📄 chemistry

SERS Chemical Tongue for Rapid Detection of Chili Oils Assisted with a Convolutional Neural Network

This study presents a rapid and accurate detection system for chili oils that integrates a novel SERS chemical tongue using three carbonized polymer dot-functionalized magnetic nanoparticles as universal receptors with a specialized convolutional neural network to achieve superior identification and quantification of capsaicin in both pure and complex compounded matrices.

Original authors: Peng Wang, Cun-Fang Zhang, Yu Qin, Xiao-Lan Wei

Published 2026-07-27

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Peng Wang, Cun-Fang Zhang, Yu Qin, Xiao-Lan Wei

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a busy food market, trying to figure out exactly what's inside a jar of spicy chili oil. Is it made with cheap oil? Did they use the right kind of chili? Is it actually spicy enough? In the world of chemistry, this is a tricky puzzle because oil is thick, dark, and full of different ingredients that all get mixed together. Scientists have a special tool called a "chemical tongue" to solve this. Think of it like a super-smart nose that doesn't just smell one thing, but tastes the whole mix at once. It uses a technique called SERS (Surface-Enhanced Raman Scattering), which is like shining a super-powerful flashlight on molecules to make them glow with a unique color pattern, almost like a fingerprint. Usually, these tools are great for water or alcohol, but they get confused and messy when faced with thick, greasy oil. That's why researchers are always looking for a way to make these chemical tongues work better in the real, oily world of our kitchens.

This paper introduces a brand-new, high-tech "chemical tongue" designed specifically to taste and analyze chili oil. The team, led by Peng Wang and colleagues, built a system that combines three special sensors with a smart computer brain. Instead of using flimsy sensors that might fall apart in oil, they created tiny, magnetic nanoparticles coated with "carbonized polymer dots" (CPDs). You can think of these dots as super-strong glue that locks the sensors in place, so they don't get washed away by the oil. These three sensors act like three different taste buds: one reacts to how oily the mix is, another to the spicy heat, and the third to the overall flavor profile. When they touch the chili oil, they each send back a unique signal, creating a complex "super-spectrum" or a giant fingerprint of the oil's contents.

To make sense of all these signals, the researchers didn't just use a simple calculator; they built a custom "Convolutional Neural Network" (CNN), which is a type of artificial intelligence that acts like a detective. This AI was trained to look at the combined fingerprints from the three sensors and figure out two things at once: what kind of chili oil it is (qualitative) and exactly how much capsaicin (the chemical that makes things spicy) is inside (quantitative). The results were impressive. For pure chili oils, the AI got it right 100% of the time. Even for complicated mixtures with different oils, chilies, and additives, it still got it right 99.63% of the time. When it came to measuring the exact amount of spice, the AI was much more accurate than older methods, finding the spice levels with a detection limit as low as 9.1 ng/mL. The study suggests that by locking the sensors in place with these carbon dots and using a smart AI to read the results, we can finally get a fast, reliable, and on-the-spot way to check the quality of our favorite spicy condiments.

Technical Summary: SERS Chemical Tongue for Rapid Detection of Chili Oils Assisted with a Convolutional Neural Network

Problem Statement
Chili oil is a globally significant condiment, yet its quality assessment is hindered by a complex matrix comprising various oil bases, chili varieties, and auxiliary ingredients. Traditional detection methods for capsaicinoids (the primary pungency agents) rely on chromatographic separation, which requires expensive instrumentation and extensive pretreatment, making them unsuitable for rapid, on-site analysis. While chemical sensor arrays ("chemical tongues") offer a potential alternative, existing Surface-Enhanced Raman Scattering (SERS) sensors often utilize physically adsorbed or self-assembled small-molecule receptors. These receptors suffer from ligand desorption and poor stability in non-polar, oil-rich matrices. Furthermore, previous SERS chemical tongue applications have largely been limited to water or alcohol systems (e.g., Baijiu), with no reported success in the highly viscous and interference-prone environment of pure chili oil. Additionally, existing machine learning approaches for SERS data often employ generic architectures that fail to capture the dual nature of Raman spectra (sharp peaks and regional trends) or perform only single tasks (classification or regression) rather than integrated analysis.

Methodology
The authors developed a hybrid platform combining a novel material-based SERS chemical tongue with a tailored Convolutional Neural Network (CNN).

