SERS Fluctuations Become Concentration Fingerprints for Ultrasensitive Analysis

A statistical strategy for surface-enhanced Raman scattering, or SERS, converted apparently random signal fluctuations into concentration-specific fingerprints that could support quantitative analysis across very low analyte concentrations. Researchers at Capital Normal University and the Henan Academy of Sciences in China used the shapes of complete Raman intensity distributions, rather than a single peak or the average of many peaks, to train a compact neural network that assigned predefined concentration levels. The experimental study, published in Opto-Electronic Advances, reported 100% identification accuracy across five concentrations of the model dye rhodamine 6G, from 1 × 10⁻¹¹ to 1 × 10⁻⁷ M. Tests with other chemical and biological samples also produced high classification accuracy, although the results remain demonstrations on selected analytes, matrices and laboratory substrates rather than evidence of universal performance.

SERS can amplify molecular Raman signals by many orders of magnitude by placing analytes near gold or silver nanostructures, where localized electromagnetic fields create especially intense “hot spots” in nanoscale gaps. This sensitivity makes the technique potentially useful for trace detection in areas such as food safety, environmental monitoring, drug analysis, disease diagnosis and public security. The same enhancement also makes conventional quantification difficult. At extremely low concentrations, molecules enter hot spots unpredictably, and their number, position and local enhancement vary between measurement points and over time. Spectra collected at the same nominal concentration can therefore differ sharply, while rare bright events may exert disproportionate influence on peak-height averages and calibration curves.

Digital SERS addresses part of this problem by converting measurements into binary events: a signal is classified as present or absent, and the event probability is analyzed with Poisson statistics to infer concentration. The approach has improved quantitation under single-molecule conditions, but binary classification discards amplitude, distribution width, skewness and other features of the original measurements. The new strategy instead retains the continuous SERS values and asks whether the population of fluctuations carries information about analyte concentration. To test that idea, the team collected spectra at many positions, grouped the intensities of a characteristic Raman band into histograms and used a one-dimensional residual neural network called ResNet4 to recognize patterns associated with concentration.

For the main experiment, the researchers prepared relatively uniform arrays of silver nanoparticles with the Langmuir–Blodgett method and measured SERS maps over 30 μm × 30 μm areas. Rhodamine 6G served as the model analyte. As expected, mean peak intensity generally rose with concentration, but individual measurements remained highly variable. At 1 × 10⁻⁹ M, the relative standard deviation of the 612 cm⁻¹ band was 37.9%, and calibration slopes and intercepts differed among substrate batches. Those observations illustrate the measurement variability that makes single-point or mean-intensity calibration vulnerable to outliers and manufacturing differences in the enhancing substrate.

Viewed as statistical distributions, the same mapping data revealed a more structured concentration signal. Across rhodamine 6G concentrations from 1 × 10⁻¹¹ to 1 × 10⁻⁷ M, the histograms changed continuously from strongly skewed profiles with long tails toward nearly symmetric, Gaussian-like shapes. The measured skewness declined from 3.515 to 0.086. Concentration was therefore encoded not only in average signal strength but also in the width, asymmetry and overall form of the complete intensity population. The authors described this evolving profile as a “concentration fingerprint,” a term that describes the observed distribution pattern rather than a chemically unique signature of an individual molecule.

Physical simulations supported the experimental trend. The researchers combined Poisson statistics, generalized Mie theory and Monte Carlo calculations of molecules randomly adsorbed within a plasmonic hot spot formed by two 60 nm silver nanoparticles. In the model, the number of contributing molecules increased with analyte concentration, and the simulated distributions shifted from long-tailed to nearly Gaussian in the same general manner as the measurements. A narrower 1 nm gap generated stronger electromagnetic fields and greater asymmetry at low concentration, but it did not alter the overall concentration-dependent transition. Within the study’s model, molecular number fluctuations drove the evolving distribution shape, while hot-spot strength influenced how strongly the low-concentration profile became asymmetric.

ResNet4 was trained and evaluated with mapping measurements from independent batches of SERS substrates, allowing the researchers to test whether the learned patterns extended beyond one set of samples. For the five rhodamine 6G levels, the network achieved 100% identification accuracy and performed better than the tested k-nearest-neighbor and support-vector-machine alternatives. The classification result shows that the model could distinguish the predefined concentration categories in the tested dataset; it does not by itself establish that every unknown sample would be assigned the correct absolute concentration. The comparison nevertheless supports the central proposal that retaining the shape of the intensity distribution can provide more usable information than reducing the spectrum to one intensity value or a binary event.

The approach also generalized beyond the model dye. In additional tests, it classified methamphetamine in methanol from 1 part per billion to 10 parts per million with 99.6% average accuracy. It identified thiram in bean-sprout extract with 99% accuracy across 1 × 10⁻⁸ to 1 × 10⁻⁶ M, and rosiglitazone maleate in methanol-treated artificial serum with approximately 96% accuracy across 1 × 10⁻⁷ to 1 × 10⁻⁵ M. These results broaden the evidence from a controlled nanoparticle-and-dye experiment to applications involving different analytes and sample matrices. They remain laboratory validation results, however, and do not demonstrate field readiness, performance in every biological or environmental sample, or equivalence to a standardized measurement procedure.

The immediate advance is a change in how SERS variability is used. Rather than treating variation as noise to be suppressed, the method preserves the full continuous intensity distribution and allows the neural network to learn from molecular-number fluctuations, differences in hot-spot enhancement, distribution width, skewness and overall profile. That statistical foundation may reduce dependence on rare bright events and single-spectrum calibration while avoiding a strict requirement for binary “present or absent” measurements under single-molecule conditions. The study therefore offers a promising route to more reliable ultrasensitive quantitation, but broader deployment will require testing across more substrates, instruments, operators, sample preparations and real-world samples, as well as clearer characterization of generalization, uncertainty and failure modes.

Subject of Research: Statistical analysis of SERS intensity distributions for deep-learning-assisted concentration classification across selected chemical and biological samples.

Article Title: From signal fluctuations to concentration fingerprints: Statistical SERS intensity distributions enable reliable quantitation

Article References: From signal fluctuations to concentration fingerprints: Statistical SERS intensity distributions enable reliable quantitation. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: surface-enhanced Raman scattering, SERS quantitation, intensity distributions, concentration fingerprints, deep learning, ResNet4, ultrasensitive detection, rhodamine 6G, plasmonic hot spots, chemical analysis, signal, fluctuations