AI-Powered Optical Fiber Sensor Spots Which Animal Your Gelatin Came From in Seconds

Gelatin hides in plain sight. It thickens gummy candies, stabilizes yogurts, forms the shells of capsules, and smooths countless cosmetics. Yet for millions of consumers, the invisible question of which animal it came from matters deeply: porcine-derived gelatin is prohibited in halal and kosher products, and undeclared species substitution remains a recurring form of food and pharmaceutical fraud. Now, a team of researchers in Indonesia has built a compact optical fiber sensor, paired with machine learning, that can tell bovine, porcine, and fish gelatin apart in a matter of seconds — without labels, reagents, or destroying the sample.

The study, published in Results in Optics, tackles a problem that has long frustrated analytical chemists. Gelatin is a partially hydrolyzed form of collagen, and whether it comes from cow hide, pig skin, or fish scales, the underlying protein backbone is essentially the same. The differences show up only in subtle shifts of amino-acid composition, molecular-weight distribution, and hydration behavior. Worse, the heat and chemical treatments used to extract gelatin degrade many of the species-specific markers on which conventional tests such as PCR and mass spectrometry rely. Those laboratory techniques are accurate, but they are destructive, reagent-intensive, and demand trained operators — poor fits for rapid screening at the factory gate or certification office.

The researchers, led by Wildan Panji Tresna and Ian Yulianti of Universitas Negeri Semarang, chose a different route: surface plasmon resonance, or SPR. This phenomenon occurs when light hitting a thin metal film excites collective oscillations of electrons at the metal–dielectric interface. The resonance condition is exquisitely sensitive to the refractive index of whatever sits within the evanescent field — a region of light that extends a few hundred nanometers beyond the surface — so molecules binding to or layering onto the metal shift the optical response in real time, with no fluorescent or radioactive tags needed. Implementing SPR on an optical fiber replaces the bulky prism of classical laboratory instruments with a slim, flexible probe that can simply be dipped into a sample.

The probe itself is elegantly simple. The team stripped the jacket and cladding from a section of a 400-micrometer-core multimode silica fiber, exposing the bare core, then sputter-coated it with a 50-nanometer-thick gold film — a thickness chosen experimentally because it produced the deepest and narrowest resonance dip. A silver paste mirror on the distal tip sends the light back through the sensing region, doubling the interaction length and allowing the whole probe to be interrogated through a single fiber lead. Atomic force microscopy confirmed the gold film deposited uniformly, with a surface roughness averaging just 6.7 nanometers. When the probe is immersed in a gelatin solution, a characteristic resonance appears in the 500–600 nanometer range of the reflected spectrum, and the dip migrates toward longer wavelengths as concentration rises.

Here the team confronted the central obstacle. Measured with a refractometer, the three gelatins behave almost identically: raising the concentration of any one source from 1% to 8% changes its refractive index by about 1.0 × 10⁻² RIU, while at a matched concentration the three species differ by only 1.4 × 10⁻⁴ to 2.1 × 10⁻³ RIU — roughly an order of magnitude less. Even the ranking of the sources is not stable across the concentration series. A sensor reading a single wavelength shift would therefore track concentration far more easily than identity. What actually distinguishes the sources is the shape of the entire resonance — its depth, width, and asymmetry — which reflects how each gelatin layer damps and thickens on the gold surface. Fish gelatin consistently produced the deepest, narrowest dip; bovine the shallowest, broadest.

That full-spectrum shape information is exactly what machine learning can exploit. The researchers recorded reflection spectra from 72 samples — three sources, eight concentrations, three replicates — and built a classification pipeline in which a Savitzky–Golay filter smooths the raw spectra, a scaler standardizes each wavelength channel, and principal component analysis compresses the 750-nanometer-wide window into two coordinates that retain a remarkable 98.09% of the spectral variance. A support vector machine with a radial basis function kernel then draws the decision boundaries among the three classes, using a one-versus-one voting scheme. Crucially, every data-dependent step was fitted only on training data and applied unchanged to held-out spectra, closing the subtle information-leakage pathway that can inflate results in spectral classification.

The performance was striking. On a held-out test partition of 29 spectra the classifier reached 96.55% overall accuracy, with porcine and fish gelatin identified without a single error and bovine gelatin at 90% recall — the lone miss being one bovine spectrum assigned to porcine, exactly where the two species’ optical properties overlap most. Per-class F1 scores ranged from 0.917 to a perfect 1.000. In the PCA score plot, fish gelatin forms a tight, well-separated cluster while bovine and porcine sit closer together, mirroring the spectroscopy. Fourier-transform infrared measurements of the same materials provided chemical corroboration, showing the expected collagen amide bands with the largest differences confined to the amide I region.

Most impressive is the validation design. Because concentration dominates the optical signal, a classifier could cheat by learning concentration instead of source. The team attacked this with leave-one-concentration-out cross-validation, training on seven concentrations and testing on an entirely unseen eighth, and with matched-concentration tests in which identity is the only variable left. A concentration-driven decision rule cannot survive either test, so strong performance there demonstrates that the decision boundaries follow genuinely source-specific spectral structure rather than incidental correlations with how concentrated the solution happened to be.

The authors are refreshingly candid about scope. All measurements were made on pure, laboratory-prepared gelatin in deionized water, one batch per source. Real adulteration is partial — a few percent of porcine gelatin in a bovine product — and real matrices like confectionery, capsule shells, and dairy contain sugars, fats, and surfactants that would alter both bulk refractive index and surface adsorption. Batch-to-batch variation across manufacturers has not yet been tested, and the mechanistic explanation for the resonance shape differences, while plausible, remains a hypothesis pending nanoscale characterization of the adsorbed layers by techniques such as ellipsometry or quartz-crystal microbalance.

Still, the platform’s virtues are clear: it is inexpensive, label-free, interrogated through a single fiber, and produces a full spectrum in seconds — making it practical to accumulate the larger, more varied datasets a robust deployable classifier will need. If follow-up work on mixtures, real foods, and multi-batch samples succeeds, the same gold-coated fiber that now distinguishes three proteins in a beaker could become a rapid screening tool for halal certification bodies and food-quality labs worldwide, bringing one of the food industry’s most stubborn authentication problems within reach of a probe thinner than a pencil lead.

Subject of Research: Label-free identification of the animal source of gelatin using a machine-learning-assisted optical fiber surface plasmon resonance sensor

Article Title: Optical Fiber-surface plasmon resonance (SPR) sensors with machine learning-assisted for label-free identification of animal sources-gelatine

Article References: Tresna, W. P., Yulianti, I., Putra, N. M. D., Alauhdin, M., Leonardy, J., Randa, D., & Aprilia, L. (2026). Optical Fiber-surface plasmon resonance (SPR) sensors with machine learning-assisted for label-free identification of animal sources-gelatine. Results in Optics, Article 101183. https://doi.org/10.1016/j.rio.2026.101183

Image Credits: AI Generated

DOI: 10.1016/j.rio.2026.101183

Keywords: gelatin, surface plasmon resonance, optical fiber sensor, machine learning, food authentication, halal certification, support vector machine, principal component analysis, refractive index, biosensor, porcine gelatin, FTIR spectroscopy