Sensor Fabrication (Material Innovation): The team synthesized three distinct SERS sensors using magnetic core–silver shell nanoparticles (MNP@Ag) functionalized with carbonized polymer dots (CPDs). The CPDs were generated via in-situ microwave carbonization of aromatic molecules (4-aminothiophenol, benzoic acid, and 4-hydroxythiophenol) on the silver surface. Unlike traditional methods, this process forms covalent C–N/Ag and C–O/Ag bonds, "locking" the receptors to the surface. This design ensures exceptional stability in non-polar oil matrices and prevents receptor desorption. The three sensors (MNP@Ag-CPDA, MNP@Ag-CPDAB, and MNP@Ag-CPDH) provide complementary, cross-reactive responses to different flavor components.
Data Acquisition: Raman spectra were collected from pure and compounded chili oil samples using the three sensors. The individual spectra were concatenated into a "super-spectrum" to serve as the input for the machine learning model.
Algorithm Design (CNN Innovation): A custom multi-task CNN was developed specifically for SERS spectral features. Key architectural elements include:
- Parallel Pooling: An input stage utilizing parallel max-pooling (to preserve sharp, discriminative Raman peaks) and average-pooling (to retain global spectral trends and reduce noise), followed by summation.
- Dual Output Heads: The network simultaneously performs multi-class classification (identifying oil type/chili variety) and regression (quantifying capsaicin concentration).
- Data Augmentation: To address sample scarcity, the training data was augmented with Gaussian noise, wavenumber shifts, and baseline drifts to simulate realistic measurement perturbations.
Experimental Scope: The study evaluated the system on simplified matrices (pure rapeseed oil, binary oil blends) and complex matrices (ternary blends with animal oils, various additives, and commercial samples). Performance was benchmarked against conventional machine learning models (KNN, SVM, LDA) and linear regression (PLS-R).

Key Results

Sensor Performance: The CPD-functionalized sensors demonstrated high stability and sensitivity, achieving a limit of detection (LOD) for capsaicin of 9.1 ng/mL. The sensors showed excellent batch-to-batch reproducibility (RSD < 2%) and stability over 28 days.
Qualitative Classification:
- For pure chili oils, the CNN achieved 100% identification accuracy, outperforming KNN (90.00%), SVM (94.44%), and LDA (99.26%).
- For complex compounded chili oils (varying oil types, chili varieties, and additives), the CNN maintained 99.63% accuracy and a macro F1 score of 99.63%, significantly surpassing KNN (84.44%), SVM (88.15%), and LDA (96.67%).
Quantitative Regression: In capsaicin quantification across both simplified and complex matrices, the CNN consistently outperformed Partial Least Squares Regression (PLS-R).
- In complex matrices (Series 5), the CNN achieved an $R^2$ of 0.979 with an RMSE of 0.306, compared to PLS-R's $R^2$ of 0.889 and RMSE of 0.699.
- The CNN demonstrated superior robustness against matrix interference, maintaining higher $R^2$ values and lower error rates (RMSE, MAE) than linear models.
Mechanism: The study elucidated that the sensors respond to chili oil components through specific interactions: hydrogen bonding (for hydroxyl-functionalized sensors), hydrophobic/ $\pi$ - $\pi$ interactions (for aryl-modified sensors), and laser-induced dimer formation (DMAB) promoted by flavor molecules (for amino-functionalized sensors).

Significance and Claims
The authors claim this work represents the first successful application of a SERS-based chemical tongue to a complex, 100% oil-phase food matrix. The significance of the study lies in three main advancements:

Material Stability: The introduction of covalently anchored CPDs as "universal receptors" solves the critical issue of receptor desorption in non-polar media, a limitation of previous SERS chemical tongues.
Algorithmic Tailoring: The development of a dedicated multi-task CNN with parallel pooling and dual output heads allows for the simultaneous, high-accuracy qualitative and quantitative analysis of complex food samples, outperforming both traditional chemometrics and generic machine learning models.
Practical Application: The platform enables rapid, accurate, and on-site-ready detection of chili oil quality and capsaicin content, addressing the urgent need for reliable quality assessment in the expanding global chili oil market.

The paper concludes that the synergistic integration of CPD-based universal receptors and a SERS-tailored multi-task CNN provides a feasible solution for the rapid characterization of chili oil and analogous edible oil products, with future work planned to expand the sample library and develop handheld devices for field use.

More like this