New Nanolayer Sensor Combines Plasmonics and Machine Learning to Detect Dangerous Bacteria

A team of researchers from India, Saudi Arabia, Nepal, and the United Kingdom has unveiled a theoretical surface plasmon resonance (SPR) sensor that they say can detect pathogenic bacteria with markedly higher sensitivity than comparable designs, and they have paired the optics with a machine learning model that learns how to optimize the sensor on its own. The study, published in Frontiers in Sensors, describes a multilayer architecture built from tellurium dioxide (TeO2), cadmium gallium sulfide (CdGa2S4), and the two-dimensional material aluminium antimonide (AlSb) stacked on a thin silver film, and reports a sensitivity of 269.397 degrees per refractive index unit along with a figure of merit of 180.6 per RIU, values the authors describe as a significant improvement over existing work.

The detection principle is deceptively simple. In the classic Kretschmann configuration, a prism couples p-polarized light into a thin metal film, and at a precise angle of incidence the light’s in-plane momentum matches that of surface plasmon waves, collective oscillations of electrons rippling along the metal-dielectric interface. When that resonance condition is met, energy drains from the reflected beam, producing a sharp dip in reflectance. Because the resonance angle depends exquisitely on the refractive index of whatever sits just above the metal, the presence of bacteria, which changes the local optical environment, shifts the dip in a measurable way. Even minute refractive index variations produce a noticeable displacement of the resonance minimum, enabling label-free, real-time detection without the fluorescent tags or enzymatic amplification that conventional assays require.

Traditional methods for spotting dangerous bacteria, including microbial culture, polymerase chain reaction, enzyme-linked immunosorbent assays, fluorescence-based techniques, and electrochemical approaches, are accurate but often demand complex instrumentation, skilled personnel, and lengthy sample preparation, which limits their usefulness for rapid, on-site testing. That bottleneck is why the search for optical biosensing platforms has intensified. Among optical techniques, SPR is regarded as one of the most sensitive for chemical and biological detection, and the new work aims to push that sensitivity further by engineering the stack of materials through which the evanescent electromagnetic field must pass before probing the sample.

The authors modeled their six-layer device with the transfer matrix method, an exact analytical technique for electromagnetic wave propagation through layered thin films, under angular interrogation at a wavelength of 633 nanometers. The stack begins with a BK7 prism of refractive index 1.515, chosen for its optical transparency and phase-matching capability, followed by a silver film whose complex refractive index of 0.056266 plus 4.27761i reflects the metal’s low optical loss and its knack for producing a sharp resonance curve. Above the silver sit two dielectric interlayers, one nanometer of TeO2 with a refractive index of 2.2596 and one nanometer of CdGa2S4 with a refractive index of 2.4482, and finally a 0.53-nanometer layer of AlSb with a complex refractive index of 3.8638 plus 0.0035982i, on which the sensing medium rests.

Each layer plays a distinct physical role. The tellurium dioxide both confines the electromagnetic field and shields the silver surface from oxidation, addressing a well-known weakness of silver-based plasmonic sensors, which otherwise outperform many alternatives in field confinement but degrade in air. The CdGa2S4 layer enhances plasmonic coupling and broadens the interface over which the evanescent field can interact with the analyte, lending stability to the resonance. The AlSb nanolayer, with its ultra-high surface area and strong light-matter interaction, increases the effective sensing area available for biomolecular adsorption. Together, the researchers argue, these materials create a synergistic improvement in plasmonic coupling and electromagnetic field localization that no single layer could deliver alone.

Getting the metal thickness right turned out to be critical. The team swept silver film thicknesses from 40 to 60 nanometers in 5-nanometer steps and tracked how the minimum reflectance and resonance angle responded in sensing media of refractive index 1.371 and 1.410. As the film thickened from 40 to 55 nanometers, the minimum reflectance plunged from 0.34442 to just 0.002464, a sign that nearly all the incident energy was being funneled into the surface plasmon waves. At 60 nanometers, the reflectance minimum climbed again because the thicker metal added optical losses and suppressed the evanescent field’s reach into the sample. A 55-nanometer film was therefore selected as the optimum, and it anchors all subsequent performance claims in the study.

With the metal fixed, the authors compared their full stack against four simpler configurations: a conventional BK7/Ag/analyte structure, versions with only AlSb, only TeO2, or only CdGa2S4 added, and the complete proposed design. The conventional structure managed a sensitivity of 185.620 degrees per RIU, while the partial additions lifted sensitivity to roughly 200 to 203.68 degrees per RIU. The full multilayer design reached 269.397 degrees per RIU, an improvement of approximately 45.1, 32.3, 34.4, and 32.3 percent over the four reference structures respectively, and it simultaneously posted the best quality factor at 181.04 per RIU, figure of merit at 180.6 per RIU, and combined sensitivity factor at 179.679 per RIU. In other words, the design improves sensitivity without sacrificing the resonance sharpness that determines real-world detection accuracy, a balance many competing sensors fail to strike.

