand to the metal core.

Exposure to ammonia vapor quenched the nanocluster fluorescence in a linear fashion across concentrations from 10 to 120 parts per million, with a detection limit of 2.95 parts per million. Mechanistic experiments showed that the quenching is a synergistic combination of static and dynamic pathways. Ultraviolet-visible spectroscopy revealed two new absorption bands after ammonia exposure, evidence of a non-fluorescent ground-state complex, while fluorescence lifetime measurements showed the average lifetime plummeting from 4.023 microseconds to 3.27 nanoseconds, a signature of dynamic quenching. X-ray photoelectron spectroscopy showed the nitrogen content of the material rising after exposure and shifts in the copper oxidation-state features, and electron microscopy captured the physical disaggregation of the nanocluster aggregates. The researchers attribute the dominant pathway to static quenching through formation of an ammonia-coordinated copper-silver complex.

To convert this single-signal quenching into a visually readable ratiometric response, the team combined the orange-emitting nanoclusters with fluorescein isothiocyanate, or FITC, a dye that behaves in the opposite way. In the presence of the alkaline environment created by ammonia, the lactone ring of FITC opens to its carboxylate form and its green fluorescence at 510 nanometers brightens. Drop-casting the two probes together onto circular cellulose fiber paper produced a tag that glows orange-yellow under 365 nanometer ultraviolet light but shifts progressively toward green as amine concentrations rise. Scanning electron microscopy confirmed that both components, sheet-like FITC aggregates and petal-shaped nanocluster particles, coexisted on the paper fibers. After optimizing the volume ratio of nanoclusters to FITC at 4 to 0.5, the tag displayed a clean, graduated orange-to-yellow-to-green progression.

The practical test came with fresh sand shrimp stored at 4 degrees Celsius, with a tag affixed inside the lid of a sealed Petri dish where it never touched the sample. Independent measurements of TVB-N and total viable counts classified the shrimp as fresh at days zero to one, less fresh at days two to three, and spoiled at day four. The tag’s color changes tracked these grades precisely: bright orange when fresh, pale yellow when declining, and distinctly green when spoiled. Based on this behavior, the team built a fluorescence color card allowing anyone to grade freshness at a glance. The tag also proved rugged, retaining 94.47 percent of its initial fluorescence after 30 days at 4 degrees Celsius, more than 91 percent after a month across relative humidities from roughly 33 to 90 percent, and 95.47 percent after five days of continuous ultraviolet exposure.

Visual color cards, however, remain hostage to human subjectivity, since individual differences in color perception can produce inconsistent judgments. To eliminate that variability, the researchers trained five convolutional neural network architectures on a dataset of 1,842 labeled fluorescence images spanning the three freshness categories, augmented with flipping, rotation, scaling, cropping, translation, and Gaussian noise. ResNet-34 decisively outperformed the alternatives, achieving 98.39 percent accuracy, 98.48 percent precision, 98.41 percent recall, a 98.41 percent F1-score, and an area under the receiver operating characteristic curve of 99.02 percent on held-out data. VGGNet-16 and DenseNet managed only about 75 to 76 percent accuracy, GoogLeNet 53 percent, and AlexNet a dismal 36 percent. Stratified five-fold cross-validation confirmed ResNet-34’s robustness, with accuracy of 96.81 plus or minus 0.92 percent across folds, and on an independent test set of 122 unseen images the model achieved perfect precision for fresh and spoiled samples and 95.24 percent for the intermediate category.

The optimized network was then packaged into a proof-of-concept smartphone application called FreshSense, in which a user photographs the illuminated tag and the app displays an automatic freshness prediction, with inference currently handled on a remote server. The system generalized well beyond shrimp: applied to salmon, cod, and grass carp stored at 4 degrees Celsius, it achieved classification accuracies of 93.52, 90.74, and 92.59 percent, respectively, showing that the sensing chemistry is not tied to a single species. The authors note that future work will target model compression and on-device deployment so the tool can function fully offline, along with systematic evaluation of computational costs. Selectivity testing against thirteen volatile substances produced during spoilage, including methanol, hydrogen sulfide, formaldehyde, and acetic acid, showed strong responses only to ammonia, trimethylamine, putrescine, and a mixed-gas cocktail, precisely the compounds that matter for spoilage, suggesting the platform could become a practical, intelligent fixture in cold chains, retail displays, and eventually home refrigerators.

Subject of Research: A CNN-integrated ratiometric fluorescent nanocluster tag for rapid seafood freshness monitoring at refrigerated temperatures

Article Title: CNN-integrated ratiometric fluorescent tag for rapid monitoring of seafood freshness at refrigerated temperatures

Article References: Zhang, W., Tian, Y., Chen, Y., & Wei, M. (2026). CNN-integrated ratiometric fluorescent tag for rapid monitoring of seafood freshness at refrigerated temperatures. Current Research in Food Science, 13, Article 101560. https://doi.org/10.1016/j.crfs.2026.101560

Image Credits: AI Generated

DOI: 10.1016/j.crfs.2026.101560

Keywords: seafood freshness, ratiometric fluorescent tag, aggregation-induced emission, silver-copper nanoclusters, volatile amines, ammonia detection, convolutional neural network, ResNet-34, fluorescein isothiocyanate, TVB-N, smartphone sensing, refrigerated storage