The team also swapped the 2D layer among black phosphorus, graphene, molybdenum disulfide, and tungsten disulfide to see how each material changes the picture. Black phosphorus came closest to AlSb with a sensitivity of 261.179 degrees per RIU and a quality factor of 178.036 per RIU, while tungsten disulfide actually produced the largest angular shift, at 282.538 degrees per RIU, but its broader resonance curve dragged the quality factor down to 144.152 per RIU and the figure of merit to 134.875 per RIU, meaning the resonance profile becomes too damped for high-quality sensing. AlSb emerged as the best overall performer because it combines high sensitivity with a sharp, well-defined resonance dip, and the authors note that in a 2D-material comparison the abstract reports maximum attained parameters of 285.538 degrees per RIU sensitivity, 144.152 per RIU quality factor, and 134.875 per RIU figure of merit when WS2 configurations are included in the analysis.

Electric field calculations add a biological plausibility check. The evanescent field of the optimized design penetrates 448.43 nanometers into the sensing medium, decaying to roughly 36.8 percent of its maximum intensity at that depth, which is well matched to the scale at which biomolecules, proteins, pesticides, and bacterial cells interact with a functionalized surface. The dielectric interlayers proved to modulate the field confinement in a way that strengthens the interaction between the plasmonic field and the analyte while keeping optical losses in check, so that the slightest refractive index shift in the sample translates into a large deviation of the resonance condition.

The second half of the study is where the machine learning enters, and it may prove to be the paper’s most transferable contribution. The researchers generated an augmented dataset from their numerical simulations, covering silver thickness, refractive index, resonance angle, minimum reflectance, full width at half maximum, quality factor, figure of merit, and combined sensitivity factor, and trained a Random Forest regression model of 250 decision trees with a maximum depth of ten levels to predict sensor sensitivity. Trained on 80 percent of the data and tested on the remaining 20 percent, the model showed excellent agreement between predicted and target values, with residuals clustered symmetrically near zero and learning curves confirming the absence of significant overfitting. Feature importance analysis identified the silver layer thickness as the most influential parameter, with resonance angle and refractive index close behind, a result consistent with the direct electromagnetic optimization. When the trained model was then used to predict sensitivity across the 40-to-60-nanometer thickness range, it independently identified an optimum near 55 nanometers, precisely matching the transfer matrix analysis but at a fraction of the computational cost.

The authors are candid about the work’s limitations. Everything reported so far is numerical; no device has been fabricated or experimentally characterized, and the model does not yet include bacteria-specific surface functionalization or molecular recognition elements, so selective detection of a particular bacterial species would still require a suitable bioreceptor layer. Real biological samples bring nonspecific adsorption, background refractive index fluctuations, and matrix interference from salts, proteins, and cells, while temperature variations, fabrication tolerances, surface roughness, and material absorption could all degrade the idealized response. The team plans to pursue experimental fabrication, bacteria-directed surface chemistry, and validation with real-world bio samples. If those steps succeed, the combination of a carefully engineered TeO2/CdGa2S4/AlSb plasmonic stack with a data-driven optimization loop could offer a template for a new generation of fast, label-free pathogen detectors for clinics, food safety labs, and environmental monitoring, and the Random Forest framework may let future designers find the best sensor geometry in hours of computation rather than weeks of simulation.

Subject of Research: Machine learning-optimized surface plasmon resonance refractive index sensor with TeO2/CdGa2S4/AlSb layers for pathogenic bacteria detection

Article Title: Pathogenic bacteria sensing with tellurium dioxide/cadmium gallium sulfide-assisted surface plasmonic resonance refractive index sensor: a machine learning approach

Article References: Merigala, R., Boddu, S., Vasimalla, Y., Prashanth, M. S., Chebaane, S., Alsaif, N. A. M., Khalifa, S. B., & Karki, B. (2026). Pathogenic bacteria sensing with tellurium dioxide/cadmium gallium sulfide-assisted surface plasmonic resonance refractive index sensor: a machine learning approach. Frontiers in Sensors, 7, Article 1962401. https://doi.org/10.3389/fsens.2026.1962401

Image Credits: AI Generated

DOI: 10.3389/fsens.2026.1962401

Keywords: surface plasmon resonance, pathogenic bacteria, biosensor, machine learning, Random Forest, tellurium dioxide, cadmium gallium sulfide, aluminium antimonide, refractive index sensing, Kretschmann configuration, 2D materials, nanophotonics